fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/writeback.h>
  18 #include <linux/compat.h>
  19 #include <linux/xattr.h>
  20 #include <linux/posix_acl.h>
  21 #include <linux/falloc.h>
  22 #include <linux/slab.h>
  23 #include <linux/ratelimit.h>
  24 #include <linux/btrfs.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/posix_acl_xattr.h>
  27 #include <linux/uio.h>
  28 #include <linux/magic.h>
  29 #include <linux/iversion.h>
  30 #include <linux/swap.h>
  31 #include <linux/sched/mm.h>
  32 #include <asm/unaligned.h>
  33 #include "ctree.h"
  34 #include "disk-io.h"
  35 #include "transaction.h"
  36 #include "btrfs_inode.h"
  37 #include "print-tree.h"
  38 #include "ordered-data.h"
  39 #include "xattr.h"
  40 #include "tree-log.h"
  41 #include "volumes.h"
  42 #include "compression.h"
  43 #include "locking.h"
  44 #include "free-space-cache.h"
  45 #include "inode-map.h"
  46 #include "backref.h"
  47 #include "props.h"
  48 #include "qgroup.h"
  49 #include "delalloc-space.h"
  50 #include "block-group.h"
  51
  52 struct btrfs_iget_args {
  53         struct btrfs_key *location;
  54         struct btrfs_root *root;
  55 };
  56
  57 struct btrfs_dio_data {
  58         u64 reserve;
  59         u64 unsubmitted_oe_range_start;
  60         u64 unsubmitted_oe_range_end;
  61         int overwrite;
  62 };
  63
  64 static const struct inode_operations btrfs_dir_inode_operations;
  65 static const struct inode_operations btrfs_symlink_inode_operations;
  66 static const struct inode_operations btrfs_dir_ro_inode_operations;
  67 static const struct inode_operations btrfs_special_inode_operations;
  68 static const struct inode_operations btrfs_file_inode_operations;
  69 static const struct address_space_operations btrfs_aops;
  70 static const struct file_operations btrfs_dir_file_operations;
  71 static const struct extent_io_ops btrfs_extent_io_ops;
  72
  73 static struct kmem_cache *btrfs_inode_cachep;
  74 struct kmem_cache *btrfs_trans_handle_cachep;
  75 struct kmem_cache *btrfs_path_cachep;
  76 struct kmem_cache *btrfs_free_space_cachep;
  77
  78 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  79 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  80 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  81 static noinline int cow_file_range(struct inode *inode,
  82                                    struct page *locked_page,
  83                                    u64 start, u64 end, int *page_started,
  84                                    unsigned long *nr_written, int unlock);
  85 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  86                                        u64 orig_start, u64 block_start,
  87                                        u64 block_len, u64 orig_block_len,
  88                                        u64 ram_bytes, int compress_type,
  89                                        int type);
  90
  91 static void __endio_write_update_ordered(struct inode *inode,
  92                                          const u64 offset, const u64 bytes,
  93                                          const bool uptodate);
  94
  95 /*
  96  * Cleanup all submitted ordered extents in specified range to handle errors
  97  * from the btrfs_run_delalloc_range() callback.
  98  *
  99  * NOTE: caller must ensure that when an error happens, it can not call
 100  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 101  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 102  * to be released, which we want to happen only when finishing the ordered
 103  * extent (btrfs_finish_ordered_io()).
 104  */
 105 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 106                                                  struct page *locked_page,
 107                                                  u64 offset, u64 bytes)
 108 {
 109         unsigned long index = offset >> PAGE_SHIFT;
 110         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 111         u64 page_start = page_offset(locked_page);
 112         u64 page_end = page_start + PAGE_SIZE - 1;
 113
 114         struct page *page;
 115
 116         while (index <= end_index) {
 117                 page = find_get_page(inode->i_mapping, index);
 118                 index++;
 119                 if (!page)
 120                         continue;
 121                 ClearPagePrivate2(page);
 122                 put_page(page);
 123         }
 124
 125         /*
 126          * In case this page belongs to the delalloc range being instantiated
 127          * then skip it, since the first page of a range is going to be
 128          * properly cleaned up by the caller of run_delalloc_range
 129          */
 130         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
 131                 offset += PAGE_SIZE;
 132                 bytes -= PAGE_SIZE;
 133         }
 134
 135         return __endio_write_update_ordered(inode, offset, bytes, false);
 136 }
 137
 138 static int btrfs_dirty_inode(struct inode *inode);
 139
 140 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 141 void btrfs_test_inode_set_ops(struct inode *inode)
 142 {
 143         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 144 }
 145 #endif
 146
 147 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 148                                      struct inode *inode,  struct inode *dir,
 149                                      const struct qstr *qstr)
 150 {
 151         int err;
 152
 153         err = btrfs_init_acl(trans, inode, dir);
 154         if (!err)
 155                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 156         return err;
 157 }
 158
 159 /*
 160  * this does all the hard work for inserting an inline extent into
 161  * the btree.  The caller should have done a btrfs_drop_extents so that
 162  * no overlapping inline items exist in the btree
 163  */
 164 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 165                                 struct btrfs_path *path, int extent_inserted,
 166                                 struct btrfs_root *root, struct inode *inode,
 167                                 u64 start, size_t size, size_t compressed_size,
 168                                 int compress_type,
 169                                 struct page **compressed_pages)
 170 {
 171         struct extent_buffer *leaf;
 172         struct page *page = NULL;
 173         char *kaddr;
 174         unsigned long ptr;
 175         struct btrfs_file_extent_item *ei;
 176         int ret;
 177         size_t cur_size = size;
 178         unsigned long offset;
 179
 180         ASSERT((compressed_size > 0 && compressed_pages) ||
 181                (compressed_size == 0 && !compressed_pages));
 182
 183         if (compressed_size && compressed_pages)
 184                 cur_size = compressed_size;
 185
 186         inode_add_bytes(inode, size);
 187
 188         if (!extent_inserted) {
 189                 struct btrfs_key key;
 190                 size_t datasize;
 191
 192                 key.objectid = btrfs_ino(BTRFS_I(inode));
 193                 key.offset = start;
 194                 key.type = BTRFS_EXTENT_DATA_KEY;
 195
 196                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 197                 path->leave_spinning = 1;
 198                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 199                                               datasize);
 200                 if (ret)
 201                         goto fail;
 202         }
 203         leaf = path->nodes[0];
 204         ei = btrfs_item_ptr(leaf, path->slots[0],
 205                             struct btrfs_file_extent_item);
 206         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 207         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 208         btrfs_set_file_extent_encryption(leaf, ei, 0);
 209         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 210         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 211         ptr = btrfs_file_extent_inline_start(ei);
 212
 213         if (compress_type != BTRFS_COMPRESS_NONE) {
 214                 struct page *cpage;
 215                 int i = 0;
 216                 while (compressed_size > 0) {
 217                         cpage = compressed_pages[i];
 218                         cur_size = min_t(unsigned long, compressed_size,
 219                                        PAGE_SIZE);
 220
 221                         kaddr = kmap_atomic(cpage);
 222                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 223                         kunmap_atomic(kaddr);
 224
 225                         i++;
 226                         ptr += cur_size;
 227                         compressed_size -= cur_size;
 228                 }
 229                 btrfs_set_file_extent_compression(leaf, ei,
 230                                                   compress_type);
 231         } else {
 232                 page = find_get_page(inode->i_mapping,
 233                                      start >> PAGE_SHIFT);
 234                 btrfs_set_file_extent_compression(leaf, ei, 0);
 235                 kaddr = kmap_atomic(page);
 236                 offset = offset_in_page(start);
 237                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 238                 kunmap_atomic(kaddr);
 239                 put_page(page);
 240         }
 241         btrfs_mark_buffer_dirty(leaf);
 242         btrfs_release_path(path);
 243
 244         /*
 245          * we're an inline extent, so nobody can
 246          * extend the file past i_size without locking
 247          * a page we already have locked.
 248          *
 249          * We must do any isize and inode updates
 250          * before we unlock the pages.  Otherwise we
 251          * could end up racing with unlink.
 252          */
 253         BTRFS_I(inode)->disk_i_size = inode->i_size;
 254         ret = btrfs_update_inode(trans, root, inode);
 255
 256 fail:
 257         return ret;
 258 }
 259
 260
 261 /*
 262  * conditionally insert an inline extent into the file.  This
 263  * does the checks required to make sure the data is small enough
 264  * to fit as an inline extent.
 265  */
 266 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 267                                           u64 end, size_t compressed_size,
 268                                           int compress_type,
 269                                           struct page **compressed_pages)
 270 {
 271         struct btrfs_root *root = BTRFS_I(inode)->root;
 272         struct btrfs_fs_info *fs_info = root->fs_info;
 273         struct btrfs_trans_handle *trans;
 274         u64 isize = i_size_read(inode);
 275         u64 actual_end = min(end + 1, isize);
 276         u64 inline_len = actual_end - start;
 277         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 278         u64 data_len = inline_len;
 279         int ret;
 280         struct btrfs_path *path;
 281         int extent_inserted = 0;
 282         u32 extent_item_size;
 283
 284         if (compressed_size)
 285                 data_len = compressed_size;
 286
 287         if (start > 0 ||
 288             actual_end > fs_info->sectorsize ||
 289             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 290             (!compressed_size &&
 291             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 292             end + 1 < isize ||
 293             data_len > fs_info->max_inline) {
 294                 return 1;
 295         }
 296
 297         path = btrfs_alloc_path();
 298         if (!path)
 299                 return -ENOMEM;
 300
 301         trans = btrfs_join_transaction(root);
 302         if (IS_ERR(trans)) {
 303                 btrfs_free_path(path);
 304                 return PTR_ERR(trans);
 305         }
 306         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 307
 308         if (compressed_size && compressed_pages)
 309                 extent_item_size = btrfs_file_extent_calc_inline_size(
 310                    compressed_size);
 311         else
 312                 extent_item_size = btrfs_file_extent_calc_inline_size(
 313                     inline_len);
 314
 315         ret = __btrfs_drop_extents(trans, root, inode, path,
 316                                    start, aligned_end, NULL,
 317                                    1, 1, extent_item_size, &extent_inserted);
 318         if (ret) {
 319                 btrfs_abort_transaction(trans, ret);
 320                 goto out;
 321         }
 322
 323         if (isize > actual_end)
 324                 inline_len = min_t(u64, isize, actual_end);
 325         ret = insert_inline_extent(trans, path, extent_inserted,
 326                                    root, inode, start,
 327                                    inline_len, compressed_size,
 328                                    compress_type, compressed_pages);
 329         if (ret && ret != -ENOSPC) {
 330                 btrfs_abort_transaction(trans, ret);
 331                 goto out;
 332         } else if (ret == -ENOSPC) {
 333                 ret = 1;
 334                 goto out;
 335         }
 336
 337         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 338         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 339 out:
 340         /*
 341          * Don't forget to free the reserved space, as for inlined extent
 342          * it won't count as data extent, free them directly here.
 343          * And at reserve time, it's always aligned to page size, so
 344          * just free one page here.
 345          */
 346         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 347         btrfs_free_path(path);
 348         btrfs_end_transaction(trans);
 349         return ret;
 350 }
 351
 352 struct async_extent {
 353         u64 start;
 354         u64 ram_size;
 355         u64 compressed_size;
 356         struct page **pages;
 357         unsigned long nr_pages;
 358         int compress_type;
 359         struct list_head list;
 360 };
 361
 362 struct async_chunk {
 363         struct inode *inode;
 364         struct page *locked_page;
 365         u64 start;
 366         u64 end;
 367         unsigned int write_flags;
 368         struct list_head extents;
 369         struct btrfs_work work;
 370         atomic_t *pending;
 371 };
 372
 373 struct async_cow {
 374         /* Number of chunks in flight; must be first in the structure */
 375         atomic_t num_chunks;
 376         struct async_chunk chunks[];
 377 };
 378
 379 static noinline int add_async_extent(struct async_chunk *cow,
 380                                      u64 start, u64 ram_size,
 381                                      u64 compressed_size,
 382                                      struct page **pages,
 383                                      unsigned long nr_pages,
 384                                      int compress_type)
 385 {
 386         struct async_extent *async_extent;
 387
 388         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 389         BUG_ON(!async_extent); /* -ENOMEM */
 390         async_extent->start = start;
 391         async_extent->ram_size = ram_size;
 392         async_extent->compressed_size = compressed_size;
 393         async_extent->pages = pages;
 394         async_extent->nr_pages = nr_pages;
 395         async_extent->compress_type = compress_type;
 396         list_add_tail(&async_extent->list, &cow->extents);
 397         return 0;
 398 }
 399
 400 /*
 401  * Check if the inode has flags compatible with compression
 402  */
 403 static inline bool inode_can_compress(struct inode *inode)
 404 {
 405         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW ||
 406             BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 407                 return false;
 408         return true;
 409 }
 410
 411 /*
 412  * Check if the inode needs to be submitted to compression, based on mount
 413  * options, defragmentation, properties or heuristics.
 414  */
 415 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 416 {
 417         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 418
 419         if (!inode_can_compress(inode)) {
 420                 WARN(IS_ENABLED(CONFIG_BTRFS_DEBUG),
 421                         KERN_ERR "BTRFS: unexpected compression for ino %llu\n",
 422                         btrfs_ino(BTRFS_I(inode)));
 423                 return 0;
 424         }
 425         /* force compress */
 426         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 427                 return 1;
 428         /* defrag ioctl */
 429         if (BTRFS_I(inode)->defrag_compress)
 430                 return 1;
 431         /* bad compression ratios */
 432         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 433                 return 0;
 434         if (btrfs_test_opt(fs_info, COMPRESS) ||
 435             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 436             BTRFS_I(inode)->prop_compress)
 437                 return btrfs_compress_heuristic(inode, start, end);
 438         return 0;
 439 }
 440
 441 static inline void inode_should_defrag(struct btrfs_inode *inode,
 442                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 443 {
 444         /* If this is a small write inside eof, kick off a defrag */
 445         if (num_bytes < small_write &&
 446             (start > 0 || end + 1 < inode->disk_i_size))
 447                 btrfs_add_inode_defrag(NULL, inode);
 448 }
 449
 450 /*
 451  * we create compressed extents in two phases.  The first
 452  * phase compresses a range of pages that have already been
 453  * locked (both pages and state bits are locked).
 454  *
 455  * This is done inside an ordered work queue, and the compression
 456  * is spread across many cpus.  The actual IO submission is step
 457  * two, and the ordered work queue takes care of making sure that
 458  * happens in the same order things were put onto the queue by
 459  * writepages and friends.
 460  *
 461  * If this code finds it can't get good compression, it puts an
 462  * entry onto the work queue to write the uncompressed bytes.  This
 463  * makes sure that both compressed inodes and uncompressed inodes
 464  * are written in the same order that the flusher thread sent them
 465  * down.
 466  */
 467 static noinline int compress_file_range(struct async_chunk *async_chunk)
 468 {
 469         struct inode *inode = async_chunk->inode;
 470         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 471         u64 blocksize = fs_info->sectorsize;
 472         u64 start = async_chunk->start;
 473         u64 end = async_chunk->end;
 474         u64 actual_end;
 475         int ret = 0;
 476         struct page **pages = NULL;
 477         unsigned long nr_pages;
 478         unsigned long total_compressed = 0;
 479         unsigned long total_in = 0;
 480         int i;
 481         int will_compress;
 482         int compress_type = fs_info->compress_type;
 483         int compressed_extents = 0;
 484         int redirty = 0;
 485
 486         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 487                         SZ_16K);
 488
 489         actual_end = min_t(u64, i_size_read(inode), end + 1);
 490 again:
 491         will_compress = 0;
 492         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 493         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 494         nr_pages = min_t(unsigned long, nr_pages,
 495                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 496
 497         /*
 498          * we don't want to send crud past the end of i_size through
 499          * compression, that's just a waste of CPU time.  So, if the
 500          * end of the file is before the start of our current
 501          * requested range of bytes, we bail out to the uncompressed
 502          * cleanup code that can deal with all of this.
 503          *
 504          * It isn't really the fastest way to fix things, but this is a
 505          * very uncommon corner.
 506          */
 507         if (actual_end <= start)
 508                 goto cleanup_and_bail_uncompressed;
 509
 510         total_compressed = actual_end - start;
 511
 512         /*
 513          * skip compression for a small file range(<=blocksize) that
 514          * isn't an inline extent, since it doesn't save disk space at all.
 515          */
 516         if (total_compressed <= blocksize &&
 517            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 518                 goto cleanup_and_bail_uncompressed;
 519
 520         total_compressed = min_t(unsigned long, total_compressed,
 521                         BTRFS_MAX_UNCOMPRESSED);
 522         total_in = 0;
 523         ret = 0;
 524
 525         /*
 526          * we do compression for mount -o compress and when the
 527          * inode has not been flagged as nocompress.  This flag can
 528          * change at any time if we discover bad compression ratios.
 529          */
 530         if (inode_need_compress(inode, start, end)) {
 531                 WARN_ON(pages);
 532                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 533                 if (!pages) {
 534                         /* just bail out to the uncompressed code */
 535                         nr_pages = 0;
 536                         goto cont;
 537                 }
 538
 539                 if (BTRFS_I(inode)->defrag_compress)
 540                         compress_type = BTRFS_I(inode)->defrag_compress;
 541                 else if (BTRFS_I(inode)->prop_compress)
 542                         compress_type = BTRFS_I(inode)->prop_compress;
 543
 544                 /*
 545                  * we need to call clear_page_dirty_for_io on each
 546                  * page in the range.  Otherwise applications with the file
 547                  * mmap'd can wander in and change the page contents while
 548                  * we are compressing them.
 549                  *
 550                  * If the compression fails for any reason, we set the pages
 551                  * dirty again later on.
 552                  *
 553                  * Note that the remaining part is redirtied, the start pointer
 554                  * has moved, the end is the original one.
 555                  */
 556                 if (!redirty) {
 557                         extent_range_clear_dirty_for_io(inode, start, end);
 558                         redirty = 1;
 559                 }
 560
 561                 /* Compression level is applied here and only here */
 562                 ret = btrfs_compress_pages(
 563                         compress_type | (fs_info->compress_level << 4),
 564                                            inode->i_mapping, start,
 565                                            pages,
 566                                            &nr_pages,
 567                                            &total_in,
 568                                            &total_compressed);
 569
 570                 if (!ret) {
 571                         unsigned long offset = offset_in_page(total_compressed);
 572                         struct page *page = pages[nr_pages - 1];
 573                         char *kaddr;
 574
 575                         /* zero the tail end of the last page, we might be
 576                          * sending it down to disk
 577                          */
 578                         if (offset) {
 579                                 kaddr = kmap_atomic(page);
 580                                 memset(kaddr + offset, 0,
 581                                        PAGE_SIZE - offset);
 582                                 kunmap_atomic(kaddr);
 583                         }
 584                         will_compress = 1;
 585                 }
 586         }
 587 cont:
 588         if (start == 0) {
 589                 /* lets try to make an inline extent */
 590                 if (ret || total_in < actual_end) {
 591                         /* we didn't compress the entire range, try
 592                          * to make an uncompressed inline extent.
 593                          */
 594                         ret = cow_file_range_inline(inode, start, end, 0,
 595                                                     BTRFS_COMPRESS_NONE, NULL);
 596                 } else {
 597                         /* try making a compressed inline extent */
 598                         ret = cow_file_range_inline(inode, start, end,
 599                                                     total_compressed,
 600                                                     compress_type, pages);
 601                 }
 602                 if (ret <= 0) {
 603                         unsigned long clear_flags = EXTENT_DELALLOC |
 604                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 605                                 EXTENT_DO_ACCOUNTING;
 606                         unsigned long page_error_op;
 607
 608                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 609
 610                         /*
 611                          * inline extent creation worked or returned error,
 612                          * we don't need to create any more async work items.
 613                          * Unlock and free up our temp pages.
 614                          *
 615                          * We use DO_ACCOUNTING here because we need the
 616                          * delalloc_release_metadata to be done _after_ we drop
 617                          * our outstanding extent for clearing delalloc for this
 618                          * range.
 619                          */
 620                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 621                                                      clear_flags,
 622                                                      PAGE_UNLOCK |
 623                                                      PAGE_CLEAR_DIRTY |
 624                                                      PAGE_SET_WRITEBACK |
 625                                                      page_error_op |
 626                                                      PAGE_END_WRITEBACK);
 627
 628                         for (i = 0; i < nr_pages; i++) {
 629                                 WARN_ON(pages[i]->mapping);
 630                                 put_page(pages[i]);
 631                         }
 632                         kfree(pages);
 633
 634                         return 0;
 635                 }
 636         }
 637
 638         if (will_compress) {
 639                 /*
 640                  * we aren't doing an inline extent round the compressed size
 641                  * up to a block size boundary so the allocator does sane
 642                  * things
 643                  */
 644                 total_compressed = ALIGN(total_compressed, blocksize);
 645
 646                 /*
 647                  * one last check to make sure the compression is really a
 648                  * win, compare the page count read with the blocks on disk,
 649                  * compression must free at least one sector size
 650                  */
 651                 total_in = ALIGN(total_in, PAGE_SIZE);
 652                 if (total_compressed + blocksize <= total_in) {
 653                         compressed_extents++;
 654
 655                         /*
 656                          * The async work queues will take care of doing actual
 657                          * allocation on disk for these compressed pages, and
 658                          * will submit them to the elevator.
 659                          */
 660                         add_async_extent(async_chunk, start, total_in,
 661                                         total_compressed, pages, nr_pages,
 662                                         compress_type);
 663
 664                         if (start + total_in < end) {
 665                                 start += total_in;
 666                                 pages = NULL;
 667                                 cond_resched();
 668                                 goto again;
 669                         }
 670                         return compressed_extents;
 671                 }
 672         }
 673         if (pages) {
 674                 /*
 675                  * the compression code ran but failed to make things smaller,
 676                  * free any pages it allocated and our page pointer array
 677                  */
 678                 for (i = 0; i < nr_pages; i++) {
 679                         WARN_ON(pages[i]->mapping);
 680                         put_page(pages[i]);
 681                 }
 682                 kfree(pages);
 683                 pages = NULL;
 684                 total_compressed = 0;
 685                 nr_pages = 0;
 686
 687                 /* flag the file so we don't compress in the future */
 688                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 689                     !(BTRFS_I(inode)->prop_compress)) {
 690                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 691                 }
 692         }
 693 cleanup_and_bail_uncompressed:
 694         /*
 695          * No compression, but we still need to write the pages in the file
 696          * we've been given so far.  redirty the locked page if it corresponds
 697          * to our extent and set things up for the async work queue to run
 698          * cow_file_range to do the normal delalloc dance.
 699          */
 700         if (page_offset(async_chunk->locked_page) >= start &&
 701             page_offset(async_chunk->locked_page) <= end)
 702                 __set_page_dirty_nobuffers(async_chunk->locked_page);
 703                 /* unlocked later on in the async handlers */
 704
 705         if (redirty)
 706                 extent_range_redirty_for_io(inode, start, end);
 707         add_async_extent(async_chunk, start, end - start + 1, 0, NULL, 0,
 708                          BTRFS_COMPRESS_NONE);
 709         compressed_extents++;
 710
 711         return compressed_extents;
 712 }
 713
 714 static void free_async_extent_pages(struct async_extent *async_extent)
 715 {
 716         int i;
 717
 718         if (!async_extent->pages)
 719                 return;
 720
 721         for (i = 0; i < async_extent->nr_pages; i++) {
 722                 WARN_ON(async_extent->pages[i]->mapping);
 723                 put_page(async_extent->pages[i]);
 724         }
 725         kfree(async_extent->pages);
 726         async_extent->nr_pages = 0;
 727         async_extent->pages = NULL;
 728 }
 729
 730 /*
 731  * phase two of compressed writeback.  This is the ordered portion
 732  * of the code, which only gets called in the order the work was
 733  * queued.  We walk all the async extents created by compress_file_range
 734  * and send them down to the disk.
 735  */
 736 static noinline void submit_compressed_extents(struct async_chunk *async_chunk)
 737 {
 738         struct inode *inode = async_chunk->inode;
 739         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 740         struct async_extent *async_extent;
 741         u64 alloc_hint = 0;
 742         struct btrfs_key ins;
 743         struct extent_map *em;
 744         struct btrfs_root *root = BTRFS_I(inode)->root;
 745         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 746         int ret = 0;
 747
 748 again:
 749         while (!list_empty(&async_chunk->extents)) {
 750                 async_extent = list_entry(async_chunk->extents.next,
 751                                           struct async_extent, list);
 752                 list_del(&async_extent->list);
 753
 754 retry:
 755                 lock_extent(io_tree, async_extent->start,
 756                             async_extent->start + async_extent->ram_size - 1);
 757                 /* did the compression code fall back to uncompressed IO? */
 758                 if (!async_extent->pages) {
 759                         int page_started = 0;
 760                         unsigned long nr_written = 0;
 761
 762                         /* allocate blocks */
 763                         ret = cow_file_range(inode, async_chunk->locked_page,
 764                                              async_extent->start,
 765                                              async_extent->start +
 766                                              async_extent->ram_size - 1,
 767                                              &page_started, &nr_written, 0);
 768
 769                         /* JDM XXX */
 770
 771                         /*
 772                          * if page_started, cow_file_range inserted an
 773                          * inline extent and took care of all the unlocking
 774                          * and IO for us.  Otherwise, we need to submit
 775                          * all those pages down to the drive.
 776                          */
 777                         if (!page_started && !ret)
 778                                 extent_write_locked_range(inode,
 779                                                   async_extent->start,
 780                                                   async_extent->start +
 781                                                   async_extent->ram_size - 1,
 782                                                   WB_SYNC_ALL);
 783                         else if (ret)
 784                                 unlock_page(async_chunk->locked_page);
 785                         kfree(async_extent);
 786                         cond_resched();
 787                         continue;
 788                 }
 789
 790                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 791                                            async_extent->compressed_size,
 792                                            async_extent->compressed_size,
 793                                            0, alloc_hint, &ins, 1, 1);
 794                 if (ret) {
 795                         free_async_extent_pages(async_extent);
 796
 797                         if (ret == -ENOSPC) {
 798                                 unlock_extent(io_tree, async_extent->start,
 799                                               async_extent->start +
 800                                               async_extent->ram_size - 1);
 801
 802                                 /*
 803                                  * we need to redirty the pages if we decide to
 804                                  * fallback to uncompressed IO, otherwise we
 805                                  * will not submit these pages down to lower
 806                                  * layers.
 807                                  */
 808                                 extent_range_redirty_for_io(inode,
 809                                                 async_extent->start,
 810                                                 async_extent->start +
 811                                                 async_extent->ram_size - 1);
 812
 813                                 goto retry;
 814                         }
 815                         goto out_free;
 816                 }
 817                 /*
 818                  * here we're doing allocation and writeback of the
 819                  * compressed pages
 820                  */
 821                 em = create_io_em(inode, async_extent->start,
 822                                   async_extent->ram_size, /* len */
 823                                   async_extent->start, /* orig_start */
 824                                   ins.objectid, /* block_start */
 825                                   ins.offset, /* block_len */
 826                                   ins.offset, /* orig_block_len */
 827                                   async_extent->ram_size, /* ram_bytes */
 828                                   async_extent->compress_type,
 829                                   BTRFS_ORDERED_COMPRESSED);
 830                 if (IS_ERR(em))
 831                         /* ret value is not necessary due to void function */
 832                         goto out_free_reserve;
 833                 free_extent_map(em);
 834
 835                 ret = btrfs_add_ordered_extent_compress(inode,
 836                                                 async_extent->start,
 837                                                 ins.objectid,
 838                                                 async_extent->ram_size,
 839                                                 ins.offset,
 840                                                 BTRFS_ORDERED_COMPRESSED,
 841                                                 async_extent->compress_type);
 842                 if (ret) {
 843                         btrfs_drop_extent_cache(BTRFS_I(inode),
 844                                                 async_extent->start,
 845                                                 async_extent->start +
 846                                                 async_extent->ram_size - 1, 0);
 847                         goto out_free_reserve;
 848                 }
 849                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 850
 851                 /*
 852                  * clear dirty, set writeback and unlock the pages.
 853                  */
 854                 extent_clear_unlock_delalloc(inode, async_extent->start,
 855                                 async_extent->start +
 856                                 async_extent->ram_size - 1,
 857                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 858                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 859                                 PAGE_SET_WRITEBACK);
 860                 if (btrfs_submit_compressed_write(inode,
 861                                     async_extent->start,
 862                                     async_extent->ram_size,
 863                                     ins.objectid,
 864                                     ins.offset, async_extent->pages,
 865                                     async_extent->nr_pages,
 866                                     async_chunk->write_flags)) {
 867                         struct page *p = async_extent->pages[0];
 868                         const u64 start = async_extent->start;
 869                         const u64 end = start + async_extent->ram_size - 1;
 870
 871                         p->mapping = inode->i_mapping;
 872                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
 873
 874                         p->mapping = NULL;
 875                         extent_clear_unlock_delalloc(inode, start, end,
 876                                                      NULL, 0,
 877                                                      PAGE_END_WRITEBACK |
 878                                                      PAGE_SET_ERROR);
 879                         free_async_extent_pages(async_extent);
 880                 }
 881                 alloc_hint = ins.objectid + ins.offset;
 882                 kfree(async_extent);
 883                 cond_resched();
 884         }
 885         return;
 886 out_free_reserve:
 887         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 888         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 889 out_free:
 890         extent_clear_unlock_delalloc(inode, async_extent->start,
 891                                      async_extent->start +
 892                                      async_extent->ram_size - 1,
 893                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 894                                      EXTENT_DELALLOC_NEW |
 895                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 896                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 897                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 898                                      PAGE_SET_ERROR);
 899         free_async_extent_pages(async_extent);
 900         kfree(async_extent);
 901         goto again;
 902 }
 903
 904 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 905                                       u64 num_bytes)
 906 {
 907         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 908         struct extent_map *em;
 909         u64 alloc_hint = 0;
 910
 911         read_lock(&em_tree->lock);
 912         em = search_extent_mapping(em_tree, start, num_bytes);
 913         if (em) {
 914                 /*
 915                  * if block start isn't an actual block number then find the
 916                  * first block in this inode and use that as a hint.  If that
 917                  * block is also bogus then just don't worry about it.
 918                  */
 919                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 920                         free_extent_map(em);
 921                         em = search_extent_mapping(em_tree, 0, 0);
 922                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 923                                 alloc_hint = em->block_start;
 924                         if (em)
 925                                 free_extent_map(em);
 926                 } else {
 927                         alloc_hint = em->block_start;
 928                         free_extent_map(em);
 929                 }
 930         }
 931         read_unlock(&em_tree->lock);
 932
 933         return alloc_hint;
 934 }
 935
 936 /*
 937  * when extent_io.c finds a delayed allocation range in the file,
 938  * the call backs end up in this code.  The basic idea is to
 939  * allocate extents on disk for the range, and create ordered data structs
 940  * in ram to track those extents.
 941  *
 942  * locked_page is the page that writepage had locked already.  We use
 943  * it to make sure we don't do extra locks or unlocks.
 944  *
 945  * *page_started is set to one if we unlock locked_page and do everything
 946  * required to start IO on it.  It may be clean and already done with
 947  * IO when we return.
 948  */
 949 static noinline int cow_file_range(struct inode *inode,
 950                                    struct page *locked_page,
 951                                    u64 start, u64 end, int *page_started,
 952                                    unsigned long *nr_written, int unlock)
 953 {
 954         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 955         struct btrfs_root *root = BTRFS_I(inode)->root;
 956         u64 alloc_hint = 0;
 957         u64 num_bytes;
 958         unsigned long ram_size;
 959         u64 cur_alloc_size = 0;
 960         u64 blocksize = fs_info->sectorsize;
 961         struct btrfs_key ins;
 962         struct extent_map *em;
 963         unsigned clear_bits;
 964         unsigned long page_ops;
 965         bool extent_reserved = false;
 966         int ret = 0;
 967
 968         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 969                 WARN_ON_ONCE(1);
 970                 ret = -EINVAL;
 971                 goto out_unlock;
 972         }
 973
 974         num_bytes = ALIGN(end - start + 1, blocksize);
 975         num_bytes = max(blocksize,  num_bytes);
 976         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 977
 978         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 979
 980         if (start == 0) {
 981                 /* lets try to make an inline extent */
 982                 ret = cow_file_range_inline(inode, start, end, 0,
 983                                             BTRFS_COMPRESS_NONE, NULL);
 984                 if (ret == 0) {
 985                         /*
 986                          * We use DO_ACCOUNTING here because we need the
 987                          * delalloc_release_metadata to be run _after_ we drop
 988                          * our outstanding extent for clearing delalloc for this
 989                          * range.
 990                          */
 991                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 992                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 993                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 994                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 995                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 996                                      PAGE_END_WRITEBACK);
 997                         *nr_written = *nr_written +
 998                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 999                         *page_started = 1;
1000                         goto out;
1001                 } else if (ret < 0) {
1002                         goto out_unlock;
1003                 }
1004         }
1005
1006         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1007         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1008                         start + num_bytes - 1, 0);
1009
1010         while (num_bytes > 0) {
1011                 cur_alloc_size = num_bytes;
1012                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1013                                            fs_info->sectorsize, 0, alloc_hint,
1014                                            &ins, 1, 1);
1015                 if (ret < 0)
1016                         goto out_unlock;
1017                 cur_alloc_size = ins.offset;
1018                 extent_reserved = true;
1019
1020                 ram_size = ins.offset;
1021                 em = create_io_em(inode, start, ins.offset, /* len */
1022                                   start, /* orig_start */
1023                                   ins.objectid, /* block_start */
1024                                   ins.offset, /* block_len */
1025                                   ins.offset, /* orig_block_len */
1026                                   ram_size, /* ram_bytes */
1027                                   BTRFS_COMPRESS_NONE, /* compress_type */
1028                                   BTRFS_ORDERED_REGULAR /* type */);
1029                 if (IS_ERR(em)) {
1030                         ret = PTR_ERR(em);
1031                         goto out_reserve;
1032                 }
1033                 free_extent_map(em);
1034
1035                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1036                                                ram_size, cur_alloc_size, 0);
1037                 if (ret)
1038                         goto out_drop_extent_cache;
1039
1040                 if (root->root_key.objectid ==
1041                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1042                         ret = btrfs_reloc_clone_csums(inode, start,
1043                                                       cur_alloc_size);
1044                         /*
1045                          * Only drop cache here, and process as normal.
1046                          *
1047                          * We must not allow extent_clear_unlock_delalloc()
1048                          * at out_unlock label to free meta of this ordered
1049                          * extent, as its meta should be freed by
1050                          * btrfs_finish_ordered_io().
1051                          *
1052                          * So we must continue until @start is increased to
1053                          * skip current ordered extent.
1054                          */
1055                         if (ret)
1056                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1057                                                 start + ram_size - 1, 0);
1058                 }
1059
1060                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1061
1062                 /* we're not doing compressed IO, don't unlock the first
1063                  * page (which the caller expects to stay locked), don't
1064                  * clear any dirty bits and don't set any writeback bits
1065                  *
1066                  * Do set the Private2 bit so we know this page was properly
1067                  * setup for writepage
1068                  */
1069                 page_ops = unlock ? PAGE_UNLOCK : 0;
1070                 page_ops |= PAGE_SET_PRIVATE2;
1071
1072                 extent_clear_unlock_delalloc(inode, start,
1073                                              start + ram_size - 1,
1074                                              locked_page,
1075                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1076                                              page_ops);
1077                 if (num_bytes < cur_alloc_size)
1078                         num_bytes = 0;
1079                 else
1080                         num_bytes -= cur_alloc_size;
1081                 alloc_hint = ins.objectid + ins.offset;
1082                 start += cur_alloc_size;
1083                 extent_reserved = false;
1084
1085                 /*
1086                  * btrfs_reloc_clone_csums() error, since start is increased
1087                  * extent_clear_unlock_delalloc() at out_unlock label won't
1088                  * free metadata of current ordered extent, we're OK to exit.
1089                  */
1090                 if (ret)
1091                         goto out_unlock;
1092         }
1093 out:
1094         return ret;
1095
1096 out_drop_extent_cache:
1097         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1098 out_reserve:
1099         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1100         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1101 out_unlock:
1102         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1103                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1104         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1105                 PAGE_END_WRITEBACK;
1106         /*
1107          * If we reserved an extent for our delalloc range (or a subrange) and
1108          * failed to create the respective ordered extent, then it means that
1109          * when we reserved the extent we decremented the extent's size from
1110          * the data space_info's bytes_may_use counter and incremented the
1111          * space_info's bytes_reserved counter by the same amount. We must make
1112          * sure extent_clear_unlock_delalloc() does not try to decrement again
1113          * the data space_info's bytes_may_use counter, therefore we do not pass
1114          * it the flag EXTENT_CLEAR_DATA_RESV.
1115          */
1116         if (extent_reserved) {
1117                 extent_clear_unlock_delalloc(inode, start,
1118                                              start + cur_alloc_size,
1119                                              locked_page,
1120                                              clear_bits,
1121                                              page_ops);
1122                 start += cur_alloc_size;
1123                 if (start >= end)
1124                         goto out;
1125         }
1126         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1127                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1128                                      page_ops);
1129         goto out;
1130 }
1131
1132 /*
1133  * work queue call back to started compression on a file and pages
1134  */
1135 static noinline void async_cow_start(struct btrfs_work *work)
1136 {
1137         struct async_chunk *async_chunk;
1138         int compressed_extents;
1139
1140         async_chunk = container_of(work, struct async_chunk, work);
1141
1142         compressed_extents = compress_file_range(async_chunk);
1143         if (compressed_extents == 0) {
1144                 btrfs_add_delayed_iput(async_chunk->inode);
1145                 async_chunk->inode = NULL;
1146         }
1147 }
1148
1149 /*
1150  * work queue call back to submit previously compressed pages
1151  */
1152 static noinline void async_cow_submit(struct btrfs_work *work)
1153 {
1154         struct async_chunk *async_chunk = container_of(work, struct async_chunk,
1155                                                      work);
1156         struct btrfs_fs_info *fs_info = btrfs_work_owner(work);
1157         unsigned long nr_pages;
1158
1159         nr_pages = (async_chunk->end - async_chunk->start + PAGE_SIZE) >>
1160                 PAGE_SHIFT;
1161
1162         /* atomic_sub_return implies a barrier */
1163         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1164             5 * SZ_1M)
1165                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1166
1167         /*
1168          * ->inode could be NULL if async_chunk_start has failed to compress,
1169          * in which case we don't have anything to submit, yet we need to
1170          * always adjust ->async_delalloc_pages as its paired with the init
1171          * happening in cow_file_range_async
1172          */
1173         if (async_chunk->inode)
1174                 submit_compressed_extents(async_chunk);
1175 }
1176
1177 static noinline void async_cow_free(struct btrfs_work *work)
1178 {
1179         struct async_chunk *async_chunk;
1180
1181         async_chunk = container_of(work, struct async_chunk, work);
1182         if (async_chunk->inode)
1183                 btrfs_add_delayed_iput(async_chunk->inode);
1184         /*
1185          * Since the pointer to 'pending' is at the beginning of the array of
1186          * async_chunk's, freeing it ensures the whole array has been freed.
1187          */
1188         if (atomic_dec_and_test(async_chunk->pending))
1189                 kvfree(async_chunk->pending);
1190 }
1191
1192 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1193                                 u64 start, u64 end, int *page_started,
1194                                 unsigned long *nr_written,
1195                                 unsigned int write_flags)
1196 {
1197         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1198         struct async_cow *ctx;
1199         struct async_chunk *async_chunk;
1200         unsigned long nr_pages;
1201         u64 cur_end;
1202         u64 num_chunks = DIV_ROUND_UP(end - start, SZ_512K);
1203         int i;
1204         bool should_compress;
1205         unsigned nofs_flag;
1206
1207         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1208
1209         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1210             !btrfs_test_opt(fs_info, FORCE_COMPRESS)) {
1211                 num_chunks = 1;
1212                 should_compress = false;
1213         } else {
1214                 should_compress = true;
1215         }
1216
1217         nofs_flag = memalloc_nofs_save();
1218         ctx = kvmalloc(struct_size(ctx, chunks, num_chunks), GFP_KERNEL);
1219         memalloc_nofs_restore(nofs_flag);
1220
1221         if (!ctx) {
1222                 unsigned clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC |
1223                         EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1224                         EXTENT_DO_ACCOUNTING;
1225                 unsigned long page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1226                         PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
1227                         PAGE_SET_ERROR;
1228
1229                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1230                                              clear_bits, page_ops);
1231                 return -ENOMEM;
1232         }
1233
1234         async_chunk = ctx->chunks;
1235         atomic_set(&ctx->num_chunks, num_chunks);
1236
1237         for (i = 0; i < num_chunks; i++) {
1238                 if (should_compress)
1239                         cur_end = min(end, start + SZ_512K - 1);
1240                 else
1241                         cur_end = end;
1242
1243                 /*
1244                  * igrab is called higher up in the call chain, take only the
1245                  * lightweight reference for the callback lifetime
1246                  */
1247                 ihold(inode);
1248                 async_chunk[i].pending = &ctx->num_chunks;
1249                 async_chunk[i].inode = inode;
1250                 async_chunk[i].start = start;
1251                 async_chunk[i].end = cur_end;
1252                 async_chunk[i].locked_page = locked_page;
1253                 async_chunk[i].write_flags = write_flags;
1254                 INIT_LIST_HEAD(&async_chunk[i].extents);
1255
1256                 btrfs_init_work(&async_chunk[i].work,
1257                                 btrfs_delalloc_helper,
1258                                 async_cow_start, async_cow_submit,
1259                                 async_cow_free);
1260
1261                 nr_pages = DIV_ROUND_UP(cur_end - start, PAGE_SIZE);
1262                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1263
1264                 btrfs_queue_work(fs_info->delalloc_workers, &async_chunk[i].work);
1265
1266                 *nr_written += nr_pages;
1267                 start = cur_end + 1;
1268         }
1269         *page_started = 1;
1270         return 0;
1271 }
1272
1273 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1274                                         u64 bytenr, u64 num_bytes)
1275 {
1276         int ret;
1277         struct btrfs_ordered_sum *sums;
1278         LIST_HEAD(list);
1279
1280         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1281                                        bytenr + num_bytes - 1, &list, 0);
1282         if (ret == 0 && list_empty(&list))
1283                 return 0;
1284
1285         while (!list_empty(&list)) {
1286                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1287                 list_del(&sums->list);
1288                 kfree(sums);
1289         }
1290         if (ret < 0)
1291                 return ret;
1292         return 1;
1293 }
1294
1295 /*
1296  * when nowcow writeback call back.  This checks for snapshots or COW copies
1297  * of the extents that exist in the file, and COWs the file as required.
1298  *
1299  * If no cow copies or snapshots exist, we write directly to the existing
1300  * blocks on disk
1301  */
1302 static noinline int run_delalloc_nocow(struct inode *inode,
1303                                        struct page *locked_page,
1304                                        const u64 start, const u64 end,
1305                                        int *page_started, int force,
1306                                        unsigned long *nr_written)
1307 {
1308         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1309         struct btrfs_root *root = BTRFS_I(inode)->root;
1310         struct btrfs_path *path;
1311         u64 cow_start = (u64)-1;
1312         u64 cur_offset = start;
1313         int ret;
1314         bool check_prev = true;
1315         const bool freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
1316         u64 ino = btrfs_ino(BTRFS_I(inode));
1317         bool nocow = false;
1318         u64 disk_bytenr = 0;
1319
1320         path = btrfs_alloc_path();
1321         if (!path) {
1322                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1323                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1324                                              EXTENT_DO_ACCOUNTING |
1325                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1326                                              PAGE_CLEAR_DIRTY |
1327                                              PAGE_SET_WRITEBACK |
1328                                              PAGE_END_WRITEBACK);
1329                 return -ENOMEM;
1330         }
1331
1332         while (1) {
1333                 struct btrfs_key found_key;
1334                 struct btrfs_file_extent_item *fi;
1335                 struct extent_buffer *leaf;
1336                 u64 extent_end;
1337                 u64 extent_offset;
1338                 u64 num_bytes = 0;
1339                 u64 disk_num_bytes;
1340                 u64 ram_bytes;
1341                 int extent_type;
1342
1343                 nocow = false;
1344
1345                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1346                                                cur_offset, 0);
1347                 if (ret < 0)
1348                         goto error;
1349
1350                 /*
1351                  * If there is no extent for our range when doing the initial
1352                  * search, then go back to the previous slot as it will be the
1353                  * one containing the search offset
1354                  */
1355                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1356                         leaf = path->nodes[0];
1357                         btrfs_item_key_to_cpu(leaf, &found_key,
1358                                               path->slots[0] - 1);
1359                         if (found_key.objectid == ino &&
1360                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1361                                 path->slots[0]--;
1362                 }
1363                 check_prev = false;
1364 next_slot:
1365                 /* Go to next leaf if we have exhausted the current one */
1366                 leaf = path->nodes[0];
1367                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1368                         ret = btrfs_next_leaf(root, path);
1369                         if (ret < 0) {
1370                                 if (cow_start != (u64)-1)
1371                                         cur_offset = cow_start;
1372                                 goto error;
1373                         }
1374                         if (ret > 0)
1375                                 break;
1376                         leaf = path->nodes[0];
1377                 }
1378
1379                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1380
1381                 /* Didn't find anything for our INO */
1382                 if (found_key.objectid > ino)
1383                         break;
1384                 /*
1385                  * Keep searching until we find an EXTENT_ITEM or there are no
1386                  * more extents for this inode
1387                  */
1388                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1389                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1390                         path->slots[0]++;
1391                         goto next_slot;
1392                 }
1393
1394                 /* Found key is not EXTENT_DATA_KEY or starts after req range */
1395                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1396                     found_key.offset > end)
1397                         break;
1398
1399                 /*
1400                  * If the found extent starts after requested offset, then
1401                  * adjust extent_end to be right before this extent begins
1402                  */
1403                 if (found_key.offset > cur_offset) {
1404                         extent_end = found_key.offset;
1405                         extent_type = 0;
1406                         goto out_check;
1407                 }
1408
1409                 /*
1410                  * Found extent which begins before our range and potentially
1411                  * intersect it
1412                  */
1413                 fi = btrfs_item_ptr(leaf, path->slots[0],
1414                                     struct btrfs_file_extent_item);
1415                 extent_type = btrfs_file_extent_type(leaf, fi);
1416
1417                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1418                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1419                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1420                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1421                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1422                         extent_end = found_key.offset +
1423                                 btrfs_file_extent_num_bytes(leaf, fi);
1424                         disk_num_bytes =
1425                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1426                         /*
1427                          * If extent we got ends before our range starts, skip
1428                          * to next extent
1429                          */
1430                         if (extent_end <= start) {
1431                                 path->slots[0]++;
1432                                 goto next_slot;
1433                         }
1434                         /* Skip holes */
1435                         if (disk_bytenr == 0)
1436                                 goto out_check;
1437                         /* Skip compressed/encrypted/encoded extents */
1438                         if (btrfs_file_extent_compression(leaf, fi) ||
1439                             btrfs_file_extent_encryption(leaf, fi) ||
1440                             btrfs_file_extent_other_encoding(leaf, fi))
1441                                 goto out_check;
1442                         /*
1443                          * If extent is created before the last volume's snapshot
1444                          * this implies the extent is shared, hence we can't do
1445                          * nocow. This is the same check as in
1446                          * btrfs_cross_ref_exist but without calling
1447                          * btrfs_search_slot.
1448                          */
1449                         if (!freespace_inode &&
1450                             btrfs_file_extent_generation(leaf, fi) <=
1451                             btrfs_root_last_snapshot(&root->root_item))
1452                                 goto out_check;
1453                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1454                                 goto out_check;
1455                         /* If extent is RO, we must COW it */
1456                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1457                                 goto out_check;
1458                         ret = btrfs_cross_ref_exist(root, ino,
1459                                                     found_key.offset -
1460                                                     extent_offset, disk_bytenr);
1461                         if (ret) {
1462                                 /*
1463                                  * ret could be -EIO if the above fails to read
1464                                  * metadata.
1465                                  */
1466                                 if (ret < 0) {
1467                                         if (cow_start != (u64)-1)
1468                                                 cur_offset = cow_start;
1469                                         goto error;
1470                                 }
1471
1472                                 WARN_ON_ONCE(freespace_inode);
1473                                 goto out_check;
1474                         }
1475                         disk_bytenr += extent_offset;
1476                         disk_bytenr += cur_offset - found_key.offset;
1477                         num_bytes = min(end + 1, extent_end) - cur_offset;
1478                         /*
1479                          * If there are pending snapshots for this root, we
1480                          * fall into common COW way
1481                          */
1482                         if (!freespace_inode && atomic_read(&root->snapshot_force_cow))
1483                                 goto out_check;
1484                         /*
1485                          * force cow if csum exists in the range.
1486                          * this ensure that csum for a given extent are
1487                          * either valid or do not exist.
1488                          */
1489                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1490                                                   num_bytes);
1491                         if (ret) {
1492                                 /*
1493                                  * ret could be -EIO if the above fails to read
1494                                  * metadata.
1495                                  */
1496                                 if (ret < 0) {
1497                                         if (cow_start != (u64)-1)
1498                                                 cur_offset = cow_start;
1499                                         goto error;
1500                                 }
1501                                 WARN_ON_ONCE(freespace_inode);
1502                                 goto out_check;
1503                         }
1504                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1505                                 goto out_check;
1506                         nocow = true;
1507                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1508                         extent_end = found_key.offset + ram_bytes;
1509                         extent_end = ALIGN(extent_end, fs_info->sectorsize);
1510                         /* Skip extents outside of our requested range */
1511                         if (extent_end <= start) {
1512                                 path->slots[0]++;
1513                                 goto next_slot;
1514                         }
1515                 } else {
1516                         /* If this triggers then we have a memory corruption */
1517                         BUG();
1518                 }
1519 out_check:
1520                 /*
1521                  * If nocow is false then record the beginning of the range
1522                  * that needs to be COWed
1523                  */
1524                 if (!nocow) {
1525                         if (cow_start == (u64)-1)
1526                                 cow_start = cur_offset;
1527                         cur_offset = extent_end;
1528                         if (cur_offset > end)
1529                                 break;
1530                         path->slots[0]++;
1531                         goto next_slot;
1532                 }
1533
1534                 btrfs_release_path(path);
1535
1536                 /*
1537                  * COW range from cow_start to found_key.offset - 1. As the key
1538                  * will contain the beginning of the first extent that can be
1539                  * NOCOW, following one which needs to be COW'ed
1540                  */
1541                 if (cow_start != (u64)-1) {
1542                         ret = cow_file_range(inode, locked_page,
1543                                              cow_start, found_key.offset - 1,
1544                                              page_started, nr_written, 1);
1545                         if (ret) {
1546                                 if (nocow)
1547                                         btrfs_dec_nocow_writers(fs_info,
1548                                                                 disk_bytenr);
1549                                 goto error;
1550                         }
1551                         cow_start = (u64)-1;
1552                 }
1553
1554                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1555                         u64 orig_start = found_key.offset - extent_offset;
1556                         struct extent_map *em;
1557
1558                         em = create_io_em(inode, cur_offset, num_bytes,
1559                                           orig_start,
1560                                           disk_bytenr, /* block_start */
1561                                           num_bytes, /* block_len */
1562                                           disk_num_bytes, /* orig_block_len */
1563                                           ram_bytes, BTRFS_COMPRESS_NONE,
1564                                           BTRFS_ORDERED_PREALLOC);
1565                         if (IS_ERR(em)) {
1566                                 if (nocow)
1567                                         btrfs_dec_nocow_writers(fs_info,
1568                                                                 disk_bytenr);
1569                                 ret = PTR_ERR(em);
1570                                 goto error;
1571                         }
1572                         free_extent_map(em);
1573                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1574                                                        disk_bytenr, num_bytes,
1575                                                        num_bytes,
1576                                                        BTRFS_ORDERED_PREALLOC);
1577                         if (ret) {
1578                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1579                                                         cur_offset,
1580                                                         cur_offset + num_bytes - 1,
1581                                                         0);
1582                                 goto error;
1583                         }
1584                 } else {
1585                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1586                                                        disk_bytenr, num_bytes,
1587                                                        num_bytes,
1588                                                        BTRFS_ORDERED_NOCOW);
1589                         if (ret)
1590                                 goto error;
1591                 }
1592
1593                 if (nocow)
1594                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1595                 nocow = false;
1596
1597                 if (root->root_key.objectid ==
1598                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1599                         /*
1600                          * Error handled later, as we must prevent
1601                          * extent_clear_unlock_delalloc() in error handler
1602                          * from freeing metadata of created ordered extent.
1603                          */
1604                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1605                                                       num_bytes);
1606
1607                 extent_clear_unlock_delalloc(inode, cur_offset,
1608                                              cur_offset + num_bytes - 1,
1609                                              locked_page, EXTENT_LOCKED |
1610                                              EXTENT_DELALLOC |
1611                                              EXTENT_CLEAR_DATA_RESV,
1612                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1613
1614                 cur_offset = extent_end;
1615
1616                 /*
1617                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1618                  * handler, as metadata for created ordered extent will only
1619                  * be freed by btrfs_finish_ordered_io().
1620                  */
1621                 if (ret)
1622                         goto error;
1623                 if (cur_offset > end)
1624                         break;
1625         }
1626         btrfs_release_path(path);
1627
1628         if (cur_offset <= end && cow_start == (u64)-1)
1629                 cow_start = cur_offset;
1630
1631         if (cow_start != (u64)-1) {
1632                 cur_offset = end;
1633                 ret = cow_file_range(inode, locked_page, cow_start, end,
1634                                      page_started, nr_written, 1);
1635                 if (ret)
1636                         goto error;
1637         }
1638
1639 error:
1640         if (nocow)
1641                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1642
1643         if (ret && cur_offset < end)
1644                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1645                                              locked_page, EXTENT_LOCKED |
1646                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1647                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1648                                              PAGE_CLEAR_DIRTY |
1649                                              PAGE_SET_WRITEBACK |
1650                                              PAGE_END_WRITEBACK);
1651         btrfs_free_path(path);
1652         return ret;
1653 }
1654
1655 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1656 {
1657
1658         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1659             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1660                 return 0;
1661
1662         /*
1663          * @defrag_bytes is a hint value, no spinlock held here,
1664          * if is not zero, it means the file is defragging.
1665          * Force cow if given extent needs to be defragged.
1666          */
1667         if (BTRFS_I(inode)->defrag_bytes &&
1668             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1669                            EXTENT_DEFRAG, 0, NULL))
1670                 return 1;
1671
1672         return 0;
1673 }
1674
1675 /*
1676  * Function to process delayed allocation (create CoW) for ranges which are
1677  * being touched for the first time.
1678  */
1679 int btrfs_run_delalloc_range(struct inode *inode, struct page *locked_page,
1680                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1681                 struct writeback_control *wbc)
1682 {
1683         int ret;
1684         int force_cow = need_force_cow(inode, start, end);
1685         unsigned int write_flags = wbc_to_write_flags(wbc);
1686
1687         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1688                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1689                                          page_started, 1, nr_written);
1690         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1691                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1692                                          page_started, 0, nr_written);
1693         } else if (!inode_can_compress(inode) ||
1694                    !inode_need_compress(inode, start, end)) {
1695                 ret = cow_file_range(inode, locked_page, start, end,
1696                                       page_started, nr_written, 1);
1697         } else {
1698                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1699                         &BTRFS_I(inode)->runtime_flags);
1700                 ret = cow_file_range_async(inode, locked_page, start, end,
1701                                            page_started, nr_written,
1702                                            write_flags);
1703         }
1704         if (ret)
1705                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1706                                               end - start + 1);
1707         return ret;
1708 }
1709
1710 void btrfs_split_delalloc_extent(struct inode *inode,
1711                                  struct extent_state *orig, u64 split)
1712 {
1713         u64 size;
1714
1715         /* not delalloc, ignore it */
1716         if (!(orig->state & EXTENT_DELALLOC))
1717                 return;
1718
1719         size = orig->end - orig->start + 1;
1720         if (size > BTRFS_MAX_EXTENT_SIZE) {
1721                 u32 num_extents;
1722                 u64 new_size;
1723
1724                 /*
1725                  * See the explanation in btrfs_merge_delalloc_extent, the same
1726                  * applies here, just in reverse.
1727                  */
1728                 new_size = orig->end - split + 1;
1729                 num_extents = count_max_extents(new_size);
1730                 new_size = split - orig->start;
1731                 num_extents += count_max_extents(new_size);
1732                 if (count_max_extents(size) >= num_extents)
1733                         return;
1734         }
1735
1736         spin_lock(&BTRFS_I(inode)->lock);
1737         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1738         spin_unlock(&BTRFS_I(inode)->lock);
1739 }
1740
1741 /*
1742  * Handle merged delayed allocation extents so we can keep track of new extents
1743  * that are just merged onto old extents, such as when we are doing sequential
1744  * writes, so we can properly account for the metadata space we'll need.
1745  */
1746 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1747                                  struct extent_state *other)
1748 {
1749         u64 new_size, old_size;
1750         u32 num_extents;
1751
1752         /* not delalloc, ignore it */
1753         if (!(other->state & EXTENT_DELALLOC))
1754                 return;
1755
1756         if (new->start > other->start)
1757                 new_size = new->end - other->start + 1;
1758         else
1759                 new_size = other->end - new->start + 1;
1760
1761         /* we're not bigger than the max, unreserve the space and go */
1762         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1763                 spin_lock(&BTRFS_I(inode)->lock);
1764                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1765                 spin_unlock(&BTRFS_I(inode)->lock);
1766                 return;
1767         }
1768
1769         /*
1770          * We have to add up either side to figure out how many extents were
1771          * accounted for before we merged into one big extent.  If the number of
1772          * extents we accounted for is <= the amount we need for the new range
1773          * then we can return, otherwise drop.  Think of it like this
1774          *
1775          * [ 4k][MAX_SIZE]
1776          *
1777          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1778          * need 2 outstanding extents, on one side we have 1 and the other side
1779          * we have 1 so they are == and we can return.  But in this case
1780          *
1781          * [MAX_SIZE+4k][MAX_SIZE+4k]
1782          *
1783          * Each range on their own accounts for 2 extents, but merged together
1784          * they are only 3 extents worth of accounting, so we need to drop in
1785          * this case.
1786          */
1787         old_size = other->end - other->start + 1;
1788         num_extents = count_max_extents(old_size);
1789         old_size = new->end - new->start + 1;
1790         num_extents += count_max_extents(old_size);
1791         if (count_max_extents(new_size) >= num_extents)
1792                 return;
1793
1794         spin_lock(&BTRFS_I(inode)->lock);
1795         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1796         spin_unlock(&BTRFS_I(inode)->lock);
1797 }
1798
1799 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1800                                       struct inode *inode)
1801 {
1802         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1803
1804         spin_lock(&root->delalloc_lock);
1805         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1806                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1807                               &root->delalloc_inodes);
1808                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1809                         &BTRFS_I(inode)->runtime_flags);
1810                 root->nr_delalloc_inodes++;
1811                 if (root->nr_delalloc_inodes == 1) {
1812                         spin_lock(&fs_info->delalloc_root_lock);
1813                         BUG_ON(!list_empty(&root->delalloc_root));
1814                         list_add_tail(&root->delalloc_root,
1815                                       &fs_info->delalloc_roots);
1816                         spin_unlock(&fs_info->delalloc_root_lock);
1817                 }
1818         }
1819         spin_unlock(&root->delalloc_lock);
1820 }
1821
1822
1823 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1824                                 struct btrfs_inode *inode)
1825 {
1826         struct btrfs_fs_info *fs_info = root->fs_info;
1827
1828         if (!list_empty(&inode->delalloc_inodes)) {
1829                 list_del_init(&inode->delalloc_inodes);
1830                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1831                           &inode->runtime_flags);
1832                 root->nr_delalloc_inodes--;
1833                 if (!root->nr_delalloc_inodes) {
1834                         ASSERT(list_empty(&root->delalloc_inodes));
1835                         spin_lock(&fs_info->delalloc_root_lock);
1836                         BUG_ON(list_empty(&root->delalloc_root));
1837                         list_del_init(&root->delalloc_root);
1838                         spin_unlock(&fs_info->delalloc_root_lock);
1839                 }
1840         }
1841 }
1842
1843 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1844                                      struct btrfs_inode *inode)
1845 {
1846         spin_lock(&root->delalloc_lock);
1847         __btrfs_del_delalloc_inode(root, inode);
1848         spin_unlock(&root->delalloc_lock);
1849 }
1850
1851 /*
1852  * Properly track delayed allocation bytes in the inode and to maintain the
1853  * list of inodes that have pending delalloc work to be done.
1854  */
1855 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1856                                unsigned *bits)
1857 {
1858         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1859
1860         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1861                 WARN_ON(1);
1862         /*
1863          * set_bit and clear bit hooks normally require _irqsave/restore
1864          * but in this case, we are only testing for the DELALLOC
1865          * bit, which is only set or cleared with irqs on
1866          */
1867         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1868                 struct btrfs_root *root = BTRFS_I(inode)->root;
1869                 u64 len = state->end + 1 - state->start;
1870                 u32 num_extents = count_max_extents(len);
1871                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1872
1873                 spin_lock(&BTRFS_I(inode)->lock);
1874                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1875                 spin_unlock(&BTRFS_I(inode)->lock);
1876
1877                 /* For sanity tests */
1878                 if (btrfs_is_testing(fs_info))
1879                         return;
1880
1881                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1882                                          fs_info->delalloc_batch);
1883                 spin_lock(&BTRFS_I(inode)->lock);
1884                 BTRFS_I(inode)->delalloc_bytes += len;
1885                 if (*bits & EXTENT_DEFRAG)
1886                         BTRFS_I(inode)->defrag_bytes += len;
1887                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1888                                          &BTRFS_I(inode)->runtime_flags))
1889                         btrfs_add_delalloc_inodes(root, inode);
1890                 spin_unlock(&BTRFS_I(inode)->lock);
1891         }
1892
1893         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1894             (*bits & EXTENT_DELALLOC_NEW)) {
1895                 spin_lock(&BTRFS_I(inode)->lock);
1896                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1897                         state->start;
1898                 spin_unlock(&BTRFS_I(inode)->lock);
1899         }
1900 }
1901
1902 /*
1903  * Once a range is no longer delalloc this function ensures that proper
1904  * accounting happens.
1905  */
1906 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
1907                                  struct extent_state *state, unsigned *bits)
1908 {
1909         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
1910         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
1911         u64 len = state->end + 1 - state->start;
1912         u32 num_extents = count_max_extents(len);
1913
1914         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1915                 spin_lock(&inode->lock);
1916                 inode->defrag_bytes -= len;
1917                 spin_unlock(&inode->lock);
1918         }
1919
1920         /*
1921          * set_bit and clear bit hooks normally require _irqsave/restore
1922          * but in this case, we are only testing for the DELALLOC
1923          * bit, which is only set or cleared with irqs on
1924          */
1925         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1926                 struct btrfs_root *root = inode->root;
1927                 bool do_list = !btrfs_is_free_space_inode(inode);
1928
1929                 spin_lock(&inode->lock);
1930                 btrfs_mod_outstanding_extents(inode, -num_extents);
1931                 spin_unlock(&inode->lock);
1932
1933                 /*
1934                  * We don't reserve metadata space for space cache inodes so we
1935                  * don't need to call delalloc_release_metadata if there is an
1936                  * error.
1937                  */
1938                 if (*bits & EXTENT_CLEAR_META_RESV &&
1939                     root != fs_info->tree_root)
1940                         btrfs_delalloc_release_metadata(inode, len, false);
1941
1942                 /* For sanity tests. */
1943                 if (btrfs_is_testing(fs_info))
1944                         return;
1945
1946                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1947                     do_list && !(state->state & EXTENT_NORESERVE) &&
1948                     (*bits & EXTENT_CLEAR_DATA_RESV))
1949                         btrfs_free_reserved_data_space_noquota(
1950                                         &inode->vfs_inode,
1951                                         state->start, len);
1952
1953                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1954                                          fs_info->delalloc_batch);
1955                 spin_lock(&inode->lock);
1956                 inode->delalloc_bytes -= len;
1957                 if (do_list && inode->delalloc_bytes == 0 &&
1958                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1959                                         &inode->runtime_flags))
1960                         btrfs_del_delalloc_inode(root, inode);
1961                 spin_unlock(&inode->lock);
1962         }
1963
1964         if ((state->state & EXTENT_DELALLOC_NEW) &&
1965             (*bits & EXTENT_DELALLOC_NEW)) {
1966                 spin_lock(&inode->lock);
1967                 ASSERT(inode->new_delalloc_bytes >= len);
1968                 inode->new_delalloc_bytes -= len;
1969                 spin_unlock(&inode->lock);
1970         }
1971 }
1972
1973 /*
1974  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
1975  * in a chunk's stripe. This function ensures that bios do not span a
1976  * stripe/chunk
1977  *
1978  * @page - The page we are about to add to the bio
1979  * @size - size we want to add to the bio
1980  * @bio - bio we want to ensure is smaller than a stripe
1981  * @bio_flags - flags of the bio
1982  *
1983  * return 1 if page cannot be added to the bio
1984  * return 0 if page can be added to the bio
1985  * return error otherwise
1986  */
1987 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
1988                              unsigned long bio_flags)
1989 {
1990         struct inode *inode = page->mapping->host;
1991         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1992         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1993         u64 length = 0;
1994         u64 map_length;
1995         int ret;
1996         struct btrfs_io_geometry geom;
1997
1998         if (bio_flags & EXTENT_BIO_COMPRESSED)
1999                 return 0;
2000
2001         length = bio->bi_iter.bi_size;
2002         map_length = length;
2003         ret = btrfs_get_io_geometry(fs_info, btrfs_op(bio), logical, map_length,
2004                                     &geom);
2005         if (ret < 0)
2006                 return ret;
2007
2008         if (geom.len < length + size)
2009                 return 1;
2010         return 0;
2011 }
2012
2013 /*
2014  * in order to insert checksums into the metadata in large chunks,
2015  * we wait until bio submission time.   All the pages in the bio are
2016  * checksummed and sums are attached onto the ordered extent record.
2017  *
2018  * At IO completion time the cums attached on the ordered extent record
2019  * are inserted into the btree
2020  */
2021 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
2022                                     u64 bio_offset)
2023 {
2024         struct inode *inode = private_data;
2025         blk_status_t ret = 0;
2026
2027         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2028         BUG_ON(ret); /* -ENOMEM */
2029         return 0;
2030 }
2031
2032 /*
2033  * extent_io.c submission hook. This does the right thing for csum calculation
2034  * on write, or reading the csums from the tree before a read.
2035  *
2036  * Rules about async/sync submit,
2037  * a) read:                             sync submit
2038  *
2039  * b) write without checksum:           sync submit
2040  *
2041  * c) write with checksum:
2042  *    c-1) if bio is issued by fsync:   sync submit
2043  *         (sync_writers != 0)
2044  *
2045  *    c-2) if root is reloc root:       sync submit
2046  *         (only in case of buffered IO)
2047  *
2048  *    c-3) otherwise:                   async submit
2049  */
2050 static blk_status_t btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
2051                                           int mirror_num,
2052                                           unsigned long bio_flags)
2053
2054 {
2055         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2056         struct btrfs_root *root = BTRFS_I(inode)->root;
2057         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2058         blk_status_t ret = 0;
2059         int skip_sum;
2060         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2061
2062         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2063
2064         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2065                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2066
2067         if (bio_op(bio) != REQ_OP_WRITE) {
2068                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2069                 if (ret)
2070                         goto out;
2071
2072                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2073                         ret = btrfs_submit_compressed_read(inode, bio,
2074                                                            mirror_num,
2075                                                            bio_flags);
2076                         goto out;
2077                 } else if (!skip_sum) {
2078                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2079                         if (ret)
2080                                 goto out;
2081                 }
2082                 goto mapit;
2083         } else if (async && !skip_sum) {
2084                 /* csum items have already been cloned */
2085                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2086                         goto mapit;
2087                 /* we're doing a write, do the async checksumming */
2088                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2089                                           0, inode, btrfs_submit_bio_start);
2090                 goto out;
2091         } else if (!skip_sum) {
2092                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2093                 if (ret)
2094                         goto out;
2095         }
2096
2097 mapit:
2098         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2099
2100 out:
2101         if (ret) {
2102                 bio->bi_status = ret;
2103                 bio_endio(bio);
2104         }
2105         return ret;
2106 }
2107
2108 /*
2109  * given a list of ordered sums record them in the inode.  This happens
2110  * at IO completion time based on sums calculated at bio submission time.
2111  */
2112 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2113                              struct inode *inode, struct list_head *list)
2114 {
2115         struct btrfs_ordered_sum *sum;
2116         int ret;
2117
2118         list_for_each_entry(sum, list, list) {
2119                 trans->adding_csums = true;
2120                 ret = btrfs_csum_file_blocks(trans,
2121                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2122                 trans->adding_csums = false;
2123                 if (ret)
2124                         return ret;
2125         }
2126         return 0;
2127 }
2128
2129 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2130                               unsigned int extra_bits,
2131                               struct extent_state **cached_state)
2132 {
2133         WARN_ON(PAGE_ALIGNED(end));
2134         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2135                                    extra_bits, cached_state);
2136 }
2137
2138 /* see btrfs_writepage_start_hook for details on why this is required */
2139 struct btrfs_writepage_fixup {
2140         struct page *page;
2141         struct btrfs_work work;
2142 };
2143
2144 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2145 {
2146         struct btrfs_writepage_fixup *fixup;
2147         struct btrfs_ordered_extent *ordered;
2148         struct extent_state *cached_state = NULL;
2149         struct extent_changeset *data_reserved = NULL;
2150         struct page *page;
2151         struct inode *inode;
2152         u64 page_start;
2153         u64 page_end;
2154         int ret;
2155
2156         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2157         page = fixup->page;
2158 again:
2159         lock_page(page);
2160         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2161                 ClearPageChecked(page);
2162                 goto out_page;
2163         }
2164
2165         inode = page->mapping->host;
2166         page_start = page_offset(page);
2167         page_end = page_offset(page) + PAGE_SIZE - 1;
2168
2169         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2170                          &cached_state);
2171
2172         /* already ordered? We're done */
2173         if (PagePrivate2(page))
2174                 goto out;
2175
2176         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2177                                         PAGE_SIZE);
2178         if (ordered) {
2179                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2180                                      page_end, &cached_state);
2181                 unlock_page(page);
2182                 btrfs_start_ordered_extent(inode, ordered, 1);
2183                 btrfs_put_ordered_extent(ordered);
2184                 goto again;
2185         }
2186
2187         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2188                                            PAGE_SIZE);
2189         if (ret) {
2190                 mapping_set_error(page->mapping, ret);
2191                 end_extent_writepage(page, ret, page_start, page_end);
2192                 ClearPageChecked(page);
2193                 goto out;
2194          }
2195
2196         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2197                                         &cached_state);
2198         if (ret) {
2199                 mapping_set_error(page->mapping, ret);
2200                 end_extent_writepage(page, ret, page_start, page_end);
2201                 ClearPageChecked(page);
2202                 goto out;
2203         }
2204
2205         ClearPageChecked(page);
2206         set_page_dirty(page);
2207         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, false);
2208 out:
2209         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2210                              &cached_state);
2211 out_page:
2212         unlock_page(page);
2213         put_page(page);
2214         kfree(fixup);
2215         extent_changeset_free(data_reserved);
2216 }
2217
2218 /*
2219  * There are a few paths in the higher layers of the kernel that directly
2220  * set the page dirty bit without asking the filesystem if it is a
2221  * good idea.  This causes problems because we want to make sure COW
2222  * properly happens and the data=ordered rules are followed.
2223  *
2224  * In our case any range that doesn't have the ORDERED bit set
2225  * hasn't been properly setup for IO.  We kick off an async process
2226  * to fix it up.  The async helper will wait for ordered extents, set
2227  * the delalloc bit and make it safe to write the page.
2228  */
2229 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2230 {
2231         struct inode *inode = page->mapping->host;
2232         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2233         struct btrfs_writepage_fixup *fixup;
2234
2235         /* this page is properly in the ordered list */
2236         if (TestClearPagePrivate2(page))
2237                 return 0;
2238
2239         if (PageChecked(page))
2240                 return -EAGAIN;
2241
2242         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2243         if (!fixup)
2244                 return -EAGAIN;
2245
2246         SetPageChecked(page);
2247         get_page(page);
2248         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2249                         btrfs_writepage_fixup_worker, NULL, NULL);
2250         fixup->page = page;
2251         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2252         return -EBUSY;
2253 }
2254
2255 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2256                                        struct inode *inode, u64 file_pos,
2257                                        u64 disk_bytenr, u64 disk_num_bytes,
2258                                        u64 num_bytes, u64 ram_bytes,
2259                                        u8 compression, u8 encryption,
2260                                        u16 other_encoding, int extent_type)
2261 {
2262         struct btrfs_root *root = BTRFS_I(inode)->root;
2263         struct btrfs_file_extent_item *fi;
2264         struct btrfs_path *path;
2265         struct extent_buffer *leaf;
2266         struct btrfs_key ins;
2267         u64 qg_released;
2268         int extent_inserted = 0;
2269         int ret;
2270
2271         path = btrfs_alloc_path();
2272         if (!path)
2273                 return -ENOMEM;
2274
2275         /*
2276          * we may be replacing one extent in the tree with another.
2277          * The new extent is pinned in the extent map, and we don't want
2278          * to drop it from the cache until it is completely in the btree.
2279          *
2280          * So, tell btrfs_drop_extents to leave this extent in the cache.
2281          * the caller is expected to unpin it and allow it to be merged
2282          * with the others.
2283          */
2284         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2285                                    file_pos + num_bytes, NULL, 0,
2286                                    1, sizeof(*fi), &extent_inserted);
2287         if (ret)
2288                 goto out;
2289
2290         if (!extent_inserted) {
2291                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2292                 ins.offset = file_pos;
2293                 ins.type = BTRFS_EXTENT_DATA_KEY;
2294
2295                 path->leave_spinning = 1;
2296                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2297                                               sizeof(*fi));
2298                 if (ret)
2299                         goto out;
2300         }
2301         leaf = path->nodes[0];
2302         fi = btrfs_item_ptr(leaf, path->slots[0],
2303                             struct btrfs_file_extent_item);
2304         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2305         btrfs_set_file_extent_type(leaf, fi, extent_type);
2306         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2307         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2308         btrfs_set_file_extent_offset(leaf, fi, 0);
2309         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2310         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2311         btrfs_set_file_extent_compression(leaf, fi, compression);
2312         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2313         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2314
2315         btrfs_mark_buffer_dirty(leaf);
2316         btrfs_release_path(path);
2317
2318         inode_add_bytes(inode, num_bytes);
2319
2320         ins.objectid = disk_bytenr;
2321         ins.offset = disk_num_bytes;
2322         ins.type = BTRFS_EXTENT_ITEM_KEY;
2323
2324         /*
2325          * Release the reserved range from inode dirty range map, as it is
2326          * already moved into delayed_ref_head
2327          */
2328         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2329         if (ret < 0)
2330                 goto out;
2331         qg_released = ret;
2332         ret = btrfs_alloc_reserved_file_extent(trans, root,
2333                                                btrfs_ino(BTRFS_I(inode)),
2334                                                file_pos, qg_released, &ins);
2335 out:
2336         btrfs_free_path(path);
2337
2338         return ret;
2339 }
2340
2341 /* snapshot-aware defrag */
2342 struct sa_defrag_extent_backref {
2343         struct rb_node node;
2344         struct old_sa_defrag_extent *old;
2345         u64 root_id;
2346         u64 inum;
2347         u64 file_pos;
2348         u64 extent_offset;
2349         u64 num_bytes;
2350         u64 generation;
2351 };
2352
2353 struct old_sa_defrag_extent {
2354         struct list_head list;
2355         struct new_sa_defrag_extent *new;
2356
2357         u64 extent_offset;
2358         u64 bytenr;
2359         u64 offset;
2360         u64 len;
2361         int count;
2362 };
2363
2364 struct new_sa_defrag_extent {
2365         struct rb_root root;
2366         struct list_head head;
2367         struct btrfs_path *path;
2368         struct inode *inode;
2369         u64 file_pos;
2370         u64 len;
2371         u64 bytenr;
2372         u64 disk_len;
2373         u8 compress_type;
2374 };
2375
2376 static int backref_comp(struct sa_defrag_extent_backref *b1,
2377                         struct sa_defrag_extent_backref *b2)
2378 {
2379         if (b1->root_id < b2->root_id)
2380                 return -1;
2381         else if (b1->root_id > b2->root_id)
2382                 return 1;
2383
2384         if (b1->inum < b2->inum)
2385                 return -1;
2386         else if (b1->inum > b2->inum)
2387                 return 1;
2388
2389         if (b1->file_pos < b2->file_pos)
2390                 return -1;
2391         else if (b1->file_pos > b2->file_pos)
2392                 return 1;
2393
2394         /*
2395          * [------------------------------] ===> (a range of space)
2396          *     |<--->|   |<---->| =============> (fs/file tree A)
2397          * |<---------------------------->| ===> (fs/file tree B)
2398          *
2399          * A range of space can refer to two file extents in one tree while
2400          * refer to only one file extent in another tree.
2401          *
2402          * So we may process a disk offset more than one time(two extents in A)
2403          * and locate at the same extent(one extent in B), then insert two same
2404          * backrefs(both refer to the extent in B).
2405          */
2406         return 0;
2407 }
2408
2409 static void backref_insert(struct rb_root *root,
2410                            struct sa_defrag_extent_backref *backref)
2411 {
2412         struct rb_node **p = &root->rb_node;
2413         struct rb_node *parent = NULL;
2414         struct sa_defrag_extent_backref *entry;
2415         int ret;
2416
2417         while (*p) {
2418                 parent = *p;
2419                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2420
2421                 ret = backref_comp(backref, entry);
2422                 if (ret < 0)
2423                         p = &(*p)->rb_left;
2424                 else
2425                         p = &(*p)->rb_right;
2426         }
2427
2428         rb_link_node(&backref->node, parent, p);
2429         rb_insert_color(&backref->node, root);
2430 }
2431
2432 /*
2433  * Note the backref might has changed, and in this case we just return 0.
2434  */
2435 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2436                                        void *ctx)
2437 {
2438         struct btrfs_file_extent_item *extent;
2439         struct old_sa_defrag_extent *old = ctx;
2440         struct new_sa_defrag_extent *new = old->new;
2441         struct btrfs_path *path = new->path;
2442         struct btrfs_key key;
2443         struct btrfs_root *root;
2444         struct sa_defrag_extent_backref *backref;
2445         struct extent_buffer *leaf;
2446         struct inode *inode = new->inode;
2447         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2448         int slot;
2449         int ret;
2450         u64 extent_offset;
2451         u64 num_bytes;
2452
2453         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2454             inum == btrfs_ino(BTRFS_I(inode)))
2455                 return 0;
2456
2457         key.objectid = root_id;
2458         key.type = BTRFS_ROOT_ITEM_KEY;
2459         key.offset = (u64)-1;
2460
2461         root = btrfs_read_fs_root_no_name(fs_info, &key);
2462         if (IS_ERR(root)) {
2463                 if (PTR_ERR(root) == -ENOENT)
2464                         return 0;
2465                 WARN_ON(1);
2466                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2467                          inum, offset, root_id);
2468                 return PTR_ERR(root);
2469         }
2470
2471         key.objectid = inum;
2472         key.type = BTRFS_EXTENT_DATA_KEY;
2473         if (offset > (u64)-1 << 32)
2474                 key.offset = 0;
2475         else
2476                 key.offset = offset;
2477
2478         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2479         if (WARN_ON(ret < 0))
2480                 return ret;
2481         ret = 0;
2482
2483         while (1) {
2484                 cond_resched();
2485
2486                 leaf = path->nodes[0];
2487                 slot = path->slots[0];
2488
2489                 if (slot >= btrfs_header_nritems(leaf)) {
2490                         ret = btrfs_next_leaf(root, path);
2491                         if (ret < 0) {
2492                                 goto out;
2493                         } else if (ret > 0) {
2494                                 ret = 0;
2495                                 goto out;
2496                         }
2497                         continue;
2498                 }
2499
2500                 path->slots[0]++;
2501
2502                 btrfs_item_key_to_cpu(leaf, &key, slot);
2503
2504                 if (key.objectid > inum)
2505                         goto out;
2506
2507                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2508                         continue;
2509
2510                 extent = btrfs_item_ptr(leaf, slot,
2511                                         struct btrfs_file_extent_item);
2512
2513                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2514                         continue;
2515
2516                 /*
2517                  * 'offset' refers to the exact key.offset,
2518                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2519                  * (key.offset - extent_offset).
2520                  */
2521                 if (key.offset != offset)
2522                         continue;
2523
2524                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2525                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2526
2527                 if (extent_offset >= old->extent_offset + old->offset +
2528                     old->len || extent_offset + num_bytes <=
2529                     old->extent_offset + old->offset)
2530                         continue;
2531                 break;
2532         }
2533
2534         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2535         if (!backref) {
2536                 ret = -ENOENT;
2537                 goto out;
2538         }
2539
2540         backref->root_id = root_id;
2541         backref->inum = inum;
2542         backref->file_pos = offset;
2543         backref->num_bytes = num_bytes;
2544         backref->extent_offset = extent_offset;
2545         backref->generation = btrfs_file_extent_generation(leaf, extent);
2546         backref->old = old;
2547         backref_insert(&new->root, backref);
2548         old->count++;
2549 out:
2550         btrfs_release_path(path);
2551         WARN_ON(ret);
2552         return ret;
2553 }
2554
2555 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2556                                    struct new_sa_defrag_extent *new)
2557 {
2558         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2559         struct old_sa_defrag_extent *old, *tmp;
2560         int ret;
2561
2562         new->path = path;
2563
2564         list_for_each_entry_safe(old, tmp, &new->head, list) {
2565                 ret = iterate_inodes_from_logical(old->bytenr +
2566                                                   old->extent_offset, fs_info,
2567                                                   path, record_one_backref,
2568                                                   old, false);
2569                 if (ret < 0 && ret != -ENOENT)
2570                         return false;
2571
2572                 /* no backref to be processed for this extent */
2573                 if (!old->count) {
2574                         list_del(&old->list);
2575                         kfree(old);
2576                 }
2577         }
2578
2579         if (list_empty(&new->head))
2580                 return false;
2581
2582         return true;
2583 }
2584
2585 static int relink_is_mergable(struct extent_buffer *leaf,
2586                               struct btrfs_file_extent_item *fi,
2587                               struct new_sa_defrag_extent *new)
2588 {
2589         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2590                 return 0;
2591
2592         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2593                 return 0;
2594
2595         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2596                 return 0;
2597
2598         if (btrfs_file_extent_encryption(leaf, fi) ||
2599             btrfs_file_extent_other_encoding(leaf, fi))
2600                 return 0;
2601
2602         return 1;
2603 }
2604
2605 /*
2606  * Note the backref might has changed, and in this case we just return 0.
2607  */
2608 static noinline int relink_extent_backref(struct btrfs_path *path,
2609                                  struct sa_defrag_extent_backref *prev,
2610                                  struct sa_defrag_extent_backref *backref)
2611 {
2612         struct btrfs_file_extent_item *extent;
2613         struct btrfs_file_extent_item *item;
2614         struct btrfs_ordered_extent *ordered;
2615         struct btrfs_trans_handle *trans;
2616         struct btrfs_ref ref = { 0 };
2617         struct btrfs_root *root;
2618         struct btrfs_key key;
2619         struct extent_buffer *leaf;
2620         struct old_sa_defrag_extent *old = backref->old;
2621         struct new_sa_defrag_extent *new = old->new;
2622         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2623         struct inode *inode;
2624         struct extent_state *cached = NULL;
2625         int ret = 0;
2626         u64 start;
2627         u64 len;
2628         u64 lock_start;
2629         u64 lock_end;
2630         bool merge = false;
2631         int index;
2632
2633         if (prev && prev->root_id == backref->root_id &&
2634             prev->inum == backref->inum &&
2635             prev->file_pos + prev->num_bytes == backref->file_pos)
2636                 merge = true;
2637
2638         /* step 1: get root */
2639         key.objectid = backref->root_id;
2640         key.type = BTRFS_ROOT_ITEM_KEY;
2641         key.offset = (u64)-1;
2642
2643         index = srcu_read_lock(&fs_info->subvol_srcu);
2644
2645         root = btrfs_read_fs_root_no_name(fs_info, &key);
2646         if (IS_ERR(root)) {
2647                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2648                 if (PTR_ERR(root) == -ENOENT)
2649                         return 0;
2650                 return PTR_ERR(root);
2651         }
2652
2653         if (btrfs_root_readonly(root)) {
2654                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2655                 return 0;
2656         }
2657
2658         /* step 2: get inode */
2659         key.objectid = backref->inum;
2660         key.type = BTRFS_INODE_ITEM_KEY;
2661         key.offset = 0;
2662
2663         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2664         if (IS_ERR(inode)) {
2665                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2666                 return 0;
2667         }
2668
2669         srcu_read_unlock(&fs_info->subvol_srcu, index);
2670
2671         /* step 3: relink backref */
2672         lock_start = backref->file_pos;
2673         lock_end = backref->file_pos + backref->num_bytes - 1;
2674         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2675                          &cached);
2676
2677         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2678         if (ordered) {
2679                 btrfs_put_ordered_extent(ordered);
2680                 goto out_unlock;
2681         }
2682
2683         trans = btrfs_join_transaction(root);
2684         if (IS_ERR(trans)) {
2685                 ret = PTR_ERR(trans);
2686                 goto out_unlock;
2687         }
2688
2689         key.objectid = backref->inum;
2690         key.type = BTRFS_EXTENT_DATA_KEY;
2691         key.offset = backref->file_pos;
2692
2693         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2694         if (ret < 0) {
2695                 goto out_free_path;
2696         } else if (ret > 0) {
2697                 ret = 0;
2698                 goto out_free_path;
2699         }
2700
2701         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2702                                 struct btrfs_file_extent_item);
2703
2704         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2705             backref->generation)
2706                 goto out_free_path;
2707
2708         btrfs_release_path(path);
2709
2710         start = backref->file_pos;
2711         if (backref->extent_offset < old->extent_offset + old->offset)
2712                 start += old->extent_offset + old->offset -
2713                          backref->extent_offset;
2714
2715         len = min(backref->extent_offset + backref->num_bytes,
2716                   old->extent_offset + old->offset + old->len);
2717         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2718
2719         ret = btrfs_drop_extents(trans, root, inode, start,
2720                                  start + len, 1);
2721         if (ret)
2722                 goto out_free_path;
2723 again:
2724         key.objectid = btrfs_ino(BTRFS_I(inode));
2725         key.type = BTRFS_EXTENT_DATA_KEY;
2726         key.offset = start;
2727
2728         path->leave_spinning = 1;
2729         if (merge) {
2730                 struct btrfs_file_extent_item *fi;
2731                 u64 extent_len;
2732                 struct btrfs_key found_key;
2733
2734                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2735                 if (ret < 0)
2736                         goto out_free_path;
2737
2738                 path->slots[0]--;
2739                 leaf = path->nodes[0];
2740                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2741
2742                 fi = btrfs_item_ptr(leaf, path->slots[0],
2743                                     struct btrfs_file_extent_item);
2744                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2745
2746                 if (extent_len + found_key.offset == start &&
2747                     relink_is_mergable(leaf, fi, new)) {
2748                         btrfs_set_file_extent_num_bytes(leaf, fi,
2749                                                         extent_len + len);
2750                         btrfs_mark_buffer_dirty(leaf);
2751                         inode_add_bytes(inode, len);
2752
2753                         ret = 1;
2754                         goto out_free_path;
2755                 } else {
2756                         merge = false;
2757                         btrfs_release_path(path);
2758                         goto again;
2759                 }
2760         }
2761
2762         ret = btrfs_insert_empty_item(trans, root, path, &key,
2763                                         sizeof(*extent));
2764         if (ret) {
2765                 btrfs_abort_transaction(trans, ret);
2766                 goto out_free_path;
2767         }
2768
2769         leaf = path->nodes[0];
2770         item = btrfs_item_ptr(leaf, path->slots[0],
2771                                 struct btrfs_file_extent_item);
2772         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2773         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2774         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2775         btrfs_set_file_extent_num_bytes(leaf, item, len);
2776         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2777         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2778         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2779         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2780         btrfs_set_file_extent_encryption(leaf, item, 0);
2781         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2782
2783         btrfs_mark_buffer_dirty(leaf);
2784         inode_add_bytes(inode, len);
2785         btrfs_release_path(path);
2786
2787         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new->bytenr,
2788                                new->disk_len, 0);
2789         btrfs_init_data_ref(&ref, backref->root_id, backref->inum,
2790                             new->file_pos);  /* start - extent_offset */
2791         ret = btrfs_inc_extent_ref(trans, &ref);
2792         if (ret) {
2793                 btrfs_abort_transaction(trans, ret);
2794                 goto out_free_path;
2795         }
2796
2797         ret = 1;
2798 out_free_path:
2799         btrfs_release_path(path);
2800         path->leave_spinning = 0;
2801         btrfs_end_transaction(trans);
2802 out_unlock:
2803         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2804                              &cached);
2805         iput(inode);
2806         return ret;
2807 }
2808
2809 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2810 {
2811         struct old_sa_defrag_extent *old, *tmp;
2812
2813         if (!new)
2814                 return;
2815
2816         list_for_each_entry_safe(old, tmp, &new->head, list) {
2817                 kfree(old);
2818         }
2819         kfree(new);
2820 }
2821
2822 static void relink_file_extents(struct new_sa_defrag_extent *new)
2823 {
2824         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2825         struct btrfs_path *path;
2826         struct sa_defrag_extent_backref *backref;
2827         struct sa_defrag_extent_backref *prev = NULL;
2828         struct rb_node *node;
2829         int ret;
2830
2831         path = btrfs_alloc_path();
2832         if (!path)
2833                 return;
2834
2835         if (!record_extent_backrefs(path, new)) {
2836                 btrfs_free_path(path);
2837                 goto out;
2838         }
2839         btrfs_release_path(path);
2840
2841         while (1) {
2842                 node = rb_first(&new->root);
2843                 if (!node)
2844                         break;
2845                 rb_erase(node, &new->root);
2846
2847                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2848
2849                 ret = relink_extent_backref(path, prev, backref);
2850                 WARN_ON(ret < 0);
2851
2852                 kfree(prev);
2853
2854                 if (ret == 1)
2855                         prev = backref;
2856                 else
2857                         prev = NULL;
2858                 cond_resched();
2859         }
2860         kfree(prev);
2861
2862         btrfs_free_path(path);
2863 out:
2864         free_sa_defrag_extent(new);
2865
2866         atomic_dec(&fs_info->defrag_running);
2867         wake_up(&fs_info->transaction_wait);
2868 }
2869
2870 static struct new_sa_defrag_extent *
2871 record_old_file_extents(struct inode *inode,
2872                         struct btrfs_ordered_extent *ordered)
2873 {
2874         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2875         struct btrfs_root *root = BTRFS_I(inode)->root;
2876         struct btrfs_path *path;
2877         struct btrfs_key key;
2878         struct old_sa_defrag_extent *old;
2879         struct new_sa_defrag_extent *new;
2880         int ret;
2881
2882         new = kmalloc(sizeof(*new), GFP_NOFS);
2883         if (!new)
2884                 return NULL;
2885
2886         new->inode = inode;
2887         new->file_pos = ordered->file_offset;
2888         new->len = ordered->len;
2889         new->bytenr = ordered->start;
2890         new->disk_len = ordered->disk_len;
2891         new->compress_type = ordered->compress_type;
2892         new->root = RB_ROOT;
2893         INIT_LIST_HEAD(&new->head);
2894
2895         path = btrfs_alloc_path();
2896         if (!path)
2897                 goto out_kfree;
2898
2899         key.objectid = btrfs_ino(BTRFS_I(inode));
2900         key.type = BTRFS_EXTENT_DATA_KEY;
2901         key.offset = new->file_pos;
2902
2903         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2904         if (ret < 0)
2905                 goto out_free_path;
2906         if (ret > 0 && path->slots[0] > 0)
2907                 path->slots[0]--;
2908
2909         /* find out all the old extents for the file range */
2910         while (1) {
2911                 struct btrfs_file_extent_item *extent;
2912                 struct extent_buffer *l;
2913                 int slot;
2914                 u64 num_bytes;
2915                 u64 offset;
2916                 u64 end;
2917                 u64 disk_bytenr;
2918                 u64 extent_offset;
2919
2920                 l = path->nodes[0];
2921                 slot = path->slots[0];
2922
2923                 if (slot >= btrfs_header_nritems(l)) {
2924                         ret = btrfs_next_leaf(root, path);
2925                         if (ret < 0)
2926                                 goto out_free_path;
2927                         else if (ret > 0)
2928                                 break;
2929                         continue;
2930                 }
2931
2932                 btrfs_item_key_to_cpu(l, &key, slot);
2933
2934                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2935                         break;
2936                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2937                         break;
2938                 if (key.offset >= new->file_pos + new->len)
2939                         break;
2940
2941                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2942
2943                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2944                 if (key.offset + num_bytes < new->file_pos)
2945                         goto next;
2946
2947                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2948                 if (!disk_bytenr)
2949                         goto next;
2950
2951                 extent_offset = btrfs_file_extent_offset(l, extent);
2952
2953                 old = kmalloc(sizeof(*old), GFP_NOFS);
2954                 if (!old)
2955                         goto out_free_path;
2956
2957                 offset = max(new->file_pos, key.offset);
2958                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2959
2960                 old->bytenr = disk_bytenr;
2961                 old->extent_offset = extent_offset;
2962                 old->offset = offset - key.offset;
2963                 old->len = end - offset;
2964                 old->new = new;
2965                 old->count = 0;
2966                 list_add_tail(&old->list, &new->head);
2967 next:
2968                 path->slots[0]++;
2969                 cond_resched();
2970         }
2971
2972         btrfs_free_path(path);
2973         atomic_inc(&fs_info->defrag_running);
2974
2975         return new;
2976
2977 out_free_path:
2978         btrfs_free_path(path);
2979 out_kfree:
2980         free_sa_defrag_extent(new);
2981         return NULL;
2982 }
2983
2984 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2985                                          u64 start, u64 len)
2986 {
2987         struct btrfs_block_group_cache *cache;
2988
2989         cache = btrfs_lookup_block_group(fs_info, start);
2990         ASSERT(cache);
2991
2992         spin_lock(&cache->lock);
2993         cache->delalloc_bytes -= len;
2994         spin_unlock(&cache->lock);
2995
2996         btrfs_put_block_group(cache);
2997 }
2998
2999 /* as ordered data IO finishes, this gets called so we can finish
3000  * an ordered extent if the range of bytes in the file it covers are
3001  * fully written.
3002  */
3003 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
3004 {
3005         struct inode *inode = ordered_extent->inode;
3006         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3007         struct btrfs_root *root = BTRFS_I(inode)->root;
3008         struct btrfs_trans_handle *trans = NULL;
3009         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3010         struct extent_state *cached_state = NULL;
3011         struct new_sa_defrag_extent *new = NULL;
3012         int compress_type = 0;
3013         int ret = 0;
3014         u64 logical_len = ordered_extent->len;
3015         bool nolock;
3016         bool truncated = false;
3017         bool range_locked = false;
3018         bool clear_new_delalloc_bytes = false;
3019         bool clear_reserved_extent = true;
3020
3021         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3022             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
3023             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
3024                 clear_new_delalloc_bytes = true;
3025
3026         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
3027
3028         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
3029                 ret = -EIO;
3030                 goto out;
3031         }
3032
3033         btrfs_free_io_failure_record(BTRFS_I(inode),
3034                         ordered_extent->file_offset,
3035                         ordered_extent->file_offset +
3036                         ordered_extent->len - 1);
3037
3038         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
3039                 truncated = true;
3040                 logical_len = ordered_extent->truncated_len;
3041                 /* Truncated the entire extent, don't bother adding */
3042                 if (!logical_len)
3043                         goto out;
3044         }
3045
3046         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
3047                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
3048
3049                 /*
3050                  * For mwrite(mmap + memset to write) case, we still reserve
3051                  * space for NOCOW range.
3052                  * As NOCOW won't cause a new delayed ref, just free the space
3053                  */
3054                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3055                                        ordered_extent->len);
3056                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3057                 if (nolock)
3058                         trans = btrfs_join_transaction_nolock(root);
3059                 else
3060                         trans = btrfs_join_transaction(root);
3061                 if (IS_ERR(trans)) {
3062                         ret = PTR_ERR(trans);
3063                         trans = NULL;
3064                         goto out;
3065                 }
3066                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3067                 ret = btrfs_update_inode_fallback(trans, root, inode);
3068                 if (ret) /* -ENOMEM or corruption */
3069                         btrfs_abort_transaction(trans, ret);
3070                 goto out;
3071         }
3072
3073         range_locked = true;
3074         lock_extent_bits(io_tree, ordered_extent->file_offset,
3075                          ordered_extent->file_offset + ordered_extent->len - 1,
3076                          &cached_state);
3077
3078         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3079                         ordered_extent->file_offset + ordered_extent->len - 1,
3080                         EXTENT_DEFRAG, 0, cached_state);
3081         if (ret) {
3082                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3083                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3084                         /* the inode is shared */
3085                         new = record_old_file_extents(inode, ordered_extent);
3086
3087                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3088                         ordered_extent->file_offset + ordered_extent->len - 1,
3089                         EXTENT_DEFRAG, 0, 0, &cached_state);
3090         }
3091
3092         if (nolock)
3093                 trans = btrfs_join_transaction_nolock(root);
3094         else
3095                 trans = btrfs_join_transaction(root);
3096         if (IS_ERR(trans)) {
3097                 ret = PTR_ERR(trans);
3098                 trans = NULL;
3099                 goto out;
3100         }
3101
3102         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3103
3104         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3105                 compress_type = ordered_extent->compress_type;
3106         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3107                 BUG_ON(compress_type);
3108                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3109                                        ordered_extent->len);
3110                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3111                                                 ordered_extent->file_offset,
3112                                                 ordered_extent->file_offset +
3113                                                 logical_len);
3114         } else {
3115                 BUG_ON(root == fs_info->tree_root);
3116                 ret = insert_reserved_file_extent(trans, inode,
3117                                                 ordered_extent->file_offset,
3118                                                 ordered_extent->start,
3119                                                 ordered_extent->disk_len,
3120                                                 logical_len, logical_len,
3121                                                 compress_type, 0, 0,
3122                                                 BTRFS_FILE_EXTENT_REG);
3123                 if (!ret) {
3124                         clear_reserved_extent = false;
3125                         btrfs_release_delalloc_bytes(fs_info,
3126                                                      ordered_extent->start,
3127                                                      ordered_extent->disk_len);
3128                 }
3129         }
3130         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3131                            ordered_extent->file_offset, ordered_extent->len,
3132                            trans->transid);
3133         if (ret < 0) {
3134                 btrfs_abort_transaction(trans, ret);
3135                 goto out;
3136         }
3137
3138         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3139         if (ret) {
3140                 btrfs_abort_transaction(trans, ret);
3141                 goto out;
3142         }
3143
3144         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3145         ret = btrfs_update_inode_fallback(trans, root, inode);
3146         if (ret) { /* -ENOMEM or corruption */
3147                 btrfs_abort_transaction(trans, ret);
3148                 goto out;
3149         }
3150         ret = 0;
3151 out:
3152         if (range_locked || clear_new_delalloc_bytes) {
3153                 unsigned int clear_bits = 0;
3154
3155                 if (range_locked)
3156                         clear_bits |= EXTENT_LOCKED;
3157                 if (clear_new_delalloc_bytes)
3158                         clear_bits |= EXTENT_DELALLOC_NEW;
3159                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3160                                  ordered_extent->file_offset,
3161                                  ordered_extent->file_offset +
3162                                  ordered_extent->len - 1,
3163                                  clear_bits,
3164                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3165                                  0, &cached_state);
3166         }
3167
3168         if (trans)
3169                 btrfs_end_transaction(trans);
3170
3171         if (ret || truncated) {
3172                 u64 start, end;
3173
3174                 if (truncated)
3175                         start = ordered_extent->file_offset + logical_len;
3176                 else
3177                         start = ordered_extent->file_offset;
3178                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3179                 clear_extent_uptodate(io_tree, start, end, NULL);
3180
3181                 /* Drop the cache for the part of the extent we didn't write. */
3182                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3183
3184                 /*
3185                  * If the ordered extent had an IOERR or something else went
3186                  * wrong we need to return the space for this ordered extent
3187                  * back to the allocator.  We only free the extent in the
3188                  * truncated case if we didn't write out the extent at all.
3189                  *
3190                  * If we made it past insert_reserved_file_extent before we
3191                  * errored out then we don't need to do this as the accounting
3192                  * has already been done.
3193                  */
3194                 if ((ret || !logical_len) &&
3195                     clear_reserved_extent &&
3196                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3197                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3198                         btrfs_free_reserved_extent(fs_info,
3199                                                    ordered_extent->start,
3200                                                    ordered_extent->disk_len, 1);
3201         }
3202
3203
3204         /*
3205          * This needs to be done to make sure anybody waiting knows we are done
3206          * updating everything for this ordered extent.
3207          */
3208         btrfs_remove_ordered_extent(inode, ordered_extent);
3209
3210         /* for snapshot-aware defrag */
3211         if (new) {
3212                 if (ret) {
3213                         free_sa_defrag_extent(new);
3214                         atomic_dec(&fs_info->defrag_running);
3215                 } else {
3216                         relink_file_extents(new);
3217                 }
3218         }
3219
3220         /* once for us */
3221         btrfs_put_ordered_extent(ordered_extent);
3222         /* once for the tree */
3223         btrfs_put_ordered_extent(ordered_extent);
3224
3225         return ret;
3226 }
3227
3228 static void finish_ordered_fn(struct btrfs_work *work)
3229 {
3230         struct btrfs_ordered_extent *ordered_extent;
3231         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3232         btrfs_finish_ordered_io(ordered_extent);
3233 }
3234
3235 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
3236                                           u64 end, int uptodate)
3237 {
3238         struct inode *inode = page->mapping->host;
3239         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3240         struct btrfs_ordered_extent *ordered_extent = NULL;
3241         struct btrfs_workqueue *wq;
3242         btrfs_work_func_t func;
3243
3244         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3245
3246         ClearPagePrivate2(page);
3247         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3248                                             end - start + 1, uptodate))
3249                 return;
3250
3251         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3252                 wq = fs_info->endio_freespace_worker;
3253                 func = btrfs_freespace_write_helper;
3254         } else {
3255                 wq = fs_info->endio_write_workers;
3256                 func = btrfs_endio_write_helper;
3257         }
3258
3259         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3260                         NULL);
3261         btrfs_queue_work(wq, &ordered_extent->work);
3262 }
3263
3264 static int __readpage_endio_check(struct inode *inode,
3265                                   struct btrfs_io_bio *io_bio,
3266                                   int icsum, struct page *page,
3267                                   int pgoff, u64 start, size_t len)
3268 {
3269         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3270         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
3271         char *kaddr;
3272         u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
3273         u8 *csum_expected;
3274         u8 csum[BTRFS_CSUM_SIZE];
3275
3276         csum_expected = ((u8 *)io_bio->csum) + icsum * csum_size;
3277
3278         kaddr = kmap_atomic(page);
3279         shash->tfm = fs_info->csum_shash;
3280
3281         crypto_shash_init(shash);
3282         crypto_shash_update(shash, kaddr + pgoff, len);
3283         crypto_shash_final(shash, csum);
3284
3285         if (memcmp(csum, csum_expected, csum_size))
3286                 goto zeroit;
3287
3288         kunmap_atomic(kaddr);
3289         return 0;
3290 zeroit:
3291         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3292                                     io_bio->mirror_num);
3293         memset(kaddr + pgoff, 1, len);
3294         flush_dcache_page(page);
3295         kunmap_atomic(kaddr);
3296         return -EIO;
3297 }
3298
3299 /*
3300  * when reads are done, we need to check csums to verify the data is correct
3301  * if there's a match, we allow the bio to finish.  If not, the code in
3302  * extent_io.c will try to find good copies for us.
3303  */
3304 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3305                                       u64 phy_offset, struct page *page,
3306                                       u64 start, u64 end, int mirror)
3307 {
3308         size_t offset = start - page_offset(page);
3309         struct inode *inode = page->mapping->host;
3310         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3311         struct btrfs_root *root = BTRFS_I(inode)->root;
3312
3313         if (PageChecked(page)) {
3314                 ClearPageChecked(page);
3315                 return 0;
3316         }
3317
3318         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3319                 return 0;
3320
3321         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3322             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3323                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3324                 return 0;
3325         }
3326
3327         phy_offset >>= inode->i_sb->s_blocksize_bits;
3328         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3329                                       start, (size_t)(end - start + 1));
3330 }
3331
3332 /*
3333  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3334  *
3335  * @inode: The inode we want to perform iput on
3336  *
3337  * This function uses the generic vfs_inode::i_count to track whether we should
3338  * just decrement it (in case it's > 1) or if this is the last iput then link
3339  * the inode to the delayed iput machinery. Delayed iputs are processed at
3340  * transaction commit time/superblock commit/cleaner kthread.
3341  */
3342 void btrfs_add_delayed_iput(struct inode *inode)
3343 {
3344         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3345         struct btrfs_inode *binode = BTRFS_I(inode);
3346
3347         if (atomic_add_unless(&inode->i_count, -1, 1))
3348                 return;
3349
3350         atomic_inc(&fs_info->nr_delayed_iputs);
3351         spin_lock(&fs_info->delayed_iput_lock);
3352         ASSERT(list_empty(&binode->delayed_iput));
3353         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3354         spin_unlock(&fs_info->delayed_iput_lock);
3355         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
3356                 wake_up_process(fs_info->cleaner_kthread);
3357 }
3358
3359 static void run_delayed_iput_locked(struct btrfs_fs_info *fs_info,
3360                                     struct btrfs_inode *inode)
3361 {
3362         list_del_init(&inode->delayed_iput);
3363         spin_unlock(&fs_info->delayed_iput_lock);
3364         iput(&inode->vfs_inode);
3365         if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
3366                 wake_up(&fs_info->delayed_iputs_wait);
3367         spin_lock(&fs_info->delayed_iput_lock);
3368 }
3369
3370 static void btrfs_run_delayed_iput(struct btrfs_fs_info *fs_info,
3371                                    struct btrfs_inode *inode)
3372 {
3373         if (!list_empty(&inode->delayed_iput)) {
3374                 spin_lock(&fs_info->delayed_iput_lock);
3375                 if (!list_empty(&inode->delayed_iput))
3376                         run_delayed_iput_locked(fs_info, inode);
3377                 spin_unlock(&fs_info->delayed_iput_lock);
3378         }
3379 }
3380
3381 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3382 {
3383
3384         spin_lock(&fs_info->delayed_iput_lock);
3385         while (!list_empty(&fs_info->delayed_iputs)) {
3386                 struct btrfs_inode *inode;
3387
3388                 inode = list_first_entry(&fs_info->delayed_iputs,
3389                                 struct btrfs_inode, delayed_iput);
3390                 run_delayed_iput_locked(fs_info, inode);
3391         }
3392         spin_unlock(&fs_info->delayed_iput_lock);
3393 }
3394
3395 /**
3396  * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
3397  * @fs_info - the fs_info for this fs
3398  * @return - EINTR if we were killed, 0 if nothing's pending
3399  *
3400  * This will wait on any delayed iputs that are currently running with KILLABLE
3401  * set.  Once they are all done running we will return, unless we are killed in
3402  * which case we return EINTR. This helps in user operations like fallocate etc
3403  * that might get blocked on the iputs.
3404  */
3405 int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
3406 {
3407         int ret = wait_event_killable(fs_info->delayed_iputs_wait,
3408                         atomic_read(&fs_info->nr_delayed_iputs) == 0);
3409         if (ret)
3410                 return -EINTR;
3411         return 0;
3412 }
3413
3414 /*
3415  * This creates an orphan entry for the given inode in case something goes wrong
3416  * in the middle of an unlink.
3417  */
3418 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3419                      struct btrfs_inode *inode)
3420 {
3421         int ret;
3422
3423         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3424         if (ret && ret != -EEXIST) {
3425                 btrfs_abort_transaction(trans, ret);
3426                 return ret;
3427         }
3428
3429         return 0;
3430 }
3431
3432 /*
3433  * We have done the delete so we can go ahead and remove the orphan item for
3434  * this particular inode.
3435  */
3436 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3437                             struct btrfs_inode *inode)
3438 {
3439         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3440 }
3441
3442 /*
3443  * this cleans up any orphans that may be left on the list from the last use
3444  * of this root.
3445  */
3446 int btrfs_orphan_cleanup(struct btrfs_root *root)
3447 {
3448         struct btrfs_fs_info *fs_info = root->fs_info;
3449         struct btrfs_path *path;
3450         struct extent_buffer *leaf;
3451         struct btrfs_key key, found_key;
3452         struct btrfs_trans_handle *trans;
3453         struct inode *inode;
3454         u64 last_objectid = 0;
3455         int ret = 0, nr_unlink = 0;
3456
3457         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3458                 return 0;
3459
3460         path = btrfs_alloc_path();
3461         if (!path) {
3462                 ret = -ENOMEM;
3463                 goto out;
3464         }
3465         path->reada = READA_BACK;
3466
3467         key.objectid = BTRFS_ORPHAN_OBJECTID;
3468         key.type = BTRFS_ORPHAN_ITEM_KEY;
3469         key.offset = (u64)-1;
3470
3471         while (1) {
3472                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3473                 if (ret < 0)
3474                         goto out;
3475
3476                 /*
3477                  * if ret == 0 means we found what we were searching for, which
3478                  * is weird, but possible, so only screw with path if we didn't
3479                  * find the key and see if we have stuff that matches
3480                  */
3481                 if (ret > 0) {
3482                         ret = 0;
3483                         if (path->slots[0] == 0)
3484                                 break;
3485                         path->slots[0]--;
3486                 }
3487
3488                 /* pull out the item */
3489                 leaf = path->nodes[0];
3490                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3491
3492                 /* make sure the item matches what we want */
3493                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3494                         break;
3495                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3496                         break;
3497
3498                 /* release the path since we're done with it */
3499                 btrfs_release_path(path);
3500
3501                 /*
3502                  * this is where we are basically btrfs_lookup, without the
3503                  * crossing root thing.  we store the inode number in the
3504                  * offset of the orphan item.
3505                  */
3506
3507                 if (found_key.offset == last_objectid) {
3508                         btrfs_err(fs_info,
3509                                   "Error removing orphan entry, stopping orphan cleanup");
3510                         ret = -EINVAL;
3511                         goto out;
3512                 }
3513
3514                 last_objectid = found_key.offset;
3515
3516                 found_key.objectid = found_key.offset;
3517                 found_key.type = BTRFS_INODE_ITEM_KEY;
3518                 found_key.offset = 0;
3519                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3520                 ret = PTR_ERR_OR_ZERO(inode);
3521                 if (ret && ret != -ENOENT)
3522                         goto out;
3523
3524                 if (ret == -ENOENT && root == fs_info->tree_root) {
3525                         struct btrfs_root *dead_root;
3526                         struct btrfs_fs_info *fs_info = root->fs_info;
3527                         int is_dead_root = 0;
3528
3529                         /*
3530                          * this is an orphan in the tree root. Currently these
3531                          * could come from 2 sources:
3532                          *  a) a snapshot deletion in progress
3533                          *  b) a free space cache inode
3534                          * We need to distinguish those two, as the snapshot
3535                          * orphan must not get deleted.
3536                          * find_dead_roots already ran before us, so if this
3537                          * is a snapshot deletion, we should find the root
3538                          * in the dead_roots list
3539                          */
3540                         spin_lock(&fs_info->trans_lock);
3541                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3542                                             root_list) {
3543                                 if (dead_root->root_key.objectid ==
3544                                     found_key.objectid) {
3545                                         is_dead_root = 1;
3546                                         break;
3547                                 }
3548                         }
3549                         spin_unlock(&fs_info->trans_lock);
3550                         if (is_dead_root) {
3551                                 /* prevent this orphan from being found again */
3552                                 key.offset = found_key.objectid - 1;
3553                                 continue;
3554                         }
3555
3556                 }
3557
3558                 /*
3559                  * If we have an inode with links, there are a couple of
3560                  * possibilities. Old kernels (before v3.12) used to create an
3561                  * orphan item for truncate indicating that there were possibly
3562                  * extent items past i_size that needed to be deleted. In v3.12,
3563                  * truncate was changed to update i_size in sync with the extent
3564                  * items, but the (useless) orphan item was still created. Since
3565                  * v4.18, we don't create the orphan item for truncate at all.
3566                  *
3567                  * So, this item could mean that we need to do a truncate, but
3568                  * only if this filesystem was last used on a pre-v3.12 kernel
3569                  * and was not cleanly unmounted. The odds of that are quite
3570                  * slim, and it's a pain to do the truncate now, so just delete
3571                  * the orphan item.
3572                  *
3573                  * It's also possible that this orphan item was supposed to be
3574                  * deleted but wasn't. The inode number may have been reused,
3575                  * but either way, we can delete the orphan item.
3576                  */
3577                 if (ret == -ENOENT || inode->i_nlink) {
3578                         if (!ret)
3579                                 iput(inode);
3580                         trans = btrfs_start_transaction(root, 1);
3581                         if (IS_ERR(trans)) {
3582                                 ret = PTR_ERR(trans);
3583                                 goto out;
3584                         }
3585                         btrfs_debug(fs_info, "auto deleting %Lu",
3586                                     found_key.objectid);
3587                         ret = btrfs_del_orphan_item(trans, root,
3588                                                     found_key.objectid);
3589                         btrfs_end_transaction(trans);
3590                         if (ret)
3591                                 goto out;
3592                         continue;
3593                 }
3594
3595                 nr_unlink++;
3596
3597                 /* this will do delete_inode and everything for us */
3598                 iput(inode);
3599         }
3600         /* release the path since we're done with it */
3601         btrfs_release_path(path);
3602
3603         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3604
3605         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3606                 trans = btrfs_join_transaction(root);
3607                 if (!IS_ERR(trans))
3608                         btrfs_end_transaction(trans);
3609         }
3610
3611         if (nr_unlink)
3612                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3613
3614 out:
3615         if (ret)
3616                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3617         btrfs_free_path(path);
3618         return ret;
3619 }
3620
3621 /*
3622  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3623  * don't find any xattrs, we know there can't be any acls.
3624  *
3625  * slot is the slot the inode is in, objectid is the objectid of the inode
3626  */
3627 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3628                                           int slot, u64 objectid,
3629                                           int *first_xattr_slot)
3630 {
3631         u32 nritems = btrfs_header_nritems(leaf);
3632         struct btrfs_key found_key;
3633         static u64 xattr_access = 0;
3634         static u64 xattr_default = 0;
3635         int scanned = 0;
3636
3637         if (!xattr_access) {
3638                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3639                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3640                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3641                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3642         }
3643
3644         slot++;
3645         *first_xattr_slot = -1;
3646         while (slot < nritems) {
3647                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3648
3649                 /* we found a different objectid, there must not be acls */
3650                 if (found_key.objectid != objectid)
3651                         return 0;
3652
3653                 /* we found an xattr, assume we've got an acl */
3654                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3655                         if (*first_xattr_slot == -1)
3656                                 *first_xattr_slot = slot;
3657                         if (found_key.offset == xattr_access ||
3658                             found_key.offset == xattr_default)
3659                                 return 1;
3660                 }
3661
3662                 /*
3663                  * we found a key greater than an xattr key, there can't
3664                  * be any acls later on
3665                  */
3666                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3667                         return 0;
3668
3669                 slot++;
3670                 scanned++;
3671
3672                 /*
3673                  * it goes inode, inode backrefs, xattrs, extents,
3674                  * so if there are a ton of hard links to an inode there can
3675                  * be a lot of backrefs.  Don't waste time searching too hard,
3676                  * this is just an optimization
3677                  */
3678                 if (scanned >= 8)
3679                         break;
3680         }
3681         /* we hit the end of the leaf before we found an xattr or
3682          * something larger than an xattr.  We have to assume the inode
3683          * has acls
3684          */
3685         if (*first_xattr_slot == -1)
3686                 *first_xattr_slot = slot;
3687         return 1;
3688 }
3689
3690 /*
3691  * read an inode from the btree into the in-memory inode
3692  */
3693 static int btrfs_read_locked_inode(struct inode *inode,
3694                                    struct btrfs_path *in_path)
3695 {
3696         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3697         struct btrfs_path *path = in_path;
3698         struct extent_buffer *leaf;
3699         struct btrfs_inode_item *inode_item;
3700         struct btrfs_root *root = BTRFS_I(inode)->root;
3701         struct btrfs_key location;
3702         unsigned long ptr;
3703         int maybe_acls;
3704         u32 rdev;
3705         int ret;
3706         bool filled = false;
3707         int first_xattr_slot;
3708
3709         ret = btrfs_fill_inode(inode, &rdev);
3710         if (!ret)
3711                 filled = true;
3712
3713         if (!path) {
3714                 path = btrfs_alloc_path();
3715                 if (!path)
3716                         return -ENOMEM;
3717         }
3718
3719         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3720
3721         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3722         if (ret) {
3723                 if (path != in_path)
3724                         btrfs_free_path(path);
3725                 return ret;
3726         }
3727
3728         leaf = path->nodes[0];
3729
3730         if (filled)
3731                 goto cache_index;
3732
3733         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3734                                     struct btrfs_inode_item);
3735         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3736         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3737         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3738         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3739         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3740
3741         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3742         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3743
3744         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3745         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3746
3747         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3748         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3749
3750         BTRFS_I(inode)->i_otime.tv_sec =
3751                 btrfs_timespec_sec(leaf, &inode_item->otime);
3752         BTRFS_I(inode)->i_otime.tv_nsec =
3753                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3754
3755         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3756         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3757         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3758
3759         inode_set_iversion_queried(inode,
3760                                    btrfs_inode_sequence(leaf, inode_item));
3761         inode->i_generation = BTRFS_I(inode)->generation;
3762         inode->i_rdev = 0;
3763         rdev = btrfs_inode_rdev(leaf, inode_item);
3764
3765         BTRFS_I(inode)->index_cnt = (u64)-1;
3766         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3767
3768 cache_index:
3769         /*
3770          * If we were modified in the current generation and evicted from memory
3771          * and then re-read we need to do a full sync since we don't have any
3772          * idea about which extents were modified before we were evicted from
3773          * cache.
3774          *
3775          * This is required for both inode re-read from disk and delayed inode
3776          * in delayed_nodes_tree.
3777          */
3778         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3779                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3780                         &BTRFS_I(inode)->runtime_flags);
3781
3782         /*
3783          * We don't persist the id of the transaction where an unlink operation
3784          * against the inode was last made. So here we assume the inode might
3785          * have been evicted, and therefore the exact value of last_unlink_trans
3786          * lost, and set it to last_trans to avoid metadata inconsistencies
3787          * between the inode and its parent if the inode is fsync'ed and the log
3788          * replayed. For example, in the scenario:
3789          *
3790          * touch mydir/foo
3791          * ln mydir/foo mydir/bar
3792          * sync
3793          * unlink mydir/bar
3794          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3795          * xfs_io -c fsync mydir/foo
3796          * <power failure>
3797          * mount fs, triggers fsync log replay
3798          *
3799          * We must make sure that when we fsync our inode foo we also log its
3800          * parent inode, otherwise after log replay the parent still has the
3801          * dentry with the "bar" name but our inode foo has a link count of 1
3802          * and doesn't have an inode ref with the name "bar" anymore.
3803          *
3804          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3805          * but it guarantees correctness at the expense of occasional full
3806          * transaction commits on fsync if our inode is a directory, or if our
3807          * inode is not a directory, logging its parent unnecessarily.
3808          */
3809         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3810
3811         path->slots[0]++;
3812         if (inode->i_nlink != 1 ||
3813             path->slots[0] >= btrfs_header_nritems(leaf))
3814                 goto cache_acl;
3815
3816         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3817         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3818                 goto cache_acl;
3819
3820         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3821         if (location.type == BTRFS_INODE_REF_KEY) {
3822                 struct btrfs_inode_ref *ref;
3823
3824                 ref = (struct btrfs_inode_ref *)ptr;
3825                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3826         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3827                 struct btrfs_inode_extref *extref;
3828
3829                 extref = (struct btrfs_inode_extref *)ptr;
3830                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3831                                                                      extref);
3832         }
3833 cache_acl:
3834         /*
3835          * try to precache a NULL acl entry for files that don't have
3836          * any xattrs or acls
3837          */
3838         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3839                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3840         if (first_xattr_slot != -1) {
3841                 path->slots[0] = first_xattr_slot;
3842                 ret = btrfs_load_inode_props(inode, path);
3843                 if (ret)
3844                         btrfs_err(fs_info,
3845                                   "error loading props for ino %llu (root %llu): %d",
3846                                   btrfs_ino(BTRFS_I(inode)),
3847                                   root->root_key.objectid, ret);
3848         }
3849         if (path != in_path)
3850                 btrfs_free_path(path);
3851
3852         if (!maybe_acls)
3853                 cache_no_acl(inode);
3854
3855         switch (inode->i_mode & S_IFMT) {
3856         case S_IFREG:
3857                 inode->i_mapping->a_ops = &btrfs_aops;
3858                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3859                 inode->i_fop = &btrfs_file_operations;
3860                 inode->i_op = &btrfs_file_inode_operations;
3861                 break;
3862         case S_IFDIR:
3863                 inode->i_fop = &btrfs_dir_file_operations;
3864                 inode->i_op = &btrfs_dir_inode_operations;
3865                 break;
3866         case S_IFLNK:
3867                 inode->i_op = &btrfs_symlink_inode_operations;
3868                 inode_nohighmem(inode);
3869                 inode->i_mapping->a_ops = &btrfs_aops;
3870                 break;
3871         default:
3872                 inode->i_op = &btrfs_special_inode_operations;
3873                 init_special_inode(inode, inode->i_mode, rdev);
3874                 break;
3875         }
3876
3877         btrfs_sync_inode_flags_to_i_flags(inode);
3878         return 0;
3879 }
3880
3881 /*
3882  * given a leaf and an inode, copy the inode fields into the leaf
3883  */
3884 static void fill_inode_item(struct btrfs_trans_handle *trans,
3885                             struct extent_buffer *leaf,
3886                             struct btrfs_inode_item *item,
3887                             struct inode *inode)
3888 {
3889         struct btrfs_map_token token;
3890
3891         btrfs_init_map_token(&token);
3892
3893         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3894         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3895         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3896                                    &token);
3897         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3898         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3899
3900         btrfs_set_token_timespec_sec(leaf, &item->atime,
3901                                      inode->i_atime.tv_sec, &token);
3902         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3903                                       inode->i_atime.tv_nsec, &token);
3904
3905         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3906                                      inode->i_mtime.tv_sec, &token);
3907         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3908                                       inode->i_mtime.tv_nsec, &token);
3909
3910         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3911                                      inode->i_ctime.tv_sec, &token);
3912         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3913                                       inode->i_ctime.tv_nsec, &token);
3914
3915         btrfs_set_token_timespec_sec(leaf, &item->otime,
3916                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3917         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3918                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3919
3920         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3921                                      &token);
3922         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3923                                          &token);
3924         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3925                                        &token);
3926         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3927         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3928         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3929         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3930 }
3931
3932 /*
3933  * copy everything in the in-memory inode into the btree.
3934  */
3935 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3936                                 struct btrfs_root *root, struct inode *inode)
3937 {
3938         struct btrfs_inode_item *inode_item;
3939         struct btrfs_path *path;
3940         struct extent_buffer *leaf;
3941         int ret;
3942
3943         path = btrfs_alloc_path();
3944         if (!path)
3945                 return -ENOMEM;
3946
3947         path->leave_spinning = 1;
3948         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3949                                  1);
3950         if (ret) {
3951                 if (ret > 0)
3952                         ret = -ENOENT;
3953                 goto failed;
3954         }
3955
3956         leaf = path->nodes[0];
3957         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3958                                     struct btrfs_inode_item);
3959
3960         fill_inode_item(trans, leaf, inode_item, inode);
3961         btrfs_mark_buffer_dirty(leaf);
3962         btrfs_set_inode_last_trans(trans, inode);
3963         ret = 0;
3964 failed:
3965         btrfs_free_path(path);
3966         return ret;
3967 }
3968
3969 /*
3970  * copy everything in the in-memory inode into the btree.
3971  */
3972 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3973                                 struct btrfs_root *root, struct inode *inode)
3974 {
3975         struct btrfs_fs_info *fs_info = root->fs_info;
3976         int ret;
3977
3978         /*
3979          * If the inode is a free space inode, we can deadlock during commit
3980          * if we put it into the delayed code.
3981          *
3982          * The data relocation inode should also be directly updated
3983          * without delay
3984          */
3985         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3986             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3987             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3988                 btrfs_update_root_times(trans, root);
3989
3990                 ret = btrfs_delayed_update_inode(trans, root, inode);
3991                 if (!ret)
3992                         btrfs_set_inode_last_trans(trans, inode);
3993                 return ret;
3994         }
3995
3996         return btrfs_update_inode_item(trans, root, inode);
3997 }
3998
3999 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4000                                          struct btrfs_root *root,
4001                                          struct inode *inode)
4002 {
4003         int ret;
4004
4005         ret = btrfs_update_inode(trans, root, inode);
4006         if (ret == -ENOSPC)
4007                 return btrfs_update_inode_item(trans, root, inode);
4008         return ret;
4009 }
4010
4011 /*
4012  * unlink helper that gets used here in inode.c and in the tree logging
4013  * recovery code.  It remove a link in a directory with a given name, and
4014  * also drops the back refs in the inode to the directory
4015  */
4016 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4017                                 struct btrfs_root *root,
4018                                 struct btrfs_inode *dir,
4019                                 struct btrfs_inode *inode,
4020                                 const char *name, int name_len)
4021 {
4022         struct btrfs_fs_info *fs_info = root->fs_info;
4023         struct btrfs_path *path;
4024         int ret = 0;
4025         struct btrfs_dir_item *di;
4026         u64 index;
4027         u64 ino = btrfs_ino(inode);
4028         u64 dir_ino = btrfs_ino(dir);
4029
4030         path = btrfs_alloc_path();
4031         if (!path) {
4032                 ret = -ENOMEM;
4033                 goto out;
4034         }
4035
4036         path->leave_spinning = 1;
4037         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4038                                     name, name_len, -1);
4039         if (IS_ERR_OR_NULL(di)) {
4040                 ret = di ? PTR_ERR(di) : -ENOENT;
4041                 goto err;
4042         }
4043         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4044         if (ret)
4045                 goto err;
4046         btrfs_release_path(path);
4047
4048         /*
4049          * If we don't have dir index, we have to get it by looking up
4050          * the inode ref, since we get the inode ref, remove it directly,
4051          * it is unnecessary to do delayed deletion.
4052          *
4053          * But if we have dir index, needn't search inode ref to get it.
4054          * Since the inode ref is close to the inode item, it is better
4055          * that we delay to delete it, and just do this deletion when
4056          * we update the inode item.
4057          */
4058         if (inode->dir_index) {
4059                 ret = btrfs_delayed_delete_inode_ref(inode);
4060                 if (!ret) {
4061                         index = inode->dir_index;
4062                         goto skip_backref;
4063                 }
4064         }
4065
4066         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4067                                   dir_ino, &index);
4068         if (ret) {
4069                 btrfs_info(fs_info,
4070                         "failed to delete reference to %.*s, inode %llu parent %llu",
4071                         name_len, name, ino, dir_ino);
4072                 btrfs_abort_transaction(trans, ret);
4073                 goto err;
4074         }
4075 skip_backref:
4076         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
4077         if (ret) {
4078                 btrfs_abort_transaction(trans, ret);
4079                 goto err;
4080         }
4081
4082         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4083                         dir_ino);
4084         if (ret != 0 && ret != -ENOENT) {
4085                 btrfs_abort_transaction(trans, ret);
4086                 goto err;
4087         }
4088
4089         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4090                         index);
4091         if (ret == -ENOENT)
4092                 ret = 0;
4093         else if (ret)
4094                 btrfs_abort_transaction(trans, ret);
4095
4096         /*
4097          * If we have a pending delayed iput we could end up with the final iput
4098          * being run in btrfs-cleaner context.  If we have enough of these built
4099          * up we can end up burning a lot of time in btrfs-cleaner without any
4100          * way to throttle the unlinks.  Since we're currently holding a ref on
4101          * the inode we can run the delayed iput here without any issues as the
4102          * final iput won't be done until after we drop the ref we're currently
4103          * holding.
4104          */
4105         btrfs_run_delayed_iput(fs_info, inode);
4106 err:
4107         btrfs_free_path(path);
4108         if (ret)
4109                 goto out;
4110
4111         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4112         inode_inc_iversion(&inode->vfs_inode);
4113         inode_inc_iversion(&dir->vfs_inode);
4114         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4115                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4116         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4117 out:
4118         return ret;
4119 }
4120
4121 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4122                        struct btrfs_root *root,
4123                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4124                        const char *name, int name_len)
4125 {
4126         int ret;
4127         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4128         if (!ret) {
4129                 drop_nlink(&inode->vfs_inode);
4130                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4131         }
4132         return ret;
4133 }
4134
4135 /*
4136  * helper to start transaction for unlink and rmdir.
4137  *
4138  * unlink and rmdir are special in btrfs, they do not always free space, so
4139  * if we cannot make our reservations the normal way try and see if there is
4140  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4141  * allow the unlink to occur.
4142  */
4143 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4144 {
4145         struct btrfs_root *root = BTRFS_I(dir)->root;
4146
4147         /*
4148          * 1 for the possible orphan item
4149          * 1 for the dir item
4150          * 1 for the dir index
4151          * 1 for the inode ref
4152          * 1 for the inode
4153          */
4154         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4155 }
4156
4157 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4158 {
4159         struct btrfs_root *root = BTRFS_I(dir)->root;
4160         struct btrfs_trans_handle *trans;
4161         struct inode *inode = d_inode(dentry);
4162         int ret;
4163
4164         trans = __unlink_start_trans(dir);
4165         if (IS_ERR(trans))
4166                 return PTR_ERR(trans);
4167
4168         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4169                         0);
4170
4171         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4172                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4173                         dentry->d_name.len);
4174         if (ret)
4175                 goto out;
4176
4177         if (inode->i_nlink == 0) {
4178                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4179                 if (ret)
4180                         goto out;
4181         }
4182
4183 out:
4184         btrfs_end_transaction(trans);
4185         btrfs_btree_balance_dirty(root->fs_info);
4186         return ret;
4187 }
4188
4189 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4190                                struct inode *dir, u64 objectid,
4191                                const char *name, int name_len)
4192 {
4193         struct btrfs_root *root = BTRFS_I(dir)->root;
4194         struct btrfs_path *path;
4195         struct extent_buffer *leaf;
4196         struct btrfs_dir_item *di;
4197         struct btrfs_key key;
4198         u64 index;
4199         int ret;
4200         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4201
4202         path = btrfs_alloc_path();
4203         if (!path)
4204                 return -ENOMEM;
4205
4206         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4207                                    name, name_len, -1);
4208         if (IS_ERR_OR_NULL(di)) {
4209                 ret = di ? PTR_ERR(di) : -ENOENT;
4210                 goto out;
4211         }
4212
4213         leaf = path->nodes[0];
4214         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4215         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4216         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4217         if (ret) {
4218                 btrfs_abort_transaction(trans, ret);
4219                 goto out;
4220         }
4221         btrfs_release_path(path);
4222
4223         ret = btrfs_del_root_ref(trans, objectid, root->root_key.objectid,
4224                                  dir_ino, &index, name, name_len);
4225         if (ret < 0) {
4226                 if (ret != -ENOENT) {
4227                         btrfs_abort_transaction(trans, ret);
4228                         goto out;
4229                 }
4230                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4231                                                  name, name_len);
4232                 if (IS_ERR_OR_NULL(di)) {
4233                         if (!di)
4234                                 ret = -ENOENT;
4235                         else
4236                                 ret = PTR_ERR(di);
4237                         btrfs_abort_transaction(trans, ret);
4238                         goto out;
4239                 }
4240
4241                 leaf = path->nodes[0];
4242                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4243                 index = key.offset;
4244         }
4245         btrfs_release_path(path);
4246
4247         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4248         if (ret) {
4249                 btrfs_abort_transaction(trans, ret);
4250                 goto out;
4251         }
4252
4253         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4254         inode_inc_iversion(dir);
4255         dir->i_mtime = dir->i_ctime = current_time(dir);
4256         ret = btrfs_update_inode_fallback(trans, root, dir);
4257         if (ret)
4258                 btrfs_abort_transaction(trans, ret);
4259 out:
4260         btrfs_free_path(path);
4261         return ret;
4262 }
4263
4264 /*
4265  * Helper to check if the subvolume references other subvolumes or if it's
4266  * default.
4267  */
4268 static noinline int may_destroy_subvol(struct btrfs_root *root)
4269 {
4270         struct btrfs_fs_info *fs_info = root->fs_info;
4271         struct btrfs_path *path;
4272         struct btrfs_dir_item *di;
4273         struct btrfs_key key;
4274         u64 dir_id;
4275         int ret;
4276
4277         path = btrfs_alloc_path();
4278         if (!path)
4279                 return -ENOMEM;
4280
4281         /* Make sure this root isn't set as the default subvol */
4282         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4283         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4284                                    dir_id, "default", 7, 0);
4285         if (di && !IS_ERR(di)) {
4286                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4287                 if (key.objectid == root->root_key.objectid) {
4288                         ret = -EPERM;
4289                         btrfs_err(fs_info,
4290                                   "deleting default subvolume %llu is not allowed",
4291                                   key.objectid);
4292                         goto out;
4293                 }
4294                 btrfs_release_path(path);
4295         }
4296
4297         key.objectid = root->root_key.objectid;
4298         key.type = BTRFS_ROOT_REF_KEY;
4299         key.offset = (u64)-1;
4300
4301         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4302         if (ret < 0)
4303                 goto out;
4304         BUG_ON(ret == 0);
4305
4306         ret = 0;
4307         if (path->slots[0] > 0) {
4308                 path->slots[0]--;
4309                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4310                 if (key.objectid == root->root_key.objectid &&
4311                     key.type == BTRFS_ROOT_REF_KEY)
4312                         ret = -ENOTEMPTY;
4313         }
4314 out:
4315         btrfs_free_path(path);
4316         return ret;
4317 }
4318
4319 /* Delete all dentries for inodes belonging to the root */
4320 static void btrfs_prune_dentries(struct btrfs_root *root)
4321 {
4322         struct btrfs_fs_info *fs_info = root->fs_info;
4323         struct rb_node *node;
4324         struct rb_node *prev;
4325         struct btrfs_inode *entry;
4326         struct inode *inode;
4327         u64 objectid = 0;
4328
4329         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4330                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4331
4332         spin_lock(&root->inode_lock);
4333 again:
4334         node = root->inode_tree.rb_node;
4335         prev = NULL;
4336         while (node) {
4337                 prev = node;
4338                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4339
4340                 if (objectid < btrfs_ino(entry))
4341                         node = node->rb_left;
4342                 else if (objectid > btrfs_ino(entry))
4343                         node = node->rb_right;
4344                 else
4345                         break;
4346         }
4347         if (!node) {
4348                 while (prev) {
4349                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4350                         if (objectid <= btrfs_ino(entry)) {
4351                                 node = prev;
4352                                 break;
4353                         }
4354                         prev = rb_next(prev);
4355                 }
4356         }
4357         while (node) {
4358                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4359                 objectid = btrfs_ino(entry) + 1;
4360                 inode = igrab(&entry->vfs_inode);
4361                 if (inode) {
4362                         spin_unlock(&root->inode_lock);
4363                         if (atomic_read(&inode->i_count) > 1)
4364                                 d_prune_aliases(inode);
4365                         /*
4366                          * btrfs_drop_inode will have it removed from the inode
4367                          * cache when its usage count hits zero.
4368                          */
4369                         iput(inode);
4370                         cond_resched();
4371                         spin_lock(&root->inode_lock);
4372                         goto again;
4373                 }
4374
4375                 if (cond_resched_lock(&root->inode_lock))
4376                         goto again;
4377
4378                 node = rb_next(node);
4379         }
4380         spin_unlock(&root->inode_lock);
4381 }
4382
4383 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4384 {
4385         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4386         struct btrfs_root *root = BTRFS_I(dir)->root;
4387         struct inode *inode = d_inode(dentry);
4388         struct btrfs_root *dest = BTRFS_I(inode)->root;
4389         struct btrfs_trans_handle *trans;
4390         struct btrfs_block_rsv block_rsv;
4391         u64 root_flags;
4392         int ret;
4393         int err;
4394
4395         /*
4396          * Don't allow to delete a subvolume with send in progress. This is
4397          * inside the inode lock so the error handling that has to drop the bit
4398          * again is not run concurrently.
4399          */
4400         spin_lock(&dest->root_item_lock);
4401         if (dest->send_in_progress) {
4402                 spin_unlock(&dest->root_item_lock);
4403                 btrfs_warn(fs_info,
4404                            "attempt to delete subvolume %llu during send",
4405                            dest->root_key.objectid);
4406                 return -EPERM;
4407         }
4408         root_flags = btrfs_root_flags(&dest->root_item);
4409         btrfs_set_root_flags(&dest->root_item,
4410                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4411         spin_unlock(&dest->root_item_lock);
4412
4413         down_write(&fs_info->subvol_sem);
4414
4415         err = may_destroy_subvol(dest);
4416         if (err)
4417                 goto out_up_write;
4418
4419         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4420         /*
4421          * One for dir inode,
4422          * two for dir entries,
4423          * two for root ref/backref.
4424          */
4425         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4426         if (err)
4427                 goto out_up_write;
4428
4429         trans = btrfs_start_transaction(root, 0);
4430         if (IS_ERR(trans)) {
4431                 err = PTR_ERR(trans);
4432                 goto out_release;
4433         }
4434         trans->block_rsv = &block_rsv;
4435         trans->bytes_reserved = block_rsv.size;
4436
4437         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4438
4439         ret = btrfs_unlink_subvol(trans, dir, dest->root_key.objectid,
4440                                   dentry->d_name.name, dentry->d_name.len);
4441         if (ret) {
4442                 err = ret;
4443                 btrfs_abort_transaction(trans, ret);
4444                 goto out_end_trans;
4445         }
4446
4447         btrfs_record_root_in_trans(trans, dest);
4448
4449         memset(&dest->root_item.drop_progress, 0,
4450                 sizeof(dest->root_item.drop_progress));
4451         dest->root_item.drop_level = 0;
4452         btrfs_set_root_refs(&dest->root_item, 0);
4453
4454         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4455                 ret = btrfs_insert_orphan_item(trans,
4456                                         fs_info->tree_root,
4457                                         dest->root_key.objectid);
4458                 if (ret) {
4459                         btrfs_abort_transaction(trans, ret);
4460                         err = ret;
4461                         goto out_end_trans;
4462                 }
4463         }
4464
4465         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4466                                   BTRFS_UUID_KEY_SUBVOL,
4467                                   dest->root_key.objectid);
4468         if (ret && ret != -ENOENT) {
4469                 btrfs_abort_transaction(trans, ret);
4470                 err = ret;
4471                 goto out_end_trans;
4472         }
4473         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4474                 ret = btrfs_uuid_tree_remove(trans,
4475                                           dest->root_item.received_uuid,
4476                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4477                                           dest->root_key.objectid);
4478                 if (ret && ret != -ENOENT) {
4479                         btrfs_abort_transaction(trans, ret);
4480                         err = ret;
4481                         goto out_end_trans;
4482                 }
4483         }
4484
4485 out_end_trans:
4486         trans->block_rsv = NULL;
4487         trans->bytes_reserved = 0;
4488         ret = btrfs_end_transaction(trans);
4489         if (ret && !err)
4490                 err = ret;
4491         inode->i_flags |= S_DEAD;
4492 out_release:
4493         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4494 out_up_write:
4495         up_write(&fs_info->subvol_sem);
4496         if (err) {
4497                 spin_lock(&dest->root_item_lock);
4498                 root_flags = btrfs_root_flags(&dest->root_item);
4499                 btrfs_set_root_flags(&dest->root_item,
4500                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4501                 spin_unlock(&dest->root_item_lock);
4502         } else {
4503                 d_invalidate(dentry);
4504                 btrfs_prune_dentries(dest);
4505                 ASSERT(dest->send_in_progress == 0);
4506
4507                 /* the last ref */
4508                 if (dest->ino_cache_inode) {
4509                         iput(dest->ino_cache_inode);
4510                         dest->ino_cache_inode = NULL;
4511                 }
4512         }
4513
4514         return err;
4515 }
4516
4517 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4518 {
4519         struct inode *inode = d_inode(dentry);
4520         int err = 0;
4521         struct btrfs_root *root = BTRFS_I(dir)->root;
4522         struct btrfs_trans_handle *trans;
4523         u64 last_unlink_trans;
4524
4525         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4526                 return -ENOTEMPTY;
4527         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4528                 return btrfs_delete_subvolume(dir, dentry);
4529
4530         trans = __unlink_start_trans(dir);
4531         if (IS_ERR(trans))
4532                 return PTR_ERR(trans);
4533
4534         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4535                 err = btrfs_unlink_subvol(trans, dir,
4536                                           BTRFS_I(inode)->location.objectid,
4537                                           dentry->d_name.name,
4538                                           dentry->d_name.len);
4539                 goto out;
4540         }
4541
4542         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4543         if (err)
4544                 goto out;
4545
4546         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4547
4548         /* now the directory is empty */
4549         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4550                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4551                         dentry->d_name.len);
4552         if (!err) {
4553                 btrfs_i_size_write(BTRFS_I(inode), 0);
4554                 /*
4555                  * Propagate the last_unlink_trans value of the deleted dir to
4556                  * its parent directory. This is to prevent an unrecoverable
4557                  * log tree in the case we do something like this:
4558                  * 1) create dir foo
4559                  * 2) create snapshot under dir foo
4560                  * 3) delete the snapshot
4561                  * 4) rmdir foo
4562                  * 5) mkdir foo
4563                  * 6) fsync foo or some file inside foo
4564                  */
4565                 if (last_unlink_trans >= trans->transid)
4566                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4567         }
4568 out:
4569         btrfs_end_transaction(trans);
4570         btrfs_btree_balance_dirty(root->fs_info);
4571
4572         return err;
4573 }
4574
4575 /*
4576  * Return this if we need to call truncate_block for the last bit of the
4577  * truncate.
4578  */
4579 #define NEED_TRUNCATE_BLOCK 1
4580
4581 /*
4582  * this can truncate away extent items, csum items and directory items.
4583  * It starts at a high offset and removes keys until it can't find
4584  * any higher than new_size
4585  *
4586  * csum items that cross the new i_size are truncated to the new size
4587  * as well.
4588  *
4589  * min_type is the minimum key type to truncate down to.  If set to 0, this
4590  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4591  */
4592 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4593                                struct btrfs_root *root,
4594                                struct inode *inode,
4595                                u64 new_size, u32 min_type)
4596 {
4597         struct btrfs_fs_info *fs_info = root->fs_info;
4598         struct btrfs_path *path;
4599         struct extent_buffer *leaf;
4600         struct btrfs_file_extent_item *fi;
4601         struct btrfs_key key;
4602         struct btrfs_key found_key;
4603         u64 extent_start = 0;
4604         u64 extent_num_bytes = 0;
4605         u64 extent_offset = 0;
4606         u64 item_end = 0;
4607         u64 last_size = new_size;
4608         u32 found_type = (u8)-1;
4609         int found_extent;
4610         int del_item;
4611         int pending_del_nr = 0;
4612         int pending_del_slot = 0;
4613         int extent_type = -1;
4614         int ret;
4615         u64 ino = btrfs_ino(BTRFS_I(inode));
4616         u64 bytes_deleted = 0;
4617         bool be_nice = false;
4618         bool should_throttle = false;
4619
4620         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4621
4622         /*
4623          * for non-free space inodes and ref cows, we want to back off from
4624          * time to time
4625          */
4626         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4627             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4628                 be_nice = true;
4629
4630         path = btrfs_alloc_path();
4631         if (!path)
4632                 return -ENOMEM;
4633         path->reada = READA_BACK;
4634
4635         /*
4636          * We want to drop from the next block forward in case this new size is
4637          * not block aligned since we will be keeping the last block of the
4638          * extent just the way it is.
4639          */
4640         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4641             root == fs_info->tree_root)
4642                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4643                                         fs_info->sectorsize),
4644                                         (u64)-1, 0);
4645
4646         /*
4647          * This function is also used to drop the items in the log tree before
4648          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4649          * it is used to drop the logged items. So we shouldn't kill the delayed
4650          * items.
4651          */
4652         if (min_type == 0 && root == BTRFS_I(inode)->root)
4653                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4654
4655         key.objectid = ino;
4656         key.offset = (u64)-1;
4657         key.type = (u8)-1;
4658
4659 search_again:
4660         /*
4661          * with a 16K leaf size and 128MB extents, you can actually queue
4662          * up a huge file in a single leaf.  Most of the time that
4663          * bytes_deleted is > 0, it will be huge by the time we get here
4664          */
4665         if (be_nice && bytes_deleted > SZ_32M &&
4666             btrfs_should_end_transaction(trans)) {
4667                 ret = -EAGAIN;
4668                 goto out;
4669         }
4670
4671         path->leave_spinning = 1;
4672         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4673         if (ret < 0)
4674                 goto out;
4675
4676         if (ret > 0) {
4677                 ret = 0;
4678                 /* there are no items in the tree for us to truncate, we're
4679                  * done
4680                  */
4681                 if (path->slots[0] == 0)
4682                         goto out;
4683                 path->slots[0]--;
4684         }
4685
4686         while (1) {
4687                 fi = NULL;
4688                 leaf = path->nodes[0];
4689                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4690                 found_type = found_key.type;
4691
4692                 if (found_key.objectid != ino)
4693                         break;
4694
4695                 if (found_type < min_type)
4696                         break;
4697
4698                 item_end = found_key.offset;
4699                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4700                         fi = btrfs_item_ptr(leaf, path->slots[0],
4701                                             struct btrfs_file_extent_item);
4702                         extent_type = btrfs_file_extent_type(leaf, fi);
4703                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4704                                 item_end +=
4705                                     btrfs_file_extent_num_bytes(leaf, fi);
4706
4707                                 trace_btrfs_truncate_show_fi_regular(
4708                                         BTRFS_I(inode), leaf, fi,
4709                                         found_key.offset);
4710                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4711                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4712                                                                         fi);
4713
4714                                 trace_btrfs_truncate_show_fi_inline(
4715                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4716                                         found_key.offset);
4717                         }
4718                         item_end--;
4719                 }
4720                 if (found_type > min_type) {
4721                         del_item = 1;
4722                 } else {
4723                         if (item_end < new_size)
4724                                 break;
4725                         if (found_key.offset >= new_size)
4726                                 del_item = 1;
4727                         else
4728                                 del_item = 0;
4729                 }
4730                 found_extent = 0;
4731                 /* FIXME, shrink the extent if the ref count is only 1 */
4732                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4733                         goto delete;
4734
4735                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4736                         u64 num_dec;
4737                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4738                         if (!del_item) {
4739                                 u64 orig_num_bytes =
4740                                         btrfs_file_extent_num_bytes(leaf, fi);
4741                                 extent_num_bytes = ALIGN(new_size -
4742                                                 found_key.offset,
4743                                                 fs_info->sectorsize);
4744                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4745                                                          extent_num_bytes);
4746                                 num_dec = (orig_num_bytes -
4747                                            extent_num_bytes);
4748                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4749                                              &root->state) &&
4750                                     extent_start != 0)
4751                                         inode_sub_bytes(inode, num_dec);
4752                                 btrfs_mark_buffer_dirty(leaf);
4753                         } else {
4754                                 extent_num_bytes =
4755                                         btrfs_file_extent_disk_num_bytes(leaf,
4756                                                                          fi);
4757                                 extent_offset = found_key.offset -
4758                                         btrfs_file_extent_offset(leaf, fi);
4759
4760                                 /* FIXME blocksize != 4096 */
4761                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4762                                 if (extent_start != 0) {
4763                                         found_extent = 1;
4764                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4765                                                      &root->state))
4766                                                 inode_sub_bytes(inode, num_dec);
4767                                 }
4768                         }
4769                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4770                         /*
4771                          * we can't truncate inline items that have had
4772                          * special encodings
4773                          */
4774                         if (!del_item &&
4775                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4776                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4777                             btrfs_file_extent_compression(leaf, fi) == 0) {
4778                                 u32 size = (u32)(new_size - found_key.offset);
4779
4780                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4781                                 size = btrfs_file_extent_calc_inline_size(size);
4782                                 btrfs_truncate_item(path, size, 1);
4783                         } else if (!del_item) {
4784                                 /*
4785                                  * We have to bail so the last_size is set to
4786                                  * just before this extent.
4787                                  */
4788                                 ret = NEED_TRUNCATE_BLOCK;
4789                                 break;
4790                         }
4791
4792                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4793                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4794                 }
4795 delete:
4796                 if (del_item)
4797                         last_size = found_key.offset;
4798                 else
4799                         last_size = new_size;
4800                 if (del_item) {
4801                         if (!pending_del_nr) {
4802                                 /* no pending yet, add ourselves */
4803                                 pending_del_slot = path->slots[0];
4804                                 pending_del_nr = 1;
4805                         } else if (pending_del_nr &&
4806                                    path->slots[0] + 1 == pending_del_slot) {
4807                                 /* hop on the pending chunk */
4808                                 pending_del_nr++;
4809                                 pending_del_slot = path->slots[0];
4810                         } else {
4811                                 BUG();
4812                         }
4813                 } else {
4814                         break;
4815                 }
4816                 should_throttle = false;
4817
4818                 if (found_extent &&
4819                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4820                      root == fs_info->tree_root)) {
4821                         struct btrfs_ref ref = { 0 };
4822
4823                         btrfs_set_path_blocking(path);
4824                         bytes_deleted += extent_num_bytes;
4825
4826                         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF,
4827                                         extent_start, extent_num_bytes, 0);
4828                         ref.real_root = root->root_key.objectid;
4829                         btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
4830                                         ino, extent_offset);
4831                         ret = btrfs_free_extent(trans, &ref);
4832                         if (ret) {
4833                                 btrfs_abort_transaction(trans, ret);
4834                                 break;
4835                         }
4836                         if (be_nice) {
4837                                 if (btrfs_should_throttle_delayed_refs(trans))
4838                                         should_throttle = true;
4839                         }
4840                 }
4841
4842                 if (found_type == BTRFS_INODE_ITEM_KEY)
4843                         break;
4844
4845                 if (path->slots[0] == 0 ||
4846                     path->slots[0] != pending_del_slot ||
4847                     should_throttle) {
4848                         if (pending_del_nr) {
4849                                 ret = btrfs_del_items(trans, root, path,
4850                                                 pending_del_slot,
4851                                                 pending_del_nr);
4852                                 if (ret) {
4853                                         btrfs_abort_transaction(trans, ret);
4854                                         break;
4855                                 }
4856                                 pending_del_nr = 0;
4857                         }
4858                         btrfs_release_path(path);
4859
4860                         /*
4861                          * We can generate a lot of delayed refs, so we need to
4862                          * throttle every once and a while and make sure we're
4863                          * adding enough space to keep up with the work we are
4864                          * generating.  Since we hold a transaction here we
4865                          * can't flush, and we don't want to FLUSH_LIMIT because
4866                          * we could have generated too many delayed refs to
4867                          * actually allocate, so just bail if we're short and
4868                          * let the normal reservation dance happen higher up.
4869                          */
4870                         if (should_throttle) {
4871                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4872                                                         BTRFS_RESERVE_NO_FLUSH);
4873                                 if (ret) {
4874                                         ret = -EAGAIN;
4875                                         break;
4876                                 }
4877                         }
4878                         goto search_again;
4879                 } else {
4880                         path->slots[0]--;
4881                 }
4882         }
4883 out:
4884         if (ret >= 0 && pending_del_nr) {
4885                 int err;
4886
4887                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4888                                       pending_del_nr);
4889                 if (err) {
4890                         btrfs_abort_transaction(trans, err);
4891                         ret = err;
4892                 }
4893         }
4894         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4895                 ASSERT(last_size >= new_size);
4896                 if (!ret && last_size > new_size)
4897                         last_size = new_size;
4898                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4899         }
4900
4901         btrfs_free_path(path);
4902         return ret;
4903 }
4904
4905 /*
4906  * btrfs_truncate_block - read, zero a chunk and write a block
4907  * @inode - inode that we're zeroing
4908  * @from - the offset to start zeroing
4909  * @len - the length to zero, 0 to zero the entire range respective to the
4910  *      offset
4911  * @front - zero up to the offset instead of from the offset on
4912  *
4913  * This will find the block for the "from" offset and cow the block and zero the
4914  * part we want to zero.  This is used with truncate and hole punching.
4915  */
4916 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4917                         int front)
4918 {
4919         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4920         struct address_space *mapping = inode->i_mapping;
4921         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4922         struct btrfs_ordered_extent *ordered;
4923         struct extent_state *cached_state = NULL;
4924         struct extent_changeset *data_reserved = NULL;
4925         char *kaddr;
4926         u32 blocksize = fs_info->sectorsize;
4927         pgoff_t index = from >> PAGE_SHIFT;
4928         unsigned offset = from & (blocksize - 1);
4929         struct page *page;
4930         gfp_t mask = btrfs_alloc_write_mask(mapping);
4931         int ret = 0;
4932         u64 block_start;
4933         u64 block_end;
4934
4935         if (IS_ALIGNED(offset, blocksize) &&
4936             (!len || IS_ALIGNED(len, blocksize)))
4937                 goto out;
4938
4939         block_start = round_down(from, blocksize);
4940         block_end = block_start + blocksize - 1;
4941
4942         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4943                                            block_start, blocksize);
4944         if (ret)
4945                 goto out;
4946
4947 again:
4948         page = find_or_create_page(mapping, index, mask);
4949         if (!page) {
4950                 btrfs_delalloc_release_space(inode, data_reserved,
4951                                              block_start, blocksize, true);
4952                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, true);
4953                 ret = -ENOMEM;
4954                 goto out;
4955         }
4956
4957         if (!PageUptodate(page)) {
4958                 ret = btrfs_readpage(NULL, page);
4959                 lock_page(page);
4960                 if (page->mapping != mapping) {
4961                         unlock_page(page);
4962                         put_page(page);
4963                         goto again;
4964                 }
4965                 if (!PageUptodate(page)) {
4966                         ret = -EIO;
4967                         goto out_unlock;
4968                 }
4969         }
4970         wait_on_page_writeback(page);
4971
4972         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4973         set_page_extent_mapped(page);
4974
4975         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4976         if (ordered) {
4977                 unlock_extent_cached(io_tree, block_start, block_end,
4978                                      &cached_state);
4979                 unlock_page(page);
4980                 put_page(page);
4981                 btrfs_start_ordered_extent(inode, ordered, 1);
4982                 btrfs_put_ordered_extent(ordered);
4983                 goto again;
4984         }
4985
4986         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4987                           EXTENT_DIRTY | EXTENT_DELALLOC |
4988                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4989                           0, 0, &cached_state);
4990
4991         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4992                                         &cached_state);
4993         if (ret) {
4994                 unlock_extent_cached(io_tree, block_start, block_end,
4995                                      &cached_state);
4996                 goto out_unlock;
4997         }
4998
4999         if (offset != blocksize) {
5000                 if (!len)
5001                         len = blocksize - offset;
5002                 kaddr = kmap(page);
5003                 if (front)
5004                         memset(kaddr + (block_start - page_offset(page)),
5005                                 0, offset);
5006                 else
5007                         memset(kaddr + (block_start - page_offset(page)) +  offset,
5008                                 0, len);
5009                 flush_dcache_page(page);
5010                 kunmap(page);
5011         }
5012         ClearPageChecked(page);
5013         set_page_dirty(page);
5014         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
5015
5016 out_unlock:
5017         if (ret)
5018                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
5019                                              blocksize, true);
5020         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize, (ret != 0));
5021         unlock_page(page);
5022         put_page(page);
5023 out:
5024         extent_changeset_free(data_reserved);
5025         return ret;
5026 }
5027
5028 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
5029                              u64 offset, u64 len)
5030 {
5031         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5032         struct btrfs_trans_handle *trans;
5033         int ret;
5034
5035         /*
5036          * Still need to make sure the inode looks like it's been updated so
5037          * that any holes get logged if we fsync.
5038          */
5039         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
5040                 BTRFS_I(inode)->last_trans = fs_info->generation;
5041                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
5042                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
5043                 return 0;
5044         }
5045
5046         /*
5047          * 1 - for the one we're dropping
5048          * 1 - for the one we're adding
5049          * 1 - for updating the inode.
5050          */
5051         trans = btrfs_start_transaction(root, 3);
5052         if (IS_ERR(trans))
5053                 return PTR_ERR(trans);
5054
5055         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
5056         if (ret) {
5057                 btrfs_abort_transaction(trans, ret);
5058                 btrfs_end_transaction(trans);
5059                 return ret;
5060         }
5061
5062         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
5063                         offset, 0, 0, len, 0, len, 0, 0, 0);
5064         if (ret)
5065                 btrfs_abort_transaction(trans, ret);
5066         else
5067                 btrfs_update_inode(trans, root, inode);
5068         btrfs_end_transaction(trans);
5069         return ret;
5070 }
5071
5072 /*
5073  * This function puts in dummy file extents for the area we're creating a hole
5074  * for.  So if we are truncating this file to a larger size we need to insert
5075  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
5076  * the range between oldsize and size
5077  */
5078 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
5079 {
5080         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5081         struct btrfs_root *root = BTRFS_I(inode)->root;
5082         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5083         struct extent_map *em = NULL;
5084         struct extent_state *cached_state = NULL;
5085         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5086         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
5087         u64 block_end = ALIGN(size, fs_info->sectorsize);
5088         u64 last_byte;
5089         u64 cur_offset;
5090         u64 hole_size;
5091         int err = 0;
5092
5093         /*
5094          * If our size started in the middle of a block we need to zero out the
5095          * rest of the block before we expand the i_size, otherwise we could
5096          * expose stale data.
5097          */
5098         err = btrfs_truncate_block(inode, oldsize, 0, 0);
5099         if (err)
5100                 return err;
5101
5102         if (size <= hole_start)
5103                 return 0;
5104
5105         btrfs_lock_and_flush_ordered_range(io_tree, BTRFS_I(inode), hole_start,
5106                                            block_end - 1, &cached_state);
5107         cur_offset = hole_start;
5108         while (1) {
5109                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5110                                 block_end - cur_offset, 0);
5111                 if (IS_ERR(em)) {
5112                         err = PTR_ERR(em);
5113                         em = NULL;
5114                         break;
5115                 }
5116                 last_byte = min(extent_map_end(em), block_end);
5117                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5118                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5119                         struct extent_map *hole_em;
5120                         hole_size = last_byte - cur_offset;
5121
5122                         err = maybe_insert_hole(root, inode, cur_offset,
5123                                                 hole_size);
5124                         if (err)
5125                                 break;
5126                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5127                                                 cur_offset + hole_size - 1, 0);
5128                         hole_em = alloc_extent_map();
5129                         if (!hole_em) {
5130                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5131                                         &BTRFS_I(inode)->runtime_flags);
5132                                 goto next;
5133                         }
5134                         hole_em->start = cur_offset;
5135                         hole_em->len = hole_size;
5136                         hole_em->orig_start = cur_offset;
5137
5138                         hole_em->block_start = EXTENT_MAP_HOLE;
5139                         hole_em->block_len = 0;
5140                         hole_em->orig_block_len = 0;
5141                         hole_em->ram_bytes = hole_size;
5142                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5143                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5144                         hole_em->generation = fs_info->generation;
5145
5146                         while (1) {
5147                                 write_lock(&em_tree->lock);
5148                                 err = add_extent_mapping(em_tree, hole_em, 1);
5149                                 write_unlock(&em_tree->lock);
5150                                 if (err != -EEXIST)
5151                                         break;
5152                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5153                                                         cur_offset,
5154                                                         cur_offset +
5155                                                         hole_size - 1, 0);
5156                         }
5157                         free_extent_map(hole_em);
5158                 }
5159 next:
5160                 free_extent_map(em);
5161                 em = NULL;
5162                 cur_offset = last_byte;
5163                 if (cur_offset >= block_end)
5164                         break;
5165         }
5166         free_extent_map(em);
5167         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5168         return err;
5169 }
5170
5171 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5172 {
5173         struct btrfs_root *root = BTRFS_I(inode)->root;
5174         struct btrfs_trans_handle *trans;
5175         loff_t oldsize = i_size_read(inode);
5176         loff_t newsize = attr->ia_size;
5177         int mask = attr->ia_valid;
5178         int ret;
5179
5180         /*
5181          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5182          * special case where we need to update the times despite not having
5183          * these flags set.  For all other operations the VFS set these flags
5184          * explicitly if it wants a timestamp update.
5185          */
5186         if (newsize != oldsize) {
5187                 inode_inc_iversion(inode);
5188                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5189                         inode->i_ctime = inode->i_mtime =
5190                                 current_time(inode);
5191         }
5192
5193         if (newsize > oldsize) {
5194                 /*
5195                  * Don't do an expanding truncate while snapshotting is ongoing.
5196                  * This is to ensure the snapshot captures a fully consistent
5197                  * state of this file - if the snapshot captures this expanding
5198                  * truncation, it must capture all writes that happened before
5199                  * this truncation.
5200                  */
5201                 btrfs_wait_for_snapshot_creation(root);
5202                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5203                 if (ret) {
5204                         btrfs_end_write_no_snapshotting(root);
5205                         return ret;
5206                 }
5207
5208                 trans = btrfs_start_transaction(root, 1);
5209                 if (IS_ERR(trans)) {
5210                         btrfs_end_write_no_snapshotting(root);
5211                         return PTR_ERR(trans);
5212                 }
5213
5214                 i_size_write(inode, newsize);
5215                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5216                 pagecache_isize_extended(inode, oldsize, newsize);
5217                 ret = btrfs_update_inode(trans, root, inode);
5218                 btrfs_end_write_no_snapshotting(root);
5219                 btrfs_end_transaction(trans);
5220         } else {
5221
5222                 /*
5223                  * We're truncating a file that used to have good data down to
5224                  * zero. Make sure it gets into the ordered flush list so that
5225                  * any new writes get down to disk quickly.
5226                  */
5227                 if (newsize == 0)
5228                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5229                                 &BTRFS_I(inode)->runtime_flags);
5230
5231                 truncate_setsize(inode, newsize);
5232
5233                 /* Disable nonlocked read DIO to avoid the endless truncate */
5234                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5235                 inode_dio_wait(inode);
5236                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5237
5238                 ret = btrfs_truncate(inode, newsize == oldsize);
5239                 if (ret && inode->i_nlink) {
5240                         int err;
5241
5242                         /*
5243                          * Truncate failed, so fix up the in-memory size. We
5244                          * adjusted disk_i_size down as we removed extents, so
5245                          * wait for disk_i_size to be stable and then update the
5246                          * in-memory size to match.
5247                          */
5248                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5249                         if (err)
5250                                 return err;
5251                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5252                 }
5253         }
5254
5255         return ret;
5256 }
5257
5258 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5259 {
5260         struct inode *inode = d_inode(dentry);
5261         struct btrfs_root *root = BTRFS_I(inode)->root;
5262         int err;
5263
5264         if (btrfs_root_readonly(root))
5265                 return -EROFS;
5266
5267         err = setattr_prepare(dentry, attr);
5268         if (err)
5269                 return err;
5270
5271         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5272                 err = btrfs_setsize(inode, attr);
5273                 if (err)
5274                         return err;
5275         }
5276
5277         if (attr->ia_valid) {
5278                 setattr_copy(inode, attr);
5279                 inode_inc_iversion(inode);
5280                 err = btrfs_dirty_inode(inode);
5281
5282                 if (!err && attr->ia_valid & ATTR_MODE)
5283                         err = posix_acl_chmod(inode, inode->i_mode);
5284         }
5285
5286         return err;
5287 }
5288
5289 /*
5290  * While truncating the inode pages during eviction, we get the VFS calling
5291  * btrfs_invalidatepage() against each page of the inode. This is slow because
5292  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5293  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5294  * extent_state structures over and over, wasting lots of time.
5295  *
5296  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5297  * those expensive operations on a per page basis and do only the ordered io
5298  * finishing, while we release here the extent_map and extent_state structures,
5299  * without the excessive merging and splitting.
5300  */
5301 static void evict_inode_truncate_pages(struct inode *inode)
5302 {
5303         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5304         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5305         struct rb_node *node;
5306
5307         ASSERT(inode->i_state & I_FREEING);
5308         truncate_inode_pages_final(&inode->i_data);
5309
5310         write_lock(&map_tree->lock);
5311         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
5312                 struct extent_map *em;
5313
5314                 node = rb_first_cached(&map_tree->map);
5315                 em = rb_entry(node, struct extent_map, rb_node);
5316                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5317                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5318                 remove_extent_mapping(map_tree, em);
5319                 free_extent_map(em);
5320                 if (need_resched()) {
5321                         write_unlock(&map_tree->lock);
5322                         cond_resched();
5323                         write_lock(&map_tree->lock);
5324                 }
5325         }
5326         write_unlock(&map_tree->lock);
5327
5328         /*
5329          * Keep looping until we have no more ranges in the io tree.
5330          * We can have ongoing bios started by readpages (called from readahead)
5331          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5332          * still in progress (unlocked the pages in the bio but did not yet
5333          * unlocked the ranges in the io tree). Therefore this means some
5334          * ranges can still be locked and eviction started because before
5335          * submitting those bios, which are executed by a separate task (work
5336          * queue kthread), inode references (inode->i_count) were not taken
5337          * (which would be dropped in the end io callback of each bio).
5338          * Therefore here we effectively end up waiting for those bios and
5339          * anyone else holding locked ranges without having bumped the inode's
5340          * reference count - if we don't do it, when they access the inode's
5341          * io_tree to unlock a range it may be too late, leading to an
5342          * use-after-free issue.
5343          */
5344         spin_lock(&io_tree->lock);
5345         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5346                 struct extent_state *state;
5347                 struct extent_state *cached_state = NULL;
5348                 u64 start;
5349                 u64 end;
5350                 unsigned state_flags;
5351
5352                 node = rb_first(&io_tree->state);
5353                 state = rb_entry(node, struct extent_state, rb_node);
5354                 start = state->start;
5355                 end = state->end;
5356                 state_flags = state->state;
5357                 spin_unlock(&io_tree->lock);
5358
5359                 lock_extent_bits(io_tree, start, end, &cached_state);
5360
5361                 /*
5362                  * If still has DELALLOC flag, the extent didn't reach disk,
5363                  * and its reserved space won't be freed by delayed_ref.
5364                  * So we need to free its reserved space here.
5365                  * (Refer to comment in btrfs_invalidatepage, case 2)
5366                  *
5367                  * Note, end is the bytenr of last byte, so we need + 1 here.
5368                  */
5369                 if (state_flags & EXTENT_DELALLOC)
5370                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5371
5372                 clear_extent_bit(io_tree, start, end,
5373                                  EXTENT_LOCKED | EXTENT_DIRTY |
5374                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5375                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5376
5377                 cond_resched();
5378                 spin_lock(&io_tree->lock);
5379         }
5380         spin_unlock(&io_tree->lock);
5381 }
5382
5383 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5384                                                         struct btrfs_block_rsv *rsv)
5385 {
5386         struct btrfs_fs_info *fs_info = root->fs_info;
5387         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5388         struct btrfs_trans_handle *trans;
5389         u64 delayed_refs_extra = btrfs_calc_insert_metadata_size(fs_info, 1);
5390         int ret;
5391
5392         /*
5393          * Eviction should be taking place at some place safe because of our
5394          * delayed iputs.  However the normal flushing code will run delayed
5395          * iputs, so we cannot use FLUSH_ALL otherwise we'll deadlock.
5396          *
5397          * We reserve the delayed_refs_extra here again because we can't use
5398          * btrfs_start_transaction(root, 0) for the same deadlocky reason as
5399          * above.  We reserve our extra bit here because we generate a ton of
5400          * delayed refs activity by truncating.
5401          *
5402          * If we cannot make our reservation we'll attempt to steal from the
5403          * global reserve, because we really want to be able to free up space.
5404          */
5405         ret = btrfs_block_rsv_refill(root, rsv, rsv->size + delayed_refs_extra,
5406                                      BTRFS_RESERVE_FLUSH_EVICT);
5407         if (ret) {
5408                 /*
5409                  * Try to steal from the global reserve if there is space for
5410                  * it.
5411                  */
5412                 if (btrfs_check_space_for_delayed_refs(fs_info) ||
5413                     btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0)) {
5414                         btrfs_warn(fs_info,
5415                                    "could not allocate space for delete; will truncate on mount");
5416                         return ERR_PTR(-ENOSPC);
5417                 }
5418                 delayed_refs_extra = 0;
5419         }
5420
5421         trans = btrfs_join_transaction(root);
5422         if (IS_ERR(trans))
5423                 return trans;
5424
5425         if (delayed_refs_extra) {
5426                 trans->block_rsv = &fs_info->trans_block_rsv;
5427                 trans->bytes_reserved = delayed_refs_extra;
5428                 btrfs_block_rsv_migrate(rsv, trans->block_rsv,
5429                                         delayed_refs_extra, 1);
5430         }
5431         return trans;
5432 }
5433
5434 void btrfs_evict_inode(struct inode *inode)
5435 {
5436         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5437         struct btrfs_trans_handle *trans;
5438         struct btrfs_root *root = BTRFS_I(inode)->root;
5439         struct btrfs_block_rsv *rsv;
5440         int ret;
5441
5442         trace_btrfs_inode_evict(inode);
5443
5444         if (!root) {
5445                 clear_inode(inode);
5446                 return;
5447         }
5448
5449         evict_inode_truncate_pages(inode);
5450
5451         if (inode->i_nlink &&
5452             ((btrfs_root_refs(&root->root_item) != 0 &&
5453               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5454              btrfs_is_free_space_inode(BTRFS_I(inode))))
5455                 goto no_delete;
5456
5457         if (is_bad_inode(inode))
5458                 goto no_delete;
5459
5460         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5461
5462         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5463                 goto no_delete;
5464
5465         if (inode->i_nlink > 0) {
5466                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5467                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5468                 goto no_delete;
5469         }
5470
5471         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5472         if (ret)
5473                 goto no_delete;
5474
5475         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5476         if (!rsv)
5477                 goto no_delete;
5478         rsv->size = btrfs_calc_metadata_size(fs_info, 1);
5479         rsv->failfast = 1;
5480
5481         btrfs_i_size_write(BTRFS_I(inode), 0);
5482
5483         while (1) {
5484                 trans = evict_refill_and_join(root, rsv);
5485                 if (IS_ERR(trans))
5486                         goto free_rsv;
5487
5488                 trans->block_rsv = rsv;
5489
5490                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5491                 trans->block_rsv = &fs_info->trans_block_rsv;
5492                 btrfs_end_transaction(trans);
5493                 btrfs_btree_balance_dirty(fs_info);
5494                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5495                         goto free_rsv;
5496                 else if (!ret)
5497                         break;
5498         }
5499
5500         /*
5501          * Errors here aren't a big deal, it just means we leave orphan items in
5502          * the tree. They will be cleaned up on the next mount. If the inode
5503          * number gets reused, cleanup deletes the orphan item without doing
5504          * anything, and unlink reuses the existing orphan item.
5505          *
5506          * If it turns out that we are dropping too many of these, we might want
5507          * to add a mechanism for retrying these after a commit.
5508          */
5509         trans = evict_refill_and_join(root, rsv);
5510         if (!IS_ERR(trans)) {
5511                 trans->block_rsv = rsv;
5512                 btrfs_orphan_del(trans, BTRFS_I(inode));
5513                 trans->block_rsv = &fs_info->trans_block_rsv;
5514                 btrfs_end_transaction(trans);
5515         }
5516
5517         if (!(root == fs_info->tree_root ||
5518               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5519                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5520
5521 free_rsv:
5522         btrfs_free_block_rsv(fs_info, rsv);
5523 no_delete:
5524         /*
5525          * If we didn't successfully delete, the orphan item will still be in
5526          * the tree and we'll retry on the next mount. Again, we might also want
5527          * to retry these periodically in the future.
5528          */
5529         btrfs_remove_delayed_node(BTRFS_I(inode));
5530         clear_inode(inode);
5531 }
5532
5533 /*
5534  * Return the key found in the dir entry in the location pointer, fill @type
5535  * with BTRFS_FT_*, and return 0.
5536  *
5537  * If no dir entries were found, returns -ENOENT.
5538  * If found a corrupted location in dir entry, returns -EUCLEAN.
5539  */
5540 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5541                                struct btrfs_key *location, u8 *type)
5542 {
5543         const char *name = dentry->d_name.name;
5544         int namelen = dentry->d_name.len;
5545         struct btrfs_dir_item *di;
5546         struct btrfs_path *path;
5547         struct btrfs_root *root = BTRFS_I(dir)->root;
5548         int ret = 0;
5549
5550         path = btrfs_alloc_path();
5551         if (!path)
5552                 return -ENOMEM;
5553
5554         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5555                         name, namelen, 0);
5556         if (IS_ERR_OR_NULL(di)) {
5557                 ret = di ? PTR_ERR(di) : -ENOENT;
5558                 goto out;
5559         }
5560
5561         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5562         if (location->type != BTRFS_INODE_ITEM_KEY &&
5563             location->type != BTRFS_ROOT_ITEM_KEY) {
5564                 ret = -EUCLEAN;
5565                 btrfs_warn(root->fs_info,
5566 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5567                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5568                            location->objectid, location->type, location->offset);
5569         }
5570         if (!ret)
5571                 *type = btrfs_dir_type(path->nodes[0], di);
5572 out:
5573         btrfs_free_path(path);
5574         return ret;
5575 }
5576
5577 /*
5578  * when we hit a tree root in a directory, the btrfs part of the inode
5579  * needs to be changed to reflect the root directory of the tree root.  This
5580  * is kind of like crossing a mount point.
5581  */
5582 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5583                                     struct inode *dir,
5584                                     struct dentry *dentry,
5585                                     struct btrfs_key *location,
5586                                     struct btrfs_root **sub_root)
5587 {
5588         struct btrfs_path *path;
5589         struct btrfs_root *new_root;
5590         struct btrfs_root_ref *ref;
5591         struct extent_buffer *leaf;
5592         struct btrfs_key key;
5593         int ret;
5594         int err = 0;
5595
5596         path = btrfs_alloc_path();
5597         if (!path) {
5598                 err = -ENOMEM;
5599                 goto out;
5600         }
5601
5602         err = -ENOENT;
5603         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5604         key.type = BTRFS_ROOT_REF_KEY;
5605         key.offset = location->objectid;
5606
5607         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5608         if (ret) {
5609                 if (ret < 0)
5610                         err = ret;
5611                 goto out;
5612         }
5613
5614         leaf = path->nodes[0];
5615         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5616         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5617             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5618                 goto out;
5619
5620         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5621                                    (unsigned long)(ref + 1),
5622                                    dentry->d_name.len);
5623         if (ret)
5624                 goto out;
5625
5626         btrfs_release_path(path);
5627
5628         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5629         if (IS_ERR(new_root)) {
5630                 err = PTR_ERR(new_root);
5631                 goto out;
5632         }
5633
5634         *sub_root = new_root;
5635         location->objectid = btrfs_root_dirid(&new_root->root_item);
5636         location->type = BTRFS_INODE_ITEM_KEY;
5637         location->offset = 0;
5638         err = 0;
5639 out:
5640         btrfs_free_path(path);
5641         return err;
5642 }
5643
5644 static void inode_tree_add(struct inode *inode)
5645 {
5646         struct btrfs_root *root = BTRFS_I(inode)->root;
5647         struct btrfs_inode *entry;
5648         struct rb_node **p;
5649         struct rb_node *parent;
5650         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5651         u64 ino = btrfs_ino(BTRFS_I(inode));
5652
5653         if (inode_unhashed(inode))
5654                 return;
5655         parent = NULL;
5656         spin_lock(&root->inode_lock);
5657         p = &root->inode_tree.rb_node;
5658         while (*p) {
5659                 parent = *p;
5660                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5661
5662                 if (ino < btrfs_ino(entry))
5663                         p = &parent->rb_left;
5664                 else if (ino > btrfs_ino(entry))
5665                         p = &parent->rb_right;
5666                 else {
5667                         WARN_ON(!(entry->vfs_inode.i_state &
5668                                   (I_WILL_FREE | I_FREEING)));
5669                         rb_replace_node(parent, new, &root->inode_tree);
5670                         RB_CLEAR_NODE(parent);
5671                         spin_unlock(&root->inode_lock);
5672                         return;
5673                 }
5674         }
5675         rb_link_node(new, parent, p);
5676         rb_insert_color(new, &root->inode_tree);
5677         spin_unlock(&root->inode_lock);
5678 }
5679
5680 static void inode_tree_del(struct inode *inode)
5681 {
5682         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5683         struct btrfs_root *root = BTRFS_I(inode)->root;
5684         int empty = 0;
5685
5686         spin_lock(&root->inode_lock);
5687         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5688                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5689                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5690                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5691         }
5692         spin_unlock(&root->inode_lock);
5693
5694         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5695                 synchronize_srcu(&fs_info->subvol_srcu);
5696                 spin_lock(&root->inode_lock);
5697                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5698                 spin_unlock(&root->inode_lock);
5699                 if (empty)
5700                         btrfs_add_dead_root(root);
5701         }
5702 }
5703
5704
5705 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5706 {
5707         struct btrfs_iget_args *args = p;
5708         inode->i_ino = args->location->objectid;
5709         memcpy(&BTRFS_I(inode)->location, args->location,
5710                sizeof(*args->location));
5711         BTRFS_I(inode)->root = args->root;
5712         return 0;
5713 }
5714
5715 static int btrfs_find_actor(struct inode *inode, void *opaque)
5716 {
5717         struct btrfs_iget_args *args = opaque;
5718         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5719                 args->root == BTRFS_I(inode)->root;
5720 }
5721
5722 static struct inode *btrfs_iget_locked(struct super_block *s,
5723                                        struct btrfs_key *location,
5724                                        struct btrfs_root *root)
5725 {
5726         struct inode *inode;
5727         struct btrfs_iget_args args;
5728         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5729
5730         args.location = location;
5731         args.root = root;
5732
5733         inode = iget5_locked(s, hashval, btrfs_find_actor,
5734                              btrfs_init_locked_inode,
5735                              (void *)&args);
5736         return inode;
5737 }
5738
5739 /* Get an inode object given its location and corresponding root.
5740  * Returns in *is_new if the inode was read from disk
5741  */
5742 struct inode *btrfs_iget_path(struct super_block *s, struct btrfs_key *location,
5743                               struct btrfs_root *root, int *new,
5744                               struct btrfs_path *path)
5745 {
5746         struct inode *inode;
5747
5748         inode = btrfs_iget_locked(s, location, root);
5749         if (!inode)
5750                 return ERR_PTR(-ENOMEM);
5751
5752         if (inode->i_state & I_NEW) {
5753                 int ret;
5754
5755                 ret = btrfs_read_locked_inode(inode, path);
5756                 if (!ret) {
5757                         inode_tree_add(inode);
5758                         unlock_new_inode(inode);
5759                         if (new)
5760                                 *new = 1;
5761                 } else {
5762                         iget_failed(inode);
5763                         /*
5764                          * ret > 0 can come from btrfs_search_slot called by
5765                          * btrfs_read_locked_inode, this means the inode item
5766                          * was not found.
5767                          */
5768                         if (ret > 0)
5769                                 ret = -ENOENT;
5770                         inode = ERR_PTR(ret);
5771                 }
5772         }
5773
5774         return inode;
5775 }
5776
5777 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5778                          struct btrfs_root *root, int *new)
5779 {
5780         return btrfs_iget_path(s, location, root, new, NULL);
5781 }
5782
5783 static struct inode *new_simple_dir(struct super_block *s,
5784                                     struct btrfs_key *key,
5785                                     struct btrfs_root *root)
5786 {
5787         struct inode *inode = new_inode(s);
5788
5789         if (!inode)
5790                 return ERR_PTR(-ENOMEM);
5791
5792         BTRFS_I(inode)->root = root;
5793         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5794         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5795
5796         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5797         inode->i_op = &btrfs_dir_ro_inode_operations;
5798         inode->i_opflags &= ~IOP_XATTR;
5799         inode->i_fop = &simple_dir_operations;
5800         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5801         inode->i_mtime = current_time(inode);
5802         inode->i_atime = inode->i_mtime;
5803         inode->i_ctime = inode->i_mtime;
5804         BTRFS_I(inode)->i_otime = inode->i_mtime;
5805
5806         return inode;
5807 }
5808
5809 static inline u8 btrfs_inode_type(struct inode *inode)
5810 {
5811         /*
5812          * Compile-time asserts that generic FT_* types still match
5813          * BTRFS_FT_* types
5814          */
5815         BUILD_BUG_ON(BTRFS_FT_UNKNOWN != FT_UNKNOWN);
5816         BUILD_BUG_ON(BTRFS_FT_REG_FILE != FT_REG_FILE);
5817         BUILD_BUG_ON(BTRFS_FT_DIR != FT_DIR);
5818         BUILD_BUG_ON(BTRFS_FT_CHRDEV != FT_CHRDEV);
5819         BUILD_BUG_ON(BTRFS_FT_BLKDEV != FT_BLKDEV);
5820         BUILD_BUG_ON(BTRFS_FT_FIFO != FT_FIFO);
5821         BUILD_BUG_ON(BTRFS_FT_SOCK != FT_SOCK);
5822         BUILD_BUG_ON(BTRFS_FT_SYMLINK != FT_SYMLINK);
5823
5824         return fs_umode_to_ftype(inode->i_mode);
5825 }
5826
5827 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5828 {
5829         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5830         struct inode *inode;
5831         struct btrfs_root *root = BTRFS_I(dir)->root;
5832         struct btrfs_root *sub_root = root;
5833         struct btrfs_key location;
5834         u8 di_type = 0;
5835         int index;
5836         int ret = 0;
5837
5838         if (dentry->d_name.len > BTRFS_NAME_LEN)
5839                 return ERR_PTR(-ENAMETOOLONG);
5840
5841         ret = btrfs_inode_by_name(dir, dentry, &location, &di_type);
5842         if (ret < 0)
5843                 return ERR_PTR(ret);
5844
5845         if (location.type == BTRFS_INODE_ITEM_KEY) {
5846                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5847                 if (IS_ERR(inode))
5848                         return inode;
5849
5850                 /* Do extra check against inode mode with di_type */
5851                 if (btrfs_inode_type(inode) != di_type) {
5852                         btrfs_crit(fs_info,
5853 "inode mode mismatch with dir: inode mode=0%o btrfs type=%u dir type=%u",
5854                                   inode->i_mode, btrfs_inode_type(inode),
5855                                   di_type);
5856                         iput(inode);
5857                         return ERR_PTR(-EUCLEAN);
5858                 }
5859                 return inode;
5860         }
5861
5862         index = srcu_read_lock(&fs_info->subvol_srcu);
5863         ret = fixup_tree_root_location(fs_info, dir, dentry,
5864                                        &location, &sub_root);
5865         if (ret < 0) {
5866                 if (ret != -ENOENT)
5867                         inode = ERR_PTR(ret);
5868                 else
5869                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5870         } else {
5871                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5872         }
5873         srcu_read_unlock(&fs_info->subvol_srcu, index);
5874
5875         if (!IS_ERR(inode) && root != sub_root) {
5876                 down_read(&fs_info->cleanup_work_sem);
5877                 if (!sb_rdonly(inode->i_sb))
5878                         ret = btrfs_orphan_cleanup(sub_root);
5879                 up_read(&fs_info->cleanup_work_sem);
5880                 if (ret) {
5881                         iput(inode);
5882                         inode = ERR_PTR(ret);
5883                 }
5884         }
5885
5886         return inode;
5887 }
5888
5889 static int btrfs_dentry_delete(const struct dentry *dentry)
5890 {
5891         struct btrfs_root *root;
5892         struct inode *inode = d_inode(dentry);
5893
5894         if (!inode && !IS_ROOT(dentry))
5895                 inode = d_inode(dentry->d_parent);
5896
5897         if (inode) {
5898                 root = BTRFS_I(inode)->root;
5899                 if (btrfs_root_refs(&root->root_item) == 0)
5900                         return 1;
5901
5902                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5903                         return 1;
5904         }
5905         return 0;
5906 }
5907
5908 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5909                                    unsigned int flags)
5910 {
5911         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5912
5913         if (inode == ERR_PTR(-ENOENT))
5914                 inode = NULL;
5915         return d_splice_alias(inode, dentry);
5916 }
5917
5918 /*
5919  * All this infrastructure exists because dir_emit can fault, and we are holding
5920  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5921  * our information into that, and then dir_emit from the buffer.  This is
5922  * similar to what NFS does, only we don't keep the buffer around in pagecache
5923  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5924  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5925  * tree lock.
5926  */
5927 static int btrfs_opendir(struct inode *inode, struct file *file)
5928 {
5929         struct btrfs_file_private *private;
5930
5931         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5932         if (!private)
5933                 return -ENOMEM;
5934         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5935         if (!private->filldir_buf) {
5936                 kfree(private);
5937                 return -ENOMEM;
5938         }
5939         file->private_data = private;
5940         return 0;
5941 }
5942
5943 struct dir_entry {
5944         u64 ino;
5945         u64 offset;
5946         unsigned type;
5947         int name_len;
5948 };
5949
5950 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5951 {
5952         while (entries--) {
5953                 struct dir_entry *entry = addr;
5954                 char *name = (char *)(entry + 1);
5955
5956                 ctx->pos = get_unaligned(&entry->offset);
5957                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5958                                          get_unaligned(&entry->ino),
5959                                          get_unaligned(&entry->type)))
5960                         return 1;
5961                 addr += sizeof(struct dir_entry) +
5962                         get_unaligned(&entry->name_len);
5963                 ctx->pos++;
5964         }
5965         return 0;
5966 }
5967
5968 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5969 {
5970         struct inode *inode = file_inode(file);
5971         struct btrfs_root *root = BTRFS_I(inode)->root;
5972         struct btrfs_file_private *private = file->private_data;
5973         struct btrfs_dir_item *di;
5974         struct btrfs_key key;
5975         struct btrfs_key found_key;
5976         struct btrfs_path *path;
5977         void *addr;
5978         struct list_head ins_list;
5979         struct list_head del_list;
5980         int ret;
5981         struct extent_buffer *leaf;
5982         int slot;
5983         char *name_ptr;
5984         int name_len;
5985         int entries = 0;
5986         int total_len = 0;
5987         bool put = false;
5988         struct btrfs_key location;
5989
5990         if (!dir_emit_dots(file, ctx))
5991                 return 0;
5992
5993         path = btrfs_alloc_path();
5994         if (!path)
5995                 return -ENOMEM;
5996
5997         addr = private->filldir_buf;
5998         path->reada = READA_FORWARD;
5999
6000         INIT_LIST_HEAD(&ins_list);
6001         INIT_LIST_HEAD(&del_list);
6002         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
6003
6004 again:
6005         key.type = BTRFS_DIR_INDEX_KEY;
6006         key.offset = ctx->pos;
6007         key.objectid = btrfs_ino(BTRFS_I(inode));
6008
6009         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6010         if (ret < 0)
6011                 goto err;
6012
6013         while (1) {
6014                 struct dir_entry *entry;
6015
6016                 leaf = path->nodes[0];
6017                 slot = path->slots[0];
6018                 if (slot >= btrfs_header_nritems(leaf)) {
6019                         ret = btrfs_next_leaf(root, path);
6020                         if (ret < 0)
6021                                 goto err;
6022                         else if (ret > 0)
6023                                 break;
6024                         continue;
6025                 }
6026
6027                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6028
6029                 if (found_key.objectid != key.objectid)
6030                         break;
6031                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
6032                         break;
6033                 if (found_key.offset < ctx->pos)
6034                         goto next;
6035                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
6036                         goto next;
6037                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
6038                 name_len = btrfs_dir_name_len(leaf, di);
6039                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6040                     PAGE_SIZE) {
6041                         btrfs_release_path(path);
6042                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6043                         if (ret)
6044                                 goto nopos;
6045                         addr = private->filldir_buf;
6046                         entries = 0;
6047                         total_len = 0;
6048                         goto again;
6049                 }
6050
6051                 entry = addr;
6052                 put_unaligned(name_len, &entry->name_len);
6053                 name_ptr = (char *)(entry + 1);
6054                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6055                                    name_len);
6056                 put_unaligned(fs_ftype_to_dtype(btrfs_dir_type(leaf, di)),
6057                                 &entry->type);
6058                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6059                 put_unaligned(location.objectid, &entry->ino);
6060                 put_unaligned(found_key.offset, &entry->offset);
6061                 entries++;
6062                 addr += sizeof(struct dir_entry) + name_len;
6063                 total_len += sizeof(struct dir_entry) + name_len;
6064 next:
6065                 path->slots[0]++;
6066         }
6067         btrfs_release_path(path);
6068
6069         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6070         if (ret)
6071                 goto nopos;
6072
6073         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6074         if (ret)
6075                 goto nopos;
6076
6077         /*
6078          * Stop new entries from being returned after we return the last
6079          * entry.
6080          *
6081          * New directory entries are assigned a strictly increasing
6082          * offset.  This means that new entries created during readdir
6083          * are *guaranteed* to be seen in the future by that readdir.
6084          * This has broken buggy programs which operate on names as
6085          * they're returned by readdir.  Until we re-use freed offsets
6086          * we have this hack to stop new entries from being returned
6087          * under the assumption that they'll never reach this huge
6088          * offset.
6089          *
6090          * This is being careful not to overflow 32bit loff_t unless the
6091          * last entry requires it because doing so has broken 32bit apps
6092          * in the past.
6093          */
6094         if (ctx->pos >= INT_MAX)
6095                 ctx->pos = LLONG_MAX;
6096         else
6097                 ctx->pos = INT_MAX;
6098 nopos:
6099         ret = 0;
6100 err:
6101         if (put)
6102                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6103         btrfs_free_path(path);
6104         return ret;
6105 }
6106
6107 /*
6108  * This is somewhat expensive, updating the tree every time the
6109  * inode changes.  But, it is most likely to find the inode in cache.
6110  * FIXME, needs more benchmarking...there are no reasons other than performance
6111  * to keep or drop this code.
6112  */
6113 static int btrfs_dirty_inode(struct inode *inode)
6114 {
6115         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6116         struct btrfs_root *root = BTRFS_I(inode)->root;
6117         struct btrfs_trans_handle *trans;
6118         int ret;
6119
6120         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6121                 return 0;
6122
6123         trans = btrfs_join_transaction(root);
6124         if (IS_ERR(trans))
6125                 return PTR_ERR(trans);
6126
6127         ret = btrfs_update_inode(trans, root, inode);
6128         if (ret && ret == -ENOSPC) {
6129                 /* whoops, lets try again with the full transaction */
6130                 btrfs_end_transaction(trans);
6131                 trans = btrfs_start_transaction(root, 1);
6132                 if (IS_ERR(trans))
6133                         return PTR_ERR(trans);
6134
6135                 ret = btrfs_update_inode(trans, root, inode);
6136         }
6137         btrfs_end_transaction(trans);
6138         if (BTRFS_I(inode)->delayed_node)
6139                 btrfs_balance_delayed_items(fs_info);
6140
6141         return ret;
6142 }
6143
6144 /*
6145  * This is a copy of file_update_time.  We need this so we can return error on
6146  * ENOSPC for updating the inode in the case of file write and mmap writes.
6147  */
6148 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
6149                              int flags)
6150 {
6151         struct btrfs_root *root = BTRFS_I(inode)->root;
6152         bool dirty = flags & ~S_VERSION;
6153
6154         if (btrfs_root_readonly(root))
6155                 return -EROFS;
6156
6157         if (flags & S_VERSION)
6158                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6159         if (flags & S_CTIME)
6160                 inode->i_ctime = *now;
6161         if (flags & S_MTIME)
6162                 inode->i_mtime = *now;
6163         if (flags & S_ATIME)
6164                 inode->i_atime = *now;
6165         return dirty ? btrfs_dirty_inode(inode) : 0;
6166 }
6167
6168 /*
6169  * find the highest existing sequence number in a directory
6170  * and then set the in-memory index_cnt variable to reflect
6171  * free sequence numbers
6172  */
6173 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6174 {
6175         struct btrfs_root *root = inode->root;
6176         struct btrfs_key key, found_key;
6177         struct btrfs_path *path;
6178         struct extent_buffer *leaf;
6179         int ret;
6180
6181         key.objectid = btrfs_ino(inode);
6182         key.type = BTRFS_DIR_INDEX_KEY;
6183         key.offset = (u64)-1;
6184
6185         path = btrfs_alloc_path();
6186         if (!path)
6187                 return -ENOMEM;
6188
6189         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6190         if (ret < 0)
6191                 goto out;
6192         /* FIXME: we should be able to handle this */
6193         if (ret == 0)
6194                 goto out;
6195         ret = 0;
6196
6197         /*
6198          * MAGIC NUMBER EXPLANATION:
6199          * since we search a directory based on f_pos we have to start at 2
6200          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6201          * else has to start at 2
6202          */
6203         if (path->slots[0] == 0) {
6204                 inode->index_cnt = 2;
6205                 goto out;
6206         }
6207
6208         path->slots[0]--;
6209
6210         leaf = path->nodes[0];
6211         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6212
6213         if (found_key.objectid != btrfs_ino(inode) ||
6214             found_key.type != BTRFS_DIR_INDEX_KEY) {
6215                 inode->index_cnt = 2;
6216                 goto out;
6217         }
6218
6219         inode->index_cnt = found_key.offset + 1;
6220 out:
6221         btrfs_free_path(path);
6222         return ret;
6223 }
6224
6225 /*
6226  * helper to find a free sequence number in a given directory.  This current
6227  * code is very simple, later versions will do smarter things in the btree
6228  */
6229 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6230 {
6231         int ret = 0;
6232
6233         if (dir->index_cnt == (u64)-1) {
6234                 ret = btrfs_inode_delayed_dir_index_count(dir);
6235                 if (ret) {
6236                         ret = btrfs_set_inode_index_count(dir);
6237                         if (ret)
6238                                 return ret;
6239                 }
6240         }
6241
6242         *index = dir->index_cnt;
6243         dir->index_cnt++;
6244
6245         return ret;
6246 }
6247
6248 static int btrfs_insert_inode_locked(struct inode *inode)
6249 {
6250         struct btrfs_iget_args args;
6251         args.location = &BTRFS_I(inode)->location;
6252         args.root = BTRFS_I(inode)->root;
6253
6254         return insert_inode_locked4(inode,
6255                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6256                    btrfs_find_actor, &args);
6257 }
6258
6259 /*
6260  * Inherit flags from the parent inode.
6261  *
6262  * Currently only the compression flags and the cow flags are inherited.
6263  */
6264 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6265 {
6266         unsigned int flags;
6267
6268         if (!dir)
6269                 return;
6270
6271         flags = BTRFS_I(dir)->flags;
6272
6273         if (flags & BTRFS_INODE_NOCOMPRESS) {
6274                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6275                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6276         } else if (flags & BTRFS_INODE_COMPRESS) {
6277                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6278                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6279         }
6280
6281         if (flags & BTRFS_INODE_NODATACOW) {
6282                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6283                 if (S_ISREG(inode->i_mode))
6284                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6285         }
6286
6287         btrfs_sync_inode_flags_to_i_flags(inode);
6288 }
6289
6290 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6291                                      struct btrfs_root *root,
6292                                      struct inode *dir,
6293                                      const char *name, int name_len,
6294                                      u64 ref_objectid, u64 objectid,
6295                                      umode_t mode, u64 *index)
6296 {
6297         struct btrfs_fs_info *fs_info = root->fs_info;
6298         struct inode *inode;
6299         struct btrfs_inode_item *inode_item;
6300         struct btrfs_key *location;
6301         struct btrfs_path *path;
6302         struct btrfs_inode_ref *ref;
6303         struct btrfs_key key[2];
6304         u32 sizes[2];
6305         int nitems = name ? 2 : 1;
6306         unsigned long ptr;
6307         int ret;
6308
6309         path = btrfs_alloc_path();
6310         if (!path)
6311                 return ERR_PTR(-ENOMEM);
6312
6313         inode = new_inode(fs_info->sb);
6314         if (!inode) {
6315                 btrfs_free_path(path);
6316                 return ERR_PTR(-ENOMEM);
6317         }
6318
6319         /*
6320          * O_TMPFILE, set link count to 0, so that after this point,
6321          * we fill in an inode item with the correct link count.
6322          */
6323         if (!name)
6324                 set_nlink(inode, 0);
6325
6326         /*
6327          * we have to initialize this early, so we can reclaim the inode
6328          * number if we fail afterwards in this function.
6329          */
6330         inode->i_ino = objectid;
6331
6332         if (dir && name) {
6333                 trace_btrfs_inode_request(dir);
6334
6335                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6336                 if (ret) {
6337                         btrfs_free_path(path);
6338                         iput(inode);
6339                         return ERR_PTR(ret);
6340                 }
6341         } else if (dir) {
6342                 *index = 0;
6343         }
6344         /*
6345          * index_cnt is ignored for everything but a dir,
6346          * btrfs_set_inode_index_count has an explanation for the magic
6347          * number
6348          */
6349         BTRFS_I(inode)->index_cnt = 2;
6350         BTRFS_I(inode)->dir_index = *index;
6351         BTRFS_I(inode)->root = root;
6352         BTRFS_I(inode)->generation = trans->transid;
6353         inode->i_generation = BTRFS_I(inode)->generation;
6354
6355         /*
6356          * We could have gotten an inode number from somebody who was fsynced
6357          * and then removed in this same transaction, so let's just set full
6358          * sync since it will be a full sync anyway and this will blow away the
6359          * old info in the log.
6360          */
6361         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6362
6363         key[0].objectid = objectid;
6364         key[0].type = BTRFS_INODE_ITEM_KEY;
6365         key[0].offset = 0;
6366
6367         sizes[0] = sizeof(struct btrfs_inode_item);
6368
6369         if (name) {
6370                 /*
6371                  * Start new inodes with an inode_ref. This is slightly more
6372                  * efficient for small numbers of hard links since they will
6373                  * be packed into one item. Extended refs will kick in if we
6374                  * add more hard links than can fit in the ref item.
6375                  */
6376                 key[1].objectid = objectid;
6377                 key[1].type = BTRFS_INODE_REF_KEY;
6378                 key[1].offset = ref_objectid;
6379
6380                 sizes[1] = name_len + sizeof(*ref);
6381         }
6382
6383         location = &BTRFS_I(inode)->location;
6384         location->objectid = objectid;
6385         location->offset = 0;
6386         location->type = BTRFS_INODE_ITEM_KEY;
6387
6388         ret = btrfs_insert_inode_locked(inode);
6389         if (ret < 0) {
6390                 iput(inode);
6391                 goto fail;
6392         }
6393
6394         path->leave_spinning = 1;
6395         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6396         if (ret != 0)
6397                 goto fail_unlock;
6398
6399         inode_init_owner(inode, dir, mode);
6400         inode_set_bytes(inode, 0);
6401
6402         inode->i_mtime = current_time(inode);
6403         inode->i_atime = inode->i_mtime;
6404         inode->i_ctime = inode->i_mtime;
6405         BTRFS_I(inode)->i_otime = inode->i_mtime;
6406
6407         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6408                                   struct btrfs_inode_item);
6409         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6410                              sizeof(*inode_item));
6411         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6412
6413         if (name) {
6414                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6415                                      struct btrfs_inode_ref);
6416                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6417                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6418                 ptr = (unsigned long)(ref + 1);
6419                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6420         }
6421
6422         btrfs_mark_buffer_dirty(path->nodes[0]);
6423         btrfs_free_path(path);
6424
6425         btrfs_inherit_iflags(inode, dir);
6426
6427         if (S_ISREG(mode)) {
6428                 if (btrfs_test_opt(fs_info, NODATASUM))
6429                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6430                 if (btrfs_test_opt(fs_info, NODATACOW))
6431                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6432                                 BTRFS_INODE_NODATASUM;
6433         }
6434
6435         inode_tree_add(inode);
6436
6437         trace_btrfs_inode_new(inode);
6438         btrfs_set_inode_last_trans(trans, inode);
6439
6440         btrfs_update_root_times(trans, root);
6441
6442         ret = btrfs_inode_inherit_props(trans, inode, dir);
6443         if (ret)
6444                 btrfs_err(fs_info,
6445                           "error inheriting props for ino %llu (root %llu): %d",
6446                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6447
6448         return inode;
6449
6450 fail_unlock:
6451         discard_new_inode(inode);
6452 fail:
6453         if (dir && name)
6454                 BTRFS_I(dir)->index_cnt--;
6455         btrfs_free_path(path);
6456         return ERR_PTR(ret);
6457 }
6458
6459 /*
6460  * utility function to add 'inode' into 'parent_inode' with
6461  * a give name and a given sequence number.
6462  * if 'add_backref' is true, also insert a backref from the
6463  * inode to the parent directory.
6464  */
6465 int btrfs_add_link(struct btrfs_trans_handle *trans,
6466                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6467                    const char *name, int name_len, int add_backref, u64 index)
6468 {
6469         int ret = 0;
6470         struct btrfs_key key;
6471         struct btrfs_root *root = parent_inode->root;
6472         u64 ino = btrfs_ino(inode);
6473         u64 parent_ino = btrfs_ino(parent_inode);
6474
6475         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6476                 memcpy(&key, &inode->root->root_key, sizeof(key));
6477         } else {
6478                 key.objectid = ino;
6479                 key.type = BTRFS_INODE_ITEM_KEY;
6480                 key.offset = 0;
6481         }
6482
6483         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6484                 ret = btrfs_add_root_ref(trans, key.objectid,
6485                                          root->root_key.objectid, parent_ino,
6486                                          index, name, name_len);
6487         } else if (add_backref) {
6488                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6489                                              parent_ino, index);
6490         }
6491
6492         /* Nothing to clean up yet */
6493         if (ret)
6494                 return ret;
6495
6496         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6497                                     btrfs_inode_type(&inode->vfs_inode), index);
6498         if (ret == -EEXIST || ret == -EOVERFLOW)
6499                 goto fail_dir_item;
6500         else if (ret) {
6501                 btrfs_abort_transaction(trans, ret);
6502                 return ret;
6503         }
6504
6505         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6506                            name_len * 2);
6507         inode_inc_iversion(&parent_inode->vfs_inode);
6508         /*
6509          * If we are replaying a log tree, we do not want to update the mtime
6510          * and ctime of the parent directory with the current time, since the
6511          * log replay procedure is responsible for setting them to their correct
6512          * values (the ones it had when the fsync was done).
6513          */
6514         if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
6515                 struct timespec64 now = current_time(&parent_inode->vfs_inode);
6516
6517                 parent_inode->vfs_inode.i_mtime = now;
6518                 parent_inode->vfs_inode.i_ctime = now;
6519         }
6520         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6521         if (ret)
6522                 btrfs_abort_transaction(trans, ret);
6523         return ret;
6524
6525 fail_dir_item:
6526         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6527                 u64 local_index;
6528                 int err;
6529                 err = btrfs_del_root_ref(trans, key.objectid,
6530                                          root->root_key.objectid, parent_ino,
6531                                          &local_index, name, name_len);
6532                 if (err)
6533                         btrfs_abort_transaction(trans, err);
6534         } else if (add_backref) {
6535                 u64 local_index;
6536                 int err;
6537
6538                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6539                                           ino, parent_ino, &local_index);
6540                 if (err)
6541                         btrfs_abort_transaction(trans, err);
6542         }
6543
6544         /* Return the original error code */
6545         return ret;
6546 }
6547
6548 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6549                             struct btrfs_inode *dir, struct dentry *dentry,
6550                             struct btrfs_inode *inode, int backref, u64 index)
6551 {
6552         int err = btrfs_add_link(trans, dir, inode,
6553                                  dentry->d_name.name, dentry->d_name.len,
6554                                  backref, index);
6555         if (err > 0)
6556                 err = -EEXIST;
6557         return err;
6558 }
6559
6560 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6561                         umode_t mode, dev_t rdev)
6562 {
6563         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6564         struct btrfs_trans_handle *trans;
6565         struct btrfs_root *root = BTRFS_I(dir)->root;
6566         struct inode *inode = NULL;
6567         int err;
6568         u64 objectid;
6569         u64 index = 0;
6570
6571         /*
6572          * 2 for inode item and ref
6573          * 2 for dir items
6574          * 1 for xattr if selinux is on
6575          */
6576         trans = btrfs_start_transaction(root, 5);
6577         if (IS_ERR(trans))
6578                 return PTR_ERR(trans);
6579
6580         err = btrfs_find_free_ino(root, &objectid);
6581         if (err)
6582                 goto out_unlock;
6583
6584         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6585                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6586                         mode, &index);
6587         if (IS_ERR(inode)) {
6588                 err = PTR_ERR(inode);
6589                 inode = NULL;
6590                 goto out_unlock;
6591         }
6592
6593         /*
6594         * If the active LSM wants to access the inode during
6595         * d_instantiate it needs these. Smack checks to see
6596         * if the filesystem supports xattrs by looking at the
6597         * ops vector.
6598         */
6599         inode->i_op = &btrfs_special_inode_operations;
6600         init_special_inode(inode, inode->i_mode, rdev);
6601
6602         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6603         if (err)
6604                 goto out_unlock;
6605
6606         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6607                         0, index);
6608         if (err)
6609                 goto out_unlock;
6610
6611         btrfs_update_inode(trans, root, inode);
6612         d_instantiate_new(dentry, inode);
6613
6614 out_unlock:
6615         btrfs_end_transaction(trans);
6616         btrfs_btree_balance_dirty(fs_info);
6617         if (err && inode) {
6618                 inode_dec_link_count(inode);
6619                 discard_new_inode(inode);
6620         }
6621         return err;
6622 }
6623
6624 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6625                         umode_t mode, bool excl)
6626 {
6627         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6628         struct btrfs_trans_handle *trans;
6629         struct btrfs_root *root = BTRFS_I(dir)->root;
6630         struct inode *inode = NULL;
6631         int err;
6632         u64 objectid;
6633         u64 index = 0;
6634
6635         /*
6636          * 2 for inode item and ref
6637          * 2 for dir items
6638          * 1 for xattr if selinux is on
6639          */
6640         trans = btrfs_start_transaction(root, 5);
6641         if (IS_ERR(trans))
6642                 return PTR_ERR(trans);
6643
6644         err = btrfs_find_free_ino(root, &objectid);
6645         if (err)
6646                 goto out_unlock;
6647
6648         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6649                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6650                         mode, &index);
6651         if (IS_ERR(inode)) {
6652                 err = PTR_ERR(inode);
6653                 inode = NULL;
6654                 goto out_unlock;
6655         }
6656         /*
6657         * If the active LSM wants to access the inode during
6658         * d_instantiate it needs these. Smack checks to see
6659         * if the filesystem supports xattrs by looking at the
6660         * ops vector.
6661         */
6662         inode->i_fop = &btrfs_file_operations;
6663         inode->i_op = &btrfs_file_inode_operations;
6664         inode->i_mapping->a_ops = &btrfs_aops;
6665
6666         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6667         if (err)
6668                 goto out_unlock;
6669
6670         err = btrfs_update_inode(trans, root, inode);
6671         if (err)
6672                 goto out_unlock;
6673
6674         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6675                         0, index);
6676         if (err)
6677                 goto out_unlock;
6678
6679         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6680         d_instantiate_new(dentry, inode);
6681
6682 out_unlock:
6683         btrfs_end_transaction(trans);
6684         if (err && inode) {
6685                 inode_dec_link_count(inode);
6686                 discard_new_inode(inode);
6687         }
6688         btrfs_btree_balance_dirty(fs_info);
6689         return err;
6690 }
6691
6692 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6693                       struct dentry *dentry)
6694 {
6695         struct btrfs_trans_handle *trans = NULL;
6696         struct btrfs_root *root = BTRFS_I(dir)->root;
6697         struct inode *inode = d_inode(old_dentry);
6698         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6699         u64 index;
6700         int err;
6701         int drop_inode = 0;
6702
6703         /* do not allow sys_link's with other subvols of the same device */
6704         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6705                 return -EXDEV;
6706
6707         if (inode->i_nlink >= BTRFS_LINK_MAX)
6708                 return -EMLINK;
6709
6710         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6711         if (err)
6712                 goto fail;
6713
6714         /*
6715          * 2 items for inode and inode ref
6716          * 2 items for dir items
6717          * 1 item for parent inode
6718          * 1 item for orphan item deletion if O_TMPFILE
6719          */
6720         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6721         if (IS_ERR(trans)) {
6722                 err = PTR_ERR(trans);
6723                 trans = NULL;
6724                 goto fail;
6725         }
6726
6727         /* There are several dir indexes for this inode, clear the cache. */
6728         BTRFS_I(inode)->dir_index = 0ULL;
6729         inc_nlink(inode);
6730         inode_inc_iversion(inode);
6731         inode->i_ctime = current_time(inode);
6732         ihold(inode);
6733         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6734
6735         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6736                         1, index);
6737
6738         if (err) {
6739                 drop_inode = 1;
6740         } else {
6741                 struct dentry *parent = dentry->d_parent;
6742                 int ret;
6743
6744                 err = btrfs_update_inode(trans, root, inode);
6745                 if (err)
6746                         goto fail;
6747                 if (inode->i_nlink == 1) {
6748                         /*
6749                          * If new hard link count is 1, it's a file created
6750                          * with open(2) O_TMPFILE flag.
6751                          */
6752                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6753                         if (err)
6754                                 goto fail;
6755                 }
6756                 d_instantiate(dentry, inode);
6757                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6758                                          true, NULL);
6759                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6760                         err = btrfs_commit_transaction(trans);
6761                         trans = NULL;
6762                 }
6763         }
6764
6765 fail:
6766         if (trans)
6767                 btrfs_end_transaction(trans);
6768         if (drop_inode) {
6769                 inode_dec_link_count(inode);
6770                 iput(inode);
6771         }
6772         btrfs_btree_balance_dirty(fs_info);
6773         return err;
6774 }
6775
6776 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6777 {
6778         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6779         struct inode *inode = NULL;
6780         struct btrfs_trans_handle *trans;
6781         struct btrfs_root *root = BTRFS_I(dir)->root;
6782         int err = 0;
6783         u64 objectid = 0;
6784         u64 index = 0;
6785
6786         /*
6787          * 2 items for inode and ref
6788          * 2 items for dir items
6789          * 1 for xattr if selinux is on
6790          */
6791         trans = btrfs_start_transaction(root, 5);
6792         if (IS_ERR(trans))
6793                 return PTR_ERR(trans);
6794
6795         err = btrfs_find_free_ino(root, &objectid);
6796         if (err)
6797                 goto out_fail;
6798
6799         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6800                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6801                         S_IFDIR | mode, &index);
6802         if (IS_ERR(inode)) {
6803                 err = PTR_ERR(inode);
6804                 inode = NULL;
6805                 goto out_fail;
6806         }
6807
6808         /* these must be set before we unlock the inode */
6809         inode->i_op = &btrfs_dir_inode_operations;
6810         inode->i_fop = &btrfs_dir_file_operations;
6811
6812         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6813         if (err)
6814                 goto out_fail;
6815
6816         btrfs_i_size_write(BTRFS_I(inode), 0);
6817         err = btrfs_update_inode(trans, root, inode);
6818         if (err)
6819                 goto out_fail;
6820
6821         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6822                         dentry->d_name.name,
6823                         dentry->d_name.len, 0, index);
6824         if (err)
6825                 goto out_fail;
6826
6827         d_instantiate_new(dentry, inode);
6828
6829 out_fail:
6830         btrfs_end_transaction(trans);
6831         if (err && inode) {
6832                 inode_dec_link_count(inode);
6833                 discard_new_inode(inode);
6834         }
6835         btrfs_btree_balance_dirty(fs_info);
6836         return err;
6837 }
6838
6839 static noinline int uncompress_inline(struct btrfs_path *path,
6840                                       struct page *page,
6841                                       size_t pg_offset, u64 extent_offset,
6842                                       struct btrfs_file_extent_item *item)
6843 {
6844         int ret;
6845         struct extent_buffer *leaf = path->nodes[0];
6846         char *tmp;
6847         size_t max_size;
6848         unsigned long inline_size;
6849         unsigned long ptr;
6850         int compress_type;
6851
6852         WARN_ON(pg_offset != 0);
6853         compress_type = btrfs_file_extent_compression(leaf, item);
6854         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6855         inline_size = btrfs_file_extent_inline_item_len(leaf,
6856                                         btrfs_item_nr(path->slots[0]));
6857         tmp = kmalloc(inline_size, GFP_NOFS);
6858         if (!tmp)
6859                 return -ENOMEM;
6860         ptr = btrfs_file_extent_inline_start(item);
6861
6862         read_extent_buffer(leaf, tmp, ptr, inline_size);
6863
6864         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6865         ret = btrfs_decompress(compress_type, tmp, page,
6866                                extent_offset, inline_size, max_size);
6867
6868         /*
6869          * decompression code contains a memset to fill in any space between the end
6870          * of the uncompressed data and the end of max_size in case the decompressed
6871          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6872          * the end of an inline extent and the beginning of the next block, so we
6873          * cover that region here.
6874          */
6875
6876         if (max_size + pg_offset < PAGE_SIZE) {
6877                 char *map = kmap(page);
6878                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6879                 kunmap(page);
6880         }
6881         kfree(tmp);
6882         return ret;
6883 }
6884
6885 /*
6886  * a bit scary, this does extent mapping from logical file offset to the disk.
6887  * the ugly parts come from merging extents from the disk with the in-ram
6888  * representation.  This gets more complex because of the data=ordered code,
6889  * where the in-ram extents might be locked pending data=ordered completion.
6890  *
6891  * This also copies inline extents directly into the page.
6892  */
6893 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6894                                     struct page *page,
6895                                     size_t pg_offset, u64 start, u64 len,
6896                                     int create)
6897 {
6898         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6899         int ret;
6900         int err = 0;
6901         u64 extent_start = 0;
6902         u64 extent_end = 0;
6903         u64 objectid = btrfs_ino(inode);
6904         int extent_type = -1;
6905         struct btrfs_path *path = NULL;
6906         struct btrfs_root *root = inode->root;
6907         struct btrfs_file_extent_item *item;
6908         struct extent_buffer *leaf;
6909         struct btrfs_key found_key;
6910         struct extent_map *em = NULL;
6911         struct extent_map_tree *em_tree = &inode->extent_tree;
6912         struct extent_io_tree *io_tree = &inode->io_tree;
6913         const bool new_inline = !page || create;
6914
6915         read_lock(&em_tree->lock);
6916         em = lookup_extent_mapping(em_tree, start, len);
6917         if (em)
6918                 em->bdev = fs_info->fs_devices->latest_bdev;
6919         read_unlock(&em_tree->lock);
6920
6921         if (em) {
6922                 if (em->start > start || em->start + em->len <= start)
6923                         free_extent_map(em);
6924                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6925                         free_extent_map(em);
6926                 else
6927                         goto out;
6928         }
6929         em = alloc_extent_map();
6930         if (!em) {
6931                 err = -ENOMEM;
6932                 goto out;
6933         }
6934         em->bdev = fs_info->fs_devices->latest_bdev;
6935         em->start = EXTENT_MAP_HOLE;
6936         em->orig_start = EXTENT_MAP_HOLE;
6937         em->len = (u64)-1;
6938         em->block_len = (u64)-1;
6939
6940         path = btrfs_alloc_path();
6941         if (!path) {
6942                 err = -ENOMEM;
6943                 goto out;
6944         }
6945
6946         /* Chances are we'll be called again, so go ahead and do readahead */
6947         path->reada = READA_FORWARD;
6948
6949         /*
6950          * Unless we're going to uncompress the inline extent, no sleep would
6951          * happen.
6952          */
6953         path->leave_spinning = 1;
6954
6955         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6956         if (ret < 0) {
6957                 err = ret;
6958                 goto out;
6959         } else if (ret > 0) {
6960                 if (path->slots[0] == 0)
6961                         goto not_found;
6962                 path->slots[0]--;
6963         }
6964
6965         leaf = path->nodes[0];
6966         item = btrfs_item_ptr(leaf, path->slots[0],
6967                               struct btrfs_file_extent_item);
6968         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6969         if (found_key.objectid != objectid ||
6970             found_key.type != BTRFS_EXTENT_DATA_KEY) {
6971                 /*
6972                  * If we backup past the first extent we want to move forward
6973                  * and see if there is an extent in front of us, otherwise we'll
6974                  * say there is a hole for our whole search range which can
6975                  * cause problems.
6976                  */
6977                 extent_end = start;
6978                 goto next;
6979         }
6980
6981         extent_type = btrfs_file_extent_type(leaf, item);
6982         extent_start = found_key.offset;
6983         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6984             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6985                 /* Only regular file could have regular/prealloc extent */
6986                 if (!S_ISREG(inode->vfs_inode.i_mode)) {
6987                         ret = -EUCLEAN;
6988                         btrfs_crit(fs_info,
6989                 "regular/prealloc extent found for non-regular inode %llu",
6990                                    btrfs_ino(inode));
6991                         goto out;
6992                 }
6993                 extent_end = extent_start +
6994                        btrfs_file_extent_num_bytes(leaf, item);
6995
6996                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6997                                                        extent_start);
6998         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
6999                 size_t size;
7000
7001                 size = btrfs_file_extent_ram_bytes(leaf, item);
7002                 extent_end = ALIGN(extent_start + size,
7003                                    fs_info->sectorsize);
7004
7005                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7006                                                       path->slots[0],
7007                                                       extent_start);
7008         }
7009 next:
7010         if (start >= extent_end) {
7011                 path->slots[0]++;
7012                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7013                         ret = btrfs_next_leaf(root, path);
7014                         if (ret < 0) {
7015                                 err = ret;
7016                                 goto out;
7017                         } else if (ret > 0) {
7018                                 goto not_found;
7019                         }
7020                         leaf = path->nodes[0];
7021                 }
7022                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7023                 if (found_key.objectid != objectid ||
7024                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7025                         goto not_found;
7026                 if (start + len <= found_key.offset)
7027                         goto not_found;
7028                 if (start > found_key.offset)
7029                         goto next;
7030
7031                 /* New extent overlaps with existing one */
7032                 em->start = start;
7033                 em->orig_start = start;
7034                 em->len = found_key.offset - start;
7035                 em->block_start = EXTENT_MAP_HOLE;
7036                 goto insert;
7037         }
7038
7039         btrfs_extent_item_to_extent_map(inode, path, item,
7040                         new_inline, em);
7041
7042         if (extent_type == BTRFS_FILE_EXTENT_REG ||
7043             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
7044                 goto insert;
7045         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
7046                 unsigned long ptr;
7047                 char *map;
7048                 size_t size;
7049                 size_t extent_offset;
7050                 size_t copy_size;
7051
7052                 if (new_inline)
7053                         goto out;
7054
7055                 size = btrfs_file_extent_ram_bytes(leaf, item);
7056                 extent_offset = page_offset(page) + pg_offset - extent_start;
7057                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7058                                   size - extent_offset);
7059                 em->start = extent_start + extent_offset;
7060                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7061                 em->orig_block_len = em->len;
7062                 em->orig_start = em->start;
7063                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7064
7065                 btrfs_set_path_blocking(path);
7066                 if (!PageUptodate(page)) {
7067                         if (btrfs_file_extent_compression(leaf, item) !=
7068                             BTRFS_COMPRESS_NONE) {
7069                                 ret = uncompress_inline(path, page, pg_offset,
7070                                                         extent_offset, item);
7071                                 if (ret) {
7072                                         err = ret;
7073                                         goto out;
7074                                 }
7075                         } else {
7076                                 map = kmap(page);
7077                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7078                                                    copy_size);
7079                                 if (pg_offset + copy_size < PAGE_SIZE) {
7080                                         memset(map + pg_offset + copy_size, 0,
7081                                                PAGE_SIZE - pg_offset -
7082                                                copy_size);
7083                                 }
7084                                 kunmap(page);
7085                         }
7086                         flush_dcache_page(page);
7087                 }
7088                 set_extent_uptodate(io_tree, em->start,
7089                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7090                 goto insert;
7091         }
7092 not_found:
7093         em->start = start;
7094         em->orig_start = start;
7095         em->len = len;
7096         em->block_start = EXTENT_MAP_HOLE;
7097 insert:
7098         btrfs_release_path(path);
7099         if (em->start > start || extent_map_end(em) <= start) {
7100                 btrfs_err(fs_info,
7101                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7102                           em->start, em->len, start, len);
7103                 err = -EIO;
7104                 goto out;
7105         }
7106
7107         err = 0;
7108         write_lock(&em_tree->lock);
7109         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
7110         write_unlock(&em_tree->lock);
7111 out:
7112         btrfs_free_path(path);
7113
7114         trace_btrfs_get_extent(root, inode, em);
7115
7116         if (err) {
7117                 free_extent_map(em);
7118                 return ERR_PTR(err);
7119         }
7120         BUG_ON(!em); /* Error is always set */
7121         return em;
7122 }
7123
7124 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7125                                            u64 start, u64 len)
7126 {
7127         struct extent_map *em;
7128         struct extent_map *hole_em = NULL;
7129         u64 delalloc_start = start;
7130         u64 end;
7131         u64 delalloc_len;
7132         u64 delalloc_end;
7133         int err = 0;
7134
7135         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7136         if (IS_ERR(em))
7137                 return em;
7138         /*
7139          * If our em maps to:
7140          * - a hole or
7141          * - a pre-alloc extent,
7142          * there might actually be delalloc bytes behind it.
7143          */
7144         if (em->block_start != EXTENT_MAP_HOLE &&
7145             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7146                 return em;
7147         else
7148                 hole_em = em;
7149
7150         /* check to see if we've wrapped (len == -1 or similar) */
7151         end = start + len;
7152         if (end < start)
7153                 end = (u64)-1;
7154         else
7155                 end -= 1;
7156
7157         em = NULL;
7158
7159         /* ok, we didn't find anything, lets look for delalloc */
7160         delalloc_len = count_range_bits(&inode->io_tree, &delalloc_start,
7161                                  end, len, EXTENT_DELALLOC, 1);
7162         delalloc_end = delalloc_start + delalloc_len;
7163         if (delalloc_end < delalloc_start)
7164                 delalloc_end = (u64)-1;
7165
7166         /*
7167          * We didn't find anything useful, return the original results from
7168          * get_extent()
7169          */
7170         if (delalloc_start > end || delalloc_end <= start) {
7171                 em = hole_em;
7172                 hole_em = NULL;
7173                 goto out;
7174         }
7175
7176         /*
7177          * Adjust the delalloc_start to make sure it doesn't go backwards from
7178          * the start they passed in
7179          */
7180         delalloc_start = max(start, delalloc_start);
7181         delalloc_len = delalloc_end - delalloc_start;
7182
7183         if (delalloc_len > 0) {
7184                 u64 hole_start;
7185                 u64 hole_len;
7186                 const u64 hole_end = extent_map_end(hole_em);
7187
7188                 em = alloc_extent_map();
7189                 if (!em) {
7190                         err = -ENOMEM;
7191                         goto out;
7192                 }
7193                 em->bdev = NULL;
7194
7195                 ASSERT(hole_em);
7196                 /*
7197                  * When btrfs_get_extent can't find anything it returns one
7198                  * huge hole
7199                  *
7200                  * Make sure what it found really fits our range, and adjust to
7201                  * make sure it is based on the start from the caller
7202                  */
7203                 if (hole_end <= start || hole_em->start > end) {
7204                        free_extent_map(hole_em);
7205                        hole_em = NULL;
7206                 } else {
7207                        hole_start = max(hole_em->start, start);
7208                        hole_len = hole_end - hole_start;
7209                 }
7210
7211                 if (hole_em && delalloc_start > hole_start) {
7212                         /*
7213                          * Our hole starts before our delalloc, so we have to
7214                          * return just the parts of the hole that go until the
7215                          * delalloc starts
7216                          */
7217                         em->len = min(hole_len, delalloc_start - hole_start);
7218                         em->start = hole_start;
7219                         em->orig_start = hole_start;
7220                         /*
7221                          * Don't adjust block start at all, it is fixed at
7222                          * EXTENT_MAP_HOLE
7223                          */
7224                         em->block_start = hole_em->block_start;
7225                         em->block_len = hole_len;
7226                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7227                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7228                 } else {
7229                         /*
7230                          * Hole is out of passed range or it starts after
7231                          * delalloc range
7232                          */
7233                         em->start = delalloc_start;
7234                         em->len = delalloc_len;
7235                         em->orig_start = delalloc_start;
7236                         em->block_start = EXTENT_MAP_DELALLOC;
7237                         em->block_len = delalloc_len;
7238                 }
7239         } else {
7240                 return hole_em;
7241         }
7242 out:
7243
7244         free_extent_map(hole_em);
7245         if (err) {
7246                 free_extent_map(em);
7247                 return ERR_PTR(err);
7248         }
7249         return em;
7250 }
7251
7252 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7253                                                   const u64 start,
7254                                                   const u64 len,
7255                                                   const u64 orig_start,
7256                                                   const u64 block_start,
7257                                                   const u64 block_len,
7258                                                   const u64 orig_block_len,
7259                                                   const u64 ram_bytes,
7260                                                   const int type)
7261 {
7262         struct extent_map *em = NULL;
7263         int ret;
7264
7265         if (type != BTRFS_ORDERED_NOCOW) {
7266                 em = create_io_em(inode, start, len, orig_start,
7267                                   block_start, block_len, orig_block_len,
7268                                   ram_bytes,
7269                                   BTRFS_COMPRESS_NONE, /* compress_type */
7270                                   type);
7271                 if (IS_ERR(em))
7272                         goto out;
7273         }
7274         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7275                                            len, block_len, type);
7276         if (ret) {
7277                 if (em) {
7278                         free_extent_map(em);
7279                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7280                                                 start + len - 1, 0);
7281                 }
7282                 em = ERR_PTR(ret);
7283         }
7284  out:
7285
7286         return em;
7287 }
7288
7289 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7290                                                   u64 start, u64 len)
7291 {
7292         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7293         struct btrfs_root *root = BTRFS_I(inode)->root;
7294         struct extent_map *em;
7295         struct btrfs_key ins;
7296         u64 alloc_hint;
7297         int ret;
7298
7299         alloc_hint = get_extent_allocation_hint(inode, start, len);
7300         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7301                                    0, alloc_hint, &ins, 1, 1);
7302         if (ret)
7303                 return ERR_PTR(ret);
7304
7305         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7306                                      ins.objectid, ins.offset, ins.offset,
7307                                      ins.offset, BTRFS_ORDERED_REGULAR);
7308         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7309         if (IS_ERR(em))
7310                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7311                                            ins.offset, 1);
7312
7313         return em;
7314 }
7315
7316 /*
7317  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7318  * block must be cow'd
7319  */
7320 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7321                               u64 *orig_start, u64 *orig_block_len,
7322                               u64 *ram_bytes)
7323 {
7324         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7325         struct btrfs_path *path;
7326         int ret;
7327         struct extent_buffer *leaf;
7328         struct btrfs_root *root = BTRFS_I(inode)->root;
7329         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7330         struct btrfs_file_extent_item *fi;
7331         struct btrfs_key key;
7332         u64 disk_bytenr;
7333         u64 backref_offset;
7334         u64 extent_end;
7335         u64 num_bytes;
7336         int slot;
7337         int found_type;
7338         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7339
7340         path = btrfs_alloc_path();
7341         if (!path)
7342                 return -ENOMEM;
7343
7344         ret = btrfs_lookup_file_extent(NULL, root, path,
7345                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7346         if (ret < 0)
7347                 goto out;
7348
7349         slot = path->slots[0];
7350         if (ret == 1) {
7351                 if (slot == 0) {
7352                         /* can't find the item, must cow */
7353                         ret = 0;
7354                         goto out;
7355                 }
7356                 slot--;
7357         }
7358         ret = 0;
7359         leaf = path->nodes[0];
7360         btrfs_item_key_to_cpu(leaf, &key, slot);
7361         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7362             key.type != BTRFS_EXTENT_DATA_KEY) {
7363                 /* not our file or wrong item type, must cow */
7364                 goto out;
7365         }
7366
7367         if (key.offset > offset) {
7368                 /* Wrong offset, must cow */
7369                 goto out;
7370         }
7371
7372         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7373         found_type = btrfs_file_extent_type(leaf, fi);
7374         if (found_type != BTRFS_FILE_EXTENT_REG &&
7375             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7376                 /* not a regular extent, must cow */
7377                 goto out;
7378         }
7379
7380         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7381                 goto out;
7382
7383         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7384         if (extent_end <= offset)
7385                 goto out;
7386
7387         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7388         if (disk_bytenr == 0)
7389                 goto out;
7390
7391         if (btrfs_file_extent_compression(leaf, fi) ||
7392             btrfs_file_extent_encryption(leaf, fi) ||
7393             btrfs_file_extent_other_encoding(leaf, fi))
7394                 goto out;
7395
7396         /*
7397          * Do the same check as in btrfs_cross_ref_exist but without the
7398          * unnecessary search.
7399          */
7400         if (btrfs_file_extent_generation(leaf, fi) <=
7401             btrfs_root_last_snapshot(&root->root_item))
7402                 goto out;
7403
7404         backref_offset = btrfs_file_extent_offset(leaf, fi);
7405
7406         if (orig_start) {
7407                 *orig_start = key.offset - backref_offset;
7408                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7409                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7410         }
7411
7412         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7413                 goto out;
7414
7415         num_bytes = min(offset + *len, extent_end) - offset;
7416         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7417                 u64 range_end;
7418
7419                 range_end = round_up(offset + num_bytes,
7420                                      root->fs_info->sectorsize) - 1;
7421                 ret = test_range_bit(io_tree, offset, range_end,
7422                                      EXTENT_DELALLOC, 0, NULL);
7423                 if (ret) {
7424                         ret = -EAGAIN;
7425                         goto out;
7426                 }
7427         }
7428
7429         btrfs_release_path(path);
7430
7431         /*
7432          * look for other files referencing this extent, if we
7433          * find any we must cow
7434          */
7435
7436         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7437                                     key.offset - backref_offset, disk_bytenr);
7438         if (ret) {
7439                 ret = 0;
7440                 goto out;
7441         }
7442
7443         /*
7444          * adjust disk_bytenr and num_bytes to cover just the bytes
7445          * in this extent we are about to write.  If there
7446          * are any csums in that range we have to cow in order
7447          * to keep the csums correct
7448          */
7449         disk_bytenr += backref_offset;
7450         disk_bytenr += offset - key.offset;
7451         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7452                 goto out;
7453         /*
7454          * all of the above have passed, it is safe to overwrite this extent
7455          * without cow
7456          */
7457         *len = num_bytes;
7458         ret = 1;
7459 out:
7460         btrfs_free_path(path);
7461         return ret;
7462 }
7463
7464 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7465                               struct extent_state **cached_state, int writing)
7466 {
7467         struct btrfs_ordered_extent *ordered;
7468         int ret = 0;
7469
7470         while (1) {
7471                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7472                                  cached_state);
7473                 /*
7474                  * We're concerned with the entire range that we're going to be
7475                  * doing DIO to, so we need to make sure there's no ordered
7476                  * extents in this range.
7477                  */
7478                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7479                                                      lockend - lockstart + 1);
7480
7481                 /*
7482                  * We need to make sure there are no buffered pages in this
7483                  * range either, we could have raced between the invalidate in
7484                  * generic_file_direct_write and locking the extent.  The
7485                  * invalidate needs to happen so that reads after a write do not
7486                  * get stale data.
7487                  */
7488                 if (!ordered &&
7489                     (!writing || !filemap_range_has_page(inode->i_mapping,
7490                                                          lockstart, lockend)))
7491                         break;
7492
7493                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7494                                      cached_state);
7495
7496                 if (ordered) {
7497                         /*
7498                          * If we are doing a DIO read and the ordered extent we
7499                          * found is for a buffered write, we can not wait for it
7500                          * to complete and retry, because if we do so we can
7501                          * deadlock with concurrent buffered writes on page
7502                          * locks. This happens only if our DIO read covers more
7503                          * than one extent map, if at this point has already
7504                          * created an ordered extent for a previous extent map
7505                          * and locked its range in the inode's io tree, and a
7506                          * concurrent write against that previous extent map's
7507                          * range and this range started (we unlock the ranges
7508                          * in the io tree only when the bios complete and
7509                          * buffered writes always lock pages before attempting
7510                          * to lock range in the io tree).
7511                          */
7512                         if (writing ||
7513                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7514                                 btrfs_start_ordered_extent(inode, ordered, 1);
7515                         else
7516                                 ret = -ENOTBLK;
7517                         btrfs_put_ordered_extent(ordered);
7518                 } else {
7519                         /*
7520                          * We could trigger writeback for this range (and wait
7521                          * for it to complete) and then invalidate the pages for
7522                          * this range (through invalidate_inode_pages2_range()),
7523                          * but that can lead us to a deadlock with a concurrent
7524                          * call to readpages() (a buffered read or a defrag call
7525                          * triggered a readahead) on a page lock due to an
7526                          * ordered dio extent we created before but did not have
7527                          * yet a corresponding bio submitted (whence it can not
7528                          * complete), which makes readpages() wait for that
7529                          * ordered extent to complete while holding a lock on
7530                          * that page.
7531                          */
7532                         ret = -ENOTBLK;
7533                 }
7534
7535                 if (ret)
7536                         break;
7537
7538                 cond_resched();
7539         }
7540
7541         return ret;
7542 }
7543
7544 /* The callers of this must take lock_extent() */
7545 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7546                                        u64 orig_start, u64 block_start,
7547                                        u64 block_len, u64 orig_block_len,
7548                                        u64 ram_bytes, int compress_type,
7549                                        int type)
7550 {
7551         struct extent_map_tree *em_tree;
7552         struct extent_map *em;
7553         struct btrfs_root *root = BTRFS_I(inode)->root;
7554         int ret;
7555
7556         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7557                type == BTRFS_ORDERED_COMPRESSED ||
7558                type == BTRFS_ORDERED_NOCOW ||
7559                type == BTRFS_ORDERED_REGULAR);
7560
7561         em_tree = &BTRFS_I(inode)->extent_tree;
7562         em = alloc_extent_map();
7563         if (!em)
7564                 return ERR_PTR(-ENOMEM);
7565
7566         em->start = start;
7567         em->orig_start = orig_start;
7568         em->len = len;
7569         em->block_len = block_len;
7570         em->block_start = block_start;
7571         em->bdev = root->fs_info->fs_devices->latest_bdev;
7572         em->orig_block_len = orig_block_len;
7573         em->ram_bytes = ram_bytes;
7574         em->generation = -1;
7575         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7576         if (type == BTRFS_ORDERED_PREALLOC) {
7577                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7578         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7579                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7580                 em->compress_type = compress_type;
7581         }
7582
7583         do {
7584                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7585                                 em->start + em->len - 1, 0);
7586                 write_lock(&em_tree->lock);
7587                 ret = add_extent_mapping(em_tree, em, 1);
7588                 write_unlock(&em_tree->lock);
7589                 /*
7590                  * The caller has taken lock_extent(), who could race with us
7591                  * to add em?
7592                  */
7593         } while (ret == -EEXIST);
7594
7595         if (ret) {
7596                 free_extent_map(em);
7597                 return ERR_PTR(ret);
7598         }
7599
7600         /* em got 2 refs now, callers needs to do free_extent_map once. */
7601         return em;
7602 }
7603
7604
7605 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7606                                         struct buffer_head *bh_result,
7607                                         struct inode *inode,
7608                                         u64 start, u64 len)
7609 {
7610         if (em->block_start == EXTENT_MAP_HOLE ||
7611                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7612                 return -ENOENT;
7613
7614         len = min(len, em->len - (start - em->start));
7615
7616         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7617                 inode->i_blkbits;
7618         bh_result->b_size = len;
7619         bh_result->b_bdev = em->bdev;
7620         set_buffer_mapped(bh_result);
7621
7622         return 0;
7623 }
7624
7625 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7626                                          struct buffer_head *bh_result,
7627                                          struct inode *inode,
7628                                          struct btrfs_dio_data *dio_data,
7629                                          u64 start, u64 len)
7630 {
7631         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7632         struct extent_map *em = *map;
7633         int ret = 0;
7634
7635         /*
7636          * We don't allocate a new extent in the following cases
7637          *
7638          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7639          * existing extent.
7640          * 2) The extent is marked as PREALLOC. We're good to go here and can
7641          * just use the extent.
7642          *
7643          */
7644         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7645             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7646              em->block_start != EXTENT_MAP_HOLE)) {
7647                 int type;
7648                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7649
7650                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7651                         type = BTRFS_ORDERED_PREALLOC;
7652                 else
7653                         type = BTRFS_ORDERED_NOCOW;
7654                 len = min(len, em->len - (start - em->start));
7655                 block_start = em->block_start + (start - em->start);
7656
7657                 if (can_nocow_extent(inode, start, &len, &orig_start,
7658                                      &orig_block_len, &ram_bytes) == 1 &&
7659                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7660                         struct extent_map *em2;
7661
7662                         em2 = btrfs_create_dio_extent(inode, start, len,
7663                                                       orig_start, block_start,
7664                                                       len, orig_block_len,
7665                                                       ram_bytes, type);
7666                         btrfs_dec_nocow_writers(fs_info, block_start);
7667                         if (type == BTRFS_ORDERED_PREALLOC) {
7668                                 free_extent_map(em);
7669                                 *map = em = em2;
7670                         }
7671
7672                         if (em2 && IS_ERR(em2)) {
7673                                 ret = PTR_ERR(em2);
7674                                 goto out;
7675                         }
7676                         /*
7677                          * For inode marked NODATACOW or extent marked PREALLOC,
7678                          * use the existing or preallocated extent, so does not
7679                          * need to adjust btrfs_space_info's bytes_may_use.
7680                          */
7681                         btrfs_free_reserved_data_space_noquota(inode, start,
7682                                                                len);
7683                         goto skip_cow;
7684                 }
7685         }
7686
7687         /* this will cow the extent */
7688         len = bh_result->b_size;
7689         free_extent_map(em);
7690         *map = em = btrfs_new_extent_direct(inode, start, len);
7691         if (IS_ERR(em)) {
7692                 ret = PTR_ERR(em);
7693                 goto out;
7694         }
7695
7696         len = min(len, em->len - (start - em->start));
7697
7698 skip_cow:
7699         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7700                 inode->i_blkbits;
7701         bh_result->b_size = len;
7702         bh_result->b_bdev = em->bdev;
7703         set_buffer_mapped(bh_result);
7704
7705         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7706                 set_buffer_new(bh_result);
7707
7708         /*
7709          * Need to update the i_size under the extent lock so buffered
7710          * readers will get the updated i_size when we unlock.
7711          */
7712         if (!dio_data->overwrite && start + len > i_size_read(inode))
7713                 i_size_write(inode, start + len);
7714
7715         WARN_ON(dio_data->reserve < len);
7716         dio_data->reserve -= len;
7717         dio_data->unsubmitted_oe_range_end = start + len;
7718         current->journal_info = dio_data;
7719 out:
7720         return ret;
7721 }
7722
7723 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7724                                    struct buffer_head *bh_result, int create)
7725 {
7726         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7727         struct extent_map *em;
7728         struct extent_state *cached_state = NULL;
7729         struct btrfs_dio_data *dio_data = NULL;
7730         u64 start = iblock << inode->i_blkbits;
7731         u64 lockstart, lockend;
7732         u64 len = bh_result->b_size;
7733         int unlock_bits = EXTENT_LOCKED;
7734         int ret = 0;
7735
7736         if (create)
7737                 unlock_bits |= EXTENT_DIRTY;
7738         else
7739                 len = min_t(u64, len, fs_info->sectorsize);
7740
7741         lockstart = start;
7742         lockend = start + len - 1;
7743
7744         if (current->journal_info) {
7745                 /*
7746                  * Need to pull our outstanding extents and set journal_info to NULL so
7747                  * that anything that needs to check if there's a transaction doesn't get
7748                  * confused.
7749                  */
7750                 dio_data = current->journal_info;
7751                 current->journal_info = NULL;
7752         }
7753
7754         /*
7755          * If this errors out it's because we couldn't invalidate pagecache for
7756          * this range and we need to fallback to buffered.
7757          */
7758         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7759                                create)) {
7760                 ret = -ENOTBLK;
7761                 goto err;
7762         }
7763
7764         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7765         if (IS_ERR(em)) {
7766                 ret = PTR_ERR(em);
7767                 goto unlock_err;
7768         }
7769
7770         /*
7771          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7772          * io.  INLINE is special, and we could probably kludge it in here, but
7773          * it's still buffered so for safety lets just fall back to the generic
7774          * buffered path.
7775          *
7776          * For COMPRESSED we _have_ to read the entire extent in so we can
7777          * decompress it, so there will be buffering required no matter what we
7778          * do, so go ahead and fallback to buffered.
7779          *
7780          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7781          * to buffered IO.  Don't blame me, this is the price we pay for using
7782          * the generic code.
7783          */
7784         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7785             em->block_start == EXTENT_MAP_INLINE) {
7786                 free_extent_map(em);
7787                 ret = -ENOTBLK;
7788                 goto unlock_err;
7789         }
7790
7791         if (create) {
7792                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7793                                                     dio_data, start, len);
7794                 if (ret < 0)
7795                         goto unlock_err;
7796
7797                 /* clear and unlock the entire range */
7798                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7799                                  unlock_bits, 1, 0, &cached_state);
7800         } else {
7801                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7802                                                    start, len);
7803                 /* Can be negative only if we read from a hole */
7804                 if (ret < 0) {
7805                         ret = 0;
7806                         free_extent_map(em);
7807                         goto unlock_err;
7808                 }
7809                 /*
7810                  * We need to unlock only the end area that we aren't using.
7811                  * The rest is going to be unlocked by the endio routine.
7812                  */
7813                 lockstart = start + bh_result->b_size;
7814                 if (lockstart < lockend) {
7815                         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7816                                          lockend, unlock_bits, 1, 0,
7817                                          &cached_state);
7818                 } else {
7819                         free_extent_state(cached_state);
7820                 }
7821         }
7822
7823         free_extent_map(em);
7824
7825         return 0;
7826
7827 unlock_err:
7828         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7829                          unlock_bits, 1, 0, &cached_state);
7830 err:
7831         if (dio_data)
7832                 current->journal_info = dio_data;
7833         return ret;
7834 }
7835
7836 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7837                                                  struct bio *bio,
7838                                                  int mirror_num)
7839 {
7840         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7841         blk_status_t ret;
7842
7843         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7844
7845         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7846         if (ret)
7847                 return ret;
7848
7849         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7850
7851         return ret;
7852 }
7853
7854 static int btrfs_check_dio_repairable(struct inode *inode,
7855                                       struct bio *failed_bio,
7856                                       struct io_failure_record *failrec,
7857                                       int failed_mirror)
7858 {
7859         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7860         int num_copies;
7861
7862         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7863         if (num_copies == 1) {
7864                 /*
7865                  * we only have a single copy of the data, so don't bother with
7866                  * all the retry and error correction code that follows. no
7867                  * matter what the error is, it is very likely to persist.
7868                  */
7869                 btrfs_debug(fs_info,
7870                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7871                         num_copies, failrec->this_mirror, failed_mirror);
7872                 return 0;
7873         }
7874
7875         failrec->failed_mirror = failed_mirror;
7876         failrec->this_mirror++;
7877         if (failrec->this_mirror == failed_mirror)
7878                 failrec->this_mirror++;
7879
7880         if (failrec->this_mirror > num_copies) {
7881                 btrfs_debug(fs_info,
7882                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7883                         num_copies, failrec->this_mirror, failed_mirror);
7884                 return 0;
7885         }
7886
7887         return 1;
7888 }
7889
7890 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7891                                    struct page *page, unsigned int pgoff,
7892                                    u64 start, u64 end, int failed_mirror,
7893                                    bio_end_io_t *repair_endio, void *repair_arg)
7894 {
7895         struct io_failure_record *failrec;
7896         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7897         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7898         struct bio *bio;
7899         int isector;
7900         unsigned int read_mode = 0;
7901         int segs;
7902         int ret;
7903         blk_status_t status;
7904         struct bio_vec bvec;
7905
7906         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7907
7908         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7909         if (ret)
7910                 return errno_to_blk_status(ret);
7911
7912         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7913                                          failed_mirror);
7914         if (!ret) {
7915                 free_io_failure(failure_tree, io_tree, failrec);
7916                 return BLK_STS_IOERR;
7917         }
7918
7919         segs = bio_segments(failed_bio);
7920         bio_get_first_bvec(failed_bio, &bvec);
7921         if (segs > 1 ||
7922             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7923                 read_mode |= REQ_FAILFAST_DEV;
7924
7925         isector = start - btrfs_io_bio(failed_bio)->logical;
7926         isector >>= inode->i_sb->s_blocksize_bits;
7927         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7928                                 pgoff, isector, repair_endio, repair_arg);
7929         bio->bi_opf = REQ_OP_READ | read_mode;
7930
7931         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7932                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7933                     read_mode, failrec->this_mirror, failrec->in_validation);
7934
7935         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7936         if (status) {
7937                 free_io_failure(failure_tree, io_tree, failrec);
7938                 bio_put(bio);
7939         }
7940
7941         return status;
7942 }
7943
7944 struct btrfs_retry_complete {
7945         struct completion done;
7946         struct inode *inode;
7947         u64 start;
7948         int uptodate;
7949 };
7950
7951 static void btrfs_retry_endio_nocsum(struct bio *bio)
7952 {
7953         struct btrfs_retry_complete *done = bio->bi_private;
7954         struct inode *inode = done->inode;
7955         struct bio_vec *bvec;
7956         struct extent_io_tree *io_tree, *failure_tree;
7957         struct bvec_iter_all iter_all;
7958
7959         if (bio->bi_status)
7960                 goto end;
7961
7962         ASSERT(bio->bi_vcnt == 1);
7963         io_tree = &BTRFS_I(inode)->io_tree;
7964         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7965         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7966
7967         done->uptodate = 1;
7968         ASSERT(!bio_flagged(bio, BIO_CLONED));
7969         bio_for_each_segment_all(bvec, bio, iter_all)
7970                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7971                                  io_tree, done->start, bvec->bv_page,
7972                                  btrfs_ino(BTRFS_I(inode)), 0);
7973 end:
7974         complete(&done->done);
7975         bio_put(bio);
7976 }
7977
7978 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7979                                                 struct btrfs_io_bio *io_bio)
7980 {
7981         struct btrfs_fs_info *fs_info;
7982         struct bio_vec bvec;
7983         struct bvec_iter iter;
7984         struct btrfs_retry_complete done;
7985         u64 start;
7986         unsigned int pgoff;
7987         u32 sectorsize;
7988         int nr_sectors;
7989         blk_status_t ret;
7990         blk_status_t err = BLK_STS_OK;
7991
7992         fs_info = BTRFS_I(inode)->root->fs_info;
7993         sectorsize = fs_info->sectorsize;
7994
7995         start = io_bio->logical;
7996         done.inode = inode;
7997         io_bio->bio.bi_iter = io_bio->iter;
7998
7999         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8000                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8001                 pgoff = bvec.bv_offset;
8002
8003 next_block_or_try_again:
8004                 done.uptodate = 0;
8005                 done.start = start;
8006                 init_completion(&done.done);
8007
8008                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8009                                 pgoff, start, start + sectorsize - 1,
8010                                 io_bio->mirror_num,
8011                                 btrfs_retry_endio_nocsum, &done);
8012                 if (ret) {
8013                         err = ret;
8014                         goto next;
8015                 }
8016
8017                 wait_for_completion_io(&done.done);
8018
8019                 if (!done.uptodate) {
8020                         /* We might have another mirror, so try again */
8021                         goto next_block_or_try_again;
8022                 }
8023
8024 next:
8025                 start += sectorsize;
8026
8027                 nr_sectors--;
8028                 if (nr_sectors) {
8029                         pgoff += sectorsize;
8030                         ASSERT(pgoff < PAGE_SIZE);
8031                         goto next_block_or_try_again;
8032                 }
8033         }
8034
8035         return err;
8036 }
8037
8038 static void btrfs_retry_endio(struct bio *bio)
8039 {
8040         struct btrfs_retry_complete *done = bio->bi_private;
8041         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8042         struct extent_io_tree *io_tree, *failure_tree;
8043         struct inode *inode = done->inode;
8044         struct bio_vec *bvec;
8045         int uptodate;
8046         int ret;
8047         int i = 0;
8048         struct bvec_iter_all iter_all;
8049
8050         if (bio->bi_status)
8051                 goto end;
8052
8053         uptodate = 1;
8054
8055         ASSERT(bio->bi_vcnt == 1);
8056         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
8057
8058         io_tree = &BTRFS_I(inode)->io_tree;
8059         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8060
8061         ASSERT(!bio_flagged(bio, BIO_CLONED));
8062         bio_for_each_segment_all(bvec, bio, iter_all) {
8063                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8064                                              bvec->bv_offset, done->start,
8065                                              bvec->bv_len);
8066                 if (!ret)
8067                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8068                                          failure_tree, io_tree, done->start,
8069                                          bvec->bv_page,
8070                                          btrfs_ino(BTRFS_I(inode)),
8071                                          bvec->bv_offset);
8072                 else
8073                         uptodate = 0;
8074                 i++;
8075         }
8076
8077         done->uptodate = uptodate;
8078 end:
8079         complete(&done->done);
8080         bio_put(bio);
8081 }
8082
8083 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8084                 struct btrfs_io_bio *io_bio, blk_status_t err)
8085 {
8086         struct btrfs_fs_info *fs_info;
8087         struct bio_vec bvec;
8088         struct bvec_iter iter;
8089         struct btrfs_retry_complete done;
8090         u64 start;
8091         u64 offset = 0;
8092         u32 sectorsize;
8093         int nr_sectors;
8094         unsigned int pgoff;
8095         int csum_pos;
8096         bool uptodate = (err == 0);
8097         int ret;
8098         blk_status_t status;
8099
8100         fs_info = BTRFS_I(inode)->root->fs_info;
8101         sectorsize = fs_info->sectorsize;
8102
8103         err = BLK_STS_OK;
8104         start = io_bio->logical;
8105         done.inode = inode;
8106         io_bio->bio.bi_iter = io_bio->iter;
8107
8108         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8109                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8110
8111                 pgoff = bvec.bv_offset;
8112 next_block:
8113                 if (uptodate) {
8114                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8115                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8116                                         bvec.bv_page, pgoff, start, sectorsize);
8117                         if (likely(!ret))
8118                                 goto next;
8119                 }
8120 try_again:
8121                 done.uptodate = 0;
8122                 done.start = start;
8123                 init_completion(&done.done);
8124
8125                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8126                                         pgoff, start, start + sectorsize - 1,
8127                                         io_bio->mirror_num, btrfs_retry_endio,
8128                                         &done);
8129                 if (status) {
8130                         err = status;
8131                         goto next;
8132                 }
8133
8134                 wait_for_completion_io(&done.done);
8135
8136                 if (!done.uptodate) {
8137                         /* We might have another mirror, so try again */
8138                         goto try_again;
8139                 }
8140 next:
8141                 offset += sectorsize;
8142                 start += sectorsize;
8143
8144                 ASSERT(nr_sectors);
8145
8146                 nr_sectors--;
8147                 if (nr_sectors) {
8148                         pgoff += sectorsize;
8149                         ASSERT(pgoff < PAGE_SIZE);
8150                         goto next_block;
8151                 }
8152         }
8153
8154         return err;
8155 }
8156
8157 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8158                 struct btrfs_io_bio *io_bio, blk_status_t err)
8159 {
8160         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8161
8162         if (skip_csum) {
8163                 if (unlikely(err))
8164                         return __btrfs_correct_data_nocsum(inode, io_bio);
8165                 else
8166                         return BLK_STS_OK;
8167         } else {
8168                 return __btrfs_subio_endio_read(inode, io_bio, err);
8169         }
8170 }
8171
8172 static void btrfs_endio_direct_read(struct bio *bio)
8173 {
8174         struct btrfs_dio_private *dip = bio->bi_private;
8175         struct inode *inode = dip->inode;
8176         struct bio *dio_bio;
8177         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8178         blk_status_t err = bio->bi_status;
8179
8180         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8181                 err = btrfs_subio_endio_read(inode, io_bio, err);
8182
8183         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8184                       dip->logical_offset + dip->bytes - 1);
8185         dio_bio = dip->dio_bio;
8186
8187         kfree(dip);
8188
8189         dio_bio->bi_status = err;
8190         dio_end_io(dio_bio);
8191         btrfs_io_bio_free_csum(io_bio);
8192         bio_put(bio);
8193 }
8194
8195 static void __endio_write_update_ordered(struct inode *inode,
8196                                          const u64 offset, const u64 bytes,
8197                                          const bool uptodate)
8198 {
8199         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8200         struct btrfs_ordered_extent *ordered = NULL;
8201         struct btrfs_workqueue *wq;
8202         btrfs_work_func_t func;
8203         u64 ordered_offset = offset;
8204         u64 ordered_bytes = bytes;
8205         u64 last_offset;
8206
8207         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8208                 wq = fs_info->endio_freespace_worker;
8209                 func = btrfs_freespace_write_helper;
8210         } else {
8211                 wq = fs_info->endio_write_workers;
8212                 func = btrfs_endio_write_helper;
8213         }
8214
8215         while (ordered_offset < offset + bytes) {
8216                 last_offset = ordered_offset;
8217                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8218                                                            &ordered_offset,
8219                                                            ordered_bytes,
8220                                                            uptodate)) {
8221                         btrfs_init_work(&ordered->work, func,
8222                                         finish_ordered_fn,
8223                                         NULL, NULL);
8224                         btrfs_queue_work(wq, &ordered->work);
8225                 }
8226                 /*
8227                  * If btrfs_dec_test_ordered_pending does not find any ordered
8228                  * extent in the range, we can exit.
8229                  */
8230                 if (ordered_offset == last_offset)
8231                         return;
8232                 /*
8233                  * Our bio might span multiple ordered extents. In this case
8234                  * we keep going until we have accounted the whole dio.
8235                  */
8236                 if (ordered_offset < offset + bytes) {
8237                         ordered_bytes = offset + bytes - ordered_offset;
8238                         ordered = NULL;
8239                 }
8240         }
8241 }
8242
8243 static void btrfs_endio_direct_write(struct bio *bio)
8244 {
8245         struct btrfs_dio_private *dip = bio->bi_private;
8246         struct bio *dio_bio = dip->dio_bio;
8247
8248         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8249                                      dip->bytes, !bio->bi_status);
8250
8251         kfree(dip);
8252
8253         dio_bio->bi_status = bio->bi_status;
8254         dio_end_io(dio_bio);
8255         bio_put(bio);
8256 }
8257
8258 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8259                                     struct bio *bio, u64 offset)
8260 {
8261         struct inode *inode = private_data;
8262         blk_status_t ret;
8263         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8264         BUG_ON(ret); /* -ENOMEM */
8265         return 0;
8266 }
8267
8268 static void btrfs_end_dio_bio(struct bio *bio)
8269 {
8270         struct btrfs_dio_private *dip = bio->bi_private;
8271         blk_status_t err = bio->bi_status;
8272
8273         if (err)
8274                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8275                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8276                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8277                            bio->bi_opf,
8278                            (unsigned long long)bio->bi_iter.bi_sector,
8279                            bio->bi_iter.bi_size, err);
8280
8281         if (dip->subio_endio)
8282                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8283
8284         if (err) {
8285                 /*
8286                  * We want to perceive the errors flag being set before
8287                  * decrementing the reference count. We don't need a barrier
8288                  * since atomic operations with a return value are fully
8289                  * ordered as per atomic_t.txt
8290                  */
8291                 dip->errors = 1;
8292         }
8293
8294         /* if there are more bios still pending for this dio, just exit */
8295         if (!atomic_dec_and_test(&dip->pending_bios))
8296                 goto out;
8297
8298         if (dip->errors) {
8299                 bio_io_error(dip->orig_bio);
8300         } else {
8301                 dip->dio_bio->bi_status = BLK_STS_OK;
8302                 bio_endio(dip->orig_bio);
8303         }
8304 out:
8305         bio_put(bio);
8306 }
8307
8308 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8309                                                  struct btrfs_dio_private *dip,
8310                                                  struct bio *bio,
8311                                                  u64 file_offset)
8312 {
8313         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8314         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8315         blk_status_t ret;
8316
8317         /*
8318          * We load all the csum data we need when we submit
8319          * the first bio to reduce the csum tree search and
8320          * contention.
8321          */
8322         if (dip->logical_offset == file_offset) {
8323                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8324                                                 file_offset);
8325                 if (ret)
8326                         return ret;
8327         }
8328
8329         if (bio == dip->orig_bio)
8330                 return 0;
8331
8332         file_offset -= dip->logical_offset;
8333         file_offset >>= inode->i_sb->s_blocksize_bits;
8334         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8335
8336         return 0;
8337 }
8338
8339 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8340                 struct inode *inode, u64 file_offset, int async_submit)
8341 {
8342         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8343         struct btrfs_dio_private *dip = bio->bi_private;
8344         bool write = bio_op(bio) == REQ_OP_WRITE;
8345         blk_status_t ret;
8346
8347         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8348         if (async_submit)
8349                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8350
8351         if (!write) {
8352                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8353                 if (ret)
8354                         goto err;
8355         }
8356
8357         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8358                 goto map;
8359
8360         if (write && async_submit) {
8361                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8362                                           file_offset, inode,
8363                                           btrfs_submit_bio_start_direct_io);
8364                 goto err;
8365         } else if (write) {
8366                 /*
8367                  * If we aren't doing async submit, calculate the csum of the
8368                  * bio now.
8369                  */
8370                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8371                 if (ret)
8372                         goto err;
8373         } else {
8374                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8375                                                      file_offset);
8376                 if (ret)
8377                         goto err;
8378         }
8379 map:
8380         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8381 err:
8382         return ret;
8383 }
8384
8385 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8386 {
8387         struct inode *inode = dip->inode;
8388         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8389         struct bio *bio;
8390         struct bio *orig_bio = dip->orig_bio;
8391         u64 start_sector = orig_bio->bi_iter.bi_sector;
8392         u64 file_offset = dip->logical_offset;
8393         int async_submit = 0;
8394         u64 submit_len;
8395         int clone_offset = 0;
8396         int clone_len;
8397         int ret;
8398         blk_status_t status;
8399         struct btrfs_io_geometry geom;
8400
8401         submit_len = orig_bio->bi_iter.bi_size;
8402         ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8403                                     start_sector << 9, submit_len, &geom);
8404         if (ret)
8405                 return -EIO;
8406
8407         if (geom.len >= submit_len) {
8408                 bio = orig_bio;
8409                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8410                 goto submit;
8411         }
8412
8413         /* async crcs make it difficult to collect full stripe writes. */
8414         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8415                 async_submit = 0;
8416         else
8417                 async_submit = 1;
8418
8419         /* bio split */
8420         ASSERT(geom.len <= INT_MAX);
8421         atomic_inc(&dip->pending_bios);
8422         do {
8423                 clone_len = min_t(int, submit_len, geom.len);
8424
8425                 /*
8426                  * This will never fail as it's passing GPF_NOFS and
8427                  * the allocation is backed by btrfs_bioset.
8428                  */
8429                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8430                                               clone_len);
8431                 bio->bi_private = dip;
8432                 bio->bi_end_io = btrfs_end_dio_bio;
8433                 btrfs_io_bio(bio)->logical = file_offset;
8434
8435                 ASSERT(submit_len >= clone_len);
8436                 submit_len -= clone_len;
8437                 if (submit_len == 0)
8438                         break;
8439
8440                 /*
8441                  * Increase the count before we submit the bio so we know
8442                  * the end IO handler won't happen before we increase the
8443                  * count. Otherwise, the dip might get freed before we're
8444                  * done setting it up.
8445                  */
8446                 atomic_inc(&dip->pending_bios);
8447
8448                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8449                                                 async_submit);
8450                 if (status) {
8451                         bio_put(bio);
8452                         atomic_dec(&dip->pending_bios);
8453                         goto out_err;
8454                 }
8455
8456                 clone_offset += clone_len;
8457                 start_sector += clone_len >> 9;
8458                 file_offset += clone_len;
8459
8460                 ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8461                                       start_sector << 9, submit_len, &geom);
8462                 if (ret)
8463                         goto out_err;
8464         } while (submit_len > 0);
8465
8466 submit:
8467         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8468         if (!status)
8469                 return 0;
8470
8471         bio_put(bio);
8472 out_err:
8473         dip->errors = 1;
8474         /*
8475          * Before atomic variable goto zero, we must  make sure dip->errors is
8476          * perceived to be set. This ordering is ensured by the fact that an
8477          * atomic operations with a return value are fully ordered as per
8478          * atomic_t.txt
8479          */
8480         if (atomic_dec_and_test(&dip->pending_bios))
8481                 bio_io_error(dip->orig_bio);
8482
8483         /* bio_end_io() will handle error, so we needn't return it */
8484         return 0;
8485 }
8486
8487 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8488                                 loff_t file_offset)
8489 {
8490         struct btrfs_dio_private *dip = NULL;
8491         struct bio *bio = NULL;
8492         struct btrfs_io_bio *io_bio;
8493         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8494         int ret = 0;
8495
8496         bio = btrfs_bio_clone(dio_bio);
8497
8498         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8499         if (!dip) {
8500                 ret = -ENOMEM;
8501                 goto free_ordered;
8502         }
8503
8504         dip->private = dio_bio->bi_private;
8505         dip->inode = inode;
8506         dip->logical_offset = file_offset;
8507         dip->bytes = dio_bio->bi_iter.bi_size;
8508         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8509         bio->bi_private = dip;
8510         dip->orig_bio = bio;
8511         dip->dio_bio = dio_bio;
8512         atomic_set(&dip->pending_bios, 0);
8513         io_bio = btrfs_io_bio(bio);
8514         io_bio->logical = file_offset;
8515
8516         if (write) {
8517                 bio->bi_end_io = btrfs_endio_direct_write;
8518         } else {
8519                 bio->bi_end_io = btrfs_endio_direct_read;
8520                 dip->subio_endio = btrfs_subio_endio_read;
8521         }
8522
8523         /*
8524          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8525          * even if we fail to submit a bio, because in such case we do the
8526          * corresponding error handling below and it must not be done a second
8527          * time by btrfs_direct_IO().
8528          */
8529         if (write) {
8530                 struct btrfs_dio_data *dio_data = current->journal_info;
8531
8532                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8533                         dip->bytes;
8534                 dio_data->unsubmitted_oe_range_start =
8535                         dio_data->unsubmitted_oe_range_end;
8536         }
8537
8538         ret = btrfs_submit_direct_hook(dip);
8539         if (!ret)
8540                 return;
8541
8542         btrfs_io_bio_free_csum(io_bio);
8543
8544 free_ordered:
8545         /*
8546          * If we arrived here it means either we failed to submit the dip
8547          * or we either failed to clone the dio_bio or failed to allocate the
8548          * dip. If we cloned the dio_bio and allocated the dip, we can just
8549          * call bio_endio against our io_bio so that we get proper resource
8550          * cleanup if we fail to submit the dip, otherwise, we must do the
8551          * same as btrfs_endio_direct_[write|read] because we can't call these
8552          * callbacks - they require an allocated dip and a clone of dio_bio.
8553          */
8554         if (bio && dip) {
8555                 bio_io_error(bio);
8556                 /*
8557                  * The end io callbacks free our dip, do the final put on bio
8558                  * and all the cleanup and final put for dio_bio (through
8559                  * dio_end_io()).
8560                  */
8561                 dip = NULL;
8562                 bio = NULL;
8563         } else {
8564                 if (write)
8565                         __endio_write_update_ordered(inode,
8566                                                 file_offset,
8567                                                 dio_bio->bi_iter.bi_size,
8568                                                 false);
8569                 else
8570                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8571                               file_offset + dio_bio->bi_iter.bi_size - 1);
8572
8573                 dio_bio->bi_status = BLK_STS_IOERR;
8574                 /*
8575                  * Releases and cleans up our dio_bio, no need to bio_put()
8576                  * nor bio_endio()/bio_io_error() against dio_bio.
8577                  */
8578                 dio_end_io(dio_bio);
8579         }
8580         if (bio)
8581                 bio_put(bio);
8582         kfree(dip);
8583 }
8584
8585 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8586                                const struct iov_iter *iter, loff_t offset)
8587 {
8588         int seg;
8589         int i;
8590         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8591         ssize_t retval = -EINVAL;
8592
8593         if (offset & blocksize_mask)
8594                 goto out;
8595
8596         if (iov_iter_alignment(iter) & blocksize_mask)
8597                 goto out;
8598
8599         /* If this is a write we don't need to check anymore */
8600         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8601                 return 0;
8602         /*
8603          * Check to make sure we don't have duplicate iov_base's in this
8604          * iovec, if so return EINVAL, otherwise we'll get csum errors
8605          * when reading back.
8606          */
8607         for (seg = 0; seg < iter->nr_segs; seg++) {
8608                 for (i = seg + 1; i < iter->nr_segs; i++) {
8609                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8610                                 goto out;
8611                 }
8612         }
8613         retval = 0;
8614 out:
8615         return retval;
8616 }
8617
8618 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8619 {
8620         struct file *file = iocb->ki_filp;
8621         struct inode *inode = file->f_mapping->host;
8622         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8623         struct btrfs_dio_data dio_data = { 0 };
8624         struct extent_changeset *data_reserved = NULL;
8625         loff_t offset = iocb->ki_pos;
8626         size_t count = 0;
8627         int flags = 0;
8628         bool wakeup = true;
8629         bool relock = false;
8630         ssize_t ret;
8631
8632         if (check_direct_IO(fs_info, iter, offset))
8633                 return 0;
8634
8635         inode_dio_begin(inode);
8636
8637         /*
8638          * The generic stuff only does filemap_write_and_wait_range, which
8639          * isn't enough if we've written compressed pages to this area, so
8640          * we need to flush the dirty pages again to make absolutely sure
8641          * that any outstanding dirty pages are on disk.
8642          */
8643         count = iov_iter_count(iter);
8644         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8645                      &BTRFS_I(inode)->runtime_flags))
8646                 filemap_fdatawrite_range(inode->i_mapping, offset,
8647                                          offset + count - 1);
8648
8649         if (iov_iter_rw(iter) == WRITE) {
8650                 /*
8651                  * If the write DIO is beyond the EOF, we need update
8652                  * the isize, but it is protected by i_mutex. So we can
8653                  * not unlock the i_mutex at this case.
8654                  */
8655                 if (offset + count <= inode->i_size) {
8656                         dio_data.overwrite = 1;
8657                         inode_unlock(inode);
8658                         relock = true;
8659                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8660                         ret = -EAGAIN;
8661                         goto out;
8662                 }
8663                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8664                                                    offset, count);
8665                 if (ret)
8666                         goto out;
8667
8668                 /*
8669                  * We need to know how many extents we reserved so that we can
8670                  * do the accounting properly if we go over the number we
8671                  * originally calculated.  Abuse current->journal_info for this.
8672                  */
8673                 dio_data.reserve = round_up(count,
8674                                             fs_info->sectorsize);
8675                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8676                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8677                 current->journal_info = &dio_data;
8678                 down_read(&BTRFS_I(inode)->dio_sem);
8679         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8680                                      &BTRFS_I(inode)->runtime_flags)) {
8681                 inode_dio_end(inode);
8682                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8683                 wakeup = false;
8684         }
8685
8686         ret = __blockdev_direct_IO(iocb, inode,
8687                                    fs_info->fs_devices->latest_bdev,
8688                                    iter, btrfs_get_blocks_direct, NULL,
8689                                    btrfs_submit_direct, flags);
8690         if (iov_iter_rw(iter) == WRITE) {
8691                 up_read(&BTRFS_I(inode)->dio_sem);
8692                 current->journal_info = NULL;
8693                 if (ret < 0 && ret != -EIOCBQUEUED) {
8694                         if (dio_data.reserve)
8695                                 btrfs_delalloc_release_space(inode, data_reserved,
8696                                         offset, dio_data.reserve, true);
8697                         /*
8698                          * On error we might have left some ordered extents
8699                          * without submitting corresponding bios for them, so
8700                          * cleanup them up to avoid other tasks getting them
8701                          * and waiting for them to complete forever.
8702                          */
8703                         if (dio_data.unsubmitted_oe_range_start <
8704                             dio_data.unsubmitted_oe_range_end)
8705                                 __endio_write_update_ordered(inode,
8706                                         dio_data.unsubmitted_oe_range_start,
8707                                         dio_data.unsubmitted_oe_range_end -
8708                                         dio_data.unsubmitted_oe_range_start,
8709                                         false);
8710                 } else if (ret >= 0 && (size_t)ret < count)
8711                         btrfs_delalloc_release_space(inode, data_reserved,
8712                                         offset, count - (size_t)ret, true);
8713                 btrfs_delalloc_release_extents(BTRFS_I(inode), count, false);
8714         }
8715 out:
8716         if (wakeup)
8717                 inode_dio_end(inode);
8718         if (relock)
8719                 inode_lock(inode);
8720
8721         extent_changeset_free(data_reserved);
8722         return ret;
8723 }
8724
8725 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8726
8727 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8728                 __u64 start, __u64 len)
8729 {
8730         int     ret;
8731
8732         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8733         if (ret)
8734                 return ret;
8735
8736         return extent_fiemap(inode, fieinfo, start, len);
8737 }
8738
8739 int btrfs_readpage(struct file *file, struct page *page)
8740 {
8741         struct extent_io_tree *tree;
8742         tree = &BTRFS_I(page->mapping->host)->io_tree;
8743         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8744 }
8745
8746 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8747 {
8748         struct inode *inode = page->mapping->host;
8749         int ret;
8750
8751         if (current->flags & PF_MEMALLOC) {
8752                 redirty_page_for_writepage(wbc, page);
8753                 unlock_page(page);
8754                 return 0;
8755         }
8756
8757         /*
8758          * If we are under memory pressure we will call this directly from the
8759          * VM, we need to make sure we have the inode referenced for the ordered
8760          * extent.  If not just return like we didn't do anything.
8761          */
8762         if (!igrab(inode)) {
8763                 redirty_page_for_writepage(wbc, page);
8764                 return AOP_WRITEPAGE_ACTIVATE;
8765         }
8766         ret = extent_write_full_page(page, wbc);
8767         btrfs_add_delayed_iput(inode);
8768         return ret;
8769 }
8770
8771 static int btrfs_writepages(struct address_space *mapping,
8772                             struct writeback_control *wbc)
8773 {
8774         return extent_writepages(mapping, wbc);
8775 }
8776
8777 static int
8778 btrfs_readpages(struct file *file, struct address_space *mapping,
8779                 struct list_head *pages, unsigned nr_pages)
8780 {
8781         return extent_readpages(mapping, pages, nr_pages);
8782 }
8783
8784 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8785 {
8786         int ret = try_release_extent_mapping(page, gfp_flags);
8787         if (ret == 1) {
8788                 ClearPagePrivate(page);
8789                 set_page_private(page, 0);
8790                 put_page(page);
8791         }
8792         return ret;
8793 }
8794
8795 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8796 {
8797         if (PageWriteback(page) || PageDirty(page))
8798                 return 0;
8799         return __btrfs_releasepage(page, gfp_flags);
8800 }
8801
8802 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8803                                  unsigned int length)
8804 {
8805         struct inode *inode = page->mapping->host;
8806         struct extent_io_tree *tree;
8807         struct btrfs_ordered_extent *ordered;
8808         struct extent_state *cached_state = NULL;
8809         u64 page_start = page_offset(page);
8810         u64 page_end = page_start + PAGE_SIZE - 1;
8811         u64 start;
8812         u64 end;
8813         int inode_evicting = inode->i_state & I_FREEING;
8814
8815         /*
8816          * we have the page locked, so new writeback can't start,
8817          * and the dirty bit won't be cleared while we are here.
8818          *
8819          * Wait for IO on this page so that we can safely clear
8820          * the PagePrivate2 bit and do ordered accounting
8821          */
8822         wait_on_page_writeback(page);
8823
8824         tree = &BTRFS_I(inode)->io_tree;
8825         if (offset) {
8826                 btrfs_releasepage(page, GFP_NOFS);
8827                 return;
8828         }
8829
8830         if (!inode_evicting)
8831                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8832 again:
8833         start = page_start;
8834         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8835                                         page_end - start + 1);
8836         if (ordered) {
8837                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8838                 /*
8839                  * IO on this page will never be started, so we need
8840                  * to account for any ordered extents now
8841                  */
8842                 if (!inode_evicting)
8843                         clear_extent_bit(tree, start, end,
8844                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8845                                          EXTENT_DELALLOC_NEW |
8846                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8847                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8848                 /*
8849                  * whoever cleared the private bit is responsible
8850                  * for the finish_ordered_io
8851                  */
8852                 if (TestClearPagePrivate2(page)) {
8853                         struct btrfs_ordered_inode_tree *tree;
8854                         u64 new_len;
8855
8856                         tree = &BTRFS_I(inode)->ordered_tree;
8857
8858                         spin_lock_irq(&tree->lock);
8859                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8860                         new_len = start - ordered->file_offset;
8861                         if (new_len < ordered->truncated_len)
8862                                 ordered->truncated_len = new_len;
8863                         spin_unlock_irq(&tree->lock);
8864
8865                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8866                                                            start,
8867                                                            end - start + 1, 1))
8868                                 btrfs_finish_ordered_io(ordered);
8869                 }
8870                 btrfs_put_ordered_extent(ordered);
8871                 if (!inode_evicting) {
8872                         cached_state = NULL;
8873                         lock_extent_bits(tree, start, end,
8874                                          &cached_state);
8875                 }
8876
8877                 start = end + 1;
8878                 if (start < page_end)
8879                         goto again;
8880         }
8881
8882         /*
8883          * Qgroup reserved space handler
8884          * Page here will be either
8885          * 1) Already written to disk
8886          *    In this case, its reserved space is released from data rsv map
8887          *    and will be freed by delayed_ref handler finally.
8888          *    So even we call qgroup_free_data(), it won't decrease reserved
8889          *    space.
8890          * 2) Not written to disk
8891          *    This means the reserved space should be freed here. However,
8892          *    if a truncate invalidates the page (by clearing PageDirty)
8893          *    and the page is accounted for while allocating extent
8894          *    in btrfs_check_data_free_space() we let delayed_ref to
8895          *    free the entire extent.
8896          */
8897         if (PageDirty(page))
8898                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8899         if (!inode_evicting) {
8900                 clear_extent_bit(tree, page_start, page_end,
8901                                  EXTENT_LOCKED | EXTENT_DIRTY |
8902                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8903                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8904                                  &cached_state);
8905
8906                 __btrfs_releasepage(page, GFP_NOFS);
8907         }
8908
8909         ClearPageChecked(page);
8910         if (PagePrivate(page)) {
8911                 ClearPagePrivate(page);
8912                 set_page_private(page, 0);
8913                 put_page(page);
8914         }
8915 }
8916
8917 /*
8918  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8919  * called from a page fault handler when a page is first dirtied. Hence we must
8920  * be careful to check for EOF conditions here. We set the page up correctly
8921  * for a written page which means we get ENOSPC checking when writing into
8922  * holes and correct delalloc and unwritten extent mapping on filesystems that
8923  * support these features.
8924  *
8925  * We are not allowed to take the i_mutex here so we have to play games to
8926  * protect against truncate races as the page could now be beyond EOF.  Because
8927  * truncate_setsize() writes the inode size before removing pages, once we have
8928  * the page lock we can determine safely if the page is beyond EOF. If it is not
8929  * beyond EOF, then the page is guaranteed safe against truncation until we
8930  * unlock the page.
8931  */
8932 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8933 {
8934         struct page *page = vmf->page;
8935         struct inode *inode = file_inode(vmf->vma->vm_file);
8936         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8937         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8938         struct btrfs_ordered_extent *ordered;
8939         struct extent_state *cached_state = NULL;
8940         struct extent_changeset *data_reserved = NULL;
8941         char *kaddr;
8942         unsigned long zero_start;
8943         loff_t size;
8944         vm_fault_t ret;
8945         int ret2;
8946         int reserved = 0;
8947         u64 reserved_space;
8948         u64 page_start;
8949         u64 page_end;
8950         u64 end;
8951
8952         reserved_space = PAGE_SIZE;
8953
8954         sb_start_pagefault(inode->i_sb);
8955         page_start = page_offset(page);
8956         page_end = page_start + PAGE_SIZE - 1;
8957         end = page_end;
8958
8959         /*
8960          * Reserving delalloc space after obtaining the page lock can lead to
8961          * deadlock. For example, if a dirty page is locked by this function
8962          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8963          * dirty page write out, then the btrfs_writepage() function could
8964          * end up waiting indefinitely to get a lock on the page currently
8965          * being processed by btrfs_page_mkwrite() function.
8966          */
8967         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8968                                            reserved_space);
8969         if (!ret2) {
8970                 ret2 = file_update_time(vmf->vma->vm_file);
8971                 reserved = 1;
8972         }
8973         if (ret2) {
8974                 ret = vmf_error(ret2);
8975                 if (reserved)
8976                         goto out;
8977                 goto out_noreserve;
8978         }
8979
8980         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8981 again:
8982         lock_page(page);
8983         size = i_size_read(inode);
8984
8985         if ((page->mapping != inode->i_mapping) ||
8986             (page_start >= size)) {
8987                 /* page got truncated out from underneath us */
8988                 goto out_unlock;
8989         }
8990         wait_on_page_writeback(page);
8991
8992         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8993         set_page_extent_mapped(page);
8994
8995         /*
8996          * we can't set the delalloc bits if there are pending ordered
8997          * extents.  Drop our locks and wait for them to finish
8998          */
8999         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9000                         PAGE_SIZE);
9001         if (ordered) {
9002                 unlock_extent_cached(io_tree, page_start, page_end,
9003                                      &cached_state);
9004                 unlock_page(page);
9005                 btrfs_start_ordered_extent(inode, ordered, 1);
9006                 btrfs_put_ordered_extent(ordered);
9007                 goto again;
9008         }
9009
9010         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9011                 reserved_space = round_up(size - page_start,
9012                                           fs_info->sectorsize);
9013                 if (reserved_space < PAGE_SIZE) {
9014                         end = page_start + reserved_space - 1;
9015                         btrfs_delalloc_release_space(inode, data_reserved,
9016                                         page_start, PAGE_SIZE - reserved_space,
9017                                         true);
9018                 }
9019         }
9020
9021         /*
9022          * page_mkwrite gets called when the page is firstly dirtied after it's
9023          * faulted in, but write(2) could also dirty a page and set delalloc
9024          * bits, thus in this case for space account reason, we still need to
9025          * clear any delalloc bits within this page range since we have to
9026          * reserve data&meta space before lock_page() (see above comments).
9027          */
9028         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9029                           EXTENT_DIRTY | EXTENT_DELALLOC |
9030                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9031                           0, 0, &cached_state);
9032
9033         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9034                                         &cached_state);
9035         if (ret2) {
9036                 unlock_extent_cached(io_tree, page_start, page_end,
9037                                      &cached_state);
9038                 ret = VM_FAULT_SIGBUS;
9039                 goto out_unlock;
9040         }
9041         ret2 = 0;
9042
9043         /* page is wholly or partially inside EOF */
9044         if (page_start + PAGE_SIZE > size)
9045                 zero_start = offset_in_page(size);
9046         else
9047                 zero_start = PAGE_SIZE;
9048
9049         if (zero_start != PAGE_SIZE) {
9050                 kaddr = kmap(page);
9051                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9052                 flush_dcache_page(page);
9053                 kunmap(page);
9054         }
9055         ClearPageChecked(page);
9056         set_page_dirty(page);
9057         SetPageUptodate(page);
9058
9059         BTRFS_I(inode)->last_trans = fs_info->generation;
9060         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9061         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9062
9063         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
9064
9065         if (!ret2) {
9066                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, true);
9067                 sb_end_pagefault(inode->i_sb);
9068                 extent_changeset_free(data_reserved);
9069                 return VM_FAULT_LOCKED;
9070         }
9071
9072 out_unlock:
9073         unlock_page(page);
9074 out:
9075         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, (ret != 0));
9076         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9077                                      reserved_space, (ret != 0));
9078 out_noreserve:
9079         sb_end_pagefault(inode->i_sb);
9080         extent_changeset_free(data_reserved);
9081         return ret;
9082 }
9083
9084 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
9085 {
9086         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9087         struct btrfs_root *root = BTRFS_I(inode)->root;
9088         struct btrfs_block_rsv *rsv;
9089         int ret;
9090         struct btrfs_trans_handle *trans;
9091         u64 mask = fs_info->sectorsize - 1;
9092         u64 min_size = btrfs_calc_metadata_size(fs_info, 1);
9093
9094         if (!skip_writeback) {
9095                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9096                                                (u64)-1);
9097                 if (ret)
9098                         return ret;
9099         }
9100
9101         /*
9102          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
9103          * things going on here:
9104          *
9105          * 1) We need to reserve space to update our inode.
9106          *
9107          * 2) We need to have something to cache all the space that is going to
9108          * be free'd up by the truncate operation, but also have some slack
9109          * space reserved in case it uses space during the truncate (thank you
9110          * very much snapshotting).
9111          *
9112          * And we need these to be separate.  The fact is we can use a lot of
9113          * space doing the truncate, and we have no earthly idea how much space
9114          * we will use, so we need the truncate reservation to be separate so it
9115          * doesn't end up using space reserved for updating the inode.  We also
9116          * need to be able to stop the transaction and start a new one, which
9117          * means we need to be able to update the inode several times, and we
9118          * have no idea of knowing how many times that will be, so we can't just
9119          * reserve 1 item for the entirety of the operation, so that has to be
9120          * done separately as well.
9121          *
9122          * So that leaves us with
9123          *
9124          * 1) rsv - for the truncate reservation, which we will steal from the
9125          * transaction reservation.
9126          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
9127          * updating the inode.
9128          */
9129         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9130         if (!rsv)
9131                 return -ENOMEM;
9132         rsv->size = min_size;
9133         rsv->failfast = 1;
9134
9135         /*
9136          * 1 for the truncate slack space
9137          * 1 for updating the inode.
9138          */
9139         trans = btrfs_start_transaction(root, 2);
9140         if (IS_ERR(trans)) {
9141                 ret = PTR_ERR(trans);
9142                 goto out;
9143         }
9144
9145         /* Migrate the slack space for the truncate to our reserve */
9146         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9147                                       min_size, false);
9148         BUG_ON(ret);
9149
9150         /*
9151          * So if we truncate and then write and fsync we normally would just
9152          * write the extents that changed, which is a problem if we need to
9153          * first truncate that entire inode.  So set this flag so we write out
9154          * all of the extents in the inode to the sync log so we're completely
9155          * safe.
9156          */
9157         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9158         trans->block_rsv = rsv;
9159
9160         while (1) {
9161                 ret = btrfs_truncate_inode_items(trans, root, inode,
9162                                                  inode->i_size,
9163                                                  BTRFS_EXTENT_DATA_KEY);
9164                 trans->block_rsv = &fs_info->trans_block_rsv;
9165                 if (ret != -ENOSPC && ret != -EAGAIN)
9166                         break;
9167
9168                 ret = btrfs_update_inode(trans, root, inode);
9169                 if (ret)
9170                         break;
9171
9172                 btrfs_end_transaction(trans);
9173                 btrfs_btree_balance_dirty(fs_info);
9174
9175                 trans = btrfs_start_transaction(root, 2);
9176                 if (IS_ERR(trans)) {
9177                         ret = PTR_ERR(trans);
9178                         trans = NULL;
9179                         break;
9180                 }
9181
9182                 btrfs_block_rsv_release(fs_info, rsv, -1);
9183                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9184                                               rsv, min_size, false);
9185                 BUG_ON(ret);    /* shouldn't happen */
9186                 trans->block_rsv = rsv;
9187         }
9188
9189         /*
9190          * We can't call btrfs_truncate_block inside a trans handle as we could
9191          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9192          * we've truncated everything except the last little bit, and can do
9193          * btrfs_truncate_block and then update the disk_i_size.
9194          */
9195         if (ret == NEED_TRUNCATE_BLOCK) {
9196                 btrfs_end_transaction(trans);
9197                 btrfs_btree_balance_dirty(fs_info);
9198
9199                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9200                 if (ret)
9201                         goto out;
9202                 trans = btrfs_start_transaction(root, 1);
9203                 if (IS_ERR(trans)) {
9204                         ret = PTR_ERR(trans);
9205                         goto out;
9206                 }
9207                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9208         }
9209
9210         if (trans) {
9211                 int ret2;
9212
9213                 trans->block_rsv = &fs_info->trans_block_rsv;
9214                 ret2 = btrfs_update_inode(trans, root, inode);
9215                 if (ret2 && !ret)
9216                         ret = ret2;
9217
9218                 ret2 = btrfs_end_transaction(trans);
9219                 if (ret2 && !ret)
9220                         ret = ret2;
9221                 btrfs_btree_balance_dirty(fs_info);
9222         }
9223 out:
9224         btrfs_free_block_rsv(fs_info, rsv);
9225
9226         return ret;
9227 }
9228
9229 /*
9230  * create a new subvolume directory/inode (helper for the ioctl).
9231  */
9232 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9233                              struct btrfs_root *new_root,
9234                              struct btrfs_root *parent_root,
9235                              u64 new_dirid)
9236 {
9237         struct inode *inode;
9238         int err;
9239         u64 index = 0;
9240
9241         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9242                                 new_dirid, new_dirid,
9243                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9244                                 &index);
9245         if (IS_ERR(inode))
9246                 return PTR_ERR(inode);
9247         inode->i_op = &btrfs_dir_inode_operations;
9248         inode->i_fop = &btrfs_dir_file_operations;
9249
9250         set_nlink(inode, 1);
9251         btrfs_i_size_write(BTRFS_I(inode), 0);
9252         unlock_new_inode(inode);
9253
9254         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9255         if (err)
9256                 btrfs_err(new_root->fs_info,
9257                           "error inheriting subvolume %llu properties: %d",
9258                           new_root->root_key.objectid, err);
9259
9260         err = btrfs_update_inode(trans, new_root, inode);
9261
9262         iput(inode);
9263         return err;
9264 }
9265
9266 struct inode *btrfs_alloc_inode(struct super_block *sb)
9267 {
9268         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9269         struct btrfs_inode *ei;
9270         struct inode *inode;
9271
9272         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9273         if (!ei)
9274                 return NULL;
9275
9276         ei->root = NULL;
9277         ei->generation = 0;
9278         ei->last_trans = 0;
9279         ei->last_sub_trans = 0;
9280         ei->logged_trans = 0;
9281         ei->delalloc_bytes = 0;
9282         ei->new_delalloc_bytes = 0;
9283         ei->defrag_bytes = 0;
9284         ei->disk_i_size = 0;
9285         ei->flags = 0;
9286         ei->csum_bytes = 0;
9287         ei->index_cnt = (u64)-1;
9288         ei->dir_index = 0;
9289         ei->last_unlink_trans = 0;
9290         ei->last_log_commit = 0;
9291
9292         spin_lock_init(&ei->lock);
9293         ei->outstanding_extents = 0;
9294         if (sb->s_magic != BTRFS_TEST_MAGIC)
9295                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9296                                               BTRFS_BLOCK_RSV_DELALLOC);
9297         ei->runtime_flags = 0;
9298         ei->prop_compress = BTRFS_COMPRESS_NONE;
9299         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9300
9301         ei->delayed_node = NULL;
9302
9303         ei->i_otime.tv_sec = 0;
9304         ei->i_otime.tv_nsec = 0;
9305
9306         inode = &ei->vfs_inode;
9307         extent_map_tree_init(&ei->extent_tree);
9308         extent_io_tree_init(fs_info, &ei->io_tree, IO_TREE_INODE_IO, inode);
9309         extent_io_tree_init(fs_info, &ei->io_failure_tree,
9310                             IO_TREE_INODE_IO_FAILURE, inode);
9311         ei->io_tree.track_uptodate = true;
9312         ei->io_failure_tree.track_uptodate = true;
9313         atomic_set(&ei->sync_writers, 0);
9314         mutex_init(&ei->log_mutex);
9315         mutex_init(&ei->delalloc_mutex);
9316         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9317         INIT_LIST_HEAD(&ei->delalloc_inodes);
9318         INIT_LIST_HEAD(&ei->delayed_iput);
9319         RB_CLEAR_NODE(&ei->rb_node);
9320         init_rwsem(&ei->dio_sem);
9321
9322         return inode;
9323 }
9324
9325 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9326 void btrfs_test_destroy_inode(struct inode *inode)
9327 {
9328         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9329         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9330 }
9331 #endif
9332
9333 void btrfs_free_inode(struct inode *inode)
9334 {
9335         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9336 }
9337
9338 void btrfs_destroy_inode(struct inode *inode)
9339 {
9340         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9341         struct btrfs_ordered_extent *ordered;
9342         struct btrfs_root *root = BTRFS_I(inode)->root;
9343
9344         WARN_ON(!hlist_empty(&inode->i_dentry));
9345         WARN_ON(inode->i_data.nrpages);
9346         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9347         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9348         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9349         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9350         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9351         WARN_ON(BTRFS_I(inode)->csum_bytes);
9352         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9353
9354         /*
9355          * This can happen where we create an inode, but somebody else also
9356          * created the same inode and we need to destroy the one we already
9357          * created.
9358          */
9359         if (!root)
9360                 return;
9361
9362         while (1) {
9363                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9364                 if (!ordered)
9365                         break;
9366                 else {
9367                         btrfs_err(fs_info,
9368                                   "found ordered extent %llu %llu on inode cleanup",
9369                                   ordered->file_offset, ordered->len);
9370                         btrfs_remove_ordered_extent(inode, ordered);
9371                         btrfs_put_ordered_extent(ordered);
9372                         btrfs_put_ordered_extent(ordered);
9373                 }
9374         }
9375         btrfs_qgroup_check_reserved_leak(inode);
9376         inode_tree_del(inode);
9377         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9378 }
9379
9380 int btrfs_drop_inode(struct inode *inode)
9381 {
9382         struct btrfs_root *root = BTRFS_I(inode)->root;
9383
9384         if (root == NULL)
9385                 return 1;
9386
9387         /* the snap/subvol tree is on deleting */
9388         if (btrfs_root_refs(&root->root_item) == 0)
9389                 return 1;
9390         else
9391                 return generic_drop_inode(inode);
9392 }
9393
9394 static void init_once(void *foo)
9395 {
9396         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9397
9398         inode_init_once(&ei->vfs_inode);
9399 }
9400
9401 void __cold btrfs_destroy_cachep(void)
9402 {
9403         /*
9404          * Make sure all delayed rcu free inodes are flushed before we
9405          * destroy cache.
9406          */
9407         rcu_barrier();
9408         kmem_cache_destroy(btrfs_inode_cachep);
9409         kmem_cache_destroy(btrfs_trans_handle_cachep);
9410         kmem_cache_destroy(btrfs_path_cachep);
9411         kmem_cache_destroy(btrfs_free_space_cachep);
9412 }
9413
9414 int __init btrfs_init_cachep(void)
9415 {
9416         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9417                         sizeof(struct btrfs_inode), 0,
9418                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9419                         init_once);
9420         if (!btrfs_inode_cachep)
9421                 goto fail;
9422
9423         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9424                         sizeof(struct btrfs_trans_handle), 0,
9425                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9426         if (!btrfs_trans_handle_cachep)
9427                 goto fail;
9428
9429         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9430                         sizeof(struct btrfs_path), 0,
9431                         SLAB_MEM_SPREAD, NULL);
9432         if (!btrfs_path_cachep)
9433                 goto fail;
9434
9435         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9436                         sizeof(struct btrfs_free_space), 0,
9437                         SLAB_MEM_SPREAD, NULL);
9438         if (!btrfs_free_space_cachep)
9439                 goto fail;
9440
9441         return 0;
9442 fail:
9443         btrfs_destroy_cachep();
9444         return -ENOMEM;
9445 }
9446
9447 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9448                          u32 request_mask, unsigned int flags)
9449 {
9450         u64 delalloc_bytes;
9451         struct inode *inode = d_inode(path->dentry);
9452         u32 blocksize = inode->i_sb->s_blocksize;
9453         u32 bi_flags = BTRFS_I(inode)->flags;
9454
9455         stat->result_mask |= STATX_BTIME;
9456         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9457         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9458         if (bi_flags & BTRFS_INODE_APPEND)
9459                 stat->attributes |= STATX_ATTR_APPEND;
9460         if (bi_flags & BTRFS_INODE_COMPRESS)
9461                 stat->attributes |= STATX_ATTR_COMPRESSED;
9462         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9463                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9464         if (bi_flags & BTRFS_INODE_NODUMP)
9465                 stat->attributes |= STATX_ATTR_NODUMP;
9466
9467         stat->attributes_mask |= (STATX_ATTR_APPEND |
9468                                   STATX_ATTR_COMPRESSED |
9469                                   STATX_ATTR_IMMUTABLE |
9470                                   STATX_ATTR_NODUMP);
9471
9472         generic_fillattr(inode, stat);
9473         stat->dev = BTRFS_I(inode)->root->anon_dev;
9474
9475         spin_lock(&BTRFS_I(inode)->lock);
9476         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9477         spin_unlock(&BTRFS_I(inode)->lock);
9478         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9479                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9480         return 0;
9481 }
9482
9483 static int btrfs_rename_exchange(struct inode *old_dir,
9484                               struct dentry *old_dentry,
9485                               struct inode *new_dir,
9486                               struct dentry *new_dentry)
9487 {
9488         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9489         struct btrfs_trans_handle *trans;
9490         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9491         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9492         struct inode *new_inode = new_dentry->d_inode;
9493         struct inode *old_inode = old_dentry->d_inode;
9494         struct timespec64 ctime = current_time(old_inode);
9495         struct dentry *parent;
9496         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9497         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9498         u64 old_idx = 0;
9499         u64 new_idx = 0;
9500         u64 root_objectid;
9501         int ret;
9502         bool root_log_pinned = false;
9503         bool dest_log_pinned = false;
9504         struct btrfs_log_ctx ctx_root;
9505         struct btrfs_log_ctx ctx_dest;
9506         bool sync_log_root = false;
9507         bool sync_log_dest = false;
9508         bool commit_transaction = false;
9509
9510         /* we only allow rename subvolume link between subvolumes */
9511         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9512                 return -EXDEV;
9513
9514         btrfs_init_log_ctx(&ctx_root, old_inode);
9515         btrfs_init_log_ctx(&ctx_dest, new_inode);
9516
9517         /* close the race window with snapshot create/destroy ioctl */
9518         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9519                 down_read(&fs_info->subvol_sem);
9520         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9521                 down_read(&fs_info->subvol_sem);
9522
9523         /*
9524          * We want to reserve the absolute worst case amount of items.  So if
9525          * both inodes are subvols and we need to unlink them then that would
9526          * require 4 item modifications, but if they are both normal inodes it
9527          * would require 5 item modifications, so we'll assume their normal
9528          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9529          * should cover the worst case number of items we'll modify.
9530          */
9531         trans = btrfs_start_transaction(root, 12);
9532         if (IS_ERR(trans)) {
9533                 ret = PTR_ERR(trans);
9534                 goto out_notrans;
9535         }
9536
9537         /*
9538          * We need to find a free sequence number both in the source and
9539          * in the destination directory for the exchange.
9540          */
9541         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9542         if (ret)
9543                 goto out_fail;
9544         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9545         if (ret)
9546                 goto out_fail;
9547
9548         BTRFS_I(old_inode)->dir_index = 0ULL;
9549         BTRFS_I(new_inode)->dir_index = 0ULL;
9550
9551         /* Reference for the source. */
9552         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9553                 /* force full log commit if subvolume involved. */
9554                 btrfs_set_log_full_commit(trans);
9555         } else {
9556                 btrfs_pin_log_trans(root);
9557                 root_log_pinned = true;
9558                 ret = btrfs_insert_inode_ref(trans, dest,
9559                                              new_dentry->d_name.name,
9560                                              new_dentry->d_name.len,
9561                                              old_ino,
9562                                              btrfs_ino(BTRFS_I(new_dir)),
9563                                              old_idx);
9564                 if (ret)
9565                         goto out_fail;
9566         }
9567
9568         /* And now for the dest. */
9569         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9570                 /* force full log commit if subvolume involved. */
9571                 btrfs_set_log_full_commit(trans);
9572         } else {
9573                 btrfs_pin_log_trans(dest);
9574                 dest_log_pinned = true;
9575                 ret = btrfs_insert_inode_ref(trans, root,
9576                                              old_dentry->d_name.name,
9577                                              old_dentry->d_name.len,
9578                                              new_ino,
9579                                              btrfs_ino(BTRFS_I(old_dir)),
9580                                              new_idx);
9581                 if (ret)
9582                         goto out_fail;
9583         }
9584
9585         /* Update inode version and ctime/mtime. */
9586         inode_inc_iversion(old_dir);
9587         inode_inc_iversion(new_dir);
9588         inode_inc_iversion(old_inode);
9589         inode_inc_iversion(new_inode);
9590         old_dir->i_ctime = old_dir->i_mtime = ctime;
9591         new_dir->i_ctime = new_dir->i_mtime = ctime;
9592         old_inode->i_ctime = ctime;
9593         new_inode->i_ctime = ctime;
9594
9595         if (old_dentry->d_parent != new_dentry->d_parent) {
9596                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9597                                 BTRFS_I(old_inode), 1);
9598                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9599                                 BTRFS_I(new_inode), 1);
9600         }
9601
9602         /* src is a subvolume */
9603         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9604                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9605                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9606                                           old_dentry->d_name.name,
9607                                           old_dentry->d_name.len);
9608         } else { /* src is an inode */
9609                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9610                                            BTRFS_I(old_dentry->d_inode),
9611                                            old_dentry->d_name.name,
9612                                            old_dentry->d_name.len);
9613                 if (!ret)
9614                         ret = btrfs_update_inode(trans, root, old_inode);
9615         }
9616         if (ret) {
9617                 btrfs_abort_transaction(trans, ret);
9618                 goto out_fail;
9619         }
9620
9621         /* dest is a subvolume */
9622         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9623                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9624                 ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9625                                           new_dentry->d_name.name,
9626                                           new_dentry->d_name.len);
9627         } else { /* dest is an inode */
9628                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9629                                            BTRFS_I(new_dentry->d_inode),
9630                                            new_dentry->d_name.name,
9631                                            new_dentry->d_name.len);
9632                 if (!ret)
9633                         ret = btrfs_update_inode(trans, dest, new_inode);
9634         }
9635         if (ret) {
9636                 btrfs_abort_transaction(trans, ret);
9637                 goto out_fail;
9638         }
9639
9640         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9641                              new_dentry->d_name.name,
9642                              new_dentry->d_name.len, 0, old_idx);
9643         if (ret) {
9644                 btrfs_abort_transaction(trans, ret);
9645                 goto out_fail;
9646         }
9647
9648         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9649                              old_dentry->d_name.name,
9650                              old_dentry->d_name.len, 0, new_idx);
9651         if (ret) {
9652                 btrfs_abort_transaction(trans, ret);
9653                 goto out_fail;
9654         }
9655
9656         if (old_inode->i_nlink == 1)
9657                 BTRFS_I(old_inode)->dir_index = old_idx;
9658         if (new_inode->i_nlink == 1)
9659                 BTRFS_I(new_inode)->dir_index = new_idx;
9660
9661         if (root_log_pinned) {
9662                 parent = new_dentry->d_parent;
9663                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9664                                          BTRFS_I(old_dir), parent,
9665                                          false, &ctx_root);
9666                 if (ret == BTRFS_NEED_LOG_SYNC)
9667                         sync_log_root = true;
9668                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9669                         commit_transaction = true;
9670                 ret = 0;
9671                 btrfs_end_log_trans(root);
9672                 root_log_pinned = false;
9673         }
9674         if (dest_log_pinned) {
9675                 if (!commit_transaction) {
9676                         parent = old_dentry->d_parent;
9677                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
9678                                                  BTRFS_I(new_dir), parent,
9679                                                  false, &ctx_dest);
9680                         if (ret == BTRFS_NEED_LOG_SYNC)
9681                                 sync_log_dest = true;
9682                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
9683                                 commit_transaction = true;
9684                         ret = 0;
9685                 }
9686                 btrfs_end_log_trans(dest);
9687                 dest_log_pinned = false;
9688         }
9689 out_fail:
9690         /*
9691          * If we have pinned a log and an error happened, we unpin tasks
9692          * trying to sync the log and force them to fallback to a transaction
9693          * commit if the log currently contains any of the inodes involved in
9694          * this rename operation (to ensure we do not persist a log with an
9695          * inconsistent state for any of these inodes or leading to any
9696          * inconsistencies when replayed). If the transaction was aborted, the
9697          * abortion reason is propagated to userspace when attempting to commit
9698          * the transaction. If the log does not contain any of these inodes, we
9699          * allow the tasks to sync it.
9700          */
9701         if (ret && (root_log_pinned || dest_log_pinned)) {
9702                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9703                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9704                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9705                     (new_inode &&
9706                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9707                         btrfs_set_log_full_commit(trans);
9708
9709                 if (root_log_pinned) {
9710                         btrfs_end_log_trans(root);
9711                         root_log_pinned = false;
9712                 }
9713                 if (dest_log_pinned) {
9714                         btrfs_end_log_trans(dest);
9715                         dest_log_pinned = false;
9716                 }
9717         }
9718         if (!ret && sync_log_root && !commit_transaction) {
9719                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
9720                                      &ctx_root);
9721                 if (ret)
9722                         commit_transaction = true;
9723         }
9724         if (!ret && sync_log_dest && !commit_transaction) {
9725                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
9726                                      &ctx_dest);
9727                 if (ret)
9728                         commit_transaction = true;
9729         }
9730         if (commit_transaction) {
9731                 ret = btrfs_commit_transaction(trans);
9732         } else {
9733                 int ret2;
9734
9735                 ret2 = btrfs_end_transaction(trans);
9736                 ret = ret ? ret : ret2;
9737         }
9738 out_notrans:
9739         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9740                 up_read(&fs_info->subvol_sem);
9741         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9742                 up_read(&fs_info->subvol_sem);
9743
9744         return ret;
9745 }
9746
9747 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9748                                      struct btrfs_root *root,
9749                                      struct inode *dir,
9750                                      struct dentry *dentry)
9751 {
9752         int ret;
9753         struct inode *inode;
9754         u64 objectid;
9755         u64 index;
9756
9757         ret = btrfs_find_free_ino(root, &objectid);
9758         if (ret)
9759                 return ret;
9760
9761         inode = btrfs_new_inode(trans, root, dir,
9762                                 dentry->d_name.name,
9763                                 dentry->d_name.len,
9764                                 btrfs_ino(BTRFS_I(dir)),
9765                                 objectid,
9766                                 S_IFCHR | WHITEOUT_MODE,
9767                                 &index);
9768
9769         if (IS_ERR(inode)) {
9770                 ret = PTR_ERR(inode);
9771                 return ret;
9772         }
9773
9774         inode->i_op = &btrfs_special_inode_operations;
9775         init_special_inode(inode, inode->i_mode,
9776                 WHITEOUT_DEV);
9777
9778         ret = btrfs_init_inode_security(trans, inode, dir,
9779                                 &dentry->d_name);
9780         if (ret)
9781                 goto out;
9782
9783         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9784                                 BTRFS_I(inode), 0, index);
9785         if (ret)
9786                 goto out;
9787
9788         ret = btrfs_update_inode(trans, root, inode);
9789 out:
9790         unlock_new_inode(inode);
9791         if (ret)
9792                 inode_dec_link_count(inode);
9793         iput(inode);
9794
9795         return ret;
9796 }
9797
9798 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9799                            struct inode *new_dir, struct dentry *new_dentry,
9800                            unsigned int flags)
9801 {
9802         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9803         struct btrfs_trans_handle *trans;
9804         unsigned int trans_num_items;
9805         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9806         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9807         struct inode *new_inode = d_inode(new_dentry);
9808         struct inode *old_inode = d_inode(old_dentry);
9809         u64 index = 0;
9810         u64 root_objectid;
9811         int ret;
9812         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9813         bool log_pinned = false;
9814         struct btrfs_log_ctx ctx;
9815         bool sync_log = false;
9816         bool commit_transaction = false;
9817
9818         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9819                 return -EPERM;
9820
9821         /* we only allow rename subvolume link between subvolumes */
9822         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9823                 return -EXDEV;
9824
9825         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9826             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9827                 return -ENOTEMPTY;
9828
9829         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9830             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9831                 return -ENOTEMPTY;
9832
9833
9834         /* check for collisions, even if the  name isn't there */
9835         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9836                              new_dentry->d_name.name,
9837                              new_dentry->d_name.len);
9838
9839         if (ret) {
9840                 if (ret == -EEXIST) {
9841                         /* we shouldn't get
9842                          * eexist without a new_inode */
9843                         if (WARN_ON(!new_inode)) {
9844                                 return ret;
9845                         }
9846                 } else {
9847                         /* maybe -EOVERFLOW */
9848                         return ret;
9849                 }
9850         }
9851         ret = 0;
9852
9853         /*
9854          * we're using rename to replace one file with another.  Start IO on it
9855          * now so  we don't add too much work to the end of the transaction
9856          */
9857         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9858                 filemap_flush(old_inode->i_mapping);
9859
9860         /* close the racy window with snapshot create/destroy ioctl */
9861         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9862                 down_read(&fs_info->subvol_sem);
9863         /*
9864          * We want to reserve the absolute worst case amount of items.  So if
9865          * both inodes are subvols and we need to unlink them then that would
9866          * require 4 item modifications, but if they are both normal inodes it
9867          * would require 5 item modifications, so we'll assume they are normal
9868          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9869          * should cover the worst case number of items we'll modify.
9870          * If our rename has the whiteout flag, we need more 5 units for the
9871          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9872          * when selinux is enabled).
9873          */
9874         trans_num_items = 11;
9875         if (flags & RENAME_WHITEOUT)
9876                 trans_num_items += 5;
9877         trans = btrfs_start_transaction(root, trans_num_items);
9878         if (IS_ERR(trans)) {
9879                 ret = PTR_ERR(trans);
9880                 goto out_notrans;
9881         }
9882
9883         if (dest != root)
9884                 btrfs_record_root_in_trans(trans, dest);
9885
9886         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9887         if (ret)
9888                 goto out_fail;
9889
9890         BTRFS_I(old_inode)->dir_index = 0ULL;
9891         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9892                 /* force full log commit if subvolume involved. */
9893                 btrfs_set_log_full_commit(trans);
9894         } else {
9895                 btrfs_pin_log_trans(root);
9896                 log_pinned = true;
9897                 ret = btrfs_insert_inode_ref(trans, dest,
9898                                              new_dentry->d_name.name,
9899                                              new_dentry->d_name.len,
9900                                              old_ino,
9901                                              btrfs_ino(BTRFS_I(new_dir)), index);
9902                 if (ret)
9903                         goto out_fail;
9904         }
9905
9906         inode_inc_iversion(old_dir);
9907         inode_inc_iversion(new_dir);
9908         inode_inc_iversion(old_inode);
9909         old_dir->i_ctime = old_dir->i_mtime =
9910         new_dir->i_ctime = new_dir->i_mtime =
9911         old_inode->i_ctime = current_time(old_dir);
9912
9913         if (old_dentry->d_parent != new_dentry->d_parent)
9914                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9915                                 BTRFS_I(old_inode), 1);
9916
9917         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9918                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9919                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9920                                         old_dentry->d_name.name,
9921                                         old_dentry->d_name.len);
9922         } else {
9923                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9924                                         BTRFS_I(d_inode(old_dentry)),
9925                                         old_dentry->d_name.name,
9926                                         old_dentry->d_name.len);
9927                 if (!ret)
9928                         ret = btrfs_update_inode(trans, root, old_inode);
9929         }
9930         if (ret) {
9931                 btrfs_abort_transaction(trans, ret);
9932                 goto out_fail;
9933         }
9934
9935         if (new_inode) {
9936                 inode_inc_iversion(new_inode);
9937                 new_inode->i_ctime = current_time(new_inode);
9938                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9939                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9940                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9941                         ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9942                                                 new_dentry->d_name.name,
9943                                                 new_dentry->d_name.len);
9944                         BUG_ON(new_inode->i_nlink == 0);
9945                 } else {
9946                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9947                                                  BTRFS_I(d_inode(new_dentry)),
9948                                                  new_dentry->d_name.name,
9949                                                  new_dentry->d_name.len);
9950                 }
9951                 if (!ret && new_inode->i_nlink == 0)
9952                         ret = btrfs_orphan_add(trans,
9953                                         BTRFS_I(d_inode(new_dentry)));
9954                 if (ret) {
9955                         btrfs_abort_transaction(trans, ret);
9956                         goto out_fail;
9957                 }
9958         }
9959
9960         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9961                              new_dentry->d_name.name,
9962                              new_dentry->d_name.len, 0, index);
9963         if (ret) {
9964                 btrfs_abort_transaction(trans, ret);
9965                 goto out_fail;
9966         }
9967
9968         if (old_inode->i_nlink == 1)
9969                 BTRFS_I(old_inode)->dir_index = index;
9970
9971         if (log_pinned) {
9972                 struct dentry *parent = new_dentry->d_parent;
9973
9974                 btrfs_init_log_ctx(&ctx, old_inode);
9975                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9976                                          BTRFS_I(old_dir), parent,
9977                                          false, &ctx);
9978                 if (ret == BTRFS_NEED_LOG_SYNC)
9979                         sync_log = true;
9980                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9981                         commit_transaction = true;
9982                 ret = 0;
9983                 btrfs_end_log_trans(root);
9984                 log_pinned = false;
9985         }
9986
9987         if (flags & RENAME_WHITEOUT) {
9988                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9989                                                 old_dentry);
9990
9991                 if (ret) {
9992                         btrfs_abort_transaction(trans, ret);
9993                         goto out_fail;
9994                 }
9995         }
9996 out_fail:
9997         /*
9998          * If we have pinned the log and an error happened, we unpin tasks
9999          * trying to sync the log and force them to fallback to a transaction
10000          * commit if the log currently contains any of the inodes involved in
10001          * this rename operation (to ensure we do not persist a log with an
10002          * inconsistent state for any of these inodes or leading to any
10003          * inconsistencies when replayed). If the transaction was aborted, the
10004          * abortion reason is propagated to userspace when attempting to commit
10005          * the transaction. If the log does not contain any of these inodes, we
10006          * allow the tasks to sync it.
10007          */
10008         if (ret && log_pinned) {
10009                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10010                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10011                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10012                     (new_inode &&
10013                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10014                         btrfs_set_log_full_commit(trans);
10015
10016                 btrfs_end_log_trans(root);
10017                 log_pinned = false;
10018         }
10019         if (!ret && sync_log) {
10020                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
10021                 if (ret)
10022                         commit_transaction = true;
10023         }
10024         if (commit_transaction) {
10025                 ret = btrfs_commit_transaction(trans);
10026         } else {
10027                 int ret2;
10028
10029                 ret2 = btrfs_end_transaction(trans);
10030                 ret = ret ? ret : ret2;
10031         }
10032 out_notrans:
10033         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10034                 up_read(&fs_info->subvol_sem);
10035
10036         return ret;
10037 }
10038
10039 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10040                          struct inode *new_dir, struct dentry *new_dentry,
10041                          unsigned int flags)
10042 {
10043         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10044                 return -EINVAL;
10045
10046         if (flags & RENAME_EXCHANGE)
10047                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10048                                           new_dentry);
10049
10050         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10051 }
10052
10053 struct btrfs_delalloc_work {
10054         struct inode *inode;
10055         struct completion completion;
10056         struct list_head list;
10057         struct btrfs_work work;
10058 };
10059
10060 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10061 {
10062         struct btrfs_delalloc_work *delalloc_work;
10063         struct inode *inode;
10064
10065         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10066                                      work);
10067         inode = delalloc_work->inode;
10068         filemap_flush(inode->i_mapping);
10069         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10070                                 &BTRFS_I(inode)->runtime_flags))
10071                 filemap_flush(inode->i_mapping);
10072
10073         iput(inode);
10074         complete(&delalloc_work->completion);
10075 }
10076
10077 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
10078 {
10079         struct btrfs_delalloc_work *work;
10080
10081         work = kmalloc(sizeof(*work), GFP_NOFS);
10082         if (!work)
10083                 return NULL;
10084
10085         init_completion(&work->completion);
10086         INIT_LIST_HEAD(&work->list);
10087         work->inode = inode;
10088         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10089                         btrfs_run_delalloc_work, NULL, NULL);
10090
10091         return work;
10092 }
10093
10094 /*
10095  * some fairly slow code that needs optimization. This walks the list
10096  * of all the inodes with pending delalloc and forces them to disk.
10097  */
10098 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
10099 {
10100         struct btrfs_inode *binode;
10101         struct inode *inode;
10102         struct btrfs_delalloc_work *work, *next;
10103         struct list_head works;
10104         struct list_head splice;
10105         int ret = 0;
10106
10107         INIT_LIST_HEAD(&works);
10108         INIT_LIST_HEAD(&splice);
10109
10110         mutex_lock(&root->delalloc_mutex);
10111         spin_lock(&root->delalloc_lock);
10112         list_splice_init(&root->delalloc_inodes, &splice);
10113         while (!list_empty(&splice)) {
10114                 binode = list_entry(splice.next, struct btrfs_inode,
10115                                     delalloc_inodes);
10116
10117                 list_move_tail(&binode->delalloc_inodes,
10118                                &root->delalloc_inodes);
10119                 inode = igrab(&binode->vfs_inode);
10120                 if (!inode) {
10121                         cond_resched_lock(&root->delalloc_lock);
10122                         continue;
10123                 }
10124                 spin_unlock(&root->delalloc_lock);
10125
10126                 if (snapshot)
10127                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
10128                                 &binode->runtime_flags);
10129                 work = btrfs_alloc_delalloc_work(inode);
10130                 if (!work) {
10131                         iput(inode);
10132                         ret = -ENOMEM;
10133                         goto out;
10134                 }
10135                 list_add_tail(&work->list, &works);
10136                 btrfs_queue_work(root->fs_info->flush_workers,
10137                                  &work->work);
10138                 ret++;
10139                 if (nr != -1 && ret >= nr)
10140                         goto out;
10141                 cond_resched();
10142                 spin_lock(&root->delalloc_lock);
10143         }
10144         spin_unlock(&root->delalloc_lock);
10145
10146 out:
10147         list_for_each_entry_safe(work, next, &works, list) {
10148                 list_del_init(&work->list);
10149                 wait_for_completion(&work->completion);
10150                 kfree(work);
10151         }
10152
10153         if (!list_empty(&splice)) {
10154                 spin_lock(&root->delalloc_lock);
10155                 list_splice_tail(&splice, &root->delalloc_inodes);
10156                 spin_unlock(&root->delalloc_lock);
10157         }
10158         mutex_unlock(&root->delalloc_mutex);
10159         return ret;
10160 }
10161
10162 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
10163 {
10164         struct btrfs_fs_info *fs_info = root->fs_info;
10165         int ret;
10166
10167         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10168                 return -EROFS;
10169
10170         ret = start_delalloc_inodes(root, -1, true);
10171         if (ret > 0)
10172                 ret = 0;
10173         return ret;
10174 }
10175
10176 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
10177 {
10178         struct btrfs_root *root;
10179         struct list_head splice;
10180         int ret;
10181
10182         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10183                 return -EROFS;
10184
10185         INIT_LIST_HEAD(&splice);
10186
10187         mutex_lock(&fs_info->delalloc_root_mutex);
10188         spin_lock(&fs_info->delalloc_root_lock);
10189         list_splice_init(&fs_info->delalloc_roots, &splice);
10190         while (!list_empty(&splice) && nr) {
10191                 root = list_first_entry(&splice, struct btrfs_root,
10192                                         delalloc_root);
10193                 root = btrfs_grab_fs_root(root);
10194                 BUG_ON(!root);
10195                 list_move_tail(&root->delalloc_root,
10196                                &fs_info->delalloc_roots);
10197                 spin_unlock(&fs_info->delalloc_root_lock);
10198
10199                 ret = start_delalloc_inodes(root, nr, false);
10200                 btrfs_put_fs_root(root);
10201                 if (ret < 0)
10202                         goto out;
10203
10204                 if (nr != -1) {
10205                         nr -= ret;
10206                         WARN_ON(nr < 0);
10207                 }
10208                 spin_lock(&fs_info->delalloc_root_lock);
10209         }
10210         spin_unlock(&fs_info->delalloc_root_lock);
10211
10212         ret = 0;
10213 out:
10214         if (!list_empty(&splice)) {
10215                 spin_lock(&fs_info->delalloc_root_lock);
10216                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10217                 spin_unlock(&fs_info->delalloc_root_lock);
10218         }
10219         mutex_unlock(&fs_info->delalloc_root_mutex);
10220         return ret;
10221 }
10222
10223 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10224                          const char *symname)
10225 {
10226         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10227         struct btrfs_trans_handle *trans;
10228         struct btrfs_root *root = BTRFS_I(dir)->root;
10229         struct btrfs_path *path;
10230         struct btrfs_key key;
10231         struct inode *inode = NULL;
10232         int err;
10233         u64 objectid;
10234         u64 index = 0;
10235         int name_len;
10236         int datasize;
10237         unsigned long ptr;
10238         struct btrfs_file_extent_item *ei;
10239         struct extent_buffer *leaf;
10240
10241         name_len = strlen(symname);
10242         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10243                 return -ENAMETOOLONG;
10244
10245         /*
10246          * 2 items for inode item and ref
10247          * 2 items for dir items
10248          * 1 item for updating parent inode item
10249          * 1 item for the inline extent item
10250          * 1 item for xattr if selinux is on
10251          */
10252         trans = btrfs_start_transaction(root, 7);
10253         if (IS_ERR(trans))
10254                 return PTR_ERR(trans);
10255
10256         err = btrfs_find_free_ino(root, &objectid);
10257         if (err)
10258                 goto out_unlock;
10259
10260         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10261                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10262                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10263         if (IS_ERR(inode)) {
10264                 err = PTR_ERR(inode);
10265                 inode = NULL;
10266                 goto out_unlock;
10267         }
10268
10269         /*
10270         * If the active LSM wants to access the inode during
10271         * d_instantiate it needs these. Smack checks to see
10272         * if the filesystem supports xattrs by looking at the
10273         * ops vector.
10274         */
10275         inode->i_fop = &btrfs_file_operations;
10276         inode->i_op = &btrfs_file_inode_operations;
10277         inode->i_mapping->a_ops = &btrfs_aops;
10278         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10279
10280         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10281         if (err)
10282                 goto out_unlock;
10283
10284         path = btrfs_alloc_path();
10285         if (!path) {
10286                 err = -ENOMEM;
10287                 goto out_unlock;
10288         }
10289         key.objectid = btrfs_ino(BTRFS_I(inode));
10290         key.offset = 0;
10291         key.type = BTRFS_EXTENT_DATA_KEY;
10292         datasize = btrfs_file_extent_calc_inline_size(name_len);
10293         err = btrfs_insert_empty_item(trans, root, path, &key,
10294                                       datasize);
10295         if (err) {
10296                 btrfs_free_path(path);
10297                 goto out_unlock;
10298         }
10299         leaf = path->nodes[0];
10300         ei = btrfs_item_ptr(leaf, path->slots[0],
10301                             struct btrfs_file_extent_item);
10302         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10303         btrfs_set_file_extent_type(leaf, ei,
10304                                    BTRFS_FILE_EXTENT_INLINE);
10305         btrfs_set_file_extent_encryption(leaf, ei, 0);
10306         btrfs_set_file_extent_compression(leaf, ei, 0);
10307         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10308         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10309
10310         ptr = btrfs_file_extent_inline_start(ei);
10311         write_extent_buffer(leaf, symname, ptr, name_len);
10312         btrfs_mark_buffer_dirty(leaf);
10313         btrfs_free_path(path);
10314
10315         inode->i_op = &btrfs_symlink_inode_operations;
10316         inode_nohighmem(inode);
10317         inode_set_bytes(inode, name_len);
10318         btrfs_i_size_write(BTRFS_I(inode), name_len);
10319         err = btrfs_update_inode(trans, root, inode);
10320         /*
10321          * Last step, add directory indexes for our symlink inode. This is the
10322          * last step to avoid extra cleanup of these indexes if an error happens
10323          * elsewhere above.
10324          */
10325         if (!err)
10326                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10327                                 BTRFS_I(inode), 0, index);
10328         if (err)
10329                 goto out_unlock;
10330
10331         d_instantiate_new(dentry, inode);
10332
10333 out_unlock:
10334         btrfs_end_transaction(trans);
10335         if (err && inode) {
10336                 inode_dec_link_count(inode);
10337                 discard_new_inode(inode);
10338         }
10339         btrfs_btree_balance_dirty(fs_info);
10340         return err;
10341 }
10342
10343 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10344                                        u64 start, u64 num_bytes, u64 min_size,
10345                                        loff_t actual_len, u64 *alloc_hint,
10346                                        struct btrfs_trans_handle *trans)
10347 {
10348         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10349         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10350         struct extent_map *em;
10351         struct btrfs_root *root = BTRFS_I(inode)->root;
10352         struct btrfs_key ins;
10353         u64 cur_offset = start;
10354         u64 i_size;
10355         u64 cur_bytes;
10356         u64 last_alloc = (u64)-1;
10357         int ret = 0;
10358         bool own_trans = true;
10359         u64 end = start + num_bytes - 1;
10360
10361         if (trans)
10362                 own_trans = false;
10363         while (num_bytes > 0) {
10364                 if (own_trans) {
10365                         trans = btrfs_start_transaction(root, 3);
10366                         if (IS_ERR(trans)) {
10367                                 ret = PTR_ERR(trans);
10368                                 break;
10369                         }
10370                 }
10371
10372                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10373                 cur_bytes = max(cur_bytes, min_size);
10374                 /*
10375                  * If we are severely fragmented we could end up with really
10376                  * small allocations, so if the allocator is returning small
10377                  * chunks lets make its job easier by only searching for those
10378                  * sized chunks.
10379                  */
10380                 cur_bytes = min(cur_bytes, last_alloc);
10381                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10382                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10383                 if (ret) {
10384                         if (own_trans)
10385                                 btrfs_end_transaction(trans);
10386                         break;
10387                 }
10388                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10389
10390                 last_alloc = ins.offset;
10391                 ret = insert_reserved_file_extent(trans, inode,
10392                                                   cur_offset, ins.objectid,
10393                                                   ins.offset, ins.offset,
10394                                                   ins.offset, 0, 0, 0,
10395                                                   BTRFS_FILE_EXTENT_PREALLOC);
10396                 if (ret) {
10397                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10398                                                    ins.offset, 0);
10399                         btrfs_abort_transaction(trans, ret);
10400                         if (own_trans)
10401                                 btrfs_end_transaction(trans);
10402                         break;
10403                 }
10404
10405                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10406                                         cur_offset + ins.offset -1, 0);
10407
10408                 em = alloc_extent_map();
10409                 if (!em) {
10410                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10411                                 &BTRFS_I(inode)->runtime_flags);
10412                         goto next;
10413                 }
10414
10415                 em->start = cur_offset;
10416                 em->orig_start = cur_offset;
10417                 em->len = ins.offset;
10418                 em->block_start = ins.objectid;
10419                 em->block_len = ins.offset;
10420                 em->orig_block_len = ins.offset;
10421                 em->ram_bytes = ins.offset;
10422                 em->bdev = fs_info->fs_devices->latest_bdev;
10423                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10424                 em->generation = trans->transid;
10425
10426                 while (1) {
10427                         write_lock(&em_tree->lock);
10428                         ret = add_extent_mapping(em_tree, em, 1);
10429                         write_unlock(&em_tree->lock);
10430                         if (ret != -EEXIST)
10431                                 break;
10432                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10433                                                 cur_offset + ins.offset - 1,
10434                                                 0);
10435                 }
10436                 free_extent_map(em);
10437 next:
10438                 num_bytes -= ins.offset;
10439                 cur_offset += ins.offset;
10440                 *alloc_hint = ins.objectid + ins.offset;
10441
10442                 inode_inc_iversion(inode);
10443                 inode->i_ctime = current_time(inode);
10444                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10445                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10446                     (actual_len > inode->i_size) &&
10447                     (cur_offset > inode->i_size)) {
10448                         if (cur_offset > actual_len)
10449                                 i_size = actual_len;
10450                         else
10451                                 i_size = cur_offset;
10452                         i_size_write(inode, i_size);
10453                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10454                 }
10455
10456                 ret = btrfs_update_inode(trans, root, inode);
10457
10458                 if (ret) {
10459                         btrfs_abort_transaction(trans, ret);
10460                         if (own_trans)
10461                                 btrfs_end_transaction(trans);
10462                         break;
10463                 }
10464
10465                 if (own_trans)
10466                         btrfs_end_transaction(trans);
10467         }
10468         if (cur_offset < end)
10469                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10470                         end - cur_offset + 1);
10471         return ret;
10472 }
10473
10474 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10475                               u64 start, u64 num_bytes, u64 min_size,
10476                               loff_t actual_len, u64 *alloc_hint)
10477 {
10478         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10479                                            min_size, actual_len, alloc_hint,
10480                                            NULL);
10481 }
10482
10483 int btrfs_prealloc_file_range_trans(struct inode *inode,
10484                                     struct btrfs_trans_handle *trans, int mode,
10485                                     u64 start, u64 num_bytes, u64 min_size,
10486                                     loff_t actual_len, u64 *alloc_hint)
10487 {
10488         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10489                                            min_size, actual_len, alloc_hint, trans);
10490 }
10491
10492 static int btrfs_set_page_dirty(struct page *page)
10493 {
10494         return __set_page_dirty_nobuffers(page);
10495 }
10496
10497 static int btrfs_permission(struct inode *inode, int mask)
10498 {
10499         struct btrfs_root *root = BTRFS_I(inode)->root;
10500         umode_t mode = inode->i_mode;
10501
10502         if (mask & MAY_WRITE &&
10503             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10504                 if (btrfs_root_readonly(root))
10505                         return -EROFS;
10506                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10507                         return -EACCES;
10508         }
10509         return generic_permission(inode, mask);
10510 }
10511
10512 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10513 {
10514         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10515         struct btrfs_trans_handle *trans;
10516         struct btrfs_root *root = BTRFS_I(dir)->root;
10517         struct inode *inode = NULL;
10518         u64 objectid;
10519         u64 index;
10520         int ret = 0;
10521
10522         /*
10523          * 5 units required for adding orphan entry
10524          */
10525         trans = btrfs_start_transaction(root, 5);
10526         if (IS_ERR(trans))
10527                 return PTR_ERR(trans);
10528
10529         ret = btrfs_find_free_ino(root, &objectid);
10530         if (ret)
10531                 goto out;
10532
10533         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10534                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10535         if (IS_ERR(inode)) {
10536                 ret = PTR_ERR(inode);
10537                 inode = NULL;
10538                 goto out;
10539         }
10540
10541         inode->i_fop = &btrfs_file_operations;
10542         inode->i_op = &btrfs_file_inode_operations;
10543
10544         inode->i_mapping->a_ops = &btrfs_aops;
10545         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10546
10547         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10548         if (ret)
10549                 goto out;
10550
10551         ret = btrfs_update_inode(trans, root, inode);
10552         if (ret)
10553                 goto out;
10554         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10555         if (ret)
10556                 goto out;
10557
10558         /*
10559          * We set number of links to 0 in btrfs_new_inode(), and here we set
10560          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10561          * through:
10562          *
10563          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10564          */
10565         set_nlink(inode, 1);
10566         d_tmpfile(dentry, inode);
10567         unlock_new_inode(inode);
10568         mark_inode_dirty(inode);
10569 out:
10570         btrfs_end_transaction(trans);
10571         if (ret && inode)
10572                 discard_new_inode(inode);
10573         btrfs_btree_balance_dirty(fs_info);
10574         return ret;
10575 }
10576
10577 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10578 {
10579         struct inode *inode = tree->private_data;
10580         unsigned long index = start >> PAGE_SHIFT;
10581         unsigned long end_index = end >> PAGE_SHIFT;
10582         struct page *page;
10583
10584         while (index <= end_index) {
10585                 page = find_get_page(inode->i_mapping, index);
10586                 ASSERT(page); /* Pages should be in the extent_io_tree */
10587                 set_page_writeback(page);
10588                 put_page(page);
10589                 index++;
10590         }
10591 }
10592
10593 #ifdef CONFIG_SWAP
10594 /*
10595  * Add an entry indicating a block group or device which is pinned by a
10596  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
10597  * negative errno on failure.
10598  */
10599 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
10600                                   bool is_block_group)
10601 {
10602         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10603         struct btrfs_swapfile_pin *sp, *entry;
10604         struct rb_node **p;
10605         struct rb_node *parent = NULL;
10606
10607         sp = kmalloc(sizeof(*sp), GFP_NOFS);
10608         if (!sp)
10609                 return -ENOMEM;
10610         sp->ptr = ptr;
10611         sp->inode = inode;
10612         sp->is_block_group = is_block_group;
10613
10614         spin_lock(&fs_info->swapfile_pins_lock);
10615         p = &fs_info->swapfile_pins.rb_node;
10616         while (*p) {
10617                 parent = *p;
10618                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
10619                 if (sp->ptr < entry->ptr ||
10620                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
10621                         p = &(*p)->rb_left;
10622                 } else if (sp->ptr > entry->ptr ||
10623                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
10624                         p = &(*p)->rb_right;
10625                 } else {
10626                         spin_unlock(&fs_info->swapfile_pins_lock);
10627                         kfree(sp);
10628                         return 1;
10629                 }
10630         }
10631         rb_link_node(&sp->node, parent, p);
10632         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
10633         spin_unlock(&fs_info->swapfile_pins_lock);
10634         return 0;
10635 }
10636
10637 /* Free all of the entries pinned by this swapfile. */
10638 static void btrfs_free_swapfile_pins(struct inode *inode)
10639 {
10640         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10641         struct btrfs_swapfile_pin *sp;
10642         struct rb_node *node, *next;
10643
10644         spin_lock(&fs_info->swapfile_pins_lock);
10645         node = rb_first(&fs_info->swapfile_pins);
10646         while (node) {
10647                 next = rb_next(node);
10648                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
10649                 if (sp->inode == inode) {
10650                         rb_erase(&sp->node, &fs_info->swapfile_pins);
10651                         if (sp->is_block_group)
10652                                 btrfs_put_block_group(sp->ptr);
10653                         kfree(sp);
10654                 }
10655                 node = next;
10656         }
10657         spin_unlock(&fs_info->swapfile_pins_lock);
10658 }
10659
10660 struct btrfs_swap_info {
10661         u64 start;
10662         u64 block_start;
10663         u64 block_len;
10664         u64 lowest_ppage;
10665         u64 highest_ppage;
10666         unsigned long nr_pages;
10667         int nr_extents;
10668 };
10669
10670 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
10671                                  struct btrfs_swap_info *bsi)
10672 {
10673         unsigned long nr_pages;
10674         u64 first_ppage, first_ppage_reported, next_ppage;
10675         int ret;
10676
10677         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
10678         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
10679                                 PAGE_SIZE) >> PAGE_SHIFT;
10680
10681         if (first_ppage >= next_ppage)
10682                 return 0;
10683         nr_pages = next_ppage - first_ppage;
10684
10685         first_ppage_reported = first_ppage;
10686         if (bsi->start == 0)
10687                 first_ppage_reported++;
10688         if (bsi->lowest_ppage > first_ppage_reported)
10689                 bsi->lowest_ppage = first_ppage_reported;
10690         if (bsi->highest_ppage < (next_ppage - 1))
10691                 bsi->highest_ppage = next_ppage - 1;
10692
10693         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
10694         if (ret < 0)
10695                 return ret;
10696         bsi->nr_extents += ret;
10697         bsi->nr_pages += nr_pages;
10698         return 0;
10699 }
10700
10701 static void btrfs_swap_deactivate(struct file *file)
10702 {
10703         struct inode *inode = file_inode(file);
10704
10705         btrfs_free_swapfile_pins(inode);
10706         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
10707 }
10708
10709 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10710                                sector_t *span)
10711 {
10712         struct inode *inode = file_inode(file);
10713         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10714         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
10715         struct extent_state *cached_state = NULL;
10716         struct extent_map *em = NULL;
10717         struct btrfs_device *device = NULL;
10718         struct btrfs_swap_info bsi = {
10719                 .lowest_ppage = (sector_t)-1ULL,
10720         };
10721         int ret = 0;
10722         u64 isize;
10723         u64 start;
10724
10725         /*
10726          * If the swap file was just created, make sure delalloc is done. If the
10727          * file changes again after this, the user is doing something stupid and
10728          * we don't really care.
10729          */
10730         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
10731         if (ret)
10732                 return ret;
10733
10734         /*
10735          * The inode is locked, so these flags won't change after we check them.
10736          */
10737         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
10738                 btrfs_warn(fs_info, "swapfile must not be compressed");
10739                 return -EINVAL;
10740         }
10741         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
10742                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
10743                 return -EINVAL;
10744         }
10745         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
10746                 btrfs_warn(fs_info, "swapfile must not be checksummed");
10747                 return -EINVAL;
10748         }
10749
10750         /*
10751          * Balance or device remove/replace/resize can move stuff around from
10752          * under us. The EXCL_OP flag makes sure they aren't running/won't run
10753          * concurrently while we are mapping the swap extents, and
10754          * fs_info->swapfile_pins prevents them from running while the swap file
10755          * is active and moving the extents. Note that this also prevents a
10756          * concurrent device add which isn't actually necessary, but it's not
10757          * really worth the trouble to allow it.
10758          */
10759         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
10760                 btrfs_warn(fs_info,
10761            "cannot activate swapfile while exclusive operation is running");
10762                 return -EBUSY;
10763         }
10764         /*
10765          * Snapshots can create extents which require COW even if NODATACOW is
10766          * set. We use this counter to prevent snapshots. We must increment it
10767          * before walking the extents because we don't want a concurrent
10768          * snapshot to run after we've already checked the extents.
10769          */
10770         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
10771
10772         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
10773
10774         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
10775         start = 0;
10776         while (start < isize) {
10777                 u64 logical_block_start, physical_block_start;
10778                 struct btrfs_block_group_cache *bg;
10779                 u64 len = isize - start;
10780
10781                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
10782                 if (IS_ERR(em)) {
10783                         ret = PTR_ERR(em);
10784                         goto out;
10785                 }
10786
10787                 if (em->block_start == EXTENT_MAP_HOLE) {
10788                         btrfs_warn(fs_info, "swapfile must not have holes");
10789                         ret = -EINVAL;
10790                         goto out;
10791                 }
10792                 if (em->block_start == EXTENT_MAP_INLINE) {
10793                         /*
10794                          * It's unlikely we'll ever actually find ourselves
10795                          * here, as a file small enough to fit inline won't be
10796                          * big enough to store more than the swap header, but in
10797                          * case something changes in the future, let's catch it
10798                          * here rather than later.
10799                          */
10800                         btrfs_warn(fs_info, "swapfile must not be inline");
10801                         ret = -EINVAL;
10802                         goto out;
10803                 }
10804                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
10805                         btrfs_warn(fs_info, "swapfile must not be compressed");
10806                         ret = -EINVAL;
10807                         goto out;
10808                 }
10809
10810                 logical_block_start = em->block_start + (start - em->start);
10811                 len = min(len, em->len - (start - em->start));
10812                 free_extent_map(em);
10813                 em = NULL;
10814
10815                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
10816                 if (ret < 0) {
10817                         goto out;
10818                 } else if (ret) {
10819                         ret = 0;
10820                 } else {
10821                         btrfs_warn(fs_info,
10822                                    "swapfile must not be copy-on-write");
10823                         ret = -EINVAL;
10824                         goto out;
10825                 }
10826
10827                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
10828                 if (IS_ERR(em)) {
10829                         ret = PTR_ERR(em);
10830                         goto out;
10831                 }
10832
10833                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
10834                         btrfs_warn(fs_info,
10835                                    "swapfile must have single data profile");
10836                         ret = -EINVAL;
10837                         goto out;
10838                 }
10839
10840                 if (device == NULL) {
10841                         device = em->map_lookup->stripes[0].dev;
10842                         ret = btrfs_add_swapfile_pin(inode, device, false);
10843                         if (ret == 1)
10844                                 ret = 0;
10845                         else if (ret)
10846                                 goto out;
10847                 } else if (device != em->map_lookup->stripes[0].dev) {
10848                         btrfs_warn(fs_info, "swapfile must be on one device");
10849                         ret = -EINVAL;
10850                         goto out;
10851                 }
10852
10853                 physical_block_start = (em->map_lookup->stripes[0].physical +
10854                                         (logical_block_start - em->start));
10855                 len = min(len, em->len - (logical_block_start - em->start));
10856                 free_extent_map(em);
10857                 em = NULL;
10858
10859                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10860                 if (!bg) {
10861                         btrfs_warn(fs_info,
10862                            "could not find block group containing swapfile");
10863                         ret = -EINVAL;
10864                         goto out;
10865                 }
10866
10867                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10868                 if (ret) {
10869                         btrfs_put_block_group(bg);
10870                         if (ret == 1)
10871                                 ret = 0;
10872                         else
10873                                 goto out;
10874                 }
10875
10876                 if (bsi.block_len &&
10877                     bsi.block_start + bsi.block_len == physical_block_start) {
10878                         bsi.block_len += len;
10879                 } else {
10880                         if (bsi.block_len) {
10881                                 ret = btrfs_add_swap_extent(sis, &bsi);
10882                                 if (ret)
10883                                         goto out;
10884                         }
10885                         bsi.start = start;
10886                         bsi.block_start = physical_block_start;
10887                         bsi.block_len = len;
10888                 }
10889
10890                 start += len;
10891         }
10892
10893         if (bsi.block_len)
10894                 ret = btrfs_add_swap_extent(sis, &bsi);
10895
10896 out:
10897         if (!IS_ERR_OR_NULL(em))
10898                 free_extent_map(em);
10899
10900         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10901
10902         if (ret)
10903                 btrfs_swap_deactivate(file);
10904
10905         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10906
10907         if (ret)
10908                 return ret;
10909
10910         if (device)
10911                 sis->bdev = device->bdev;
10912         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10913         sis->max = bsi.nr_pages;
10914         sis->pages = bsi.nr_pages - 1;
10915         sis->highest_bit = bsi.nr_pages - 1;
10916         return bsi.nr_extents;
10917 }
10918 #else
10919 static void btrfs_swap_deactivate(struct file *file)
10920 {
10921 }
10922
10923 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10924                                sector_t *span)
10925 {
10926         return -EOPNOTSUPP;
10927 }
10928 #endif
10929
10930 static const struct inode_operations btrfs_dir_inode_operations = {
10931         .getattr        = btrfs_getattr,
10932         .lookup         = btrfs_lookup,
10933         .create         = btrfs_create,
10934         .unlink         = btrfs_unlink,
10935         .link           = btrfs_link,
10936         .mkdir          = btrfs_mkdir,
10937         .rmdir          = btrfs_rmdir,
10938         .rename         = btrfs_rename2,
10939         .symlink        = btrfs_symlink,
10940         .setattr        = btrfs_setattr,
10941         .mknod          = btrfs_mknod,
10942         .listxattr      = btrfs_listxattr,
10943         .permission     = btrfs_permission,
10944         .get_acl        = btrfs_get_acl,
10945         .set_acl        = btrfs_set_acl,
10946         .update_time    = btrfs_update_time,
10947         .tmpfile        = btrfs_tmpfile,
10948 };
10949 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10950         .lookup         = btrfs_lookup,
10951         .permission     = btrfs_permission,
10952         .update_time    = btrfs_update_time,
10953 };
10954
10955 static const struct file_operations btrfs_dir_file_operations = {
10956         .llseek         = generic_file_llseek,
10957         .read           = generic_read_dir,
10958         .iterate_shared = btrfs_real_readdir,
10959         .open           = btrfs_opendir,
10960         .unlocked_ioctl = btrfs_ioctl,
10961 #ifdef CONFIG_COMPAT
10962         .compat_ioctl   = btrfs_compat_ioctl,
10963 #endif
10964         .release        = btrfs_release_file,
10965         .fsync          = btrfs_sync_file,
10966 };
10967
10968 static const struct extent_io_ops btrfs_extent_io_ops = {
10969         /* mandatory callbacks */
10970         .submit_bio_hook = btrfs_submit_bio_hook,
10971         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10972 };
10973
10974 /*
10975  * btrfs doesn't support the bmap operation because swapfiles
10976  * use bmap to make a mapping of extents in the file.  They assume
10977  * these extents won't change over the life of the file and they
10978  * use the bmap result to do IO directly to the drive.
10979  *
10980  * the btrfs bmap call would return logical addresses that aren't
10981  * suitable for IO and they also will change frequently as COW
10982  * operations happen.  So, swapfile + btrfs == corruption.
10983  *
10984  * For now we're avoiding this by dropping bmap.
10985  */
10986 static const struct address_space_operations btrfs_aops = {
10987         .readpage       = btrfs_readpage,
10988         .writepage      = btrfs_writepage,
10989         .writepages     = btrfs_writepages,
10990         .readpages      = btrfs_readpages,
10991         .direct_IO      = btrfs_direct_IO,
10992         .invalidatepage = btrfs_invalidatepage,
10993         .releasepage    = btrfs_releasepage,
10994         .set_page_dirty = btrfs_set_page_dirty,
10995         .error_remove_page = generic_error_remove_page,
10996         .swap_activate  = btrfs_swap_activate,
10997         .swap_deactivate = btrfs_swap_deactivate,
10998 };
10999
11000 static const struct inode_operations btrfs_file_inode_operations = {
11001         .getattr        = btrfs_getattr,
11002         .setattr        = btrfs_setattr,
11003         .listxattr      = btrfs_listxattr,
11004         .permission     = btrfs_permission,
11005         .fiemap         = btrfs_fiemap,
11006         .get_acl        = btrfs_get_acl,
11007         .set_acl        = btrfs_set_acl,
11008         .update_time    = btrfs_update_time,
11009 };
11010 static const struct inode_operations btrfs_special_inode_operations = {
11011         .getattr        = btrfs_getattr,
11012         .setattr        = btrfs_setattr,
11013         .permission     = btrfs_permission,
11014         .listxattr      = btrfs_listxattr,
11015         .get_acl        = btrfs_get_acl,
11016         .set_acl        = btrfs_set_acl,
11017         .update_time    = btrfs_update_time,
11018 };
11019 static const struct inode_operations btrfs_symlink_inode_operations = {
11020         .get_link       = page_get_link,
11021         .getattr        = btrfs_getattr,
11022         .setattr        = btrfs_setattr,
11023         .permission     = btrfs_permission,
11024         .listxattr      = btrfs_listxattr,
11025         .update_time    = btrfs_update_time,
11026 };
11027
11028 const struct dentry_operations btrfs_dentry_operations = {
11029         .d_delete       = btrfs_dentry_delete,
11030 };