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/migrate.h>
  32 #include <linux/sched/mm.h>
  33 #include <asm/unaligned.h>
  34 #include "misc.h"
  35 #include "ctree.h"
  36 #include "disk-io.h"
  37 #include "transaction.h"
  38 #include "btrfs_inode.h"
  39 #include "print-tree.h"
  40 #include "ordered-data.h"
  41 #include "xattr.h"
  42 #include "tree-log.h"
  43 #include "volumes.h"
  44 #include "compression.h"
  45 #include "locking.h"
  46 #include "free-space-cache.h"
  47 #include "inode-map.h"
  48 #include "props.h"
  49 #include "qgroup.h"
  50 #include "delalloc-space.h"
  51 #include "block-group.h"
  52
  53 struct btrfs_iget_args {
  54         struct btrfs_key *location;
  55         struct btrfs_root *root;
  56 };
  57
  58 struct btrfs_dio_data {
  59         u64 reserve;
  60         u64 unsubmitted_oe_range_start;
  61         u64 unsubmitted_oe_range_end;
  62         int overwrite;
  63 };
  64
  65 static const struct inode_operations btrfs_dir_inode_operations;
  66 static const struct inode_operations btrfs_symlink_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 struct kmem_cache *btrfs_free_space_bitmap_cachep;
  78
  79 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  80 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  81 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  82 static noinline int cow_file_range(struct inode *inode,
  83                                    struct page *locked_page,
  84                                    u64 start, u64 end, int *page_started,
  85                                    unsigned long *nr_written, int unlock);
  86 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  87                                        u64 orig_start, u64 block_start,
  88                                        u64 block_len, u64 orig_block_len,
  89                                        u64 ram_bytes, int compress_type,
  90                                        int type);
  91
  92 static void __endio_write_update_ordered(struct inode *inode,
  93                                          const u64 offset, const u64 bytes,
  94                                          const bool uptodate);
  95
  96 /*
  97  * Cleanup all submitted ordered extents in specified range to handle errors
  98  * from the btrfs_run_delalloc_range() callback.
  99  *
 100  * NOTE: caller must ensure that when an error happens, it can not call
 101  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 102  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 103  * to be released, which we want to happen only when finishing the ordered
 104  * extent (btrfs_finish_ordered_io()).
 105  */
 106 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 107                                                  struct page *locked_page,
 108                                                  u64 offset, u64 bytes)
 109 {
 110         unsigned long index = offset >> PAGE_SHIFT;
 111         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 112         u64 page_start = page_offset(locked_page);
 113         u64 page_end = page_start + PAGE_SIZE - 1;
 114
 115         struct page *page;
 116
 117         while (index <= end_index) {
 118                 page = find_get_page(inode->i_mapping, index);
 119                 index++;
 120                 if (!page)
 121                         continue;
 122                 ClearPagePrivate2(page);
 123                 put_page(page);
 124         }
 125
 126         /*
 127          * In case this page belongs to the delalloc range being instantiated
 128          * then skip it, since the first page of a range is going to be
 129          * properly cleaned up by the caller of run_delalloc_range
 130          */
 131         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
 132                 offset += PAGE_SIZE;
 133                 bytes -= PAGE_SIZE;
 134         }
 135
 136         return __endio_write_update_ordered(inode, offset, bytes, false);
 137 }
 138
 139 static int btrfs_dirty_inode(struct inode *inode);
 140
 141 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 142 void btrfs_test_inode_set_ops(struct inode *inode)
 143 {
 144         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 145 }
 146 #endif
 147
 148 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 149                                      struct inode *inode,  struct inode *dir,
 150                                      const struct qstr *qstr)
 151 {
 152         int err;
 153
 154         err = btrfs_init_acl(trans, inode, dir);
 155         if (!err)
 156                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 157         return err;
 158 }
 159
 160 /*
 161  * this does all the hard work for inserting an inline extent into
 162  * the btree.  The caller should have done a btrfs_drop_extents so that
 163  * no overlapping inline items exist in the btree
 164  */
 165 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 166                                 struct btrfs_path *path, int extent_inserted,
 167                                 struct btrfs_root *root, struct inode *inode,
 168                                 u64 start, size_t size, size_t compressed_size,
 169                                 int compress_type,
 170                                 struct page **compressed_pages)
 171 {
 172         struct extent_buffer *leaf;
 173         struct page *page = NULL;
 174         char *kaddr;
 175         unsigned long ptr;
 176         struct btrfs_file_extent_item *ei;
 177         int ret;
 178         size_t cur_size = size;
 179         unsigned long offset;
 180
 181         ASSERT((compressed_size > 0 && compressed_pages) ||
 182                (compressed_size == 0 && !compressed_pages));
 183
 184         if (compressed_size && compressed_pages)
 185                 cur_size = compressed_size;
 186
 187         inode_add_bytes(inode, size);
 188
 189         if (!extent_inserted) {
 190                 struct btrfs_key key;
 191                 size_t datasize;
 192
 193                 key.objectid = btrfs_ino(BTRFS_I(inode));
 194                 key.offset = start;
 195                 key.type = BTRFS_EXTENT_DATA_KEY;
 196
 197                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 198                 path->leave_spinning = 1;
 199                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 200                                               datasize);
 201                 if (ret)
 202                         goto fail;
 203         }
 204         leaf = path->nodes[0];
 205         ei = btrfs_item_ptr(leaf, path->slots[0],
 206                             struct btrfs_file_extent_item);
 207         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 208         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 209         btrfs_set_file_extent_encryption(leaf, ei, 0);
 210         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 211         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 212         ptr = btrfs_file_extent_inline_start(ei);
 213
 214         if (compress_type != BTRFS_COMPRESS_NONE) {
 215                 struct page *cpage;
 216                 int i = 0;
 217                 while (compressed_size > 0) {
 218                         cpage = compressed_pages[i];
 219                         cur_size = min_t(unsigned long, compressed_size,
 220                                        PAGE_SIZE);
 221
 222                         kaddr = kmap_atomic(cpage);
 223                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 224                         kunmap_atomic(kaddr);
 225
 226                         i++;
 227                         ptr += cur_size;
 228                         compressed_size -= cur_size;
 229                 }
 230                 btrfs_set_file_extent_compression(leaf, ei,
 231                                                   compress_type);
 232         } else {
 233                 page = find_get_page(inode->i_mapping,
 234                                      start >> PAGE_SHIFT);
 235                 btrfs_set_file_extent_compression(leaf, ei, 0);
 236                 kaddr = kmap_atomic(page);
 237                 offset = offset_in_page(start);
 238                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 239                 kunmap_atomic(kaddr);
 240                 put_page(page);
 241         }
 242         btrfs_mark_buffer_dirty(leaf);
 243         btrfs_release_path(path);
 244
 245         /*
 246          * We align size to sectorsize for inline extents just for simplicity
 247          * sake.
 248          */
 249         size = ALIGN(size, root->fs_info->sectorsize);
 250         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), start, size);
 251         if (ret)
 252                 goto fail;
 253
 254         /*
 255          * we're an inline extent, so nobody can
 256          * extend the file past i_size without locking
 257          * a page we already have locked.
 258          *
 259          * We must do any isize and inode updates
 260          * before we unlock the pages.  Otherwise we
 261          * could end up racing with unlink.
 262          */
 263         BTRFS_I(inode)->disk_i_size = inode->i_size;
 264         ret = btrfs_update_inode(trans, root, inode);
 265
 266 fail:
 267         return ret;
 268 }
 269
 270
 271 /*
 272  * conditionally insert an inline extent into the file.  This
 273  * does the checks required to make sure the data is small enough
 274  * to fit as an inline extent.
 275  */
 276 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 277                                           u64 end, size_t compressed_size,
 278                                           int compress_type,
 279                                           struct page **compressed_pages)
 280 {
 281         struct btrfs_root *root = BTRFS_I(inode)->root;
 282         struct btrfs_fs_info *fs_info = root->fs_info;
 283         struct btrfs_trans_handle *trans;
 284         u64 isize = i_size_read(inode);
 285         u64 actual_end = min(end + 1, isize);
 286         u64 inline_len = actual_end - start;
 287         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 288         u64 data_len = inline_len;
 289         int ret;
 290         struct btrfs_path *path;
 291         int extent_inserted = 0;
 292         u32 extent_item_size;
 293
 294         if (compressed_size)
 295                 data_len = compressed_size;
 296
 297         if (start > 0 ||
 298             actual_end > fs_info->sectorsize ||
 299             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 300             (!compressed_size &&
 301             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 302             end + 1 < isize ||
 303             data_len > fs_info->max_inline) {
 304                 return 1;
 305         }
 306
 307         path = btrfs_alloc_path();
 308         if (!path)
 309                 return -ENOMEM;
 310
 311         trans = btrfs_join_transaction(root);
 312         if (IS_ERR(trans)) {
 313                 btrfs_free_path(path);
 314                 return PTR_ERR(trans);
 315         }
 316         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 317
 318         if (compressed_size && compressed_pages)
 319                 extent_item_size = btrfs_file_extent_calc_inline_size(
 320                    compressed_size);
 321         else
 322                 extent_item_size = btrfs_file_extent_calc_inline_size(
 323                     inline_len);
 324
 325         ret = __btrfs_drop_extents(trans, root, inode, path,
 326                                    start, aligned_end, NULL,
 327                                    1, 1, extent_item_size, &extent_inserted);
 328         if (ret) {
 329                 btrfs_abort_transaction(trans, ret);
 330                 goto out;
 331         }
 332
 333         if (isize > actual_end)
 334                 inline_len = min_t(u64, isize, actual_end);
 335         ret = insert_inline_extent(trans, path, extent_inserted,
 336                                    root, inode, start,
 337                                    inline_len, compressed_size,
 338                                    compress_type, compressed_pages);
 339         if (ret && ret != -ENOSPC) {
 340                 btrfs_abort_transaction(trans, ret);
 341                 goto out;
 342         } else if (ret == -ENOSPC) {
 343                 ret = 1;
 344                 goto out;
 345         }
 346
 347         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 348         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 349 out:
 350         /*
 351          * Don't forget to free the reserved space, as for inlined extent
 352          * it won't count as data extent, free them directly here.
 353          * And at reserve time, it's always aligned to page size, so
 354          * just free one page here.
 355          */
 356         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 357         btrfs_free_path(path);
 358         btrfs_end_transaction(trans);
 359         return ret;
 360 }
 361
 362 struct async_extent {
 363         u64 start;
 364         u64 ram_size;
 365         u64 compressed_size;
 366         struct page **pages;
 367         unsigned long nr_pages;
 368         int compress_type;
 369         struct list_head list;
 370 };
 371
 372 struct async_chunk {
 373         struct inode *inode;
 374         struct page *locked_page;
 375         u64 start;
 376         u64 end;
 377         unsigned int write_flags;
 378         struct list_head extents;
 379         struct cgroup_subsys_state *blkcg_css;
 380         struct btrfs_work work;
 381         atomic_t *pending;
 382 };
 383
 384 struct async_cow {
 385         /* Number of chunks in flight; must be first in the structure */
 386         atomic_t num_chunks;
 387         struct async_chunk chunks[];
 388 };
 389
 390 static noinline int add_async_extent(struct async_chunk *cow,
 391                                      u64 start, u64 ram_size,
 392                                      u64 compressed_size,
 393                                      struct page **pages,
 394                                      unsigned long nr_pages,
 395                                      int compress_type)
 396 {
 397         struct async_extent *async_extent;
 398
 399         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 400         BUG_ON(!async_extent); /* -ENOMEM */
 401         async_extent->start = start;
 402         async_extent->ram_size = ram_size;
 403         async_extent->compressed_size = compressed_size;
 404         async_extent->pages = pages;
 405         async_extent->nr_pages = nr_pages;
 406         async_extent->compress_type = compress_type;
 407         list_add_tail(&async_extent->list, &cow->extents);
 408         return 0;
 409 }
 410
 411 /*
 412  * Check if the inode has flags compatible with compression
 413  */
 414 static inline bool inode_can_compress(struct inode *inode)
 415 {
 416         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW ||
 417             BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 418                 return false;
 419         return true;
 420 }
 421
 422 /*
 423  * Check if the inode needs to be submitted to compression, based on mount
 424  * options, defragmentation, properties or heuristics.
 425  */
 426 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 427 {
 428         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 429
 430         if (!inode_can_compress(inode)) {
 431                 WARN(IS_ENABLED(CONFIG_BTRFS_DEBUG),
 432                         KERN_ERR "BTRFS: unexpected compression for ino %llu\n",
 433                         btrfs_ino(BTRFS_I(inode)));
 434                 return 0;
 435         }
 436         /* force compress */
 437         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 438                 return 1;
 439         /* defrag ioctl */
 440         if (BTRFS_I(inode)->defrag_compress)
 441                 return 1;
 442         /* bad compression ratios */
 443         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 444                 return 0;
 445         if (btrfs_test_opt(fs_info, COMPRESS) ||
 446             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 447             BTRFS_I(inode)->prop_compress)
 448                 return btrfs_compress_heuristic(inode, start, end);
 449         return 0;
 450 }
 451
 452 static inline void inode_should_defrag(struct btrfs_inode *inode,
 453                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 454 {
 455         /* If this is a small write inside eof, kick off a defrag */
 456         if (num_bytes < small_write &&
 457             (start > 0 || end + 1 < inode->disk_i_size))
 458                 btrfs_add_inode_defrag(NULL, inode);
 459 }
 460
 461 /*
 462  * we create compressed extents in two phases.  The first
 463  * phase compresses a range of pages that have already been
 464  * locked (both pages and state bits are locked).
 465  *
 466  * This is done inside an ordered work queue, and the compression
 467  * is spread across many cpus.  The actual IO submission is step
 468  * two, and the ordered work queue takes care of making sure that
 469  * happens in the same order things were put onto the queue by
 470  * writepages and friends.
 471  *
 472  * If this code finds it can't get good compression, it puts an
 473  * entry onto the work queue to write the uncompressed bytes.  This
 474  * makes sure that both compressed inodes and uncompressed inodes
 475  * are written in the same order that the flusher thread sent them
 476  * down.
 477  */
 478 static noinline int compress_file_range(struct async_chunk *async_chunk)
 479 {
 480         struct inode *inode = async_chunk->inode;
 481         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 482         u64 blocksize = fs_info->sectorsize;
 483         u64 start = async_chunk->start;
 484         u64 end = async_chunk->end;
 485         u64 actual_end;
 486         u64 i_size;
 487         int ret = 0;
 488         struct page **pages = NULL;
 489         unsigned long nr_pages;
 490         unsigned long total_compressed = 0;
 491         unsigned long total_in = 0;
 492         int i;
 493         int will_compress;
 494         int compress_type = fs_info->compress_type;
 495         int compressed_extents = 0;
 496         int redirty = 0;
 497
 498         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 499                         SZ_16K);
 500
 501         /*
 502          * We need to save i_size before now because it could change in between
 503          * us evaluating the size and assigning it.  This is because we lock and
 504          * unlock the page in truncate and fallocate, and then modify the i_size
 505          * later on.
 506          *
 507          * The barriers are to emulate READ_ONCE, remove that once i_size_read
 508          * does that for us.
 509          */
 510         barrier();
 511         i_size = i_size_read(inode);
 512         barrier();
 513         actual_end = min_t(u64, i_size, end + 1);
 514 again:
 515         will_compress = 0;
 516         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 517         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 518         nr_pages = min_t(unsigned long, nr_pages,
 519                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 520
 521         /*
 522          * we don't want to send crud past the end of i_size through
 523          * compression, that's just a waste of CPU time.  So, if the
 524          * end of the file is before the start of our current
 525          * requested range of bytes, we bail out to the uncompressed
 526          * cleanup code that can deal with all of this.
 527          *
 528          * It isn't really the fastest way to fix things, but this is a
 529          * very uncommon corner.
 530          */
 531         if (actual_end <= start)
 532                 goto cleanup_and_bail_uncompressed;
 533
 534         total_compressed = actual_end - start;
 535
 536         /*
 537          * skip compression for a small file range(<=blocksize) that
 538          * isn't an inline extent, since it doesn't save disk space at all.
 539          */
 540         if (total_compressed <= blocksize &&
 541            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 542                 goto cleanup_and_bail_uncompressed;
 543
 544         total_compressed = min_t(unsigned long, total_compressed,
 545                         BTRFS_MAX_UNCOMPRESSED);
 546         total_in = 0;
 547         ret = 0;
 548
 549         /*
 550          * we do compression for mount -o compress and when the
 551          * inode has not been flagged as nocompress.  This flag can
 552          * change at any time if we discover bad compression ratios.
 553          */
 554         if (inode_need_compress(inode, start, end)) {
 555                 WARN_ON(pages);
 556                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 557                 if (!pages) {
 558                         /* just bail out to the uncompressed code */
 559                         nr_pages = 0;
 560                         goto cont;
 561                 }
 562
 563                 if (BTRFS_I(inode)->defrag_compress)
 564                         compress_type = BTRFS_I(inode)->defrag_compress;
 565                 else if (BTRFS_I(inode)->prop_compress)
 566                         compress_type = BTRFS_I(inode)->prop_compress;
 567
 568                 /*
 569                  * we need to call clear_page_dirty_for_io on each
 570                  * page in the range.  Otherwise applications with the file
 571                  * mmap'd can wander in and change the page contents while
 572                  * we are compressing them.
 573                  *
 574                  * If the compression fails for any reason, we set the pages
 575                  * dirty again later on.
 576                  *
 577                  * Note that the remaining part is redirtied, the start pointer
 578                  * has moved, the end is the original one.
 579                  */
 580                 if (!redirty) {
 581                         extent_range_clear_dirty_for_io(inode, start, end);
 582                         redirty = 1;
 583                 }
 584
 585                 /* Compression level is applied here and only here */
 586                 ret = btrfs_compress_pages(
 587                         compress_type | (fs_info->compress_level << 4),
 588                                            inode->i_mapping, start,
 589                                            pages,
 590                                            &nr_pages,
 591                                            &total_in,
 592                                            &total_compressed);
 593
 594                 if (!ret) {
 595                         unsigned long offset = offset_in_page(total_compressed);
 596                         struct page *page = pages[nr_pages - 1];
 597                         char *kaddr;
 598
 599                         /* zero the tail end of the last page, we might be
 600                          * sending it down to disk
 601                          */
 602                         if (offset) {
 603                                 kaddr = kmap_atomic(page);
 604                                 memset(kaddr + offset, 0,
 605                                        PAGE_SIZE - offset);
 606                                 kunmap_atomic(kaddr);
 607                         }
 608                         will_compress = 1;
 609                 }
 610         }
 611 cont:
 612         if (start == 0) {
 613                 /* lets try to make an inline extent */
 614                 if (ret || total_in < actual_end) {
 615                         /* we didn't compress the entire range, try
 616                          * to make an uncompressed inline extent.
 617                          */
 618                         ret = cow_file_range_inline(inode, start, end, 0,
 619                                                     BTRFS_COMPRESS_NONE, NULL);
 620                 } else {
 621                         /* try making a compressed inline extent */
 622                         ret = cow_file_range_inline(inode, start, end,
 623                                                     total_compressed,
 624                                                     compress_type, pages);
 625                 }
 626                 if (ret <= 0) {
 627                         unsigned long clear_flags = EXTENT_DELALLOC |
 628                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 629                                 EXTENT_DO_ACCOUNTING;
 630                         unsigned long page_error_op;
 631
 632                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 633
 634                         /*
 635                          * inline extent creation worked or returned error,
 636                          * we don't need to create any more async work items.
 637                          * Unlock and free up our temp pages.
 638                          *
 639                          * We use DO_ACCOUNTING here because we need the
 640                          * delalloc_release_metadata to be done _after_ we drop
 641                          * our outstanding extent for clearing delalloc for this
 642                          * range.
 643                          */
 644                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 645                                                      clear_flags,
 646                                                      PAGE_UNLOCK |
 647                                                      PAGE_CLEAR_DIRTY |
 648                                                      PAGE_SET_WRITEBACK |
 649                                                      page_error_op |
 650                                                      PAGE_END_WRITEBACK);
 651
 652                         for (i = 0; i < nr_pages; i++) {
 653                                 WARN_ON(pages[i]->mapping);
 654                                 put_page(pages[i]);
 655                         }
 656                         kfree(pages);
 657
 658                         return 0;
 659                 }
 660         }
 661
 662         if (will_compress) {
 663                 /*
 664                  * we aren't doing an inline extent round the compressed size
 665                  * up to a block size boundary so the allocator does sane
 666                  * things
 667                  */
 668                 total_compressed = ALIGN(total_compressed, blocksize);
 669
 670                 /*
 671                  * one last check to make sure the compression is really a
 672                  * win, compare the page count read with the blocks on disk,
 673                  * compression must free at least one sector size
 674                  */
 675                 total_in = ALIGN(total_in, PAGE_SIZE);
 676                 if (total_compressed + blocksize <= total_in) {
 677                         compressed_extents++;
 678
 679                         /*
 680                          * The async work queues will take care of doing actual
 681                          * allocation on disk for these compressed pages, and
 682                          * will submit them to the elevator.
 683                          */
 684                         add_async_extent(async_chunk, start, total_in,
 685                                         total_compressed, pages, nr_pages,
 686                                         compress_type);
 687
 688                         if (start + total_in < end) {
 689                                 start += total_in;
 690                                 pages = NULL;
 691                                 cond_resched();
 692                                 goto again;
 693                         }
 694                         return compressed_extents;
 695                 }
 696         }
 697         if (pages) {
 698                 /*
 699                  * the compression code ran but failed to make things smaller,
 700                  * free any pages it allocated and our page pointer array
 701                  */
 702                 for (i = 0; i < nr_pages; i++) {
 703                         WARN_ON(pages[i]->mapping);
 704                         put_page(pages[i]);
 705                 }
 706                 kfree(pages);
 707                 pages = NULL;
 708                 total_compressed = 0;
 709                 nr_pages = 0;
 710
 711                 /* flag the file so we don't compress in the future */
 712                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 713                     !(BTRFS_I(inode)->prop_compress)) {
 714                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 715                 }
 716         }
 717 cleanup_and_bail_uncompressed:
 718         /*
 719          * No compression, but we still need to write the pages in the file
 720          * we've been given so far.  redirty the locked page if it corresponds
 721          * to our extent and set things up for the async work queue to run
 722          * cow_file_range to do the normal delalloc dance.
 723          */
 724         if (async_chunk->locked_page &&
 725             (page_offset(async_chunk->locked_page) >= start &&
 726              page_offset(async_chunk->locked_page)) <= end) {
 727                 __set_page_dirty_nobuffers(async_chunk->locked_page);
 728                 /* unlocked later on in the async handlers */
 729         }
 730
 731         if (redirty)
 732                 extent_range_redirty_for_io(inode, start, end);
 733         add_async_extent(async_chunk, start, end - start + 1, 0, NULL, 0,
 734                          BTRFS_COMPRESS_NONE);
 735         compressed_extents++;
 736
 737         return compressed_extents;
 738 }
 739
 740 static void free_async_extent_pages(struct async_extent *async_extent)
 741 {
 742         int i;
 743
 744         if (!async_extent->pages)
 745                 return;
 746
 747         for (i = 0; i < async_extent->nr_pages; i++) {
 748                 WARN_ON(async_extent->pages[i]->mapping);
 749                 put_page(async_extent->pages[i]);
 750         }
 751         kfree(async_extent->pages);
 752         async_extent->nr_pages = 0;
 753         async_extent->pages = NULL;
 754 }
 755
 756 /*
 757  * phase two of compressed writeback.  This is the ordered portion
 758  * of the code, which only gets called in the order the work was
 759  * queued.  We walk all the async extents created by compress_file_range
 760  * and send them down to the disk.
 761  */
 762 static noinline void submit_compressed_extents(struct async_chunk *async_chunk)
 763 {
 764         struct inode *inode = async_chunk->inode;
 765         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 766         struct async_extent *async_extent;
 767         u64 alloc_hint = 0;
 768         struct btrfs_key ins;
 769         struct extent_map *em;
 770         struct btrfs_root *root = BTRFS_I(inode)->root;
 771         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 772         int ret = 0;
 773
 774 again:
 775         while (!list_empty(&async_chunk->extents)) {
 776                 async_extent = list_entry(async_chunk->extents.next,
 777                                           struct async_extent, list);
 778                 list_del(&async_extent->list);
 779
 780 retry:
 781                 lock_extent(io_tree, async_extent->start,
 782                             async_extent->start + async_extent->ram_size - 1);
 783                 /* did the compression code fall back to uncompressed IO? */
 784                 if (!async_extent->pages) {
 785                         int page_started = 0;
 786                         unsigned long nr_written = 0;
 787
 788                         /* allocate blocks */
 789                         ret = cow_file_range(inode, async_chunk->locked_page,
 790                                              async_extent->start,
 791                                              async_extent->start +
 792                                              async_extent->ram_size - 1,
 793                                              &page_started, &nr_written, 0);
 794
 795                         /* JDM XXX */
 796
 797                         /*
 798                          * if page_started, cow_file_range inserted an
 799                          * inline extent and took care of all the unlocking
 800                          * and IO for us.  Otherwise, we need to submit
 801                          * all those pages down to the drive.
 802                          */
 803                         if (!page_started && !ret)
 804                                 extent_write_locked_range(inode,
 805                                                   async_extent->start,
 806                                                   async_extent->start +
 807                                                   async_extent->ram_size - 1,
 808                                                   WB_SYNC_ALL);
 809                         else if (ret && async_chunk->locked_page)
 810                                 unlock_page(async_chunk->locked_page);
 811                         kfree(async_extent);
 812                         cond_resched();
 813                         continue;
 814                 }
 815
 816                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 817                                            async_extent->compressed_size,
 818                                            async_extent->compressed_size,
 819                                            0, alloc_hint, &ins, 1, 1);
 820                 if (ret) {
 821                         free_async_extent_pages(async_extent);
 822
 823                         if (ret == -ENOSPC) {
 824                                 unlock_extent(io_tree, async_extent->start,
 825                                               async_extent->start +
 826                                               async_extent->ram_size - 1);
 827
 828                                 /*
 829                                  * we need to redirty the pages if we decide to
 830                                  * fallback to uncompressed IO, otherwise we
 831                                  * will not submit these pages down to lower
 832                                  * layers.
 833                                  */
 834                                 extent_range_redirty_for_io(inode,
 835                                                 async_extent->start,
 836                                                 async_extent->start +
 837                                                 async_extent->ram_size - 1);
 838
 839                                 goto retry;
 840                         }
 841                         goto out_free;
 842                 }
 843                 /*
 844                  * here we're doing allocation and writeback of the
 845                  * compressed pages
 846                  */
 847                 em = create_io_em(inode, async_extent->start,
 848                                   async_extent->ram_size, /* len */
 849                                   async_extent->start, /* orig_start */
 850                                   ins.objectid, /* block_start */
 851                                   ins.offset, /* block_len */
 852                                   ins.offset, /* orig_block_len */
 853                                   async_extent->ram_size, /* ram_bytes */
 854                                   async_extent->compress_type,
 855                                   BTRFS_ORDERED_COMPRESSED);
 856                 if (IS_ERR(em))
 857                         /* ret value is not necessary due to void function */
 858                         goto out_free_reserve;
 859                 free_extent_map(em);
 860
 861                 ret = btrfs_add_ordered_extent_compress(inode,
 862                                                 async_extent->start,
 863                                                 ins.objectid,
 864                                                 async_extent->ram_size,
 865                                                 ins.offset,
 866                                                 BTRFS_ORDERED_COMPRESSED,
 867                                                 async_extent->compress_type);
 868                 if (ret) {
 869                         btrfs_drop_extent_cache(BTRFS_I(inode),
 870                                                 async_extent->start,
 871                                                 async_extent->start +
 872                                                 async_extent->ram_size - 1, 0);
 873                         goto out_free_reserve;
 874                 }
 875                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 876
 877                 /*
 878                  * clear dirty, set writeback and unlock the pages.
 879                  */
 880                 extent_clear_unlock_delalloc(inode, async_extent->start,
 881                                 async_extent->start +
 882                                 async_extent->ram_size - 1,
 883                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 884                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 885                                 PAGE_SET_WRITEBACK);
 886                 if (btrfs_submit_compressed_write(inode,
 887                                     async_extent->start,
 888                                     async_extent->ram_size,
 889                                     ins.objectid,
 890                                     ins.offset, async_extent->pages,
 891                                     async_extent->nr_pages,
 892                                     async_chunk->write_flags,
 893                                     async_chunk->blkcg_css)) {
 894                         struct page *p = async_extent->pages[0];
 895                         const u64 start = async_extent->start;
 896                         const u64 end = start + async_extent->ram_size - 1;
 897
 898                         p->mapping = inode->i_mapping;
 899                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
 900
 901                         p->mapping = NULL;
 902                         extent_clear_unlock_delalloc(inode, start, end,
 903                                                      NULL, 0,
 904                                                      PAGE_END_WRITEBACK |
 905                                                      PAGE_SET_ERROR);
 906                         free_async_extent_pages(async_extent);
 907                 }
 908                 alloc_hint = ins.objectid + ins.offset;
 909                 kfree(async_extent);
 910                 cond_resched();
 911         }
 912         return;
 913 out_free_reserve:
 914         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 915         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 916 out_free:
 917         extent_clear_unlock_delalloc(inode, async_extent->start,
 918                                      async_extent->start +
 919                                      async_extent->ram_size - 1,
 920                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 921                                      EXTENT_DELALLOC_NEW |
 922                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 923                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 924                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 925                                      PAGE_SET_ERROR);
 926         free_async_extent_pages(async_extent);
 927         kfree(async_extent);
 928         goto again;
 929 }
 930
 931 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 932                                       u64 num_bytes)
 933 {
 934         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 935         struct extent_map *em;
 936         u64 alloc_hint = 0;
 937
 938         read_lock(&em_tree->lock);
 939         em = search_extent_mapping(em_tree, start, num_bytes);
 940         if (em) {
 941                 /*
 942                  * if block start isn't an actual block number then find the
 943                  * first block in this inode and use that as a hint.  If that
 944                  * block is also bogus then just don't worry about it.
 945                  */
 946                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 947                         free_extent_map(em);
 948                         em = search_extent_mapping(em_tree, 0, 0);
 949                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 950                                 alloc_hint = em->block_start;
 951                         if (em)
 952                                 free_extent_map(em);
 953                 } else {
 954                         alloc_hint = em->block_start;
 955                         free_extent_map(em);
 956                 }
 957         }
 958         read_unlock(&em_tree->lock);
 959
 960         return alloc_hint;
 961 }
 962
 963 /*
 964  * when extent_io.c finds a delayed allocation range in the file,
 965  * the call backs end up in this code.  The basic idea is to
 966  * allocate extents on disk for the range, and create ordered data structs
 967  * in ram to track those extents.
 968  *
 969  * locked_page is the page that writepage had locked already.  We use
 970  * it to make sure we don't do extra locks or unlocks.
 971  *
 972  * *page_started is set to one if we unlock locked_page and do everything
 973  * required to start IO on it.  It may be clean and already done with
 974  * IO when we return.
 975  */
 976 static noinline int cow_file_range(struct inode *inode,
 977                                    struct page *locked_page,
 978                                    u64 start, u64 end, int *page_started,
 979                                    unsigned long *nr_written, int unlock)
 980 {
 981         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 982         struct btrfs_root *root = BTRFS_I(inode)->root;
 983         u64 alloc_hint = 0;
 984         u64 num_bytes;
 985         unsigned long ram_size;
 986         u64 cur_alloc_size = 0;
 987         u64 blocksize = fs_info->sectorsize;
 988         struct btrfs_key ins;
 989         struct extent_map *em;
 990         unsigned clear_bits;
 991         unsigned long page_ops;
 992         bool extent_reserved = false;
 993         int ret = 0;
 994
 995         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 996                 WARN_ON_ONCE(1);
 997                 ret = -EINVAL;
 998                 goto out_unlock;
 999         }
1000
1001         num_bytes = ALIGN(end - start + 1, blocksize);
1002         num_bytes = max(blocksize,  num_bytes);
1003         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
1004
1005         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
1006
1007         if (start == 0) {
1008                 /* lets try to make an inline extent */
1009                 ret = cow_file_range_inline(inode, start, end, 0,
1010                                             BTRFS_COMPRESS_NONE, NULL);
1011                 if (ret == 0) {
1012                         /*
1013                          * We use DO_ACCOUNTING here because we need the
1014                          * delalloc_release_metadata to be run _after_ we drop
1015                          * our outstanding extent for clearing delalloc for this
1016                          * range.
1017                          */
1018                         extent_clear_unlock_delalloc(inode, start, end, NULL,
1019                                      EXTENT_LOCKED | EXTENT_DELALLOC |
1020                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1021                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1022                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1023                                      PAGE_END_WRITEBACK);
1024                         *nr_written = *nr_written +
1025                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1026                         *page_started = 1;
1027                         goto out;
1028                 } else if (ret < 0) {
1029                         goto out_unlock;
1030                 }
1031         }
1032
1033         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1034         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1035                         start + num_bytes - 1, 0);
1036
1037         while (num_bytes > 0) {
1038                 cur_alloc_size = num_bytes;
1039                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1040                                            fs_info->sectorsize, 0, alloc_hint,
1041                                            &ins, 1, 1);
1042                 if (ret < 0)
1043                         goto out_unlock;
1044                 cur_alloc_size = ins.offset;
1045                 extent_reserved = true;
1046
1047                 ram_size = ins.offset;
1048                 em = create_io_em(inode, start, ins.offset, /* len */
1049                                   start, /* orig_start */
1050                                   ins.objectid, /* block_start */
1051                                   ins.offset, /* block_len */
1052                                   ins.offset, /* orig_block_len */
1053                                   ram_size, /* ram_bytes */
1054                                   BTRFS_COMPRESS_NONE, /* compress_type */
1055                                   BTRFS_ORDERED_REGULAR /* type */);
1056                 if (IS_ERR(em)) {
1057                         ret = PTR_ERR(em);
1058                         goto out_reserve;
1059                 }
1060                 free_extent_map(em);
1061
1062                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1063                                                ram_size, cur_alloc_size, 0);
1064                 if (ret)
1065                         goto out_drop_extent_cache;
1066
1067                 if (root->root_key.objectid ==
1068                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1069                         ret = btrfs_reloc_clone_csums(inode, start,
1070                                                       cur_alloc_size);
1071                         /*
1072                          * Only drop cache here, and process as normal.
1073                          *
1074                          * We must not allow extent_clear_unlock_delalloc()
1075                          * at out_unlock label to free meta of this ordered
1076                          * extent, as its meta should be freed by
1077                          * btrfs_finish_ordered_io().
1078                          *
1079                          * So we must continue until @start is increased to
1080                          * skip current ordered extent.
1081                          */
1082                         if (ret)
1083                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1084                                                 start + ram_size - 1, 0);
1085                 }
1086
1087                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1088
1089                 /* we're not doing compressed IO, don't unlock the first
1090                  * page (which the caller expects to stay locked), don't
1091                  * clear any dirty bits and don't set any writeback bits
1092                  *
1093                  * Do set the Private2 bit so we know this page was properly
1094                  * setup for writepage
1095                  */
1096                 page_ops = unlock ? PAGE_UNLOCK : 0;
1097                 page_ops |= PAGE_SET_PRIVATE2;
1098
1099                 extent_clear_unlock_delalloc(inode, start,
1100                                              start + ram_size - 1,
1101                                              locked_page,
1102                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1103                                              page_ops);
1104                 if (num_bytes < cur_alloc_size)
1105                         num_bytes = 0;
1106                 else
1107                         num_bytes -= cur_alloc_size;
1108                 alloc_hint = ins.objectid + ins.offset;
1109                 start += cur_alloc_size;
1110                 extent_reserved = false;
1111
1112                 /*
1113                  * btrfs_reloc_clone_csums() error, since start is increased
1114                  * extent_clear_unlock_delalloc() at out_unlock label won't
1115                  * free metadata of current ordered extent, we're OK to exit.
1116                  */
1117                 if (ret)
1118                         goto out_unlock;
1119         }
1120 out:
1121         return ret;
1122
1123 out_drop_extent_cache:
1124         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1125 out_reserve:
1126         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1127         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1128 out_unlock:
1129         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1130                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1131         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1132                 PAGE_END_WRITEBACK;
1133         /*
1134          * If we reserved an extent for our delalloc range (or a subrange) and
1135          * failed to create the respective ordered extent, then it means that
1136          * when we reserved the extent we decremented the extent's size from
1137          * the data space_info's bytes_may_use counter and incremented the
1138          * space_info's bytes_reserved counter by the same amount. We must make
1139          * sure extent_clear_unlock_delalloc() does not try to decrement again
1140          * the data space_info's bytes_may_use counter, therefore we do not pass
1141          * it the flag EXTENT_CLEAR_DATA_RESV.
1142          */
1143         if (extent_reserved) {
1144                 extent_clear_unlock_delalloc(inode, start,
1145                                              start + cur_alloc_size,
1146                                              locked_page,
1147                                              clear_bits,
1148                                              page_ops);
1149                 start += cur_alloc_size;
1150                 if (start >= end)
1151                         goto out;
1152         }
1153         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1154                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1155                                      page_ops);
1156         goto out;
1157 }
1158
1159 /*
1160  * work queue call back to started compression on a file and pages
1161  */
1162 static noinline void async_cow_start(struct btrfs_work *work)
1163 {
1164         struct async_chunk *async_chunk;
1165         int compressed_extents;
1166
1167         async_chunk = container_of(work, struct async_chunk, work);
1168
1169         compressed_extents = compress_file_range(async_chunk);
1170         if (compressed_extents == 0) {
1171                 btrfs_add_delayed_iput(async_chunk->inode);
1172                 async_chunk->inode = NULL;
1173         }
1174 }
1175
1176 /*
1177  * work queue call back to submit previously compressed pages
1178  */
1179 static noinline void async_cow_submit(struct btrfs_work *work)
1180 {
1181         struct async_chunk *async_chunk = container_of(work, struct async_chunk,
1182                                                      work);
1183         struct btrfs_fs_info *fs_info = btrfs_work_owner(work);
1184         unsigned long nr_pages;
1185
1186         nr_pages = (async_chunk->end - async_chunk->start + PAGE_SIZE) >>
1187                 PAGE_SHIFT;
1188
1189         /* atomic_sub_return implies a barrier */
1190         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1191             5 * SZ_1M)
1192                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1193
1194         /*
1195          * ->inode could be NULL if async_chunk_start has failed to compress,
1196          * in which case we don't have anything to submit, yet we need to
1197          * always adjust ->async_delalloc_pages as its paired with the init
1198          * happening in cow_file_range_async
1199          */
1200         if (async_chunk->inode)
1201                 submit_compressed_extents(async_chunk);
1202 }
1203
1204 static noinline void async_cow_free(struct btrfs_work *work)
1205 {
1206         struct async_chunk *async_chunk;
1207
1208         async_chunk = container_of(work, struct async_chunk, work);
1209         if (async_chunk->inode)
1210                 btrfs_add_delayed_iput(async_chunk->inode);
1211         if (async_chunk->blkcg_css)
1212                 css_put(async_chunk->blkcg_css);
1213         /*
1214          * Since the pointer to 'pending' is at the beginning of the array of
1215          * async_chunk's, freeing it ensures the whole array has been freed.
1216          */
1217         if (atomic_dec_and_test(async_chunk->pending))
1218                 kvfree(async_chunk->pending);
1219 }
1220
1221 static int cow_file_range_async(struct inode *inode,
1222                                 struct writeback_control *wbc,
1223                                 struct page *locked_page,
1224                                 u64 start, u64 end, int *page_started,
1225                                 unsigned long *nr_written)
1226 {
1227         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1228         struct cgroup_subsys_state *blkcg_css = wbc_blkcg_css(wbc);
1229         struct async_cow *ctx;
1230         struct async_chunk *async_chunk;
1231         unsigned long nr_pages;
1232         u64 cur_end;
1233         u64 num_chunks = DIV_ROUND_UP(end - start, SZ_512K);
1234         int i;
1235         bool should_compress;
1236         unsigned nofs_flag;
1237         const unsigned int write_flags = wbc_to_write_flags(wbc);
1238
1239         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1240
1241         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1242             !btrfs_test_opt(fs_info, FORCE_COMPRESS)) {
1243                 num_chunks = 1;
1244                 should_compress = false;
1245         } else {
1246                 should_compress = true;
1247         }
1248
1249         nofs_flag = memalloc_nofs_save();
1250         ctx = kvmalloc(struct_size(ctx, chunks, num_chunks), GFP_KERNEL);
1251         memalloc_nofs_restore(nofs_flag);
1252
1253         if (!ctx) {
1254                 unsigned clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC |
1255                         EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1256                         EXTENT_DO_ACCOUNTING;
1257                 unsigned long page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1258                         PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
1259                         PAGE_SET_ERROR;
1260
1261                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1262                                              clear_bits, page_ops);
1263                 return -ENOMEM;
1264         }
1265
1266         async_chunk = ctx->chunks;
1267         atomic_set(&ctx->num_chunks, num_chunks);
1268
1269         for (i = 0; i < num_chunks; i++) {
1270                 if (should_compress)
1271                         cur_end = min(end, start + SZ_512K - 1);
1272                 else
1273                         cur_end = end;
1274
1275                 /*
1276                  * igrab is called higher up in the call chain, take only the
1277                  * lightweight reference for the callback lifetime
1278                  */
1279                 ihold(inode);
1280                 async_chunk[i].pending = &ctx->num_chunks;
1281                 async_chunk[i].inode = inode;
1282                 async_chunk[i].start = start;
1283                 async_chunk[i].end = cur_end;
1284                 async_chunk[i].write_flags = write_flags;
1285                 INIT_LIST_HEAD(&async_chunk[i].extents);
1286
1287                 /*
1288                  * The locked_page comes all the way from writepage and its
1289                  * the original page we were actually given.  As we spread
1290                  * this large delalloc region across multiple async_chunk
1291                  * structs, only the first struct needs a pointer to locked_page
1292                  *
1293                  * This way we don't need racey decisions about who is supposed
1294                  * to unlock it.
1295                  */
1296                 if (locked_page) {
1297                         /*
1298                          * Depending on the compressibility, the pages might or
1299                          * might not go through async.  We want all of them to
1300                          * be accounted against wbc once.  Let's do it here
1301                          * before the paths diverge.  wbc accounting is used
1302                          * only for foreign writeback detection and doesn't
1303                          * need full accuracy.  Just account the whole thing
1304                          * against the first page.
1305                          */
1306                         wbc_account_cgroup_owner(wbc, locked_page,
1307                                                  cur_end - start);
1308                         async_chunk[i].locked_page = locked_page;
1309                         locked_page = NULL;
1310                 } else {
1311                         async_chunk[i].locked_page = NULL;
1312                 }
1313
1314                 if (blkcg_css != blkcg_root_css) {
1315                         css_get(blkcg_css);
1316                         async_chunk[i].blkcg_css = blkcg_css;
1317                 } else {
1318                         async_chunk[i].blkcg_css = NULL;
1319                 }
1320
1321                 btrfs_init_work(&async_chunk[i].work, async_cow_start,
1322                                 async_cow_submit, async_cow_free);
1323
1324                 nr_pages = DIV_ROUND_UP(cur_end - start, PAGE_SIZE);
1325                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1326
1327                 btrfs_queue_work(fs_info->delalloc_workers, &async_chunk[i].work);
1328
1329                 *nr_written += nr_pages;
1330                 start = cur_end + 1;
1331         }
1332         *page_started = 1;
1333         return 0;
1334 }
1335
1336 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1337                                         u64 bytenr, u64 num_bytes)
1338 {
1339         int ret;
1340         struct btrfs_ordered_sum *sums;
1341         LIST_HEAD(list);
1342
1343         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1344                                        bytenr + num_bytes - 1, &list, 0);
1345         if (ret == 0 && list_empty(&list))
1346                 return 0;
1347
1348         while (!list_empty(&list)) {
1349                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1350                 list_del(&sums->list);
1351                 kfree(sums);
1352         }
1353         if (ret < 0)
1354                 return ret;
1355         return 1;
1356 }
1357
1358 /*
1359  * when nowcow writeback call back.  This checks for snapshots or COW copies
1360  * of the extents that exist in the file, and COWs the file as required.
1361  *
1362  * If no cow copies or snapshots exist, we write directly to the existing
1363  * blocks on disk
1364  */
1365 static noinline int run_delalloc_nocow(struct inode *inode,
1366                                        struct page *locked_page,
1367                                        const u64 start, const u64 end,
1368                                        int *page_started, int force,
1369                                        unsigned long *nr_written)
1370 {
1371         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1372         struct btrfs_root *root = BTRFS_I(inode)->root;
1373         struct btrfs_path *path;
1374         u64 cow_start = (u64)-1;
1375         u64 cur_offset = start;
1376         int ret;
1377         bool check_prev = true;
1378         const bool freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
1379         u64 ino = btrfs_ino(BTRFS_I(inode));
1380         bool nocow = false;
1381         u64 disk_bytenr = 0;
1382
1383         path = btrfs_alloc_path();
1384         if (!path) {
1385                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1386                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1387                                              EXTENT_DO_ACCOUNTING |
1388                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1389                                              PAGE_CLEAR_DIRTY |
1390                                              PAGE_SET_WRITEBACK |
1391                                              PAGE_END_WRITEBACK);
1392                 return -ENOMEM;
1393         }
1394
1395         while (1) {
1396                 struct btrfs_key found_key;
1397                 struct btrfs_file_extent_item *fi;
1398                 struct extent_buffer *leaf;
1399                 u64 extent_end;
1400                 u64 extent_offset;
1401                 u64 num_bytes = 0;
1402                 u64 disk_num_bytes;
1403                 u64 ram_bytes;
1404                 int extent_type;
1405
1406                 nocow = false;
1407
1408                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1409                                                cur_offset, 0);
1410                 if (ret < 0)
1411                         goto error;
1412
1413                 /*
1414                  * If there is no extent for our range when doing the initial
1415                  * search, then go back to the previous slot as it will be the
1416                  * one containing the search offset
1417                  */
1418                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1419                         leaf = path->nodes[0];
1420                         btrfs_item_key_to_cpu(leaf, &found_key,
1421                                               path->slots[0] - 1);
1422                         if (found_key.objectid == ino &&
1423                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1424                                 path->slots[0]--;
1425                 }
1426                 check_prev = false;
1427 next_slot:
1428                 /* Go to next leaf if we have exhausted the current one */
1429                 leaf = path->nodes[0];
1430                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1431                         ret = btrfs_next_leaf(root, path);
1432                         if (ret < 0) {
1433                                 if (cow_start != (u64)-1)
1434                                         cur_offset = cow_start;
1435                                 goto error;
1436                         }
1437                         if (ret > 0)
1438                                 break;
1439                         leaf = path->nodes[0];
1440                 }
1441
1442                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1443
1444                 /* Didn't find anything for our INO */
1445                 if (found_key.objectid > ino)
1446                         break;
1447                 /*
1448                  * Keep searching until we find an EXTENT_ITEM or there are no
1449                  * more extents for this inode
1450                  */
1451                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1452                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1453                         path->slots[0]++;
1454                         goto next_slot;
1455                 }
1456
1457                 /* Found key is not EXTENT_DATA_KEY or starts after req range */
1458                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1459                     found_key.offset > end)
1460                         break;
1461
1462                 /*
1463                  * If the found extent starts after requested offset, then
1464                  * adjust extent_end to be right before this extent begins
1465                  */
1466                 if (found_key.offset > cur_offset) {
1467                         extent_end = found_key.offset;
1468                         extent_type = 0;
1469                         goto out_check;
1470                 }
1471
1472                 /*
1473                  * Found extent which begins before our range and potentially
1474                  * intersect it
1475                  */
1476                 fi = btrfs_item_ptr(leaf, path->slots[0],
1477                                     struct btrfs_file_extent_item);
1478                 extent_type = btrfs_file_extent_type(leaf, fi);
1479
1480                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1481                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1482                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1483                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1484                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1485                         extent_end = found_key.offset +
1486                                 btrfs_file_extent_num_bytes(leaf, fi);
1487                         disk_num_bytes =
1488                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1489                         /*
1490                          * If the extent we got ends before our current offset,
1491                          * skip to the next extent.
1492                          */
1493                         if (extent_end <= cur_offset) {
1494                                 path->slots[0]++;
1495                                 goto next_slot;
1496                         }
1497                         /* Skip holes */
1498                         if (disk_bytenr == 0)
1499                                 goto out_check;
1500                         /* Skip compressed/encrypted/encoded extents */
1501                         if (btrfs_file_extent_compression(leaf, fi) ||
1502                             btrfs_file_extent_encryption(leaf, fi) ||
1503                             btrfs_file_extent_other_encoding(leaf, fi))
1504                                 goto out_check;
1505                         /*
1506                          * If extent is created before the last volume's snapshot
1507                          * this implies the extent is shared, hence we can't do
1508                          * nocow. This is the same check as in
1509                          * btrfs_cross_ref_exist but without calling
1510                          * btrfs_search_slot.
1511                          */
1512                         if (!freespace_inode &&
1513                             btrfs_file_extent_generation(leaf, fi) <=
1514                             btrfs_root_last_snapshot(&root->root_item))
1515                                 goto out_check;
1516                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1517                                 goto out_check;
1518                         /* If extent is RO, we must COW it */
1519                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1520                                 goto out_check;
1521                         ret = btrfs_cross_ref_exist(root, ino,
1522                                                     found_key.offset -
1523                                                     extent_offset, disk_bytenr);
1524                         if (ret) {
1525                                 /*
1526                                  * ret could be -EIO if the above fails to read
1527                                  * metadata.
1528                                  */
1529                                 if (ret < 0) {
1530                                         if (cow_start != (u64)-1)
1531                                                 cur_offset = cow_start;
1532                                         goto error;
1533                                 }
1534
1535                                 WARN_ON_ONCE(freespace_inode);
1536                                 goto out_check;
1537                         }
1538                         disk_bytenr += extent_offset;
1539                         disk_bytenr += cur_offset - found_key.offset;
1540                         num_bytes = min(end + 1, extent_end) - cur_offset;
1541                         /*
1542                          * If there are pending snapshots for this root, we
1543                          * fall into common COW way
1544                          */
1545                         if (!freespace_inode && atomic_read(&root->snapshot_force_cow))
1546                                 goto out_check;
1547                         /*
1548                          * force cow if csum exists in the range.
1549                          * this ensure that csum for a given extent are
1550                          * either valid or do not exist.
1551                          */
1552                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1553                                                   num_bytes);
1554                         if (ret) {
1555                                 /*
1556                                  * ret could be -EIO if the above fails to read
1557                                  * metadata.
1558                                  */
1559                                 if (ret < 0) {
1560                                         if (cow_start != (u64)-1)
1561                                                 cur_offset = cow_start;
1562                                         goto error;
1563                                 }
1564                                 WARN_ON_ONCE(freespace_inode);
1565                                 goto out_check;
1566                         }
1567                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1568                                 goto out_check;
1569                         nocow = true;
1570                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1571                         extent_end = found_key.offset + ram_bytes;
1572                         extent_end = ALIGN(extent_end, fs_info->sectorsize);
1573                         /* Skip extents outside of our requested range */
1574                         if (extent_end <= start) {
1575                                 path->slots[0]++;
1576                                 goto next_slot;
1577                         }
1578                 } else {
1579                         /* If this triggers then we have a memory corruption */
1580                         BUG();
1581                 }
1582 out_check:
1583                 /*
1584                  * If nocow is false then record the beginning of the range
1585                  * that needs to be COWed
1586                  */
1587                 if (!nocow) {
1588                         if (cow_start == (u64)-1)
1589                                 cow_start = cur_offset;
1590                         cur_offset = extent_end;
1591                         if (cur_offset > end)
1592                                 break;
1593                         path->slots[0]++;
1594                         goto next_slot;
1595                 }
1596
1597                 btrfs_release_path(path);
1598
1599                 /*
1600                  * COW range from cow_start to found_key.offset - 1. As the key
1601                  * will contain the beginning of the first extent that can be
1602                  * NOCOW, following one which needs to be COW'ed
1603                  */
1604                 if (cow_start != (u64)-1) {
1605                         ret = cow_file_range(inode, locked_page,
1606                                              cow_start, found_key.offset - 1,
1607                                              page_started, nr_written, 1);
1608                         if (ret) {
1609                                 if (nocow)
1610                                         btrfs_dec_nocow_writers(fs_info,
1611                                                                 disk_bytenr);
1612                                 goto error;
1613                         }
1614                         cow_start = (u64)-1;
1615                 }
1616
1617                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1618                         u64 orig_start = found_key.offset - extent_offset;
1619                         struct extent_map *em;
1620
1621                         em = create_io_em(inode, cur_offset, num_bytes,
1622                                           orig_start,
1623                                           disk_bytenr, /* block_start */
1624                                           num_bytes, /* block_len */
1625                                           disk_num_bytes, /* orig_block_len */
1626                                           ram_bytes, BTRFS_COMPRESS_NONE,
1627                                           BTRFS_ORDERED_PREALLOC);
1628                         if (IS_ERR(em)) {
1629                                 if (nocow)
1630                                         btrfs_dec_nocow_writers(fs_info,
1631                                                                 disk_bytenr);
1632                                 ret = PTR_ERR(em);
1633                                 goto error;
1634                         }
1635                         free_extent_map(em);
1636                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1637                                                        disk_bytenr, num_bytes,
1638                                                        num_bytes,
1639                                                        BTRFS_ORDERED_PREALLOC);
1640                         if (ret) {
1641                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1642                                                         cur_offset,
1643                                                         cur_offset + num_bytes - 1,
1644                                                         0);
1645                                 goto error;
1646                         }
1647                 } else {
1648                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1649                                                        disk_bytenr, num_bytes,
1650                                                        num_bytes,
1651                                                        BTRFS_ORDERED_NOCOW);
1652                         if (ret)
1653                                 goto error;
1654                 }
1655
1656                 if (nocow)
1657                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1658                 nocow = false;
1659
1660                 if (root->root_key.objectid ==
1661                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1662                         /*
1663                          * Error handled later, as we must prevent
1664                          * extent_clear_unlock_delalloc() in error handler
1665                          * from freeing metadata of created ordered extent.
1666                          */
1667                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1668                                                       num_bytes);
1669
1670                 extent_clear_unlock_delalloc(inode, cur_offset,
1671                                              cur_offset + num_bytes - 1,
1672                                              locked_page, EXTENT_LOCKED |
1673                                              EXTENT_DELALLOC |
1674                                              EXTENT_CLEAR_DATA_RESV,
1675                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1676
1677                 cur_offset = extent_end;
1678
1679                 /*
1680                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1681                  * handler, as metadata for created ordered extent will only
1682                  * be freed by btrfs_finish_ordered_io().
1683                  */
1684                 if (ret)
1685                         goto error;
1686                 if (cur_offset > end)
1687                         break;
1688         }
1689         btrfs_release_path(path);
1690
1691         if (cur_offset <= end && cow_start == (u64)-1)
1692                 cow_start = cur_offset;
1693
1694         if (cow_start != (u64)-1) {
1695                 cur_offset = end;
1696                 ret = cow_file_range(inode, locked_page, cow_start, end,
1697                                      page_started, nr_written, 1);
1698                 if (ret)
1699                         goto error;
1700         }
1701
1702 error:
1703         if (nocow)
1704                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1705
1706         if (ret && cur_offset < end)
1707                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1708                                              locked_page, EXTENT_LOCKED |
1709                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1710                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1711                                              PAGE_CLEAR_DIRTY |
1712                                              PAGE_SET_WRITEBACK |
1713                                              PAGE_END_WRITEBACK);
1714         btrfs_free_path(path);
1715         return ret;
1716 }
1717
1718 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1719 {
1720
1721         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1722             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1723                 return 0;
1724
1725         /*
1726          * @defrag_bytes is a hint value, no spinlock held here,
1727          * if is not zero, it means the file is defragging.
1728          * Force cow if given extent needs to be defragged.
1729          */
1730         if (BTRFS_I(inode)->defrag_bytes &&
1731             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1732                            EXTENT_DEFRAG, 0, NULL))
1733                 return 1;
1734
1735         return 0;
1736 }
1737
1738 /*
1739  * Function to process delayed allocation (create CoW) for ranges which are
1740  * being touched for the first time.
1741  */
1742 int btrfs_run_delalloc_range(struct inode *inode, struct page *locked_page,
1743                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1744                 struct writeback_control *wbc)
1745 {
1746         int ret;
1747         int force_cow = need_force_cow(inode, start, end);
1748
1749         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1750                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1751                                          page_started, 1, nr_written);
1752         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1753                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1754                                          page_started, 0, nr_written);
1755         } else if (!inode_can_compress(inode) ||
1756                    !inode_need_compress(inode, start, end)) {
1757                 ret = cow_file_range(inode, locked_page, start, end,
1758                                       page_started, nr_written, 1);
1759         } else {
1760                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1761                         &BTRFS_I(inode)->runtime_flags);
1762                 ret = cow_file_range_async(inode, wbc, locked_page, start, end,
1763                                            page_started, nr_written);
1764         }
1765         if (ret)
1766                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1767                                               end - start + 1);
1768         return ret;
1769 }
1770
1771 void btrfs_split_delalloc_extent(struct inode *inode,
1772                                  struct extent_state *orig, u64 split)
1773 {
1774         u64 size;
1775
1776         /* not delalloc, ignore it */
1777         if (!(orig->state & EXTENT_DELALLOC))
1778                 return;
1779
1780         size = orig->end - orig->start + 1;
1781         if (size > BTRFS_MAX_EXTENT_SIZE) {
1782                 u32 num_extents;
1783                 u64 new_size;
1784
1785                 /*
1786                  * See the explanation in btrfs_merge_delalloc_extent, the same
1787                  * applies here, just in reverse.
1788                  */
1789                 new_size = orig->end - split + 1;
1790                 num_extents = count_max_extents(new_size);
1791                 new_size = split - orig->start;
1792                 num_extents += count_max_extents(new_size);
1793                 if (count_max_extents(size) >= num_extents)
1794                         return;
1795         }
1796
1797         spin_lock(&BTRFS_I(inode)->lock);
1798         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1799         spin_unlock(&BTRFS_I(inode)->lock);
1800 }
1801
1802 /*
1803  * Handle merged delayed allocation extents so we can keep track of new extents
1804  * that are just merged onto old extents, such as when we are doing sequential
1805  * writes, so we can properly account for the metadata space we'll need.
1806  */
1807 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1808                                  struct extent_state *other)
1809 {
1810         u64 new_size, old_size;
1811         u32 num_extents;
1812
1813         /* not delalloc, ignore it */
1814         if (!(other->state & EXTENT_DELALLOC))
1815                 return;
1816
1817         if (new->start > other->start)
1818                 new_size = new->end - other->start + 1;
1819         else
1820                 new_size = other->end - new->start + 1;
1821
1822         /* we're not bigger than the max, unreserve the space and go */
1823         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1824                 spin_lock(&BTRFS_I(inode)->lock);
1825                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1826                 spin_unlock(&BTRFS_I(inode)->lock);
1827                 return;
1828         }
1829
1830         /*
1831          * We have to add up either side to figure out how many extents were
1832          * accounted for before we merged into one big extent.  If the number of
1833          * extents we accounted for is <= the amount we need for the new range
1834          * then we can return, otherwise drop.  Think of it like this
1835          *
1836          * [ 4k][MAX_SIZE]
1837          *
1838          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1839          * need 2 outstanding extents, on one side we have 1 and the other side
1840          * we have 1 so they are == and we can return.  But in this case
1841          *
1842          * [MAX_SIZE+4k][MAX_SIZE+4k]
1843          *
1844          * Each range on their own accounts for 2 extents, but merged together
1845          * they are only 3 extents worth of accounting, so we need to drop in
1846          * this case.
1847          */
1848         old_size = other->end - other->start + 1;
1849         num_extents = count_max_extents(old_size);
1850         old_size = new->end - new->start + 1;
1851         num_extents += count_max_extents(old_size);
1852         if (count_max_extents(new_size) >= num_extents)
1853                 return;
1854
1855         spin_lock(&BTRFS_I(inode)->lock);
1856         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1857         spin_unlock(&BTRFS_I(inode)->lock);
1858 }
1859
1860 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1861                                       struct inode *inode)
1862 {
1863         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1864
1865         spin_lock(&root->delalloc_lock);
1866         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1867                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1868                               &root->delalloc_inodes);
1869                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1870                         &BTRFS_I(inode)->runtime_flags);
1871                 root->nr_delalloc_inodes++;
1872                 if (root->nr_delalloc_inodes == 1) {
1873                         spin_lock(&fs_info->delalloc_root_lock);
1874                         BUG_ON(!list_empty(&root->delalloc_root));
1875                         list_add_tail(&root->delalloc_root,
1876                                       &fs_info->delalloc_roots);
1877                         spin_unlock(&fs_info->delalloc_root_lock);
1878                 }
1879         }
1880         spin_unlock(&root->delalloc_lock);
1881 }
1882
1883
1884 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1885                                 struct btrfs_inode *inode)
1886 {
1887         struct btrfs_fs_info *fs_info = root->fs_info;
1888
1889         if (!list_empty(&inode->delalloc_inodes)) {
1890                 list_del_init(&inode->delalloc_inodes);
1891                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1892                           &inode->runtime_flags);
1893                 root->nr_delalloc_inodes--;
1894                 if (!root->nr_delalloc_inodes) {
1895                         ASSERT(list_empty(&root->delalloc_inodes));
1896                         spin_lock(&fs_info->delalloc_root_lock);
1897                         BUG_ON(list_empty(&root->delalloc_root));
1898                         list_del_init(&root->delalloc_root);
1899                         spin_unlock(&fs_info->delalloc_root_lock);
1900                 }
1901         }
1902 }
1903
1904 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1905                                      struct btrfs_inode *inode)
1906 {
1907         spin_lock(&root->delalloc_lock);
1908         __btrfs_del_delalloc_inode(root, inode);
1909         spin_unlock(&root->delalloc_lock);
1910 }
1911
1912 /*
1913  * Properly track delayed allocation bytes in the inode and to maintain the
1914  * list of inodes that have pending delalloc work to be done.
1915  */
1916 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1917                                unsigned *bits)
1918 {
1919         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1920
1921         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1922                 WARN_ON(1);
1923         /*
1924          * set_bit and clear bit hooks normally require _irqsave/restore
1925          * but in this case, we are only testing for the DELALLOC
1926          * bit, which is only set or cleared with irqs on
1927          */
1928         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1929                 struct btrfs_root *root = BTRFS_I(inode)->root;
1930                 u64 len = state->end + 1 - state->start;
1931                 u32 num_extents = count_max_extents(len);
1932                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1933
1934                 spin_lock(&BTRFS_I(inode)->lock);
1935                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1936                 spin_unlock(&BTRFS_I(inode)->lock);
1937
1938                 /* For sanity tests */
1939                 if (btrfs_is_testing(fs_info))
1940                         return;
1941
1942                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1943                                          fs_info->delalloc_batch);
1944                 spin_lock(&BTRFS_I(inode)->lock);
1945                 BTRFS_I(inode)->delalloc_bytes += len;
1946                 if (*bits & EXTENT_DEFRAG)
1947                         BTRFS_I(inode)->defrag_bytes += len;
1948                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1949                                          &BTRFS_I(inode)->runtime_flags))
1950                         btrfs_add_delalloc_inodes(root, inode);
1951                 spin_unlock(&BTRFS_I(inode)->lock);
1952         }
1953
1954         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1955             (*bits & EXTENT_DELALLOC_NEW)) {
1956                 spin_lock(&BTRFS_I(inode)->lock);
1957                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1958                         state->start;
1959                 spin_unlock(&BTRFS_I(inode)->lock);
1960         }
1961 }
1962
1963 /*
1964  * Once a range is no longer delalloc this function ensures that proper
1965  * accounting happens.
1966  */
1967 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
1968                                  struct extent_state *state, unsigned *bits)
1969 {
1970         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
1971         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
1972         u64 len = state->end + 1 - state->start;
1973         u32 num_extents = count_max_extents(len);
1974
1975         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1976                 spin_lock(&inode->lock);
1977                 inode->defrag_bytes -= len;
1978                 spin_unlock(&inode->lock);
1979         }
1980
1981         /*
1982          * set_bit and clear bit hooks normally require _irqsave/restore
1983          * but in this case, we are only testing for the DELALLOC
1984          * bit, which is only set or cleared with irqs on
1985          */
1986         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1987                 struct btrfs_root *root = inode->root;
1988                 bool do_list = !btrfs_is_free_space_inode(inode);
1989
1990                 spin_lock(&inode->lock);
1991                 btrfs_mod_outstanding_extents(inode, -num_extents);
1992                 spin_unlock(&inode->lock);
1993
1994                 /*
1995                  * We don't reserve metadata space for space cache inodes so we
1996                  * don't need to call delalloc_release_metadata if there is an
1997                  * error.
1998                  */
1999                 if (*bits & EXTENT_CLEAR_META_RESV &&
2000                     root != fs_info->tree_root)
2001                         btrfs_delalloc_release_metadata(inode, len, false);
2002
2003                 /* For sanity tests. */
2004                 if (btrfs_is_testing(fs_info))
2005                         return;
2006
2007                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
2008                     do_list && !(state->state & EXTENT_NORESERVE) &&
2009                     (*bits & EXTENT_CLEAR_DATA_RESV))
2010                         btrfs_free_reserved_data_space_noquota(
2011                                         &inode->vfs_inode,
2012                                         state->start, len);
2013
2014                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
2015                                          fs_info->delalloc_batch);
2016                 spin_lock(&inode->lock);
2017                 inode->delalloc_bytes -= len;
2018                 if (do_list && inode->delalloc_bytes == 0 &&
2019                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
2020                                         &inode->runtime_flags))
2021                         btrfs_del_delalloc_inode(root, inode);
2022                 spin_unlock(&inode->lock);
2023         }
2024
2025         if ((state->state & EXTENT_DELALLOC_NEW) &&
2026             (*bits & EXTENT_DELALLOC_NEW)) {
2027                 spin_lock(&inode->lock);
2028                 ASSERT(inode->new_delalloc_bytes >= len);
2029                 inode->new_delalloc_bytes -= len;
2030                 spin_unlock(&inode->lock);
2031         }
2032 }
2033
2034 /*
2035  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
2036  * in a chunk's stripe. This function ensures that bios do not span a
2037  * stripe/chunk
2038  *
2039  * @page - The page we are about to add to the bio
2040  * @size - size we want to add to the bio
2041  * @bio - bio we want to ensure is smaller than a stripe
2042  * @bio_flags - flags of the bio
2043  *
2044  * return 1 if page cannot be added to the bio
2045  * return 0 if page can be added to the bio
2046  * return error otherwise
2047  */
2048 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
2049                              unsigned long bio_flags)
2050 {
2051         struct inode *inode = page->mapping->host;
2052         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2053         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
2054         u64 length = 0;
2055         u64 map_length;
2056         int ret;
2057         struct btrfs_io_geometry geom;
2058
2059         if (bio_flags & EXTENT_BIO_COMPRESSED)
2060                 return 0;
2061
2062         length = bio->bi_iter.bi_size;
2063         map_length = length;
2064         ret = btrfs_get_io_geometry(fs_info, btrfs_op(bio), logical, map_length,
2065                                     &geom);
2066         if (ret < 0)
2067                 return ret;
2068
2069         if (geom.len < length + size)
2070                 return 1;
2071         return 0;
2072 }
2073
2074 /*
2075  * in order to insert checksums into the metadata in large chunks,
2076  * we wait until bio submission time.   All the pages in the bio are
2077  * checksummed and sums are attached onto the ordered extent record.
2078  *
2079  * At IO completion time the cums attached on the ordered extent record
2080  * are inserted into the btree
2081  */
2082 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
2083                                     u64 bio_offset)
2084 {
2085         struct inode *inode = private_data;
2086         blk_status_t ret = 0;
2087
2088         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2089         BUG_ON(ret); /* -ENOMEM */
2090         return 0;
2091 }
2092
2093 /*
2094  * extent_io.c submission hook. This does the right thing for csum calculation
2095  * on write, or reading the csums from the tree before a read.
2096  *
2097  * Rules about async/sync submit,
2098  * a) read:                             sync submit
2099  *
2100  * b) write without checksum:           sync submit
2101  *
2102  * c) write with checksum:
2103  *    c-1) if bio is issued by fsync:   sync submit
2104  *         (sync_writers != 0)
2105  *
2106  *    c-2) if root is reloc root:       sync submit
2107  *         (only in case of buffered IO)
2108  *
2109  *    c-3) otherwise:                   async submit
2110  */
2111 static blk_status_t btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
2112                                           int mirror_num,
2113                                           unsigned long bio_flags)
2114
2115 {
2116         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2117         struct btrfs_root *root = BTRFS_I(inode)->root;
2118         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2119         blk_status_t ret = 0;
2120         int skip_sum;
2121         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2122
2123         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2124
2125         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2126                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2127
2128         if (bio_op(bio) != REQ_OP_WRITE) {
2129                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2130                 if (ret)
2131                         goto out;
2132
2133                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2134                         ret = btrfs_submit_compressed_read(inode, bio,
2135                                                            mirror_num,
2136                                                            bio_flags);
2137                         goto out;
2138                 } else if (!skip_sum) {
2139                         ret = btrfs_lookup_bio_sums(inode, bio, (u64)-1, NULL);
2140                         if (ret)
2141                                 goto out;
2142                 }
2143                 goto mapit;
2144         } else if (async && !skip_sum) {
2145                 /* csum items have already been cloned */
2146                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2147                         goto mapit;
2148                 /* we're doing a write, do the async checksumming */
2149                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2150                                           0, inode, btrfs_submit_bio_start);
2151                 goto out;
2152         } else if (!skip_sum) {
2153                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2154                 if (ret)
2155                         goto out;
2156         }
2157
2158 mapit:
2159         ret = btrfs_map_bio(fs_info, bio, mirror_num);
2160
2161 out:
2162         if (ret) {
2163                 bio->bi_status = ret;
2164                 bio_endio(bio);
2165         }
2166         return ret;
2167 }
2168
2169 /*
2170  * given a list of ordered sums record them in the inode.  This happens
2171  * at IO completion time based on sums calculated at bio submission time.
2172  */
2173 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2174                              struct inode *inode, struct list_head *list)
2175 {
2176         struct btrfs_ordered_sum *sum;
2177         int ret;
2178
2179         list_for_each_entry(sum, list, list) {
2180                 trans->adding_csums = true;
2181                 ret = btrfs_csum_file_blocks(trans,
2182                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2183                 trans->adding_csums = false;
2184                 if (ret)
2185                         return ret;
2186         }
2187         return 0;
2188 }
2189
2190 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2191                               unsigned int extra_bits,
2192                               struct extent_state **cached_state)
2193 {
2194         WARN_ON(PAGE_ALIGNED(end));
2195         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2196                                    extra_bits, cached_state);
2197 }
2198
2199 /* see btrfs_writepage_start_hook for details on why this is required */
2200 struct btrfs_writepage_fixup {
2201         struct page *page;
2202         struct inode *inode;
2203         struct btrfs_work work;
2204 };
2205
2206 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2207 {
2208         struct btrfs_writepage_fixup *fixup;
2209         struct btrfs_ordered_extent *ordered;
2210         struct extent_state *cached_state = NULL;
2211         struct extent_changeset *data_reserved = NULL;
2212         struct page *page;
2213         struct inode *inode;
2214         u64 page_start;
2215         u64 page_end;
2216         int ret = 0;
2217         bool free_delalloc_space = true;
2218
2219         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2220         page = fixup->page;
2221         inode = fixup->inode;
2222         page_start = page_offset(page);
2223         page_end = page_offset(page) + PAGE_SIZE - 1;
2224
2225         /*
2226          * This is similar to page_mkwrite, we need to reserve the space before
2227          * we take the page lock.
2228          */
2229         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2230                                            PAGE_SIZE);
2231 again:
2232         lock_page(page);
2233
2234         /*
2235          * Before we queued this fixup, we took a reference on the page.
2236          * page->mapping may go NULL, but it shouldn't be moved to a different
2237          * address space.
2238          */
2239         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2240                 /*
2241                  * Unfortunately this is a little tricky, either
2242                  *
2243                  * 1) We got here and our page had already been dealt with and
2244                  *    we reserved our space, thus ret == 0, so we need to just
2245                  *    drop our space reservation and bail.  This can happen the
2246                  *    first time we come into the fixup worker, or could happen
2247                  *    while waiting for the ordered extent.
2248                  * 2) Our page was already dealt with, but we happened to get an
2249                  *    ENOSPC above from the btrfs_delalloc_reserve_space.  In
2250                  *    this case we obviously don't have anything to release, but
2251                  *    because the page was already dealt with we don't want to
2252                  *    mark the page with an error, so make sure we're resetting
2253                  *    ret to 0.  This is why we have this check _before_ the ret
2254                  *    check, because we do not want to have a surprise ENOSPC
2255                  *    when the page was already properly dealt with.
2256                  */
2257                 if (!ret) {
2258                         btrfs_delalloc_release_extents(BTRFS_I(inode),
2259                                                        PAGE_SIZE);
2260                         btrfs_delalloc_release_space(inode, data_reserved,
2261                                                      page_start, PAGE_SIZE,
2262                                                      true);
2263                 }
2264                 ret = 0;
2265                 goto out_page;
2266         }
2267
2268         /*
2269          * We can't mess with the page state unless it is locked, so now that
2270          * it is locked bail if we failed to make our space reservation.
2271          */
2272         if (ret)
2273                 goto out_page;
2274
2275         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2276                          &cached_state);
2277
2278         /* already ordered? We're done */
2279         if (PagePrivate2(page))
2280                 goto out_reserved;
2281
2282         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2283                                         PAGE_SIZE);
2284         if (ordered) {
2285                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2286                                      page_end, &cached_state);
2287                 unlock_page(page);
2288                 btrfs_start_ordered_extent(inode, ordered, 1);
2289                 btrfs_put_ordered_extent(ordered);
2290                 goto again;
2291         }
2292
2293         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2294                                         &cached_state);
2295         if (ret)
2296                 goto out_reserved;
2297
2298         /*
2299          * Everything went as planned, we're now the owner of a dirty page with
2300          * delayed allocation bits set and space reserved for our COW
2301          * destination.
2302          *
2303          * The page was dirty when we started, nothing should have cleaned it.
2304          */
2305         BUG_ON(!PageDirty(page));
2306         free_delalloc_space = false;
2307 out_reserved:
2308         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2309         if (free_delalloc_space)
2310                 btrfs_delalloc_release_space(inode, data_reserved, page_start,
2311                                              PAGE_SIZE, true);
2312         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2313                              &cached_state);
2314 out_page:
2315         if (ret) {
2316                 /*
2317                  * We hit ENOSPC or other errors.  Update the mapping and page
2318                  * to reflect the errors and clean the page.
2319                  */
2320                 mapping_set_error(page->mapping, ret);
2321                 end_extent_writepage(page, ret, page_start, page_end);
2322                 clear_page_dirty_for_io(page);
2323                 SetPageError(page);
2324         }
2325         ClearPageChecked(page);
2326         unlock_page(page);
2327         put_page(page);
2328         kfree(fixup);
2329         extent_changeset_free(data_reserved);
2330         /*
2331          * As a precaution, do a delayed iput in case it would be the last iput
2332          * that could need flushing space. Recursing back to fixup worker would
2333          * deadlock.
2334          */
2335         btrfs_add_delayed_iput(inode);
2336 }
2337
2338 /*
2339  * There are a few paths in the higher layers of the kernel that directly
2340  * set the page dirty bit without asking the filesystem if it is a
2341  * good idea.  This causes problems because we want to make sure COW
2342  * properly happens and the data=ordered rules are followed.
2343  *
2344  * In our case any range that doesn't have the ORDERED bit set
2345  * hasn't been properly setup for IO.  We kick off an async process
2346  * to fix it up.  The async helper will wait for ordered extents, set
2347  * the delalloc bit and make it safe to write the page.
2348  */
2349 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2350 {
2351         struct inode *inode = page->mapping->host;
2352         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2353         struct btrfs_writepage_fixup *fixup;
2354
2355         /* this page is properly in the ordered list */
2356         if (TestClearPagePrivate2(page))
2357                 return 0;
2358
2359         /*
2360          * PageChecked is set below when we create a fixup worker for this page,
2361          * don't try to create another one if we're already PageChecked()
2362          *
2363          * The extent_io writepage code will redirty the page if we send back
2364          * EAGAIN.
2365          */
2366         if (PageChecked(page))
2367                 return -EAGAIN;
2368
2369         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2370         if (!fixup)
2371                 return -EAGAIN;
2372
2373         /*
2374          * We are already holding a reference to this inode from
2375          * write_cache_pages.  We need to hold it because the space reservation
2376          * takes place outside of the page lock, and we can't trust
2377          * page->mapping outside of the page lock.
2378          */
2379         ihold(inode);
2380         SetPageChecked(page);
2381         get_page(page);
2382         btrfs_init_work(&fixup->work, btrfs_writepage_fixup_worker, NULL, NULL);
2383         fixup->page = page;
2384         fixup->inode = inode;
2385         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2386
2387         return -EAGAIN;
2388 }
2389
2390 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2391                                        struct inode *inode, u64 file_pos,
2392                                        u64 disk_bytenr, u64 disk_num_bytes,
2393                                        u64 num_bytes, u64 ram_bytes,
2394                                        u8 compression, u8 encryption,
2395                                        u16 other_encoding, int extent_type)
2396 {
2397         struct btrfs_root *root = BTRFS_I(inode)->root;
2398         struct btrfs_file_extent_item *fi;
2399         struct btrfs_path *path;
2400         struct extent_buffer *leaf;
2401         struct btrfs_key ins;
2402         u64 qg_released;
2403         int extent_inserted = 0;
2404         int ret;
2405
2406         path = btrfs_alloc_path();
2407         if (!path)
2408                 return -ENOMEM;
2409
2410         /*
2411          * we may be replacing one extent in the tree with another.
2412          * The new extent is pinned in the extent map, and we don't want
2413          * to drop it from the cache until it is completely in the btree.
2414          *
2415          * So, tell btrfs_drop_extents to leave this extent in the cache.
2416          * the caller is expected to unpin it and allow it to be merged
2417          * with the others.
2418          */
2419         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2420                                    file_pos + num_bytes, NULL, 0,
2421                                    1, sizeof(*fi), &extent_inserted);
2422         if (ret)
2423                 goto out;
2424
2425         if (!extent_inserted) {
2426                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2427                 ins.offset = file_pos;
2428                 ins.type = BTRFS_EXTENT_DATA_KEY;
2429
2430                 path->leave_spinning = 1;
2431                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2432                                               sizeof(*fi));
2433                 if (ret)
2434                         goto out;
2435         }
2436         leaf = path->nodes[0];
2437         fi = btrfs_item_ptr(leaf, path->slots[0],
2438                             struct btrfs_file_extent_item);
2439         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2440         btrfs_set_file_extent_type(leaf, fi, extent_type);
2441         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2442         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2443         btrfs_set_file_extent_offset(leaf, fi, 0);
2444         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2445         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2446         btrfs_set_file_extent_compression(leaf, fi, compression);
2447         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2448         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2449
2450         btrfs_mark_buffer_dirty(leaf);
2451         btrfs_release_path(path);
2452
2453         inode_add_bytes(inode, num_bytes);
2454
2455         ins.objectid = disk_bytenr;
2456         ins.offset = disk_num_bytes;
2457         ins.type = BTRFS_EXTENT_ITEM_KEY;
2458
2459         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode), file_pos,
2460                                                 ram_bytes);
2461         if (ret)
2462                 goto out;
2463
2464         /*
2465          * Release the reserved range from inode dirty range map, as it is
2466          * already moved into delayed_ref_head
2467          */
2468         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2469         if (ret < 0)
2470                 goto out;
2471         qg_released = ret;
2472         ret = btrfs_alloc_reserved_file_extent(trans, root,
2473                                                btrfs_ino(BTRFS_I(inode)),
2474                                                file_pos, qg_released, &ins);
2475 out:
2476         btrfs_free_path(path);
2477
2478         return ret;
2479 }
2480
2481 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2482                                          u64 start, u64 len)
2483 {
2484         struct btrfs_block_group *cache;
2485
2486         cache = btrfs_lookup_block_group(fs_info, start);
2487         ASSERT(cache);
2488
2489         spin_lock(&cache->lock);
2490         cache->delalloc_bytes -= len;
2491         spin_unlock(&cache->lock);
2492
2493         btrfs_put_block_group(cache);
2494 }
2495
2496 /* as ordered data IO finishes, this gets called so we can finish
2497  * an ordered extent if the range of bytes in the file it covers are
2498  * fully written.
2499  */
2500 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2501 {
2502         struct inode *inode = ordered_extent->inode;
2503         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2504         struct btrfs_root *root = BTRFS_I(inode)->root;
2505         struct btrfs_trans_handle *trans = NULL;
2506         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2507         struct extent_state *cached_state = NULL;
2508         u64 start, end;
2509         int compress_type = 0;
2510         int ret = 0;
2511         u64 logical_len = ordered_extent->num_bytes;
2512         bool freespace_inode;
2513         bool truncated = false;
2514         bool range_locked = false;
2515         bool clear_new_delalloc_bytes = false;
2516         bool clear_reserved_extent = true;
2517         unsigned int clear_bits;
2518
2519         start = ordered_extent->file_offset;
2520         end = start + ordered_extent->num_bytes - 1;
2521
2522         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2523             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2524             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2525                 clear_new_delalloc_bytes = true;
2526
2527         freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
2528
2529         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2530                 ret = -EIO;
2531                 goto out;
2532         }
2533
2534         btrfs_free_io_failure_record(BTRFS_I(inode), start, end);
2535
2536         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2537                 truncated = true;
2538                 logical_len = ordered_extent->truncated_len;
2539                 /* Truncated the entire extent, don't bother adding */
2540                 if (!logical_len)
2541                         goto out;
2542         }
2543
2544         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2545                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2546
2547                 /*
2548                  * For mwrite(mmap + memset to write) case, we still reserve
2549                  * space for NOCOW range.
2550                  * As NOCOW won't cause a new delayed ref, just free the space
2551                  */
2552                 btrfs_qgroup_free_data(inode, NULL, start,
2553                                        ordered_extent->num_bytes);
2554                 btrfs_inode_safe_disk_i_size_write(inode, 0);
2555                 if (freespace_inode)
2556                         trans = btrfs_join_transaction_spacecache(root);
2557                 else
2558                         trans = btrfs_join_transaction(root);
2559                 if (IS_ERR(trans)) {
2560                         ret = PTR_ERR(trans);
2561                         trans = NULL;
2562                         goto out;
2563                 }
2564                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2565                 ret = btrfs_update_inode_fallback(trans, root, inode);
2566                 if (ret) /* -ENOMEM or corruption */
2567                         btrfs_abort_transaction(trans, ret);
2568                 goto out;
2569         }
2570
2571         range_locked = true;
2572         lock_extent_bits(io_tree, start, end, &cached_state);
2573
2574         if (freespace_inode)
2575                 trans = btrfs_join_transaction_spacecache(root);
2576         else
2577                 trans = btrfs_join_transaction(root);
2578         if (IS_ERR(trans)) {
2579                 ret = PTR_ERR(trans);
2580                 trans = NULL;
2581                 goto out;
2582         }
2583
2584         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2585
2586         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2587                 compress_type = ordered_extent->compress_type;
2588         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2589                 BUG_ON(compress_type);
2590                 btrfs_qgroup_free_data(inode, NULL, start,
2591                                        ordered_extent->num_bytes);
2592                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
2593                                                 ordered_extent->file_offset,
2594                                                 ordered_extent->file_offset +
2595                                                 logical_len);
2596         } else {
2597                 BUG_ON(root == fs_info->tree_root);
2598                 ret = insert_reserved_file_extent(trans, inode, start,
2599                                                 ordered_extent->disk_bytenr,
2600                                                 ordered_extent->disk_num_bytes,
2601                                                 logical_len, logical_len,
2602                                                 compress_type, 0, 0,
2603                                                 BTRFS_FILE_EXTENT_REG);
2604                 if (!ret) {
2605                         clear_reserved_extent = false;
2606                         btrfs_release_delalloc_bytes(fs_info,
2607                                                 ordered_extent->disk_bytenr,
2608                                                 ordered_extent->disk_num_bytes);
2609                 }
2610         }
2611         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2612                            ordered_extent->file_offset,
2613                            ordered_extent->num_bytes, trans->transid);
2614         if (ret < 0) {
2615                 btrfs_abort_transaction(trans, ret);
2616                 goto out;
2617         }
2618
2619         ret = add_pending_csums(trans, inode, &ordered_extent->list);
2620         if (ret) {
2621                 btrfs_abort_transaction(trans, ret);
2622                 goto out;
2623         }
2624
2625         btrfs_inode_safe_disk_i_size_write(inode, 0);
2626         ret = btrfs_update_inode_fallback(trans, root, inode);
2627         if (ret) { /* -ENOMEM or corruption */
2628                 btrfs_abort_transaction(trans, ret);
2629                 goto out;
2630         }
2631         ret = 0;
2632 out:
2633         clear_bits = EXTENT_DEFRAG;
2634         if (range_locked)
2635                 clear_bits |= EXTENT_LOCKED;
2636         if (clear_new_delalloc_bytes)
2637                 clear_bits |= EXTENT_DELALLOC_NEW;
2638         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits,
2639                          (clear_bits & EXTENT_LOCKED) ? 1 : 0, 0,
2640                          &cached_state);
2641
2642         if (trans)
2643                 btrfs_end_transaction(trans);
2644
2645         if (ret || truncated) {
2646                 u64 unwritten_start = start;
2647
2648                 if (truncated)
2649                         unwritten_start += logical_len;
2650                 clear_extent_uptodate(io_tree, unwritten_start, end, NULL);
2651
2652                 /* Drop the cache for the part of the extent we didn't write. */
2653                 btrfs_drop_extent_cache(BTRFS_I(inode), unwritten_start, end, 0);
2654
2655                 /*
2656                  * If the ordered extent had an IOERR or something else went
2657                  * wrong we need to return the space for this ordered extent
2658                  * back to the allocator.  We only free the extent in the
2659                  * truncated case if we didn't write out the extent at all.
2660                  *
2661                  * If we made it past insert_reserved_file_extent before we
2662                  * errored out then we don't need to do this as the accounting
2663                  * has already been done.
2664                  */
2665                 if ((ret || !logical_len) &&
2666                     clear_reserved_extent &&
2667                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2668                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2669                         /*
2670                          * Discard the range before returning it back to the
2671                          * free space pool
2672                          */
2673                         if (ret && btrfs_test_opt(fs_info, DISCARD_SYNC))
2674                                 btrfs_discard_extent(fs_info,
2675                                                 ordered_extent->disk_bytenr,
2676                                                 ordered_extent->disk_num_bytes,
2677                                                 NULL);
2678                         btrfs_free_reserved_extent(fs_info,
2679                                         ordered_extent->disk_bytenr,
2680                                         ordered_extent->disk_num_bytes, 1);
2681                 }
2682         }
2683
2684         /*
2685          * This needs to be done to make sure anybody waiting knows we are done
2686          * updating everything for this ordered extent.
2687          */
2688         btrfs_remove_ordered_extent(inode, ordered_extent);
2689
2690         /* once for us */
2691         btrfs_put_ordered_extent(ordered_extent);
2692         /* once for the tree */
2693         btrfs_put_ordered_extent(ordered_extent);
2694
2695         return ret;
2696 }
2697
2698 static void finish_ordered_fn(struct btrfs_work *work)
2699 {
2700         struct btrfs_ordered_extent *ordered_extent;
2701         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
2702         btrfs_finish_ordered_io(ordered_extent);
2703 }
2704
2705 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
2706                                           u64 end, int uptodate)
2707 {
2708         struct inode *inode = page->mapping->host;
2709         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2710         struct btrfs_ordered_extent *ordered_extent = NULL;
2711         struct btrfs_workqueue *wq;
2712
2713         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
2714
2715         ClearPagePrivate2(page);
2716         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
2717                                             end - start + 1, uptodate))
2718                 return;
2719
2720         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2721                 wq = fs_info->endio_freespace_worker;
2722         else
2723                 wq = fs_info->endio_write_workers;
2724
2725         btrfs_init_work(&ordered_extent->work, finish_ordered_fn, NULL, NULL);
2726         btrfs_queue_work(wq, &ordered_extent->work);
2727 }
2728
2729 static int __readpage_endio_check(struct inode *inode,
2730                                   struct btrfs_io_bio *io_bio,
2731                                   int icsum, struct page *page,
2732                                   int pgoff, u64 start, size_t len)
2733 {
2734         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2735         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2736         char *kaddr;
2737         u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2738         u8 *csum_expected;
2739         u8 csum[BTRFS_CSUM_SIZE];
2740
2741         csum_expected = ((u8 *)io_bio->csum) + icsum * csum_size;
2742
2743         kaddr = kmap_atomic(page);
2744         shash->tfm = fs_info->csum_shash;
2745
2746         crypto_shash_init(shash);
2747         crypto_shash_update(shash, kaddr + pgoff, len);
2748         crypto_shash_final(shash, csum);
2749
2750         if (memcmp(csum, csum_expected, csum_size))
2751                 goto zeroit;
2752
2753         kunmap_atomic(kaddr);
2754         return 0;
2755 zeroit:
2756         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
2757                                     io_bio->mirror_num);
2758         memset(kaddr + pgoff, 1, len);
2759         flush_dcache_page(page);
2760         kunmap_atomic(kaddr);
2761         return -EIO;
2762 }
2763
2764 /*
2765  * when reads are done, we need to check csums to verify the data is correct
2766  * if there's a match, we allow the bio to finish.  If not, the code in
2767  * extent_io.c will try to find good copies for us.
2768  */
2769 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
2770                                       u64 phy_offset, struct page *page,
2771                                       u64 start, u64 end, int mirror)
2772 {
2773         size_t offset = start - page_offset(page);
2774         struct inode *inode = page->mapping->host;
2775         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2776         struct btrfs_root *root = BTRFS_I(inode)->root;
2777
2778         if (PageChecked(page)) {
2779                 ClearPageChecked(page);
2780                 return 0;
2781         }
2782
2783         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
2784                 return 0;
2785
2786         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
2787             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
2788                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
2789                 return 0;
2790         }
2791
2792         phy_offset >>= inode->i_sb->s_blocksize_bits;
2793         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
2794                                       start, (size_t)(end - start + 1));
2795 }
2796
2797 /*
2798  * btrfs_add_delayed_iput - perform a delayed iput on @inode
2799  *
2800  * @inode: The inode we want to perform iput on
2801  *
2802  * This function uses the generic vfs_inode::i_count to track whether we should
2803  * just decrement it (in case it's > 1) or if this is the last iput then link
2804  * the inode to the delayed iput machinery. Delayed iputs are processed at
2805  * transaction commit time/superblock commit/cleaner kthread.
2806  */
2807 void btrfs_add_delayed_iput(struct inode *inode)
2808 {
2809         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2810         struct btrfs_inode *binode = BTRFS_I(inode);
2811
2812         if (atomic_add_unless(&inode->i_count, -1, 1))
2813                 return;
2814
2815         atomic_inc(&fs_info->nr_delayed_iputs);
2816         spin_lock(&fs_info->delayed_iput_lock);
2817         ASSERT(list_empty(&binode->delayed_iput));
2818         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
2819         spin_unlock(&fs_info->delayed_iput_lock);
2820         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
2821                 wake_up_process(fs_info->cleaner_kthread);
2822 }
2823
2824 static void run_delayed_iput_locked(struct btrfs_fs_info *fs_info,
2825                                     struct btrfs_inode *inode)
2826 {
2827         list_del_init(&inode->delayed_iput);
2828         spin_unlock(&fs_info->delayed_iput_lock);
2829         iput(&inode->vfs_inode);
2830         if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
2831                 wake_up(&fs_info->delayed_iputs_wait);
2832         spin_lock(&fs_info->delayed_iput_lock);
2833 }
2834
2835 static void btrfs_run_delayed_iput(struct btrfs_fs_info *fs_info,
2836                                    struct btrfs_inode *inode)
2837 {
2838         if (!list_empty(&inode->delayed_iput)) {
2839                 spin_lock(&fs_info->delayed_iput_lock);
2840                 if (!list_empty(&inode->delayed_iput))
2841                         run_delayed_iput_locked(fs_info, inode);
2842                 spin_unlock(&fs_info->delayed_iput_lock);
2843         }
2844 }
2845
2846 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
2847 {
2848
2849         spin_lock(&fs_info->delayed_iput_lock);
2850         while (!list_empty(&fs_info->delayed_iputs)) {
2851                 struct btrfs_inode *inode;
2852
2853                 inode = list_first_entry(&fs_info->delayed_iputs,
2854                                 struct btrfs_inode, delayed_iput);
2855                 run_delayed_iput_locked(fs_info, inode);
2856         }
2857         spin_unlock(&fs_info->delayed_iput_lock);
2858 }
2859
2860 /**
2861  * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
2862  * @fs_info - the fs_info for this fs
2863  * @return - EINTR if we were killed, 0 if nothing's pending
2864  *
2865  * This will wait on any delayed iputs that are currently running with KILLABLE
2866  * set.  Once they are all done running we will return, unless we are killed in
2867  * which case we return EINTR. This helps in user operations like fallocate etc
2868  * that might get blocked on the iputs.
2869  */
2870 int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
2871 {
2872         int ret = wait_event_killable(fs_info->delayed_iputs_wait,
2873                         atomic_read(&fs_info->nr_delayed_iputs) == 0);
2874         if (ret)
2875                 return -EINTR;
2876         return 0;
2877 }
2878
2879 /*
2880  * This creates an orphan entry for the given inode in case something goes wrong
2881  * in the middle of an unlink.
2882  */
2883 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
2884                      struct btrfs_inode *inode)
2885 {
2886         int ret;
2887
2888         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
2889         if (ret && ret != -EEXIST) {
2890                 btrfs_abort_transaction(trans, ret);
2891                 return ret;
2892         }
2893
2894         return 0;
2895 }
2896
2897 /*
2898  * We have done the delete so we can go ahead and remove the orphan item for
2899  * this particular inode.
2900  */
2901 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
2902                             struct btrfs_inode *inode)
2903 {
2904         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
2905 }
2906
2907 /*
2908  * this cleans up any orphans that may be left on the list from the last use
2909  * of this root.
2910  */
2911 int btrfs_orphan_cleanup(struct btrfs_root *root)
2912 {
2913         struct btrfs_fs_info *fs_info = root->fs_info;
2914         struct btrfs_path *path;
2915         struct extent_buffer *leaf;
2916         struct btrfs_key key, found_key;
2917         struct btrfs_trans_handle *trans;
2918         struct inode *inode;
2919         u64 last_objectid = 0;
2920         int ret = 0, nr_unlink = 0;
2921
2922         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
2923                 return 0;
2924
2925         path = btrfs_alloc_path();
2926         if (!path) {
2927                 ret = -ENOMEM;
2928                 goto out;
2929         }
2930         path->reada = READA_BACK;
2931
2932         key.objectid = BTRFS_ORPHAN_OBJECTID;
2933         key.type = BTRFS_ORPHAN_ITEM_KEY;
2934         key.offset = (u64)-1;
2935
2936         while (1) {
2937                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2938                 if (ret < 0)
2939                         goto out;
2940
2941                 /*
2942                  * if ret == 0 means we found what we were searching for, which
2943                  * is weird, but possible, so only screw with path if we didn't
2944                  * find the key and see if we have stuff that matches
2945                  */
2946                 if (ret > 0) {
2947                         ret = 0;
2948                         if (path->slots[0] == 0)
2949                                 break;
2950                         path->slots[0]--;
2951                 }
2952
2953                 /* pull out the item */
2954                 leaf = path->nodes[0];
2955                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2956
2957                 /* make sure the item matches what we want */
2958                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
2959                         break;
2960                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
2961                         break;
2962
2963                 /* release the path since we're done with it */
2964                 btrfs_release_path(path);
2965
2966                 /*
2967                  * this is where we are basically btrfs_lookup, without the
2968                  * crossing root thing.  we store the inode number in the
2969                  * offset of the orphan item.
2970                  */
2971
2972                 if (found_key.offset == last_objectid) {
2973                         btrfs_err(fs_info,
2974                                   "Error removing orphan entry, stopping orphan cleanup");
2975                         ret = -EINVAL;
2976                         goto out;
2977                 }
2978
2979                 last_objectid = found_key.offset;
2980
2981                 found_key.objectid = found_key.offset;
2982                 found_key.type = BTRFS_INODE_ITEM_KEY;
2983                 found_key.offset = 0;
2984                 inode = btrfs_iget(fs_info->sb, &found_key, root);
2985                 ret = PTR_ERR_OR_ZERO(inode);
2986                 if (ret && ret != -ENOENT)
2987                         goto out;
2988
2989                 if (ret == -ENOENT && root == fs_info->tree_root) {
2990                         struct btrfs_root *dead_root;
2991                         struct btrfs_fs_info *fs_info = root->fs_info;
2992                         int is_dead_root = 0;
2993
2994                         /*
2995                          * this is an orphan in the tree root. Currently these
2996                          * could come from 2 sources:
2997                          *  a) a snapshot deletion in progress
2998                          *  b) a free space cache inode
2999                          * We need to distinguish those two, as the snapshot
3000                          * orphan must not get deleted.
3001                          * find_dead_roots already ran before us, so if this
3002                          * is a snapshot deletion, we should find the root
3003                          * in the dead_roots list
3004                          */
3005                         spin_lock(&fs_info->trans_lock);
3006                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3007                                             root_list) {
3008                                 if (dead_root->root_key.objectid ==
3009                                     found_key.objectid) {
3010                                         is_dead_root = 1;
3011                                         break;
3012                                 }
3013                         }
3014                         spin_unlock(&fs_info->trans_lock);
3015                         if (is_dead_root) {
3016                                 /* prevent this orphan from being found again */
3017                                 key.offset = found_key.objectid - 1;
3018                                 continue;
3019                         }
3020
3021                 }
3022
3023                 /*
3024                  * If we have an inode with links, there are a couple of
3025                  * possibilities. Old kernels (before v3.12) used to create an
3026                  * orphan item for truncate indicating that there were possibly
3027                  * extent items past i_size that needed to be deleted. In v3.12,
3028                  * truncate was changed to update i_size in sync with the extent
3029                  * items, but the (useless) orphan item was still created. Since
3030                  * v4.18, we don't create the orphan item for truncate at all.
3031                  *
3032                  * So, this item could mean that we need to do a truncate, but
3033                  * only if this filesystem was last used on a pre-v3.12 kernel
3034                  * and was not cleanly unmounted. The odds of that are quite
3035                  * slim, and it's a pain to do the truncate now, so just delete
3036                  * the orphan item.
3037                  *
3038                  * It's also possible that this orphan item was supposed to be
3039                  * deleted but wasn't. The inode number may have been reused,
3040                  * but either way, we can delete the orphan item.
3041                  */
3042                 if (ret == -ENOENT || inode->i_nlink) {
3043                         if (!ret)
3044                                 iput(inode);
3045                         trans = btrfs_start_transaction(root, 1);
3046                         if (IS_ERR(trans)) {
3047                                 ret = PTR_ERR(trans);
3048                                 goto out;
3049                         }
3050                         btrfs_debug(fs_info, "auto deleting %Lu",
3051                                     found_key.objectid);
3052                         ret = btrfs_del_orphan_item(trans, root,
3053                                                     found_key.objectid);
3054                         btrfs_end_transaction(trans);
3055                         if (ret)
3056                                 goto out;
3057                         continue;
3058                 }
3059
3060                 nr_unlink++;
3061
3062                 /* this will do delete_inode and everything for us */
3063                 iput(inode);
3064         }
3065         /* release the path since we're done with it */
3066         btrfs_release_path(path);
3067
3068         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3069
3070         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3071                 trans = btrfs_join_transaction(root);
3072                 if (!IS_ERR(trans))
3073                         btrfs_end_transaction(trans);
3074         }
3075
3076         if (nr_unlink)
3077                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3078
3079 out:
3080         if (ret)
3081                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3082         btrfs_free_path(path);
3083         return ret;
3084 }
3085
3086 /*
3087  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3088  * don't find any xattrs, we know there can't be any acls.
3089  *
3090  * slot is the slot the inode is in, objectid is the objectid of the inode
3091  */
3092 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3093                                           int slot, u64 objectid,
3094                                           int *first_xattr_slot)
3095 {
3096         u32 nritems = btrfs_header_nritems(leaf);
3097         struct btrfs_key found_key;
3098         static u64 xattr_access = 0;
3099         static u64 xattr_default = 0;
3100         int scanned = 0;
3101
3102         if (!xattr_access) {
3103                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3104                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3105                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3106                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3107         }
3108
3109         slot++;
3110         *first_xattr_slot = -1;
3111         while (slot < nritems) {
3112                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3113
3114                 /* we found a different objectid, there must not be acls */
3115                 if (found_key.objectid != objectid)
3116                         return 0;
3117
3118                 /* we found an xattr, assume we've got an acl */
3119                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3120                         if (*first_xattr_slot == -1)
3121                                 *first_xattr_slot = slot;
3122                         if (found_key.offset == xattr_access ||
3123                             found_key.offset == xattr_default)
3124                                 return 1;
3125                 }
3126
3127                 /*
3128                  * we found a key greater than an xattr key, there can't
3129                  * be any acls later on
3130                  */
3131                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3132                         return 0;
3133
3134                 slot++;
3135                 scanned++;
3136
3137                 /*
3138                  * it goes inode, inode backrefs, xattrs, extents,
3139                  * so if there are a ton of hard links to an inode there can
3140                  * be a lot of backrefs.  Don't waste time searching too hard,
3141                  * this is just an optimization
3142                  */
3143                 if (scanned >= 8)
3144                         break;
3145         }
3146         /* we hit the end of the leaf before we found an xattr or
3147          * something larger than an xattr.  We have to assume the inode
3148          * has acls
3149          */
3150         if (*first_xattr_slot == -1)
3151                 *first_xattr_slot = slot;
3152         return 1;
3153 }
3154
3155 /*
3156  * read an inode from the btree into the in-memory inode
3157  */
3158 static int btrfs_read_locked_inode(struct inode *inode,
3159                                    struct btrfs_path *in_path)
3160 {
3161         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3162         struct btrfs_path *path = in_path;
3163         struct extent_buffer *leaf;
3164         struct btrfs_inode_item *inode_item;
3165         struct btrfs_root *root = BTRFS_I(inode)->root;
3166         struct btrfs_key location;
3167         unsigned long ptr;
3168         int maybe_acls;
3169         u32 rdev;
3170         int ret;
3171         bool filled = false;
3172         int first_xattr_slot;
3173
3174         ret = btrfs_fill_inode(inode, &rdev);
3175         if (!ret)
3176                 filled = true;
3177
3178         if (!path) {
3179                 path = btrfs_alloc_path();
3180                 if (!path)
3181                         return -ENOMEM;
3182         }
3183
3184         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3185
3186         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3187         if (ret) {
3188                 if (path != in_path)
3189                         btrfs_free_path(path);
3190                 return ret;
3191         }
3192
3193         leaf = path->nodes[0];
3194
3195         if (filled)
3196                 goto cache_index;
3197
3198         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3199                                     struct btrfs_inode_item);
3200         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3201         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3202         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3203         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3204         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3205         btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
3206                         round_up(i_size_read(inode), fs_info->sectorsize));
3207
3208         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3209         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3210
3211         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3212         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3213
3214         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3215         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3216
3217         BTRFS_I(inode)->i_otime.tv_sec =
3218                 btrfs_timespec_sec(leaf, &inode_item->otime);
3219         BTRFS_I(inode)->i_otime.tv_nsec =
3220                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3221
3222         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3223         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3224         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3225
3226         inode_set_iversion_queried(inode,
3227                                    btrfs_inode_sequence(leaf, inode_item));
3228         inode->i_generation = BTRFS_I(inode)->generation;
3229         inode->i_rdev = 0;
3230         rdev = btrfs_inode_rdev(leaf, inode_item);
3231
3232         BTRFS_I(inode)->index_cnt = (u64)-1;
3233         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3234
3235 cache_index:
3236         /*
3237          * If we were modified in the current generation and evicted from memory
3238          * and then re-read we need to do a full sync since we don't have any
3239          * idea about which extents were modified before we were evicted from
3240          * cache.
3241          *
3242          * This is required for both inode re-read from disk and delayed inode
3243          * in delayed_nodes_tree.
3244          */
3245         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3246                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3247                         &BTRFS_I(inode)->runtime_flags);
3248
3249         /*
3250          * We don't persist the id of the transaction where an unlink operation
3251          * against the inode was last made. So here we assume the inode might
3252          * have been evicted, and therefore the exact value of last_unlink_trans
3253          * lost, and set it to last_trans to avoid metadata inconsistencies
3254          * between the inode and its parent if the inode is fsync'ed and the log
3255          * replayed. For example, in the scenario:
3256          *
3257          * touch mydir/foo
3258          * ln mydir/foo mydir/bar
3259          * sync
3260          * unlink mydir/bar
3261          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3262          * xfs_io -c fsync mydir/foo
3263          * <power failure>
3264          * mount fs, triggers fsync log replay
3265          *
3266          * We must make sure that when we fsync our inode foo we also log its
3267          * parent inode, otherwise after log replay the parent still has the
3268          * dentry with the "bar" name but our inode foo has a link count of 1
3269          * and doesn't have an inode ref with the name "bar" anymore.
3270          *
3271          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3272          * but it guarantees correctness at the expense of occasional full
3273          * transaction commits on fsync if our inode is a directory, or if our
3274          * inode is not a directory, logging its parent unnecessarily.
3275          */
3276         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3277
3278         path->slots[0]++;
3279         if (inode->i_nlink != 1 ||
3280             path->slots[0] >= btrfs_header_nritems(leaf))
3281                 goto cache_acl;
3282
3283         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3284         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3285                 goto cache_acl;
3286
3287         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3288         if (location.type == BTRFS_INODE_REF_KEY) {
3289                 struct btrfs_inode_ref *ref;
3290
3291                 ref = (struct btrfs_inode_ref *)ptr;
3292                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3293         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3294                 struct btrfs_inode_extref *extref;
3295
3296                 extref = (struct btrfs_inode_extref *)ptr;
3297                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3298                                                                      extref);
3299         }
3300 cache_acl:
3301         /*
3302          * try to precache a NULL acl entry for files that don't have
3303          * any xattrs or acls
3304          */
3305         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3306                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3307         if (first_xattr_slot != -1) {
3308                 path->slots[0] = first_xattr_slot;
3309                 ret = btrfs_load_inode_props(inode, path);
3310                 if (ret)
3311                         btrfs_err(fs_info,
3312                                   "error loading props for ino %llu (root %llu): %d",
3313                                   btrfs_ino(BTRFS_I(inode)),
3314                                   root->root_key.objectid, ret);
3315         }
3316         if (path != in_path)
3317                 btrfs_free_path(path);
3318
3319         if (!maybe_acls)
3320                 cache_no_acl(inode);
3321
3322         switch (inode->i_mode & S_IFMT) {
3323         case S_IFREG:
3324                 inode->i_mapping->a_ops = &btrfs_aops;
3325                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3326                 inode->i_fop = &btrfs_file_operations;
3327                 inode->i_op = &btrfs_file_inode_operations;
3328                 break;
3329         case S_IFDIR:
3330                 inode->i_fop = &btrfs_dir_file_operations;
3331                 inode->i_op = &btrfs_dir_inode_operations;
3332                 break;
3333         case S_IFLNK:
3334                 inode->i_op = &btrfs_symlink_inode_operations;
3335                 inode_nohighmem(inode);
3336                 inode->i_mapping->a_ops = &btrfs_aops;
3337                 break;
3338         default:
3339                 inode->i_op = &btrfs_special_inode_operations;
3340                 init_special_inode(inode, inode->i_mode, rdev);
3341                 break;
3342         }
3343
3344         btrfs_sync_inode_flags_to_i_flags(inode);
3345         return 0;
3346 }
3347
3348 /*
3349  * given a leaf and an inode, copy the inode fields into the leaf
3350  */
3351 static void fill_inode_item(struct btrfs_trans_handle *trans,
3352                             struct extent_buffer *leaf,
3353                             struct btrfs_inode_item *item,
3354                             struct inode *inode)
3355 {
3356         struct btrfs_map_token token;
3357
3358         btrfs_init_map_token(&token, leaf);
3359
3360         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3361         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3362         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3363                                    &token);
3364         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3365         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3366
3367         btrfs_set_token_timespec_sec(leaf, &item->atime,
3368                                      inode->i_atime.tv_sec, &token);
3369         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3370                                       inode->i_atime.tv_nsec, &token);
3371
3372         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3373                                      inode->i_mtime.tv_sec, &token);
3374         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3375                                       inode->i_mtime.tv_nsec, &token);
3376
3377         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3378                                      inode->i_ctime.tv_sec, &token);
3379         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3380                                       inode->i_ctime.tv_nsec, &token);
3381
3382         btrfs_set_token_timespec_sec(leaf, &item->otime,
3383                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3384         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3385                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3386
3387         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3388                                      &token);
3389         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3390                                          &token);
3391         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3392                                        &token);
3393         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3394         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3395         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3396         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3397 }
3398
3399 /*
3400  * copy everything in the in-memory inode into the btree.
3401  */
3402 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3403                                 struct btrfs_root *root, struct inode *inode)
3404 {
3405         struct btrfs_inode_item *inode_item;
3406         struct btrfs_path *path;
3407         struct extent_buffer *leaf;
3408         int ret;
3409
3410         path = btrfs_alloc_path();
3411         if (!path)
3412                 return -ENOMEM;
3413
3414         path->leave_spinning = 1;
3415         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3416                                  1);
3417         if (ret) {
3418                 if (ret > 0)
3419                         ret = -ENOENT;
3420                 goto failed;
3421         }
3422
3423         leaf = path->nodes[0];
3424         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3425                                     struct btrfs_inode_item);
3426
3427         fill_inode_item(trans, leaf, inode_item, inode);
3428         btrfs_mark_buffer_dirty(leaf);
3429         btrfs_set_inode_last_trans(trans, inode);
3430         ret = 0;
3431 failed:
3432         btrfs_free_path(path);
3433         return ret;
3434 }
3435
3436 /*
3437  * copy everything in the in-memory inode into the btree.
3438  */
3439 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3440                                 struct btrfs_root *root, struct inode *inode)
3441 {
3442         struct btrfs_fs_info *fs_info = root->fs_info;
3443         int ret;
3444
3445         /*
3446          * If the inode is a free space inode, we can deadlock during commit
3447          * if we put it into the delayed code.
3448          *
3449          * The data relocation inode should also be directly updated
3450          * without delay
3451          */
3452         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3453             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3454             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3455                 btrfs_update_root_times(trans, root);
3456
3457                 ret = btrfs_delayed_update_inode(trans, root, inode);
3458                 if (!ret)
3459                         btrfs_set_inode_last_trans(trans, inode);
3460                 return ret;
3461         }
3462
3463         return btrfs_update_inode_item(trans, root, inode);
3464 }
3465
3466 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3467                                          struct btrfs_root *root,
3468                                          struct inode *inode)
3469 {
3470         int ret;
3471
3472         ret = btrfs_update_inode(trans, root, inode);
3473         if (ret == -ENOSPC)
3474                 return btrfs_update_inode_item(trans, root, inode);
3475         return ret;
3476 }
3477
3478 /*
3479  * unlink helper that gets used here in inode.c and in the tree logging
3480  * recovery code.  It remove a link in a directory with a given name, and
3481  * also drops the back refs in the inode to the directory
3482  */
3483 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3484                                 struct btrfs_root *root,
3485                                 struct btrfs_inode *dir,
3486                                 struct btrfs_inode *inode,
3487                                 const char *name, int name_len)
3488 {
3489         struct btrfs_fs_info *fs_info = root->fs_info;
3490         struct btrfs_path *path;
3491         int ret = 0;
3492         struct btrfs_dir_item *di;
3493         u64 index;
3494         u64 ino = btrfs_ino(inode);
3495         u64 dir_ino = btrfs_ino(dir);
3496
3497         path = btrfs_alloc_path();
3498         if (!path) {
3499                 ret = -ENOMEM;
3500                 goto out;
3501         }
3502
3503         path->leave_spinning = 1;
3504         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3505                                     name, name_len, -1);
3506         if (IS_ERR_OR_NULL(di)) {
3507                 ret = di ? PTR_ERR(di) : -ENOENT;
3508                 goto err;
3509         }
3510         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3511         if (ret)
3512                 goto err;
3513         btrfs_release_path(path);
3514
3515         /*
3516          * If we don't have dir index, we have to get it by looking up
3517          * the inode ref, since we get the inode ref, remove it directly,
3518          * it is unnecessary to do delayed deletion.
3519          *
3520          * But if we have dir index, needn't search inode ref to get it.
3521          * Since the inode ref is close to the inode item, it is better
3522          * that we delay to delete it, and just do this deletion when
3523          * we update the inode item.
3524          */
3525         if (inode->dir_index) {
3526                 ret = btrfs_delayed_delete_inode_ref(inode);
3527                 if (!ret) {
3528                         index = inode->dir_index;
3529                         goto skip_backref;
3530                 }
3531         }
3532
3533         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3534                                   dir_ino, &index);
3535         if (ret) {
3536                 btrfs_info(fs_info,
3537                         "failed to delete reference to %.*s, inode %llu parent %llu",
3538                         name_len, name, ino, dir_ino);
3539                 btrfs_abort_transaction(trans, ret);
3540                 goto err;
3541         }
3542 skip_backref:
3543         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
3544         if (ret) {
3545                 btrfs_abort_transaction(trans, ret);
3546                 goto err;
3547         }
3548
3549         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
3550                         dir_ino);
3551         if (ret != 0 && ret != -ENOENT) {
3552                 btrfs_abort_transaction(trans, ret);
3553                 goto err;
3554         }
3555
3556         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
3557                         index);
3558         if (ret == -ENOENT)
3559                 ret = 0;
3560         else if (ret)
3561                 btrfs_abort_transaction(trans, ret);
3562
3563         /*
3564          * If we have a pending delayed iput we could end up with the final iput
3565          * being run in btrfs-cleaner context.  If we have enough of these built
3566          * up we can end up burning a lot of time in btrfs-cleaner without any
3567          * way to throttle the unlinks.  Since we're currently holding a ref on
3568          * the inode we can run the delayed iput here without any issues as the
3569          * final iput won't be done until after we drop the ref we're currently
3570          * holding.
3571          */
3572         btrfs_run_delayed_iput(fs_info, inode);
3573 err:
3574         btrfs_free_path(path);
3575         if (ret)
3576                 goto out;
3577
3578         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
3579         inode_inc_iversion(&inode->vfs_inode);
3580         inode_inc_iversion(&dir->vfs_inode);
3581         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
3582                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
3583         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
3584 out:
3585         return ret;
3586 }
3587
3588 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3589                        struct btrfs_root *root,
3590                        struct btrfs_inode *dir, struct btrfs_inode *inode,
3591                        const char *name, int name_len)
3592 {
3593         int ret;
3594         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
3595         if (!ret) {
3596                 drop_nlink(&inode->vfs_inode);
3597                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
3598         }
3599         return ret;
3600 }
3601
3602 /*
3603  * helper to start transaction for unlink and rmdir.
3604  *
3605  * unlink and rmdir are special in btrfs, they do not always free space, so
3606  * if we cannot make our reservations the normal way try and see if there is
3607  * plenty of slack room in the global reserve to migrate, otherwise we cannot
3608  * allow the unlink to occur.
3609  */
3610 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
3611 {
3612         struct btrfs_root *root = BTRFS_I(dir)->root;
3613
3614         /*
3615          * 1 for the possible orphan item
3616          * 1 for the dir item
3617          * 1 for the dir index
3618          * 1 for the inode ref
3619          * 1 for the inode
3620          */
3621         return btrfs_start_transaction_fallback_global_rsv(root, 5);
3622 }
3623
3624 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
3625 {
3626         struct btrfs_root *root = BTRFS_I(dir)->root;
3627         struct btrfs_trans_handle *trans;
3628         struct inode *inode = d_inode(dentry);
3629         int ret;
3630
3631         trans = __unlink_start_trans(dir);
3632         if (IS_ERR(trans))
3633                 return PTR_ERR(trans);
3634
3635         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
3636                         0);
3637
3638         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
3639                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
3640                         dentry->d_name.len);
3641         if (ret)
3642                 goto out;
3643
3644         if (inode->i_nlink == 0) {
3645                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3646                 if (ret)
3647                         goto out;
3648         }
3649
3650 out:
3651         btrfs_end_transaction(trans);
3652         btrfs_btree_balance_dirty(root->fs_info);
3653         return ret;
3654 }
3655
3656 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
3657                                struct inode *dir, struct dentry *dentry)
3658 {
3659         struct btrfs_root *root = BTRFS_I(dir)->root;
3660         struct btrfs_inode *inode = BTRFS_I(d_inode(dentry));
3661         struct btrfs_path *path;
3662         struct extent_buffer *leaf;
3663         struct btrfs_dir_item *di;
3664         struct btrfs_key key;
3665         const char *name = dentry->d_name.name;
3666         int name_len = dentry->d_name.len;
3667         u64 index;
3668         int ret;
3669         u64 objectid;
3670         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
3671
3672         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID) {
3673                 objectid = inode->root->root_key.objectid;
3674         } else if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID) {
3675                 objectid = inode->location.objectid;
3676         } else {
3677                 WARN_ON(1);
3678                 return -EINVAL;
3679         }
3680
3681         path = btrfs_alloc_path();
3682         if (!path)
3683                 return -ENOMEM;
3684
3685         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3686                                    name, name_len, -1);
3687         if (IS_ERR_OR_NULL(di)) {
3688                 ret = di ? PTR_ERR(di) : -ENOENT;
3689                 goto out;
3690         }
3691
3692         leaf = path->nodes[0];
3693         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3694         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
3695         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3696         if (ret) {
3697                 btrfs_abort_transaction(trans, ret);
3698                 goto out;
3699         }
3700         btrfs_release_path(path);
3701
3702         /*
3703          * This is a placeholder inode for a subvolume we didn't have a
3704          * reference to at the time of the snapshot creation.  In the meantime
3705          * we could have renamed the real subvol link into our snapshot, so
3706          * depending on btrfs_del_root_ref to return -ENOENT here is incorret.
3707          * Instead simply lookup the dir_index_item for this entry so we can
3708          * remove it.  Otherwise we know we have a ref to the root and we can
3709          * call btrfs_del_root_ref, and it _shouldn't_ fail.
3710          */
3711         if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID) {
3712                 di = btrfs_search_dir_index_item(root, path, dir_ino,
3713                                                  name, name_len);
3714                 if (IS_ERR_OR_NULL(di)) {
3715                         if (!di)
3716                                 ret = -ENOENT;
3717                         else
3718                                 ret = PTR_ERR(di);
3719                         btrfs_abort_transaction(trans, ret);
3720                         goto out;
3721                 }
3722
3723                 leaf = path->nodes[0];
3724                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3725                 index = key.offset;
3726                 btrfs_release_path(path);
3727         } else {
3728                 ret = btrfs_del_root_ref(trans, objectid,
3729                                          root->root_key.objectid, dir_ino,
3730                                          &index, name, name_len);
3731                 if (ret) {
3732                         btrfs_abort_transaction(trans, ret);
3733                         goto out;
3734                 }
3735         }
3736
3737         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
3738         if (ret) {
3739                 btrfs_abort_transaction(trans, ret);
3740                 goto out;
3741         }
3742
3743         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
3744         inode_inc_iversion(dir);
3745         dir->i_mtime = dir->i_ctime = current_time(dir);
3746         ret = btrfs_update_inode_fallback(trans, root, dir);
3747         if (ret)
3748                 btrfs_abort_transaction(trans, ret);
3749 out:
3750         btrfs_free_path(path);
3751         return ret;
3752 }
3753
3754 /*
3755  * Helper to check if the subvolume references other subvolumes or if it's
3756  * default.
3757  */
3758 static noinline int may_destroy_subvol(struct btrfs_root *root)
3759 {
3760         struct btrfs_fs_info *fs_info = root->fs_info;
3761         struct btrfs_path *path;
3762         struct btrfs_dir_item *di;
3763         struct btrfs_key key;
3764         u64 dir_id;
3765         int ret;
3766
3767         path = btrfs_alloc_path();
3768         if (!path)
3769                 return -ENOMEM;
3770
3771         /* Make sure this root isn't set as the default subvol */
3772         dir_id = btrfs_super_root_dir(fs_info->super_copy);
3773         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
3774                                    dir_id, "default", 7, 0);
3775         if (di && !IS_ERR(di)) {
3776                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
3777                 if (key.objectid == root->root_key.objectid) {
3778                         ret = -EPERM;
3779                         btrfs_err(fs_info,
3780                                   "deleting default subvolume %llu is not allowed",
3781                                   key.objectid);
3782                         goto out;
3783                 }
3784                 btrfs_release_path(path);
3785         }
3786
3787         key.objectid = root->root_key.objectid;
3788         key.type = BTRFS_ROOT_REF_KEY;
3789         key.offset = (u64)-1;
3790
3791         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
3792         if (ret < 0)
3793                 goto out;
3794         BUG_ON(ret == 0);
3795
3796         ret = 0;
3797         if (path->slots[0] > 0) {
3798                 path->slots[0]--;
3799                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3800                 if (key.objectid == root->root_key.objectid &&
3801                     key.type == BTRFS_ROOT_REF_KEY)
3802                         ret = -ENOTEMPTY;
3803         }
3804 out:
3805         btrfs_free_path(path);
3806         return ret;
3807 }
3808
3809 /* Delete all dentries for inodes belonging to the root */
3810 static void btrfs_prune_dentries(struct btrfs_root *root)
3811 {
3812         struct btrfs_fs_info *fs_info = root->fs_info;
3813         struct rb_node *node;
3814         struct rb_node *prev;
3815         struct btrfs_inode *entry;
3816         struct inode *inode;
3817         u64 objectid = 0;
3818
3819         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3820                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
3821
3822         spin_lock(&root->inode_lock);
3823 again:
3824         node = root->inode_tree.rb_node;
3825         prev = NULL;
3826         while (node) {
3827                 prev = node;
3828                 entry = rb_entry(node, struct btrfs_inode, rb_node);
3829
3830                 if (objectid < btrfs_ino(entry))
3831                         node = node->rb_left;
3832                 else if (objectid > btrfs_ino(entry))
3833                         node = node->rb_right;
3834                 else
3835                         break;
3836         }
3837         if (!node) {
3838                 while (prev) {
3839                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
3840                         if (objectid <= btrfs_ino(entry)) {
3841                                 node = prev;
3842                                 break;
3843                         }
3844                         prev = rb_next(prev);
3845                 }
3846         }
3847         while (node) {
3848                 entry = rb_entry(node, struct btrfs_inode, rb_node);
3849                 objectid = btrfs_ino(entry) + 1;
3850                 inode = igrab(&entry->vfs_inode);
3851                 if (inode) {
3852                         spin_unlock(&root->inode_lock);
3853                         if (atomic_read(&inode->i_count) > 1)
3854                                 d_prune_aliases(inode);
3855                         /*
3856                          * btrfs_drop_inode will have it removed from the inode
3857                          * cache when its usage count hits zero.
3858                          */
3859                         iput(inode);
3860                         cond_resched();
3861                         spin_lock(&root->inode_lock);
3862                         goto again;
3863                 }
3864
3865                 if (cond_resched_lock(&root->inode_lock))
3866                         goto again;
3867
3868                 node = rb_next(node);
3869         }
3870         spin_unlock(&root->inode_lock);
3871 }
3872
3873 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
3874 {
3875         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
3876         struct btrfs_root *root = BTRFS_I(dir)->root;
3877         struct inode *inode = d_inode(dentry);
3878         struct btrfs_root *dest = BTRFS_I(inode)->root;
3879         struct btrfs_trans_handle *trans;
3880         struct btrfs_block_rsv block_rsv;
3881         u64 root_flags;
3882         int ret;
3883         int err;
3884
3885         /*
3886          * Don't allow to delete a subvolume with send in progress. This is
3887          * inside the inode lock so the error handling that has to drop the bit
3888          * again is not run concurrently.
3889          */
3890         spin_lock(&dest->root_item_lock);
3891         if (dest->send_in_progress) {
3892                 spin_unlock(&dest->root_item_lock);
3893                 btrfs_warn(fs_info,
3894                            "attempt to delete subvolume %llu during send",
3895                            dest->root_key.objectid);
3896                 return -EPERM;
3897         }
3898         root_flags = btrfs_root_flags(&dest->root_item);
3899         btrfs_set_root_flags(&dest->root_item,
3900                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
3901         spin_unlock(&dest->root_item_lock);
3902
3903         down_write(&fs_info->subvol_sem);
3904
3905         err = may_destroy_subvol(dest);
3906         if (err)
3907                 goto out_up_write;
3908
3909         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
3910         /*
3911          * One for dir inode,
3912          * two for dir entries,
3913          * two for root ref/backref.
3914          */
3915         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
3916         if (err)
3917                 goto out_up_write;
3918
3919         trans = btrfs_start_transaction(root, 0);
3920         if (IS_ERR(trans)) {
3921                 err = PTR_ERR(trans);
3922                 goto out_release;
3923         }
3924         trans->block_rsv = &block_rsv;
3925         trans->bytes_reserved = block_rsv.size;
3926
3927         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
3928
3929         ret = btrfs_unlink_subvol(trans, dir, dentry);
3930         if (ret) {
3931                 err = ret;
3932                 btrfs_abort_transaction(trans, ret);
3933                 goto out_end_trans;
3934         }
3935
3936         btrfs_record_root_in_trans(trans, dest);
3937
3938         memset(&dest->root_item.drop_progress, 0,
3939                 sizeof(dest->root_item.drop_progress));
3940         dest->root_item.drop_level = 0;
3941         btrfs_set_root_refs(&dest->root_item, 0);
3942
3943         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
3944                 ret = btrfs_insert_orphan_item(trans,
3945                                         fs_info->tree_root,
3946                                         dest->root_key.objectid);
3947                 if (ret) {
3948                         btrfs_abort_transaction(trans, ret);
3949                         err = ret;
3950                         goto out_end_trans;
3951                 }
3952         }
3953
3954         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
3955                                   BTRFS_UUID_KEY_SUBVOL,
3956                                   dest->root_key.objectid);
3957         if (ret && ret != -ENOENT) {
3958                 btrfs_abort_transaction(trans, ret);
3959                 err = ret;
3960                 goto out_end_trans;
3961         }
3962         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
3963                 ret = btrfs_uuid_tree_remove(trans,
3964                                           dest->root_item.received_uuid,
3965                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
3966                                           dest->root_key.objectid);
3967                 if (ret && ret != -ENOENT) {
3968                         btrfs_abort_transaction(trans, ret);
3969                         err = ret;
3970                         goto out_end_trans;
3971                 }
3972         }
3973
3974 out_end_trans:
3975         trans->block_rsv = NULL;
3976         trans->bytes_reserved = 0;
3977         ret = btrfs_end_transaction(trans);
3978         if (ret && !err)
3979                 err = ret;
3980         inode->i_flags |= S_DEAD;
3981 out_release:
3982         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
3983 out_up_write:
3984         up_write(&fs_info->subvol_sem);
3985         if (err) {
3986                 spin_lock(&dest->root_item_lock);
3987                 root_flags = btrfs_root_flags(&dest->root_item);
3988                 btrfs_set_root_flags(&dest->root_item,
3989                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
3990                 spin_unlock(&dest->root_item_lock);
3991         } else {
3992                 d_invalidate(dentry);
3993                 btrfs_prune_dentries(dest);
3994                 ASSERT(dest->send_in_progress == 0);
3995
3996                 /* the last ref */
3997                 if (dest->ino_cache_inode) {
3998                         iput(dest->ino_cache_inode);
3999                         dest->ino_cache_inode = NULL;
4000                 }
4001         }
4002
4003         return err;
4004 }
4005
4006 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4007 {
4008         struct inode *inode = d_inode(dentry);
4009         int err = 0;
4010         struct btrfs_root *root = BTRFS_I(dir)->root;
4011         struct btrfs_trans_handle *trans;
4012         u64 last_unlink_trans;
4013
4014         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4015                 return -ENOTEMPTY;
4016         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4017                 return btrfs_delete_subvolume(dir, dentry);
4018
4019         trans = __unlink_start_trans(dir);
4020         if (IS_ERR(trans))
4021                 return PTR_ERR(trans);
4022
4023         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4024                 err = btrfs_unlink_subvol(trans, dir, dentry);
4025                 goto out;
4026         }
4027
4028         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4029         if (err)
4030                 goto out;
4031
4032         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4033
4034         /* now the directory is empty */
4035         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4036                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4037                         dentry->d_name.len);
4038         if (!err) {
4039                 btrfs_i_size_write(BTRFS_I(inode), 0);
4040                 /*
4041                  * Propagate the last_unlink_trans value of the deleted dir to
4042                  * its parent directory. This is to prevent an unrecoverable
4043                  * log tree in the case we do something like this:
4044                  * 1) create dir foo
4045                  * 2) create snapshot under dir foo
4046                  * 3) delete the snapshot
4047                  * 4) rmdir foo
4048                  * 5) mkdir foo
4049                  * 6) fsync foo or some file inside foo
4050                  */
4051                 if (last_unlink_trans >= trans->transid)
4052                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4053         }
4054 out:
4055         btrfs_end_transaction(trans);
4056         btrfs_btree_balance_dirty(root->fs_info);
4057
4058         return err;
4059 }
4060
4061 /*
4062  * Return this if we need to call truncate_block for the last bit of the
4063  * truncate.
4064  */
4065 #define NEED_TRUNCATE_BLOCK 1
4066
4067 /*
4068  * this can truncate away extent items, csum items and directory items.
4069  * It starts at a high offset and removes keys until it can't find
4070  * any higher than new_size
4071  *
4072  * csum items that cross the new i_size are truncated to the new size
4073  * as well.
4074  *
4075  * min_type is the minimum key type to truncate down to.  If set to 0, this
4076  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4077  */
4078 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4079                                struct btrfs_root *root,
4080                                struct inode *inode,
4081                                u64 new_size, u32 min_type)
4082 {
4083         struct btrfs_fs_info *fs_info = root->fs_info;
4084         struct btrfs_path *path;
4085         struct extent_buffer *leaf;
4086         struct btrfs_file_extent_item *fi;
4087         struct btrfs_key key;
4088         struct btrfs_key found_key;
4089         u64 extent_start = 0;
4090         u64 extent_num_bytes = 0;
4091         u64 extent_offset = 0;
4092         u64 item_end = 0;
4093         u64 last_size = new_size;
4094         u32 found_type = (u8)-1;
4095         int found_extent;
4096         int del_item;
4097         int pending_del_nr = 0;
4098         int pending_del_slot = 0;
4099         int extent_type = -1;
4100         int ret;
4101         u64 ino = btrfs_ino(BTRFS_I(inode));
4102         u64 bytes_deleted = 0;
4103         bool be_nice = false;
4104         bool should_throttle = false;
4105         const u64 lock_start = ALIGN_DOWN(new_size, fs_info->sectorsize);
4106         struct extent_state *cached_state = NULL;
4107
4108         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4109
4110         /*
4111          * for non-free space inodes and ref cows, we want to back off from
4112          * time to time
4113          */
4114         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4115             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4116                 be_nice = true;
4117
4118         path = btrfs_alloc_path();
4119         if (!path)
4120                 return -ENOMEM;
4121         path->reada = READA_BACK;
4122
4123         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4124                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, (u64)-1,
4125                                  &cached_state);
4126
4127         /*
4128          * We want to drop from the next block forward in case this new size is
4129          * not block aligned since we will be keeping the last block of the
4130          * extent just the way it is.
4131          */
4132         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4133             root == fs_info->tree_root)
4134                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4135                                         fs_info->sectorsize),
4136                                         (u64)-1, 0);
4137
4138         /*
4139          * This function is also used to drop the items in the log tree before
4140          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4141          * it is used to drop the logged items. So we shouldn't kill the delayed
4142          * items.
4143          */
4144         if (min_type == 0 && root == BTRFS_I(inode)->root)
4145                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4146
4147         key.objectid = ino;
4148         key.offset = (u64)-1;
4149         key.type = (u8)-1;
4150
4151 search_again:
4152         /*
4153          * with a 16K leaf size and 128MB extents, you can actually queue
4154          * up a huge file in a single leaf.  Most of the time that
4155          * bytes_deleted is > 0, it will be huge by the time we get here
4156          */
4157         if (be_nice && bytes_deleted > SZ_32M &&
4158             btrfs_should_end_transaction(trans)) {
4159                 ret = -EAGAIN;
4160                 goto out;
4161         }
4162
4163         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4164         if (ret < 0)
4165                 goto out;
4166
4167         if (ret > 0) {
4168                 ret = 0;
4169                 /* there are no items in the tree for us to truncate, we're
4170                  * done
4171                  */
4172                 if (path->slots[0] == 0)
4173                         goto out;
4174                 path->slots[0]--;
4175         }
4176
4177         while (1) {
4178                 u64 clear_start = 0, clear_len = 0;
4179
4180                 fi = NULL;
4181                 leaf = path->nodes[0];
4182                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4183                 found_type = found_key.type;
4184
4185                 if (found_key.objectid != ino)
4186                         break;
4187
4188                 if (found_type < min_type)
4189                         break;
4190
4191                 item_end = found_key.offset;
4192                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4193                         fi = btrfs_item_ptr(leaf, path->slots[0],
4194                                             struct btrfs_file_extent_item);
4195                         extent_type = btrfs_file_extent_type(leaf, fi);
4196                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4197                                 item_end +=
4198                                     btrfs_file_extent_num_bytes(leaf, fi);
4199
4200                                 trace_btrfs_truncate_show_fi_regular(
4201                                         BTRFS_I(inode), leaf, fi,
4202                                         found_key.offset);
4203                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4204                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4205                                                                         fi);
4206
4207                                 trace_btrfs_truncate_show_fi_inline(
4208                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4209                                         found_key.offset);
4210                         }
4211                         item_end--;
4212                 }
4213                 if (found_type > min_type) {
4214                         del_item = 1;
4215                 } else {
4216                         if (item_end < new_size)
4217                                 break;
4218                         if (found_key.offset >= new_size)
4219                                 del_item = 1;
4220                         else
4221                                 del_item = 0;
4222                 }
4223                 found_extent = 0;
4224                 /* FIXME, shrink the extent if the ref count is only 1 */
4225                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4226                         goto delete;
4227
4228                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4229                         u64 num_dec;
4230
4231                         clear_start = found_key.offset;
4232                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4233                         if (!del_item) {
4234                                 u64 orig_num_bytes =
4235                                         btrfs_file_extent_num_bytes(leaf, fi);
4236                                 extent_num_bytes = ALIGN(new_size -
4237                                                 found_key.offset,
4238                                                 fs_info->sectorsize);
4239                                 clear_start = ALIGN(new_size, fs_info->sectorsize);
4240                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4241                                                          extent_num_bytes);
4242                                 num_dec = (orig_num_bytes -
4243                                            extent_num_bytes);
4244                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4245                                              &root->state) &&
4246                                     extent_start != 0)
4247                                         inode_sub_bytes(inode, num_dec);
4248                                 btrfs_mark_buffer_dirty(leaf);
4249                         } else {
4250                                 extent_num_bytes =
4251                                         btrfs_file_extent_disk_num_bytes(leaf,
4252                                                                          fi);
4253                                 extent_offset = found_key.offset -
4254                                         btrfs_file_extent_offset(leaf, fi);
4255
4256                                 /* FIXME blocksize != 4096 */
4257                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4258                                 if (extent_start != 0) {
4259                                         found_extent = 1;
4260                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4261                                                      &root->state))
4262                                                 inode_sub_bytes(inode, num_dec);
4263                                 }
4264                         }
4265                         clear_len = num_dec;
4266                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4267                         /*
4268                          * we can't truncate inline items that have had
4269                          * special encodings
4270                          */
4271                         if (!del_item &&
4272                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4273                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4274                             btrfs_file_extent_compression(leaf, fi) == 0) {
4275                                 u32 size = (u32)(new_size - found_key.offset);
4276
4277                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4278                                 size = btrfs_file_extent_calc_inline_size(size);
4279                                 btrfs_truncate_item(path, size, 1);
4280                         } else if (!del_item) {
4281                                 /*
4282                                  * We have to bail so the last_size is set to
4283                                  * just before this extent.
4284                                  */
4285                                 ret = NEED_TRUNCATE_BLOCK;
4286                                 break;
4287                         } else {
4288                                 /*
4289                                  * Inline extents are special, we just treat
4290                                  * them as a full sector worth in the file
4291                                  * extent tree just for simplicity sake.
4292                                  */
4293                                 clear_len = fs_info->sectorsize;
4294                         }
4295
4296                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4297                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4298                 }
4299 delete:
4300                 /*
4301                  * We use btrfs_truncate_inode_items() to clean up log trees for
4302                  * multiple fsyncs, and in this case we don't want to clear the
4303                  * file extent range because it's just the log.
4304                  */
4305                 if (root == BTRFS_I(inode)->root) {
4306                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
4307                                                   clear_start, clear_len);
4308                         if (ret) {
4309                                 btrfs_abort_transaction(trans, ret);
4310                                 break;
4311                         }
4312                 }
4313
4314                 if (del_item)
4315                         last_size = found_key.offset;
4316                 else
4317                         last_size = new_size;
4318                 if (del_item) {
4319                         if (!pending_del_nr) {
4320                                 /* no pending yet, add ourselves */
4321                                 pending_del_slot = path->slots[0];
4322                                 pending_del_nr = 1;
4323                         } else if (pending_del_nr &&
4324                                    path->slots[0] + 1 == pending_del_slot) {
4325                                 /* hop on the pending chunk */
4326                                 pending_del_nr++;
4327                                 pending_del_slot = path->slots[0];
4328                         } else {
4329                                 BUG();
4330                         }
4331                 } else {
4332                         break;
4333                 }
4334                 should_throttle = false;
4335
4336                 if (found_extent &&
4337                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4338                      root == fs_info->tree_root)) {
4339                         struct btrfs_ref ref = { 0 };
4340
4341                         bytes_deleted += extent_num_bytes;
4342
4343                         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF,
4344                                         extent_start, extent_num_bytes, 0);
4345                         ref.real_root = root->root_key.objectid;
4346                         btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
4347                                         ino, extent_offset);
4348                         ret = btrfs_free_extent(trans, &ref);
4349                         if (ret) {
4350                                 btrfs_abort_transaction(trans, ret);
4351                                 break;
4352                         }
4353                         if (be_nice) {
4354                                 if (btrfs_should_throttle_delayed_refs(trans))
4355                                         should_throttle = true;
4356                         }
4357                 }
4358
4359                 if (found_type == BTRFS_INODE_ITEM_KEY)
4360                         break;
4361
4362                 if (path->slots[0] == 0 ||
4363                     path->slots[0] != pending_del_slot ||
4364                     should_throttle) {
4365                         if (pending_del_nr) {
4366                                 ret = btrfs_del_items(trans, root, path,
4367                                                 pending_del_slot,
4368                                                 pending_del_nr);
4369                                 if (ret) {
4370                                         btrfs_abort_transaction(trans, ret);
4371                                         break;
4372                                 }
4373                                 pending_del_nr = 0;
4374                         }
4375                         btrfs_release_path(path);
4376
4377                         /*
4378                          * We can generate a lot of delayed refs, so we need to
4379                          * throttle every once and a while and make sure we're
4380                          * adding enough space to keep up with the work we are
4381                          * generating.  Since we hold a transaction here we
4382                          * can't flush, and we don't want to FLUSH_LIMIT because
4383                          * we could have generated too many delayed refs to
4384                          * actually allocate, so just bail if we're short and
4385                          * let the normal reservation dance happen higher up.
4386                          */
4387                         if (should_throttle) {
4388                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4389                                                         BTRFS_RESERVE_NO_FLUSH);
4390                                 if (ret) {
4391                                         ret = -EAGAIN;
4392                                         break;
4393                                 }
4394                         }
4395                         goto search_again;
4396                 } else {
4397                         path->slots[0]--;
4398                 }
4399         }
4400 out:
4401         if (ret >= 0 && pending_del_nr) {
4402                 int err;
4403
4404                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4405                                       pending_del_nr);
4406                 if (err) {
4407                         btrfs_abort_transaction(trans, err);
4408                         ret = err;
4409                 }
4410         }
4411         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4412                 ASSERT(last_size >= new_size);
4413                 if (!ret && last_size > new_size)
4414                         last_size = new_size;
4415                 btrfs_inode_safe_disk_i_size_write(inode, last_size);
4416                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start,
4417                                      (u64)-1, &cached_state);
4418         }
4419
4420         btrfs_free_path(path);
4421         return ret;
4422 }
4423
4424 /*
4425  * btrfs_truncate_block - read, zero a chunk and write a block
4426  * @inode - inode that we're zeroing
4427  * @from - the offset to start zeroing
4428  * @len - the length to zero, 0 to zero the entire range respective to the
4429  *      offset
4430  * @front - zero up to the offset instead of from the offset on
4431  *
4432  * This will find the block for the "from" offset and cow the block and zero the
4433  * part we want to zero.  This is used with truncate and hole punching.
4434  */
4435 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4436                         int front)
4437 {
4438         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4439         struct address_space *mapping = inode->i_mapping;
4440         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4441         struct btrfs_ordered_extent *ordered;
4442         struct extent_state *cached_state = NULL;
4443         struct extent_changeset *data_reserved = NULL;
4444         char *kaddr;
4445         u32 blocksize = fs_info->sectorsize;
4446         pgoff_t index = from >> PAGE_SHIFT;
4447         unsigned offset = from & (blocksize - 1);
4448         struct page *page;
4449         gfp_t mask = btrfs_alloc_write_mask(mapping);
4450         int ret = 0;
4451         u64 block_start;
4452         u64 block_end;
4453
4454         if (IS_ALIGNED(offset, blocksize) &&
4455             (!len || IS_ALIGNED(len, blocksize)))
4456                 goto out;
4457
4458         block_start = round_down(from, blocksize);
4459         block_end = block_start + blocksize - 1;
4460
4461         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4462                                            block_start, blocksize);
4463         if (ret)
4464                 goto out;
4465
4466 again:
4467         page = find_or_create_page(mapping, index, mask);
4468         if (!page) {
4469                 btrfs_delalloc_release_space(inode, data_reserved,
4470                                              block_start, blocksize, true);
4471                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4472                 ret = -ENOMEM;
4473                 goto out;
4474         }
4475
4476         if (!PageUptodate(page)) {
4477                 ret = btrfs_readpage(NULL, page);
4478                 lock_page(page);
4479                 if (page->mapping != mapping) {
4480                         unlock_page(page);
4481                         put_page(page);
4482                         goto again;
4483                 }
4484                 if (!PageUptodate(page)) {
4485                         ret = -EIO;
4486                         goto out_unlock;
4487                 }
4488         }
4489         wait_on_page_writeback(page);
4490
4491         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4492         set_page_extent_mapped(page);
4493
4494         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4495         if (ordered) {
4496                 unlock_extent_cached(io_tree, block_start, block_end,
4497                                      &cached_state);
4498                 unlock_page(page);
4499                 put_page(page);
4500                 btrfs_start_ordered_extent(inode, ordered, 1);
4501                 btrfs_put_ordered_extent(ordered);
4502                 goto again;
4503         }
4504
4505         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4506                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4507                          0, 0, &cached_state);
4508
4509         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4510                                         &cached_state);
4511         if (ret) {
4512                 unlock_extent_cached(io_tree, block_start, block_end,
4513                                      &cached_state);
4514                 goto out_unlock;
4515         }
4516
4517         if (offset != blocksize) {
4518                 if (!len)
4519                         len = blocksize - offset;
4520                 kaddr = kmap(page);
4521                 if (front)
4522                         memset(kaddr + (block_start - page_offset(page)),
4523                                 0, offset);
4524                 else
4525                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4526                                 0, len);
4527                 flush_dcache_page(page);
4528                 kunmap(page);
4529         }
4530         ClearPageChecked(page);
4531         set_page_dirty(page);
4532         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4533
4534 out_unlock:
4535         if (ret)
4536                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4537                                              blocksize, true);
4538         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4539         unlock_page(page);
4540         put_page(page);
4541 out:
4542         extent_changeset_free(data_reserved);
4543         return ret;
4544 }
4545
4546 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4547                              u64 offset, u64 len)
4548 {
4549         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4550         struct btrfs_trans_handle *trans;
4551         int ret;
4552
4553         /*
4554          * Still need to make sure the inode looks like it's been updated so
4555          * that any holes get logged if we fsync.
4556          */
4557         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4558                 BTRFS_I(inode)->last_trans = fs_info->generation;
4559                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4560                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4561                 return 0;
4562         }
4563
4564         /*
4565          * 1 - for the one we're dropping
4566          * 1 - for the one we're adding
4567          * 1 - for updating the inode.
4568          */
4569         trans = btrfs_start_transaction(root, 3);
4570         if (IS_ERR(trans))
4571                 return PTR_ERR(trans);
4572
4573         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4574         if (ret) {
4575                 btrfs_abort_transaction(trans, ret);
4576                 btrfs_end_transaction(trans);
4577                 return ret;
4578         }
4579
4580         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4581                         offset, 0, 0, len, 0, len, 0, 0, 0);
4582         if (ret)
4583                 btrfs_abort_transaction(trans, ret);
4584         else
4585                 btrfs_update_inode(trans, root, inode);
4586         btrfs_end_transaction(trans);
4587         return ret;
4588 }
4589
4590 /*
4591  * This function puts in dummy file extents for the area we're creating a hole
4592  * for.  So if we are truncating this file to a larger size we need to insert
4593  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4594  * the range between oldsize and size
4595  */
4596 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4597 {
4598         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4599         struct btrfs_root *root = BTRFS_I(inode)->root;
4600         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4601         struct extent_map *em = NULL;
4602         struct extent_state *cached_state = NULL;
4603         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4604         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4605         u64 block_end = ALIGN(size, fs_info->sectorsize);
4606         u64 last_byte;
4607         u64 cur_offset;
4608         u64 hole_size;
4609         int err = 0;
4610
4611         /*
4612          * If our size started in the middle of a block we need to zero out the
4613          * rest of the block before we expand the i_size, otherwise we could
4614          * expose stale data.
4615          */
4616         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4617         if (err)
4618                 return err;
4619
4620         if (size <= hole_start)
4621                 return 0;
4622
4623         btrfs_lock_and_flush_ordered_range(BTRFS_I(inode), hole_start,
4624                                            block_end - 1, &cached_state);
4625         cur_offset = hole_start;
4626         while (1) {
4627                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4628                                       block_end - cur_offset);
4629                 if (IS_ERR(em)) {
4630                         err = PTR_ERR(em);
4631                         em = NULL;
4632                         break;
4633                 }
4634                 last_byte = min(extent_map_end(em), block_end);
4635                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4636                 hole_size = last_byte - cur_offset;
4637
4638                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4639                         struct extent_map *hole_em;
4640
4641                         err = maybe_insert_hole(root, inode, cur_offset,
4642                                                 hole_size);
4643                         if (err)
4644                                 break;
4645
4646                         err = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
4647                                                         cur_offset, hole_size);
4648                         if (err)
4649                                 break;
4650
4651                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4652                                                 cur_offset + hole_size - 1, 0);
4653                         hole_em = alloc_extent_map();
4654                         if (!hole_em) {
4655                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4656                                         &BTRFS_I(inode)->runtime_flags);
4657                                 goto next;
4658                         }
4659                         hole_em->start = cur_offset;
4660                         hole_em->len = hole_size;
4661                         hole_em->orig_start = cur_offset;
4662
4663                         hole_em->block_start = EXTENT_MAP_HOLE;
4664                         hole_em->block_len = 0;
4665                         hole_em->orig_block_len = 0;
4666                         hole_em->ram_bytes = hole_size;
4667                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4668                         hole_em->generation = fs_info->generation;
4669
4670                         while (1) {
4671                                 write_lock(&em_tree->lock);
4672                                 err = add_extent_mapping(em_tree, hole_em, 1);
4673                                 write_unlock(&em_tree->lock);
4674                                 if (err != -EEXIST)
4675                                         break;
4676                                 btrfs_drop_extent_cache(BTRFS_I(inode),
4677                                                         cur_offset,
4678                                                         cur_offset +
4679                                                         hole_size - 1, 0);
4680                         }
4681                         free_extent_map(hole_em);
4682                 } else {
4683                         err = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
4684                                                         cur_offset, hole_size);
4685                         if (err)
4686                                 break;
4687                 }
4688 next:
4689                 free_extent_map(em);
4690                 em = NULL;
4691                 cur_offset = last_byte;
4692                 if (cur_offset >= block_end)
4693                         break;
4694         }
4695         free_extent_map(em);
4696         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
4697         return err;
4698 }
4699
4700 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4701 {
4702         struct btrfs_root *root = BTRFS_I(inode)->root;
4703         struct btrfs_trans_handle *trans;
4704         loff_t oldsize = i_size_read(inode);
4705         loff_t newsize = attr->ia_size;
4706         int mask = attr->ia_valid;
4707         int ret;
4708
4709         /*
4710          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4711          * special case where we need to update the times despite not having
4712          * these flags set.  For all other operations the VFS set these flags
4713          * explicitly if it wants a timestamp update.
4714          */
4715         if (newsize != oldsize) {
4716                 inode_inc_iversion(inode);
4717                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4718                         inode->i_ctime = inode->i_mtime =
4719                                 current_time(inode);
4720         }
4721
4722         if (newsize > oldsize) {
4723                 /*
4724                  * Don't do an expanding truncate while snapshotting is ongoing.
4725                  * This is to ensure the snapshot captures a fully consistent
4726                  * state of this file - if the snapshot captures this expanding
4727                  * truncation, it must capture all writes that happened before
4728                  * this truncation.
4729                  */
4730                 btrfs_drew_write_lock(&root->snapshot_lock);
4731                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4732                 if (ret) {
4733                         btrfs_drew_write_unlock(&root->snapshot_lock);
4734                         return ret;
4735                 }
4736
4737                 trans = btrfs_start_transaction(root, 1);
4738                 if (IS_ERR(trans)) {
4739                         btrfs_drew_write_unlock(&root->snapshot_lock);
4740                         return PTR_ERR(trans);
4741                 }
4742
4743                 i_size_write(inode, newsize);
4744                 btrfs_inode_safe_disk_i_size_write(inode, 0);
4745                 pagecache_isize_extended(inode, oldsize, newsize);
4746                 ret = btrfs_update_inode(trans, root, inode);
4747                 btrfs_drew_write_unlock(&root->snapshot_lock);
4748                 btrfs_end_transaction(trans);
4749         } else {
4750
4751                 /*
4752                  * We're truncating a file that used to have good data down to
4753                  * zero. Make sure it gets into the ordered flush list so that
4754                  * any new writes get down to disk quickly.
4755                  */
4756                 if (newsize == 0)
4757                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4758                                 &BTRFS_I(inode)->runtime_flags);
4759
4760                 truncate_setsize(inode, newsize);
4761
4762                 /* Disable nonlocked read DIO to avoid the endless truncate */
4763                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
4764                 inode_dio_wait(inode);
4765                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
4766
4767                 ret = btrfs_truncate(inode, newsize == oldsize);
4768                 if (ret && inode->i_nlink) {
4769                         int err;
4770
4771                         /*
4772                          * Truncate failed, so fix up the in-memory size. We
4773                          * adjusted disk_i_size down as we removed extents, so
4774                          * wait for disk_i_size to be stable and then update the
4775                          * in-memory size to match.
4776                          */
4777                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
4778                         if (err)
4779                                 return err;
4780                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
4781                 }
4782         }
4783
4784         return ret;
4785 }
4786
4787 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
4788 {
4789         struct inode *inode = d_inode(dentry);
4790         struct btrfs_root *root = BTRFS_I(inode)->root;
4791         int err;
4792
4793         if (btrfs_root_readonly(root))
4794                 return -EROFS;
4795
4796         err = setattr_prepare(dentry, attr);
4797         if (err)
4798                 return err;
4799
4800         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
4801                 err = btrfs_setsize(inode, attr);
4802                 if (err)
4803                         return err;
4804         }
4805
4806         if (attr->ia_valid) {
4807                 setattr_copy(inode, attr);
4808                 inode_inc_iversion(inode);
4809                 err = btrfs_dirty_inode(inode);
4810
4811                 if (!err && attr->ia_valid & ATTR_MODE)
4812                         err = posix_acl_chmod(inode, inode->i_mode);
4813         }
4814
4815         return err;
4816 }
4817
4818 /*
4819  * While truncating the inode pages during eviction, we get the VFS calling
4820  * btrfs_invalidatepage() against each page of the inode. This is slow because
4821  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
4822  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
4823  * extent_state structures over and over, wasting lots of time.
4824  *
4825  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
4826  * those expensive operations on a per page basis and do only the ordered io
4827  * finishing, while we release here the extent_map and extent_state structures,
4828  * without the excessive merging and splitting.
4829  */
4830 static void evict_inode_truncate_pages(struct inode *inode)
4831 {
4832         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4833         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
4834         struct rb_node *node;
4835
4836         ASSERT(inode->i_state & I_FREEING);
4837         truncate_inode_pages_final(&inode->i_data);
4838
4839         write_lock(&map_tree->lock);
4840         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
4841                 struct extent_map *em;
4842
4843                 node = rb_first_cached(&map_tree->map);
4844                 em = rb_entry(node, struct extent_map, rb_node);
4845                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
4846                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
4847                 remove_extent_mapping(map_tree, em);
4848                 free_extent_map(em);
4849                 if (need_resched()) {
4850                         write_unlock(&map_tree->lock);
4851                         cond_resched();
4852                         write_lock(&map_tree->lock);
4853                 }
4854         }
4855         write_unlock(&map_tree->lock);
4856
4857         /*
4858          * Keep looping until we have no more ranges in the io tree.
4859          * We can have ongoing bios started by readpages (called from readahead)
4860          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
4861          * still in progress (unlocked the pages in the bio but did not yet
4862          * unlocked the ranges in the io tree). Therefore this means some
4863          * ranges can still be locked and eviction started because before
4864          * submitting those bios, which are executed by a separate task (work
4865          * queue kthread), inode references (inode->i_count) were not taken
4866          * (which would be dropped in the end io callback of each bio).
4867          * Therefore here we effectively end up waiting for those bios and
4868          * anyone else holding locked ranges without having bumped the inode's
4869          * reference count - if we don't do it, when they access the inode's
4870          * io_tree to unlock a range it may be too late, leading to an
4871          * use-after-free issue.
4872          */
4873         spin_lock(&io_tree->lock);
4874         while (!RB_EMPTY_ROOT(&io_tree->state)) {
4875                 struct extent_state *state;
4876                 struct extent_state *cached_state = NULL;
4877                 u64 start;
4878                 u64 end;
4879                 unsigned state_flags;
4880
4881                 node = rb_first(&io_tree->state);
4882                 state = rb_entry(node, struct extent_state, rb_node);
4883                 start = state->start;
4884                 end = state->end;
4885                 state_flags = state->state;
4886                 spin_unlock(&io_tree->lock);
4887
4888                 lock_extent_bits(io_tree, start, end, &cached_state);
4889
4890                 /*
4891                  * If still has DELALLOC flag, the extent didn't reach disk,
4892                  * and its reserved space won't be freed by delayed_ref.
4893                  * So we need to free its reserved space here.
4894                  * (Refer to comment in btrfs_invalidatepage, case 2)
4895                  *
4896                  * Note, end is the bytenr of last byte, so we need + 1 here.
4897                  */
4898                 if (state_flags & EXTENT_DELALLOC)
4899                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
4900
4901                 clear_extent_bit(io_tree, start, end,
4902                                  EXTENT_LOCKED | EXTENT_DELALLOC |
4903                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
4904                                  &cached_state);
4905
4906                 cond_resched();
4907                 spin_lock(&io_tree->lock);
4908         }
4909         spin_unlock(&io_tree->lock);
4910 }
4911
4912 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
4913                                                         struct btrfs_block_rsv *rsv)
4914 {
4915         struct btrfs_fs_info *fs_info = root->fs_info;
4916         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4917         struct btrfs_trans_handle *trans;
4918         u64 delayed_refs_extra = btrfs_calc_insert_metadata_size(fs_info, 1);
4919         int ret;
4920
4921         /*
4922          * Eviction should be taking place at some place safe because of our
4923          * delayed iputs.  However the normal flushing code will run delayed
4924          * iputs, so we cannot use FLUSH_ALL otherwise we'll deadlock.
4925          *
4926          * We reserve the delayed_refs_extra here again because we can't use
4927          * btrfs_start_transaction(root, 0) for the same deadlocky reason as
4928          * above.  We reserve our extra bit here because we generate a ton of
4929          * delayed refs activity by truncating.
4930          *
4931          * If we cannot make our reservation we'll attempt to steal from the
4932          * global reserve, because we really want to be able to free up space.
4933          */
4934         ret = btrfs_block_rsv_refill(root, rsv, rsv->size + delayed_refs_extra,
4935                                      BTRFS_RESERVE_FLUSH_EVICT);
4936         if (ret) {
4937                 /*
4938                  * Try to steal from the global reserve if there is space for
4939                  * it.
4940                  */
4941                 if (btrfs_check_space_for_delayed_refs(fs_info) ||
4942                     btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0)) {
4943                         btrfs_warn(fs_info,
4944                                    "could not allocate space for delete; will truncate on mount");
4945                         return ERR_PTR(-ENOSPC);
4946                 }
4947                 delayed_refs_extra = 0;
4948         }
4949
4950         trans = btrfs_join_transaction(root);
4951         if (IS_ERR(trans))
4952                 return trans;
4953
4954         if (delayed_refs_extra) {
4955                 trans->block_rsv = &fs_info->trans_block_rsv;
4956                 trans->bytes_reserved = delayed_refs_extra;
4957                 btrfs_block_rsv_migrate(rsv, trans->block_rsv,
4958                                         delayed_refs_extra, 1);
4959         }
4960         return trans;
4961 }
4962
4963 void btrfs_evict_inode(struct inode *inode)
4964 {
4965         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4966         struct btrfs_trans_handle *trans;
4967         struct btrfs_root *root = BTRFS_I(inode)->root;
4968         struct btrfs_block_rsv *rsv;
4969         int ret;
4970
4971         trace_btrfs_inode_evict(inode);
4972
4973         if (!root) {
4974                 clear_inode(inode);
4975                 return;
4976         }
4977
4978         evict_inode_truncate_pages(inode);
4979
4980         if (inode->i_nlink &&
4981             ((btrfs_root_refs(&root->root_item) != 0 &&
4982               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
4983              btrfs_is_free_space_inode(BTRFS_I(inode))))
4984                 goto no_delete;
4985
4986         if (is_bad_inode(inode))
4987                 goto no_delete;
4988
4989         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
4990
4991         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
4992                 goto no_delete;
4993
4994         if (inode->i_nlink > 0) {
4995                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
4996                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
4997                 goto no_delete;
4998         }
4999
5000         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5001         if (ret)
5002                 goto no_delete;
5003
5004         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5005         if (!rsv)
5006                 goto no_delete;
5007         rsv->size = btrfs_calc_metadata_size(fs_info, 1);
5008         rsv->failfast = 1;
5009
5010         btrfs_i_size_write(BTRFS_I(inode), 0);
5011
5012         while (1) {
5013                 trans = evict_refill_and_join(root, rsv);
5014                 if (IS_ERR(trans))
5015                         goto free_rsv;
5016
5017                 trans->block_rsv = rsv;
5018
5019                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5020                 trans->block_rsv = &fs_info->trans_block_rsv;
5021                 btrfs_end_transaction(trans);
5022                 btrfs_btree_balance_dirty(fs_info);
5023                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5024                         goto free_rsv;
5025                 else if (!ret)
5026                         break;
5027         }
5028
5029         /*
5030          * Errors here aren't a big deal, it just means we leave orphan items in
5031          * the tree. They will be cleaned up on the next mount. If the inode
5032          * number gets reused, cleanup deletes the orphan item without doing
5033          * anything, and unlink reuses the existing orphan item.
5034          *
5035          * If it turns out that we are dropping too many of these, we might want
5036          * to add a mechanism for retrying these after a commit.
5037          */
5038         trans = evict_refill_and_join(root, rsv);
5039         if (!IS_ERR(trans)) {
5040                 trans->block_rsv = rsv;
5041                 btrfs_orphan_del(trans, BTRFS_I(inode));
5042                 trans->block_rsv = &fs_info->trans_block_rsv;
5043                 btrfs_end_transaction(trans);
5044         }
5045
5046         if (!(root == fs_info->tree_root ||
5047               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5048                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5049
5050 free_rsv:
5051         btrfs_free_block_rsv(fs_info, rsv);
5052 no_delete:
5053         /*
5054          * If we didn't successfully delete, the orphan item will still be in
5055          * the tree and we'll retry on the next mount. Again, we might also want
5056          * to retry these periodically in the future.
5057          */
5058         btrfs_remove_delayed_node(BTRFS_I(inode));
5059         clear_inode(inode);
5060 }
5061
5062 /*
5063  * Return the key found in the dir entry in the location pointer, fill @type
5064  * with BTRFS_FT_*, and return 0.
5065  *
5066  * If no dir entries were found, returns -ENOENT.
5067  * If found a corrupted location in dir entry, returns -EUCLEAN.
5068  */
5069 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5070                                struct btrfs_key *location, u8 *type)
5071 {
5072         const char *name = dentry->d_name.name;
5073         int namelen = dentry->d_name.len;
5074         struct btrfs_dir_item *di;
5075         struct btrfs_path *path;
5076         struct btrfs_root *root = BTRFS_I(dir)->root;
5077         int ret = 0;
5078
5079         path = btrfs_alloc_path();
5080         if (!path)
5081                 return -ENOMEM;
5082
5083         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5084                         name, namelen, 0);
5085         if (IS_ERR_OR_NULL(di)) {
5086                 ret = di ? PTR_ERR(di) : -ENOENT;
5087                 goto out;
5088         }
5089
5090         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5091         if (location->type != BTRFS_INODE_ITEM_KEY &&
5092             location->type != BTRFS_ROOT_ITEM_KEY) {
5093                 ret = -EUCLEAN;
5094                 btrfs_warn(root->fs_info,
5095 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5096                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5097                            location->objectid, location->type, location->offset);
5098         }
5099         if (!ret)
5100                 *type = btrfs_dir_type(path->nodes[0], di);
5101 out:
5102         btrfs_free_path(path);
5103         return ret;
5104 }
5105
5106 /*
5107  * when we hit a tree root in a directory, the btrfs part of the inode
5108  * needs to be changed to reflect the root directory of the tree root.  This
5109  * is kind of like crossing a mount point.
5110  */
5111 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5112                                     struct inode *dir,
5113                                     struct dentry *dentry,
5114                                     struct btrfs_key *location,
5115                                     struct btrfs_root **sub_root)
5116 {
5117         struct btrfs_path *path;
5118         struct btrfs_root *new_root;
5119         struct btrfs_root_ref *ref;
5120         struct extent_buffer *leaf;
5121         struct btrfs_key key;
5122         int ret;
5123         int err = 0;
5124
5125         path = btrfs_alloc_path();
5126         if (!path) {
5127                 err = -ENOMEM;
5128                 goto out;
5129         }
5130
5131         err = -ENOENT;
5132         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5133         key.type = BTRFS_ROOT_REF_KEY;
5134         key.offset = location->objectid;
5135
5136         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5137         if (ret) {
5138                 if (ret < 0)
5139                         err = ret;
5140                 goto out;
5141         }
5142
5143         leaf = path->nodes[0];
5144         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5145         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5146             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5147                 goto out;
5148
5149         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5150                                    (unsigned long)(ref + 1),
5151                                    dentry->d_name.len);
5152         if (ret)
5153                 goto out;
5154
5155         btrfs_release_path(path);
5156
5157         new_root = btrfs_get_fs_root(fs_info, location, true);
5158         if (IS_ERR(new_root)) {
5159                 err = PTR_ERR(new_root);
5160                 goto out;
5161         }
5162
5163         *sub_root = new_root;
5164         location->objectid = btrfs_root_dirid(&new_root->root_item);
5165         location->type = BTRFS_INODE_ITEM_KEY;
5166         location->offset = 0;
5167         err = 0;
5168 out:
5169         btrfs_free_path(path);
5170         return err;
5171 }
5172
5173 static void inode_tree_add(struct inode *inode)
5174 {
5175         struct btrfs_root *root = BTRFS_I(inode)->root;
5176         struct btrfs_inode *entry;
5177         struct rb_node **p;
5178         struct rb_node *parent;
5179         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5180         u64 ino = btrfs_ino(BTRFS_I(inode));
5181
5182         if (inode_unhashed(inode))
5183                 return;
5184         parent = NULL;
5185         spin_lock(&root->inode_lock);
5186         p = &root->inode_tree.rb_node;
5187         while (*p) {
5188                 parent = *p;
5189                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5190
5191                 if (ino < btrfs_ino(entry))
5192                         p = &parent->rb_left;
5193                 else if (ino > btrfs_ino(entry))
5194                         p = &parent->rb_right;
5195                 else {
5196                         WARN_ON(!(entry->vfs_inode.i_state &
5197                                   (I_WILL_FREE | I_FREEING)));
5198                         rb_replace_node(parent, new, &root->inode_tree);
5199                         RB_CLEAR_NODE(parent);
5200                         spin_unlock(&root->inode_lock);
5201                         return;
5202                 }
5203         }
5204         rb_link_node(new, parent, p);
5205         rb_insert_color(new, &root->inode_tree);
5206         spin_unlock(&root->inode_lock);
5207 }
5208
5209 static void inode_tree_del(struct inode *inode)
5210 {
5211         struct btrfs_root *root = BTRFS_I(inode)->root;
5212         int empty = 0;
5213
5214         spin_lock(&root->inode_lock);
5215         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5216                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5217                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5218                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5219         }
5220         spin_unlock(&root->inode_lock);
5221
5222         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5223                 spin_lock(&root->inode_lock);
5224                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5225                 spin_unlock(&root->inode_lock);
5226                 if (empty)
5227                         btrfs_add_dead_root(root);
5228         }
5229 }
5230
5231
5232 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5233 {
5234         struct btrfs_iget_args *args = p;
5235         inode->i_ino = args->location->objectid;
5236         memcpy(&BTRFS_I(inode)->location, args->location,
5237                sizeof(*args->location));
5238         BTRFS_I(inode)->root = btrfs_grab_root(args->root);
5239         BUG_ON(args->root && !BTRFS_I(inode)->root);
5240         return 0;
5241 }
5242
5243 static int btrfs_find_actor(struct inode *inode, void *opaque)
5244 {
5245         struct btrfs_iget_args *args = opaque;
5246         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5247                 args->root == BTRFS_I(inode)->root;
5248 }
5249
5250 static struct inode *btrfs_iget_locked(struct super_block *s,
5251                                        struct btrfs_key *location,
5252                                        struct btrfs_root *root)
5253 {
5254         struct inode *inode;
5255         struct btrfs_iget_args args;
5256         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5257
5258         args.location = location;
5259         args.root = root;
5260
5261         inode = iget5_locked(s, hashval, btrfs_find_actor,
5262                              btrfs_init_locked_inode,
5263                              (void *)&args);
5264         return inode;
5265 }
5266
5267 /*
5268  * Get an inode object given its location and corresponding root.
5269  * Path can be preallocated to prevent recursing back to iget through
5270  * allocator. NULL is also valid but may require an additional allocation
5271  * later.
5272  */
5273 struct inode *btrfs_iget_path(struct super_block *s, struct btrfs_key *location,
5274                               struct btrfs_root *root, struct btrfs_path *path)
5275 {
5276         struct inode *inode;
5277
5278         inode = btrfs_iget_locked(s, location, root);
5279         if (!inode)
5280                 return ERR_PTR(-ENOMEM);
5281
5282         if (inode->i_state & I_NEW) {
5283                 int ret;
5284
5285                 ret = btrfs_read_locked_inode(inode, path);
5286                 if (!ret) {
5287                         inode_tree_add(inode);
5288                         unlock_new_inode(inode);
5289                 } else {
5290                         iget_failed(inode);
5291                         /*
5292                          * ret > 0 can come from btrfs_search_slot called by
5293                          * btrfs_read_locked_inode, this means the inode item
5294                          * was not found.
5295                          */
5296                         if (ret > 0)
5297                                 ret = -ENOENT;
5298                         inode = ERR_PTR(ret);
5299                 }
5300         }
5301
5302         return inode;
5303 }
5304
5305 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5306                          struct btrfs_root *root)
5307 {
5308         return btrfs_iget_path(s, location, root, NULL);
5309 }
5310
5311 static struct inode *new_simple_dir(struct super_block *s,
5312                                     struct btrfs_key *key,
5313                                     struct btrfs_root *root)
5314 {
5315         struct inode *inode = new_inode(s);
5316
5317         if (!inode)
5318                 return ERR_PTR(-ENOMEM);
5319
5320         BTRFS_I(inode)->root = btrfs_grab_root(root);
5321         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5322         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5323
5324         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5325         /*
5326          * We only need lookup, the rest is read-only and there's no inode
5327          * associated with the dentry
5328          */
5329         inode->i_op = &simple_dir_inode_operations;
5330         inode->i_opflags &= ~IOP_XATTR;
5331         inode->i_fop = &simple_dir_operations;
5332         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5333         inode->i_mtime = current_time(inode);
5334         inode->i_atime = inode->i_mtime;
5335         inode->i_ctime = inode->i_mtime;
5336         BTRFS_I(inode)->i_otime = inode->i_mtime;
5337
5338         return inode;
5339 }
5340
5341 static inline u8 btrfs_inode_type(struct inode *inode)
5342 {
5343         /*
5344          * Compile-time asserts that generic FT_* types still match
5345          * BTRFS_FT_* types
5346          */
5347         BUILD_BUG_ON(BTRFS_FT_UNKNOWN != FT_UNKNOWN);
5348         BUILD_BUG_ON(BTRFS_FT_REG_FILE != FT_REG_FILE);
5349         BUILD_BUG_ON(BTRFS_FT_DIR != FT_DIR);
5350         BUILD_BUG_ON(BTRFS_FT_CHRDEV != FT_CHRDEV);
5351         BUILD_BUG_ON(BTRFS_FT_BLKDEV != FT_BLKDEV);
5352         BUILD_BUG_ON(BTRFS_FT_FIFO != FT_FIFO);
5353         BUILD_BUG_ON(BTRFS_FT_SOCK != FT_SOCK);
5354         BUILD_BUG_ON(BTRFS_FT_SYMLINK != FT_SYMLINK);
5355
5356         return fs_umode_to_ftype(inode->i_mode);
5357 }
5358
5359 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5360 {
5361         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5362         struct inode *inode;
5363         struct btrfs_root *root = BTRFS_I(dir)->root;
5364         struct btrfs_root *sub_root = root;
5365         struct btrfs_key location;
5366         u8 di_type = 0;
5367         int ret = 0;
5368
5369         if (dentry->d_name.len > BTRFS_NAME_LEN)
5370                 return ERR_PTR(-ENAMETOOLONG);
5371
5372         ret = btrfs_inode_by_name(dir, dentry, &location, &di_type);
5373         if (ret < 0)
5374                 return ERR_PTR(ret);
5375
5376         if (location.type == BTRFS_INODE_ITEM_KEY) {
5377                 inode = btrfs_iget(dir->i_sb, &location, root);
5378                 if (IS_ERR(inode))
5379                         return inode;
5380
5381                 /* Do extra check against inode mode with di_type */
5382                 if (btrfs_inode_type(inode) != di_type) {
5383                         btrfs_crit(fs_info,
5384 "inode mode mismatch with dir: inode mode=0%o btrfs type=%u dir type=%u",
5385                                   inode->i_mode, btrfs_inode_type(inode),
5386                                   di_type);
5387                         iput(inode);
5388                         return ERR_PTR(-EUCLEAN);
5389                 }
5390                 return inode;
5391         }
5392
5393         ret = fixup_tree_root_location(fs_info, dir, dentry,
5394                                        &location, &sub_root);
5395         if (ret < 0) {
5396                 if (ret != -ENOENT)
5397                         inode = ERR_PTR(ret);
5398                 else
5399                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5400         } else {
5401                 inode = btrfs_iget(dir->i_sb, &location, sub_root);
5402         }
5403         if (root != sub_root)
5404                 btrfs_put_root(sub_root);
5405
5406         if (!IS_ERR(inode) && root != sub_root) {
5407                 down_read(&fs_info->cleanup_work_sem);
5408                 if (!sb_rdonly(inode->i_sb))
5409                         ret = btrfs_orphan_cleanup(sub_root);
5410                 up_read(&fs_info->cleanup_work_sem);
5411                 if (ret) {
5412                         iput(inode);
5413                         inode = ERR_PTR(ret);
5414                 }
5415         }
5416
5417         return inode;
5418 }
5419
5420 static int btrfs_dentry_delete(const struct dentry *dentry)
5421 {
5422         struct btrfs_root *root;
5423         struct inode *inode = d_inode(dentry);
5424
5425         if (!inode && !IS_ROOT(dentry))
5426                 inode = d_inode(dentry->d_parent);
5427
5428         if (inode) {
5429                 root = BTRFS_I(inode)->root;
5430                 if (btrfs_root_refs(&root->root_item) == 0)
5431                         return 1;
5432
5433                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5434                         return 1;
5435         }
5436         return 0;
5437 }
5438
5439 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5440                                    unsigned int flags)
5441 {
5442         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5443
5444         if (inode == ERR_PTR(-ENOENT))
5445                 inode = NULL;
5446         return d_splice_alias(inode, dentry);
5447 }
5448
5449 /*
5450  * All this infrastructure exists because dir_emit can fault, and we are holding
5451  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5452  * our information into that, and then dir_emit from the buffer.  This is
5453  * similar to what NFS does, only we don't keep the buffer around in pagecache
5454  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5455  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5456  * tree lock.
5457  */
5458 static int btrfs_opendir(struct inode *inode, struct file *file)
5459 {
5460         struct btrfs_file_private *private;
5461
5462         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5463         if (!private)
5464                 return -ENOMEM;
5465         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5466         if (!private->filldir_buf) {
5467                 kfree(private);
5468                 return -ENOMEM;
5469         }
5470         file->private_data = private;
5471         return 0;
5472 }
5473
5474 struct dir_entry {
5475         u64 ino;
5476         u64 offset;
5477         unsigned type;
5478         int name_len;
5479 };
5480
5481 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5482 {
5483         while (entries--) {
5484                 struct dir_entry *entry = addr;
5485                 char *name = (char *)(entry + 1);
5486
5487                 ctx->pos = get_unaligned(&entry->offset);
5488                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
5489                                          get_unaligned(&entry->ino),
5490                                          get_unaligned(&entry->type)))
5491                         return 1;
5492                 addr += sizeof(struct dir_entry) +
5493                         get_unaligned(&entry->name_len);
5494                 ctx->pos++;
5495         }
5496         return 0;
5497 }
5498
5499 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5500 {
5501         struct inode *inode = file_inode(file);
5502         struct btrfs_root *root = BTRFS_I(inode)->root;
5503         struct btrfs_file_private *private = file->private_data;
5504         struct btrfs_dir_item *di;
5505         struct btrfs_key key;
5506         struct btrfs_key found_key;
5507         struct btrfs_path *path;
5508         void *addr;
5509         struct list_head ins_list;
5510         struct list_head del_list;
5511         int ret;
5512         struct extent_buffer *leaf;
5513         int slot;
5514         char *name_ptr;
5515         int name_len;
5516         int entries = 0;
5517         int total_len = 0;
5518         bool put = false;
5519         struct btrfs_key location;
5520
5521         if (!dir_emit_dots(file, ctx))
5522                 return 0;
5523
5524         path = btrfs_alloc_path();
5525         if (!path)
5526                 return -ENOMEM;
5527
5528         addr = private->filldir_buf;
5529         path->reada = READA_FORWARD;
5530
5531         INIT_LIST_HEAD(&ins_list);
5532         INIT_LIST_HEAD(&del_list);
5533         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5534
5535 again:
5536         key.type = BTRFS_DIR_INDEX_KEY;
5537         key.offset = ctx->pos;
5538         key.objectid = btrfs_ino(BTRFS_I(inode));
5539
5540         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5541         if (ret < 0)
5542                 goto err;
5543
5544         while (1) {
5545                 struct dir_entry *entry;
5546
5547                 leaf = path->nodes[0];
5548                 slot = path->slots[0];
5549                 if (slot >= btrfs_header_nritems(leaf)) {
5550                         ret = btrfs_next_leaf(root, path);
5551                         if (ret < 0)
5552                                 goto err;
5553                         else if (ret > 0)
5554                                 break;
5555                         continue;
5556                 }
5557
5558                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5559
5560                 if (found_key.objectid != key.objectid)
5561                         break;
5562                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5563                         break;
5564                 if (found_key.offset < ctx->pos)
5565                         goto next;
5566                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5567                         goto next;
5568                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5569                 name_len = btrfs_dir_name_len(leaf, di);
5570                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5571                     PAGE_SIZE) {
5572                         btrfs_release_path(path);
5573                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5574                         if (ret)
5575                                 goto nopos;
5576                         addr = private->filldir_buf;
5577                         entries = 0;
5578                         total_len = 0;
5579                         goto again;
5580                 }
5581
5582                 entry = addr;
5583                 put_unaligned(name_len, &entry->name_len);
5584                 name_ptr = (char *)(entry + 1);
5585                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5586                                    name_len);
5587                 put_unaligned(fs_ftype_to_dtype(btrfs_dir_type(leaf, di)),
5588                                 &entry->type);
5589                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5590                 put_unaligned(location.objectid, &entry->ino);
5591                 put_unaligned(found_key.offset, &entry->offset);
5592                 entries++;
5593                 addr += sizeof(struct dir_entry) + name_len;
5594                 total_len += sizeof(struct dir_entry) + name_len;
5595 next:
5596                 path->slots[0]++;
5597         }
5598         btrfs_release_path(path);
5599
5600         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5601         if (ret)
5602                 goto nopos;
5603
5604         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5605         if (ret)
5606                 goto nopos;
5607
5608         /*
5609          * Stop new entries from being returned after we return the last
5610          * entry.
5611          *
5612          * New directory entries are assigned a strictly increasing
5613          * offset.  This means that new entries created during readdir
5614          * are *guaranteed* to be seen in the future by that readdir.
5615          * This has broken buggy programs which operate on names as
5616          * they're returned by readdir.  Until we re-use freed offsets
5617          * we have this hack to stop new entries from being returned
5618          * under the assumption that they'll never reach this huge
5619          * offset.
5620          *
5621          * This is being careful not to overflow 32bit loff_t unless the
5622          * last entry requires it because doing so has broken 32bit apps
5623          * in the past.
5624          */
5625         if (ctx->pos >= INT_MAX)
5626                 ctx->pos = LLONG_MAX;
5627         else
5628                 ctx->pos = INT_MAX;
5629 nopos:
5630         ret = 0;
5631 err:
5632         if (put)
5633                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5634         btrfs_free_path(path);
5635         return ret;
5636 }
5637
5638 /*
5639  * This is somewhat expensive, updating the tree every time the
5640  * inode changes.  But, it is most likely to find the inode in cache.
5641  * FIXME, needs more benchmarking...there are no reasons other than performance
5642  * to keep or drop this code.
5643  */
5644 static int btrfs_dirty_inode(struct inode *inode)
5645 {
5646         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5647         struct btrfs_root *root = BTRFS_I(inode)->root;
5648         struct btrfs_trans_handle *trans;
5649         int ret;
5650
5651         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5652                 return 0;
5653
5654         trans = btrfs_join_transaction(root);
5655         if (IS_ERR(trans))
5656                 return PTR_ERR(trans);
5657
5658         ret = btrfs_update_inode(trans, root, inode);
5659         if (ret && ret == -ENOSPC) {
5660                 /* whoops, lets try again with the full transaction */
5661                 btrfs_end_transaction(trans);
5662                 trans = btrfs_start_transaction(root, 1);
5663                 if (IS_ERR(trans))
5664                         return PTR_ERR(trans);
5665
5666                 ret = btrfs_update_inode(trans, root, inode);
5667         }
5668         btrfs_end_transaction(trans);
5669         if (BTRFS_I(inode)->delayed_node)
5670                 btrfs_balance_delayed_items(fs_info);
5671
5672         return ret;
5673 }
5674
5675 /*
5676  * This is a copy of file_update_time.  We need this so we can return error on
5677  * ENOSPC for updating the inode in the case of file write and mmap writes.
5678  */
5679 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
5680                              int flags)
5681 {
5682         struct btrfs_root *root = BTRFS_I(inode)->root;
5683         bool dirty = flags & ~S_VERSION;
5684
5685         if (btrfs_root_readonly(root))
5686                 return -EROFS;
5687
5688         if (flags & S_VERSION)
5689                 dirty |= inode_maybe_inc_iversion(inode, dirty);
5690         if (flags & S_CTIME)
5691                 inode->i_ctime = *now;
5692         if (flags & S_MTIME)
5693                 inode->i_mtime = *now;
5694         if (flags & S_ATIME)
5695                 inode->i_atime = *now;
5696         return dirty ? btrfs_dirty_inode(inode) : 0;
5697 }
5698
5699 /*
5700  * find the highest existing sequence number in a directory
5701  * and then set the in-memory index_cnt variable to reflect
5702  * free sequence numbers
5703  */
5704 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
5705 {
5706         struct btrfs_root *root = inode->root;
5707         struct btrfs_key key, found_key;
5708         struct btrfs_path *path;
5709         struct extent_buffer *leaf;
5710         int ret;
5711
5712         key.objectid = btrfs_ino(inode);
5713         key.type = BTRFS_DIR_INDEX_KEY;
5714         key.offset = (u64)-1;
5715
5716         path = btrfs_alloc_path();
5717         if (!path)
5718                 return -ENOMEM;
5719
5720         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5721         if (ret < 0)
5722                 goto out;
5723         /* FIXME: we should be able to handle this */
5724         if (ret == 0)
5725                 goto out;
5726         ret = 0;
5727
5728         /*
5729          * MAGIC NUMBER EXPLANATION:
5730          * since we search a directory based on f_pos we have to start at 2
5731          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
5732          * else has to start at 2
5733          */
5734         if (path->slots[0] == 0) {
5735                 inode->index_cnt = 2;
5736                 goto out;
5737         }
5738
5739         path->slots[0]--;
5740
5741         leaf = path->nodes[0];
5742         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5743
5744         if (found_key.objectid != btrfs_ino(inode) ||
5745             found_key.type != BTRFS_DIR_INDEX_KEY) {
5746                 inode->index_cnt = 2;
5747                 goto out;
5748         }
5749
5750         inode->index_cnt = found_key.offset + 1;
5751 out:
5752         btrfs_free_path(path);
5753         return ret;
5754 }
5755
5756 /*
5757  * helper to find a free sequence number in a given directory.  This current
5758  * code is very simple, later versions will do smarter things in the btree
5759  */
5760 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
5761 {
5762         int ret = 0;
5763
5764         if (dir->index_cnt == (u64)-1) {
5765                 ret = btrfs_inode_delayed_dir_index_count(dir);
5766                 if (ret) {
5767                         ret = btrfs_set_inode_index_count(dir);
5768                         if (ret)
5769                                 return ret;
5770                 }
5771         }
5772
5773         *index = dir->index_cnt;
5774         dir->index_cnt++;
5775
5776         return ret;
5777 }
5778
5779 static int btrfs_insert_inode_locked(struct inode *inode)
5780 {
5781         struct btrfs_iget_args args;
5782         args.location = &BTRFS_I(inode)->location;
5783         args.root = BTRFS_I(inode)->root;
5784
5785         return insert_inode_locked4(inode,
5786                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
5787                    btrfs_find_actor, &args);
5788 }
5789
5790 /*
5791  * Inherit flags from the parent inode.
5792  *
5793  * Currently only the compression flags and the cow flags are inherited.
5794  */
5795 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
5796 {
5797         unsigned int flags;
5798
5799         if (!dir)
5800                 return;
5801
5802         flags = BTRFS_I(dir)->flags;
5803
5804         if (flags & BTRFS_INODE_NOCOMPRESS) {
5805                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
5806                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
5807         } else if (flags & BTRFS_INODE_COMPRESS) {
5808                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
5809                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
5810         }
5811
5812         if (flags & BTRFS_INODE_NODATACOW) {
5813                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
5814                 if (S_ISREG(inode->i_mode))
5815                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
5816         }
5817
5818         btrfs_sync_inode_flags_to_i_flags(inode);
5819 }
5820
5821 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
5822                                      struct btrfs_root *root,
5823                                      struct inode *dir,
5824                                      const char *name, int name_len,
5825                                      u64 ref_objectid, u64 objectid,
5826                                      umode_t mode, u64 *index)
5827 {
5828         struct btrfs_fs_info *fs_info = root->fs_info;
5829         struct inode *inode;
5830         struct btrfs_inode_item *inode_item;
5831         struct btrfs_key *location;
5832         struct btrfs_path *path;
5833         struct btrfs_inode_ref *ref;
5834         struct btrfs_key key[2];
5835         u32 sizes[2];
5836         int nitems = name ? 2 : 1;
5837         unsigned long ptr;
5838         unsigned int nofs_flag;
5839         int ret;
5840
5841         path = btrfs_alloc_path();
5842         if (!path)
5843                 return ERR_PTR(-ENOMEM);
5844
5845         nofs_flag = memalloc_nofs_save();
5846         inode = new_inode(fs_info->sb);
5847         memalloc_nofs_restore(nofs_flag);
5848         if (!inode) {
5849                 btrfs_free_path(path);
5850                 return ERR_PTR(-ENOMEM);
5851         }
5852
5853         /*
5854          * O_TMPFILE, set link count to 0, so that after this point,
5855          * we fill in an inode item with the correct link count.
5856          */
5857         if (!name)
5858                 set_nlink(inode, 0);
5859
5860         /*
5861          * we have to initialize this early, so we can reclaim the inode
5862          * number if we fail afterwards in this function.
5863          */
5864         inode->i_ino = objectid;
5865
5866         if (dir && name) {
5867                 trace_btrfs_inode_request(dir);
5868
5869                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
5870                 if (ret) {
5871                         btrfs_free_path(path);
5872                         iput(inode);
5873                         return ERR_PTR(ret);
5874                 }
5875         } else if (dir) {
5876                 *index = 0;
5877         }
5878         /*
5879          * index_cnt is ignored for everything but a dir,
5880          * btrfs_set_inode_index_count has an explanation for the magic
5881          * number
5882          */
5883         BTRFS_I(inode)->index_cnt = 2;
5884         BTRFS_I(inode)->dir_index = *index;
5885         BTRFS_I(inode)->root = btrfs_grab_root(root);
5886         BTRFS_I(inode)->generation = trans->transid;
5887         inode->i_generation = BTRFS_I(inode)->generation;
5888
5889         /*
5890          * We could have gotten an inode number from somebody who was fsynced
5891          * and then removed in this same transaction, so let's just set full
5892          * sync since it will be a full sync anyway and this will blow away the
5893          * old info in the log.
5894          */
5895         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
5896
5897         key[0].objectid = objectid;
5898         key[0].type = BTRFS_INODE_ITEM_KEY;
5899         key[0].offset = 0;
5900
5901         sizes[0] = sizeof(struct btrfs_inode_item);
5902
5903         if (name) {
5904                 /*
5905                  * Start new inodes with an inode_ref. This is slightly more
5906                  * efficient for small numbers of hard links since they will
5907                  * be packed into one item. Extended refs will kick in if we
5908                  * add more hard links than can fit in the ref item.
5909                  */
5910                 key[1].objectid = objectid;
5911                 key[1].type = BTRFS_INODE_REF_KEY;
5912                 key[1].offset = ref_objectid;
5913
5914                 sizes[1] = name_len + sizeof(*ref);
5915         }
5916
5917         location = &BTRFS_I(inode)->location;
5918         location->objectid = objectid;
5919         location->offset = 0;
5920         location->type = BTRFS_INODE_ITEM_KEY;
5921
5922         ret = btrfs_insert_inode_locked(inode);
5923         if (ret < 0) {
5924                 iput(inode);
5925                 goto fail;
5926         }
5927
5928         path->leave_spinning = 1;
5929         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
5930         if (ret != 0)
5931                 goto fail_unlock;
5932
5933         inode_init_owner(inode, dir, mode);
5934         inode_set_bytes(inode, 0);
5935
5936         inode->i_mtime = current_time(inode);
5937         inode->i_atime = inode->i_mtime;
5938         inode->i_ctime = inode->i_mtime;
5939         BTRFS_I(inode)->i_otime = inode->i_mtime;
5940
5941         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
5942                                   struct btrfs_inode_item);
5943         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
5944                              sizeof(*inode_item));
5945         fill_inode_item(trans, path->nodes[0], inode_item, inode);
5946
5947         if (name) {
5948                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
5949                                      struct btrfs_inode_ref);
5950                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
5951                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
5952                 ptr = (unsigned long)(ref + 1);
5953                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
5954         }
5955
5956         btrfs_mark_buffer_dirty(path->nodes[0]);
5957         btrfs_free_path(path);
5958
5959         btrfs_inherit_iflags(inode, dir);
5960
5961         if (S_ISREG(mode)) {
5962                 if (btrfs_test_opt(fs_info, NODATASUM))
5963                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
5964                 if (btrfs_test_opt(fs_info, NODATACOW))
5965                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
5966                                 BTRFS_INODE_NODATASUM;
5967         }
5968
5969         inode_tree_add(inode);
5970
5971         trace_btrfs_inode_new(inode);
5972         btrfs_set_inode_last_trans(trans, inode);
5973
5974         btrfs_update_root_times(trans, root);
5975
5976         ret = btrfs_inode_inherit_props(trans, inode, dir);
5977         if (ret)
5978                 btrfs_err(fs_info,
5979                           "error inheriting props for ino %llu (root %llu): %d",
5980                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
5981
5982         return inode;
5983
5984 fail_unlock:
5985         discard_new_inode(inode);
5986 fail:
5987         if (dir && name)
5988                 BTRFS_I(dir)->index_cnt--;
5989         btrfs_free_path(path);
5990         return ERR_PTR(ret);
5991 }
5992
5993 /*
5994  * utility function to add 'inode' into 'parent_inode' with
5995  * a give name and a given sequence number.
5996  * if 'add_backref' is true, also insert a backref from the
5997  * inode to the parent directory.
5998  */
5999 int btrfs_add_link(struct btrfs_trans_handle *trans,
6000                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6001                    const char *name, int name_len, int add_backref, u64 index)
6002 {
6003         int ret = 0;
6004         struct btrfs_key key;
6005         struct btrfs_root *root = parent_inode->root;
6006         u64 ino = btrfs_ino(inode);
6007         u64 parent_ino = btrfs_ino(parent_inode);
6008
6009         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6010                 memcpy(&key, &inode->root->root_key, sizeof(key));
6011         } else {
6012                 key.objectid = ino;
6013                 key.type = BTRFS_INODE_ITEM_KEY;
6014                 key.offset = 0;
6015         }
6016
6017         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6018                 ret = btrfs_add_root_ref(trans, key.objectid,
6019                                          root->root_key.objectid, parent_ino,
6020                                          index, name, name_len);
6021         } else if (add_backref) {
6022                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6023                                              parent_ino, index);
6024         }
6025
6026         /* Nothing to clean up yet */
6027         if (ret)
6028                 return ret;
6029
6030         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6031                                     btrfs_inode_type(&inode->vfs_inode), index);
6032         if (ret == -EEXIST || ret == -EOVERFLOW)
6033                 goto fail_dir_item;
6034         else if (ret) {
6035                 btrfs_abort_transaction(trans, ret);
6036                 return ret;
6037         }
6038
6039         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6040                            name_len * 2);
6041         inode_inc_iversion(&parent_inode->vfs_inode);
6042         /*
6043          * If we are replaying a log tree, we do not want to update the mtime
6044          * and ctime of the parent directory with the current time, since the
6045          * log replay procedure is responsible for setting them to their correct
6046          * values (the ones it had when the fsync was done).
6047          */
6048         if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
6049                 struct timespec64 now = current_time(&parent_inode->vfs_inode);
6050
6051                 parent_inode->vfs_inode.i_mtime = now;
6052                 parent_inode->vfs_inode.i_ctime = now;
6053         }
6054         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6055         if (ret)
6056                 btrfs_abort_transaction(trans, ret);
6057         return ret;
6058
6059 fail_dir_item:
6060         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6061                 u64 local_index;
6062                 int err;
6063                 err = btrfs_del_root_ref(trans, key.objectid,
6064                                          root->root_key.objectid, parent_ino,
6065                                          &local_index, name, name_len);
6066                 if (err)
6067                         btrfs_abort_transaction(trans, err);
6068         } else if (add_backref) {
6069                 u64 local_index;
6070                 int err;
6071
6072                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6073                                           ino, parent_ino, &local_index);
6074                 if (err)
6075                         btrfs_abort_transaction(trans, err);
6076         }
6077
6078         /* Return the original error code */
6079         return ret;
6080 }
6081
6082 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6083                             struct btrfs_inode *dir, struct dentry *dentry,
6084                             struct btrfs_inode *inode, int backref, u64 index)
6085 {
6086         int err = btrfs_add_link(trans, dir, inode,
6087                                  dentry->d_name.name, dentry->d_name.len,
6088                                  backref, index);
6089         if (err > 0)
6090                 err = -EEXIST;
6091         return err;
6092 }
6093
6094 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6095                         umode_t mode, dev_t rdev)
6096 {
6097         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6098         struct btrfs_trans_handle *trans;
6099         struct btrfs_root *root = BTRFS_I(dir)->root;
6100         struct inode *inode = NULL;
6101         int err;
6102         u64 objectid;
6103         u64 index = 0;
6104
6105         /*
6106          * 2 for inode item and ref
6107          * 2 for dir items
6108          * 1 for xattr if selinux is on
6109          */
6110         trans = btrfs_start_transaction(root, 5);
6111         if (IS_ERR(trans))
6112                 return PTR_ERR(trans);
6113
6114         err = btrfs_find_free_ino(root, &objectid);
6115         if (err)
6116                 goto out_unlock;
6117
6118         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6119                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6120                         mode, &index);
6121         if (IS_ERR(inode)) {
6122                 err = PTR_ERR(inode);
6123                 inode = NULL;
6124                 goto out_unlock;
6125         }
6126
6127         /*
6128         * If the active LSM wants to access the inode during
6129         * d_instantiate it needs these. Smack checks to see
6130         * if the filesystem supports xattrs by looking at the
6131         * ops vector.
6132         */
6133         inode->i_op = &btrfs_special_inode_operations;
6134         init_special_inode(inode, inode->i_mode, rdev);
6135
6136         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6137         if (err)
6138                 goto out_unlock;
6139
6140         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6141                         0, index);
6142         if (err)
6143                 goto out_unlock;
6144
6145         btrfs_update_inode(trans, root, inode);
6146         d_instantiate_new(dentry, inode);
6147
6148 out_unlock:
6149         btrfs_end_transaction(trans);
6150         btrfs_btree_balance_dirty(fs_info);
6151         if (err && inode) {
6152                 inode_dec_link_count(inode);
6153                 discard_new_inode(inode);
6154         }
6155         return err;
6156 }
6157
6158 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6159                         umode_t mode, bool excl)
6160 {
6161         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6162         struct btrfs_trans_handle *trans;
6163         struct btrfs_root *root = BTRFS_I(dir)->root;
6164         struct inode *inode = NULL;
6165         int err;
6166         u64 objectid;
6167         u64 index = 0;
6168
6169         /*
6170          * 2 for inode item and ref
6171          * 2 for dir items
6172          * 1 for xattr if selinux is on
6173          */
6174         trans = btrfs_start_transaction(root, 5);
6175         if (IS_ERR(trans))
6176                 return PTR_ERR(trans);
6177
6178         err = btrfs_find_free_ino(root, &objectid);
6179         if (err)
6180                 goto out_unlock;
6181
6182         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6183                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6184                         mode, &index);
6185         if (IS_ERR(inode)) {
6186                 err = PTR_ERR(inode);
6187                 inode = NULL;
6188                 goto out_unlock;
6189         }
6190         /*
6191         * If the active LSM wants to access the inode during
6192         * d_instantiate it needs these. Smack checks to see
6193         * if the filesystem supports xattrs by looking at the
6194         * ops vector.
6195         */
6196         inode->i_fop = &btrfs_file_operations;
6197         inode->i_op = &btrfs_file_inode_operations;
6198         inode->i_mapping->a_ops = &btrfs_aops;
6199
6200         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6201         if (err)
6202                 goto out_unlock;
6203
6204         err = btrfs_update_inode(trans, root, inode);
6205         if (err)
6206                 goto out_unlock;
6207
6208         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6209                         0, index);
6210         if (err)
6211                 goto out_unlock;
6212
6213         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6214         d_instantiate_new(dentry, inode);
6215
6216 out_unlock:
6217         btrfs_end_transaction(trans);
6218         if (err && inode) {
6219                 inode_dec_link_count(inode);
6220                 discard_new_inode(inode);
6221         }
6222         btrfs_btree_balance_dirty(fs_info);
6223         return err;
6224 }
6225
6226 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6227                       struct dentry *dentry)
6228 {
6229         struct btrfs_trans_handle *trans = NULL;
6230         struct btrfs_root *root = BTRFS_I(dir)->root;
6231         struct inode *inode = d_inode(old_dentry);
6232         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6233         u64 index;
6234         int err;
6235         int drop_inode = 0;
6236
6237         /* do not allow sys_link's with other subvols of the same device */
6238         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6239                 return -EXDEV;
6240
6241         if (inode->i_nlink >= BTRFS_LINK_MAX)
6242                 return -EMLINK;
6243
6244         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6245         if (err)
6246                 goto fail;
6247
6248         /*
6249          * 2 items for inode and inode ref
6250          * 2 items for dir items
6251          * 1 item for parent inode
6252          * 1 item for orphan item deletion if O_TMPFILE
6253          */
6254         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6255         if (IS_ERR(trans)) {
6256                 err = PTR_ERR(trans);
6257                 trans = NULL;
6258                 goto fail;
6259         }
6260
6261         /* There are several dir indexes for this inode, clear the cache. */
6262         BTRFS_I(inode)->dir_index = 0ULL;
6263         inc_nlink(inode);
6264         inode_inc_iversion(inode);
6265         inode->i_ctime = current_time(inode);
6266         ihold(inode);
6267         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6268
6269         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6270                         1, index);
6271
6272         if (err) {
6273                 drop_inode = 1;
6274         } else {
6275                 struct dentry *parent = dentry->d_parent;
6276                 int ret;
6277
6278                 err = btrfs_update_inode(trans, root, inode);
6279                 if (err)
6280                         goto fail;
6281                 if (inode->i_nlink == 1) {
6282                         /*
6283                          * If new hard link count is 1, it's a file created
6284                          * with open(2) O_TMPFILE flag.
6285                          */
6286                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6287                         if (err)
6288                                 goto fail;
6289                 }
6290                 d_instantiate(dentry, inode);
6291                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6292                                          true, NULL);
6293                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6294                         err = btrfs_commit_transaction(trans);
6295                         trans = NULL;
6296                 }
6297         }
6298
6299 fail:
6300         if (trans)
6301                 btrfs_end_transaction(trans);
6302         if (drop_inode) {
6303                 inode_dec_link_count(inode);
6304                 iput(inode);
6305         }
6306         btrfs_btree_balance_dirty(fs_info);
6307         return err;
6308 }
6309
6310 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6311 {
6312         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6313         struct inode *inode = NULL;
6314         struct btrfs_trans_handle *trans;
6315         struct btrfs_root *root = BTRFS_I(dir)->root;
6316         int err = 0;
6317         u64 objectid = 0;
6318         u64 index = 0;
6319
6320         /*
6321          * 2 items for inode and ref
6322          * 2 items for dir items
6323          * 1 for xattr if selinux is on
6324          */
6325         trans = btrfs_start_transaction(root, 5);
6326         if (IS_ERR(trans))
6327                 return PTR_ERR(trans);
6328
6329         err = btrfs_find_free_ino(root, &objectid);
6330         if (err)
6331                 goto out_fail;
6332
6333         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6334                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6335                         S_IFDIR | mode, &index);
6336         if (IS_ERR(inode)) {
6337                 err = PTR_ERR(inode);
6338                 inode = NULL;
6339                 goto out_fail;
6340         }
6341
6342         /* these must be set before we unlock the inode */
6343         inode->i_op = &btrfs_dir_inode_operations;
6344         inode->i_fop = &btrfs_dir_file_operations;
6345
6346         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6347         if (err)
6348                 goto out_fail;
6349
6350         btrfs_i_size_write(BTRFS_I(inode), 0);
6351         err = btrfs_update_inode(trans, root, inode);
6352         if (err)
6353                 goto out_fail;
6354
6355         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6356                         dentry->d_name.name,
6357                         dentry->d_name.len, 0, index);
6358         if (err)
6359                 goto out_fail;
6360
6361         d_instantiate_new(dentry, inode);
6362
6363 out_fail:
6364         btrfs_end_transaction(trans);
6365         if (err && inode) {
6366                 inode_dec_link_count(inode);
6367                 discard_new_inode(inode);
6368         }
6369         btrfs_btree_balance_dirty(fs_info);
6370         return err;
6371 }
6372
6373 static noinline int uncompress_inline(struct btrfs_path *path,
6374                                       struct page *page,
6375                                       size_t pg_offset, u64 extent_offset,
6376                                       struct btrfs_file_extent_item *item)
6377 {
6378         int ret;
6379         struct extent_buffer *leaf = path->nodes[0];
6380         char *tmp;
6381         size_t max_size;
6382         unsigned long inline_size;
6383         unsigned long ptr;
6384         int compress_type;
6385
6386         WARN_ON(pg_offset != 0);
6387         compress_type = btrfs_file_extent_compression(leaf, item);
6388         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6389         inline_size = btrfs_file_extent_inline_item_len(leaf,
6390                                         btrfs_item_nr(path->slots[0]));
6391         tmp = kmalloc(inline_size, GFP_NOFS);
6392         if (!tmp)
6393                 return -ENOMEM;
6394         ptr = btrfs_file_extent_inline_start(item);
6395
6396         read_extent_buffer(leaf, tmp, ptr, inline_size);
6397
6398         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6399         ret = btrfs_decompress(compress_type, tmp, page,
6400                                extent_offset, inline_size, max_size);
6401
6402         /*
6403          * decompression code contains a memset to fill in any space between the end
6404          * of the uncompressed data and the end of max_size in case the decompressed
6405          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6406          * the end of an inline extent and the beginning of the next block, so we
6407          * cover that region here.
6408          */
6409
6410         if (max_size + pg_offset < PAGE_SIZE) {
6411                 char *map = kmap(page);
6412                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6413                 kunmap(page);
6414         }
6415         kfree(tmp);
6416         return ret;
6417 }
6418
6419 /**
6420  * btrfs_get_extent - Lookup the first extent overlapping a range in a file.
6421  * @inode:      file to search in
6422  * @page:       page to read extent data into if the extent is inline
6423  * @pg_offset:  offset into @page to copy to
6424  * @start:      file offset
6425  * @len:        length of range starting at @start
6426  *
6427  * This returns the first &struct extent_map which overlaps with the given
6428  * range, reading it from the B-tree and caching it if necessary. Note that
6429  * there may be more extents which overlap the given range after the returned
6430  * extent_map.
6431  *
6432  * If @page is not NULL and the extent is inline, this also reads the extent
6433  * data directly into the page and marks the extent up to date in the io_tree.
6434  *
6435  * Return: ERR_PTR on error, non-NULL extent_map on success.
6436  */
6437 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6438                                     struct page *page, size_t pg_offset,
6439                                     u64 start, u64 len)
6440 {
6441         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6442         int ret;
6443         int err = 0;
6444         u64 extent_start = 0;
6445         u64 extent_end = 0;
6446         u64 objectid = btrfs_ino(inode);
6447         int extent_type = -1;
6448         struct btrfs_path *path = NULL;
6449         struct btrfs_root *root = inode->root;
6450         struct btrfs_file_extent_item *item;
6451         struct extent_buffer *leaf;
6452         struct btrfs_key found_key;
6453         struct extent_map *em = NULL;
6454         struct extent_map_tree *em_tree = &inode->extent_tree;
6455         struct extent_io_tree *io_tree = &inode->io_tree;
6456
6457         read_lock(&em_tree->lock);
6458         em = lookup_extent_mapping(em_tree, start, len);
6459         read_unlock(&em_tree->lock);
6460
6461         if (em) {
6462                 if (em->start > start || em->start + em->len <= start)
6463                         free_extent_map(em);
6464                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6465                         free_extent_map(em);
6466                 else
6467                         goto out;
6468         }
6469         em = alloc_extent_map();
6470         if (!em) {
6471                 err = -ENOMEM;
6472                 goto out;
6473         }
6474         em->start = EXTENT_MAP_HOLE;
6475         em->orig_start = EXTENT_MAP_HOLE;
6476         em->len = (u64)-1;
6477         em->block_len = (u64)-1;
6478
6479         path = btrfs_alloc_path();
6480         if (!path) {
6481                 err = -ENOMEM;
6482                 goto out;
6483         }
6484
6485         /* Chances are we'll be called again, so go ahead and do readahead */
6486         path->reada = READA_FORWARD;
6487
6488         /*
6489          * Unless we're going to uncompress the inline extent, no sleep would
6490          * happen.
6491          */
6492         path->leave_spinning = 1;
6493
6494         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6495         if (ret < 0) {
6496                 err = ret;
6497                 goto out;
6498         } else if (ret > 0) {
6499                 if (path->slots[0] == 0)
6500                         goto not_found;
6501                 path->slots[0]--;
6502         }
6503
6504         leaf = path->nodes[0];
6505         item = btrfs_item_ptr(leaf, path->slots[0],
6506                               struct btrfs_file_extent_item);
6507         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6508         if (found_key.objectid != objectid ||
6509             found_key.type != BTRFS_EXTENT_DATA_KEY) {
6510                 /*
6511                  * If we backup past the first extent we want to move forward
6512                  * and see if there is an extent in front of us, otherwise we'll
6513                  * say there is a hole for our whole search range which can
6514                  * cause problems.
6515                  */
6516                 extent_end = start;
6517                 goto next;
6518         }
6519
6520         extent_type = btrfs_file_extent_type(leaf, item);
6521         extent_start = found_key.offset;
6522         extent_end = btrfs_file_extent_end(path);
6523         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6524             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6525                 /* Only regular file could have regular/prealloc extent */
6526                 if (!S_ISREG(inode->vfs_inode.i_mode)) {
6527                         ret = -EUCLEAN;
6528                         btrfs_crit(fs_info,
6529                 "regular/prealloc extent found for non-regular inode %llu",
6530                                    btrfs_ino(inode));
6531                         goto out;
6532                 }
6533                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
6534                                                        extent_start);
6535         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
6536                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
6537                                                       path->slots[0],
6538                                                       extent_start);
6539         }
6540 next:
6541         if (start >= extent_end) {
6542                 path->slots[0]++;
6543                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6544                         ret = btrfs_next_leaf(root, path);
6545                         if (ret < 0) {
6546                                 err = ret;
6547                                 goto out;
6548                         } else if (ret > 0) {
6549                                 goto not_found;
6550                         }
6551                         leaf = path->nodes[0];
6552                 }
6553                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6554                 if (found_key.objectid != objectid ||
6555                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6556                         goto not_found;
6557                 if (start + len <= found_key.offset)
6558                         goto not_found;
6559                 if (start > found_key.offset)
6560                         goto next;
6561
6562                 /* New extent overlaps with existing one */
6563                 em->start = start;
6564                 em->orig_start = start;
6565                 em->len = found_key.offset - start;
6566                 em->block_start = EXTENT_MAP_HOLE;
6567                 goto insert;
6568         }
6569
6570         btrfs_extent_item_to_extent_map(inode, path, item, !page, em);
6571
6572         if (extent_type == BTRFS_FILE_EXTENT_REG ||
6573             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
6574                 goto insert;
6575         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
6576                 unsigned long ptr;
6577                 char *map;
6578                 size_t size;
6579                 size_t extent_offset;
6580                 size_t copy_size;
6581
6582                 if (!page)
6583                         goto out;
6584
6585                 size = btrfs_file_extent_ram_bytes(leaf, item);
6586                 extent_offset = page_offset(page) + pg_offset - extent_start;
6587                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6588                                   size - extent_offset);
6589                 em->start = extent_start + extent_offset;
6590                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6591                 em->orig_block_len = em->len;
6592                 em->orig_start = em->start;
6593                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6594
6595                 btrfs_set_path_blocking(path);
6596                 if (!PageUptodate(page)) {
6597                         if (btrfs_file_extent_compression(leaf, item) !=
6598                             BTRFS_COMPRESS_NONE) {
6599                                 ret = uncompress_inline(path, page, pg_offset,
6600                                                         extent_offset, item);
6601                                 if (ret) {
6602                                         err = ret;
6603                                         goto out;
6604                                 }
6605                         } else {
6606                                 map = kmap(page);
6607                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6608                                                    copy_size);
6609                                 if (pg_offset + copy_size < PAGE_SIZE) {
6610                                         memset(map + pg_offset + copy_size, 0,
6611                                                PAGE_SIZE - pg_offset -
6612                                                copy_size);
6613                                 }
6614                                 kunmap(page);
6615                         }
6616                         flush_dcache_page(page);
6617                 }
6618                 set_extent_uptodate(io_tree, em->start,
6619                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6620                 goto insert;
6621         }
6622 not_found:
6623         em->start = start;
6624         em->orig_start = start;
6625         em->len = len;
6626         em->block_start = EXTENT_MAP_HOLE;
6627 insert:
6628         btrfs_release_path(path);
6629         if (em->start > start || extent_map_end(em) <= start) {
6630                 btrfs_err(fs_info,
6631                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
6632                           em->start, em->len, start, len);
6633                 err = -EIO;
6634                 goto out;
6635         }
6636
6637         err = 0;
6638         write_lock(&em_tree->lock);
6639         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
6640         write_unlock(&em_tree->lock);
6641 out:
6642         btrfs_free_path(path);
6643
6644         trace_btrfs_get_extent(root, inode, em);
6645
6646         if (err) {
6647                 free_extent_map(em);
6648                 return ERR_PTR(err);
6649         }
6650         BUG_ON(!em); /* Error is always set */
6651         return em;
6652 }
6653
6654 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
6655                                            u64 start, u64 len)
6656 {
6657         struct extent_map *em;
6658         struct extent_map *hole_em = NULL;
6659         u64 delalloc_start = start;
6660         u64 end;
6661         u64 delalloc_len;
6662         u64 delalloc_end;
6663         int err = 0;
6664
6665         em = btrfs_get_extent(inode, NULL, 0, start, len);
6666         if (IS_ERR(em))
6667                 return em;
6668         /*
6669          * If our em maps to:
6670          * - a hole or
6671          * - a pre-alloc extent,
6672          * there might actually be delalloc bytes behind it.
6673          */
6674         if (em->block_start != EXTENT_MAP_HOLE &&
6675             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6676                 return em;
6677         else
6678                 hole_em = em;
6679
6680         /* check to see if we've wrapped (len == -1 or similar) */
6681         end = start + len;
6682         if (end < start)
6683                 end = (u64)-1;
6684         else
6685                 end -= 1;
6686
6687         em = NULL;
6688
6689         /* ok, we didn't find anything, lets look for delalloc */
6690         delalloc_len = count_range_bits(&inode->io_tree, &delalloc_start,
6691                                  end, len, EXTENT_DELALLOC, 1);
6692         delalloc_end = delalloc_start + delalloc_len;
6693         if (delalloc_end < delalloc_start)
6694                 delalloc_end = (u64)-1;
6695
6696         /*
6697          * We didn't find anything useful, return the original results from
6698          * get_extent()
6699          */
6700         if (delalloc_start > end || delalloc_end <= start) {
6701                 em = hole_em;
6702                 hole_em = NULL;
6703                 goto out;
6704         }
6705
6706         /*
6707          * Adjust the delalloc_start to make sure it doesn't go backwards from
6708          * the start they passed in
6709          */
6710         delalloc_start = max(start, delalloc_start);
6711         delalloc_len = delalloc_end - delalloc_start;
6712
6713         if (delalloc_len > 0) {
6714                 u64 hole_start;
6715                 u64 hole_len;
6716                 const u64 hole_end = extent_map_end(hole_em);
6717
6718                 em = alloc_extent_map();
6719                 if (!em) {
6720                         err = -ENOMEM;
6721                         goto out;
6722                 }
6723
6724                 ASSERT(hole_em);
6725                 /*
6726                  * When btrfs_get_extent can't find anything it returns one
6727                  * huge hole
6728                  *
6729                  * Make sure what it found really fits our range, and adjust to
6730                  * make sure it is based on the start from the caller
6731                  */
6732                 if (hole_end <= start || hole_em->start > end) {
6733                        free_extent_map(hole_em);
6734                        hole_em = NULL;
6735                 } else {
6736                        hole_start = max(hole_em->start, start);
6737                        hole_len = hole_end - hole_start;
6738                 }
6739
6740                 if (hole_em && delalloc_start > hole_start) {
6741                         /*
6742                          * Our hole starts before our delalloc, so we have to
6743                          * return just the parts of the hole that go until the
6744                          * delalloc starts
6745                          */
6746                         em->len = min(hole_len, delalloc_start - hole_start);
6747                         em->start = hole_start;
6748                         em->orig_start = hole_start;
6749                         /*
6750                          * Don't adjust block start at all, it is fixed at
6751                          * EXTENT_MAP_HOLE
6752                          */
6753                         em->block_start = hole_em->block_start;
6754                         em->block_len = hole_len;
6755                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
6756                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
6757                 } else {
6758                         /*
6759                          * Hole is out of passed range or it starts after
6760                          * delalloc range
6761                          */
6762                         em->start = delalloc_start;
6763                         em->len = delalloc_len;
6764                         em->orig_start = delalloc_start;
6765                         em->block_start = EXTENT_MAP_DELALLOC;
6766                         em->block_len = delalloc_len;
6767                 }
6768         } else {
6769                 return hole_em;
6770         }
6771 out:
6772
6773         free_extent_map(hole_em);
6774         if (err) {
6775                 free_extent_map(em);
6776                 return ERR_PTR(err);
6777         }
6778         return em;
6779 }
6780
6781 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
6782                                                   const u64 start,
6783                                                   const u64 len,
6784                                                   const u64 orig_start,
6785                                                   const u64 block_start,
6786                                                   const u64 block_len,
6787                                                   const u64 orig_block_len,
6788                                                   const u64 ram_bytes,
6789                                                   const int type)
6790 {
6791         struct extent_map *em = NULL;
6792         int ret;
6793
6794         if (type != BTRFS_ORDERED_NOCOW) {
6795                 em = create_io_em(inode, start, len, orig_start,
6796                                   block_start, block_len, orig_block_len,
6797                                   ram_bytes,
6798                                   BTRFS_COMPRESS_NONE, /* compress_type */
6799                                   type);
6800                 if (IS_ERR(em))
6801                         goto out;
6802         }
6803         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
6804                                            len, block_len, type);
6805         if (ret) {
6806                 if (em) {
6807                         free_extent_map(em);
6808                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
6809                                                 start + len - 1, 0);
6810                 }
6811                 em = ERR_PTR(ret);
6812         }
6813  out:
6814
6815         return em;
6816 }
6817
6818 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
6819                                                   u64 start, u64 len)
6820 {
6821         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6822         struct btrfs_root *root = BTRFS_I(inode)->root;
6823         struct extent_map *em;
6824         struct btrfs_key ins;
6825         u64 alloc_hint;
6826         int ret;
6827
6828         alloc_hint = get_extent_allocation_hint(inode, start, len);
6829         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
6830                                    0, alloc_hint, &ins, 1, 1);
6831         if (ret)
6832                 return ERR_PTR(ret);
6833
6834         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
6835                                      ins.objectid, ins.offset, ins.offset,
6836                                      ins.offset, BTRFS_ORDERED_REGULAR);
6837         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
6838         if (IS_ERR(em))
6839                 btrfs_free_reserved_extent(fs_info, ins.objectid,
6840                                            ins.offset, 1);
6841
6842         return em;
6843 }
6844
6845 /*
6846  * returns 1 when the nocow is safe, < 1 on error, 0 if the
6847  * block must be cow'd
6848  */
6849 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
6850                               u64 *orig_start, u64 *orig_block_len,
6851                               u64 *ram_bytes)
6852 {
6853         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6854         struct btrfs_path *path;
6855         int ret;
6856         struct extent_buffer *leaf;
6857         struct btrfs_root *root = BTRFS_I(inode)->root;
6858         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6859         struct btrfs_file_extent_item *fi;
6860         struct btrfs_key key;
6861         u64 disk_bytenr;
6862         u64 backref_offset;
6863         u64 extent_end;
6864         u64 num_bytes;
6865         int slot;
6866         int found_type;
6867         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
6868
6869         path = btrfs_alloc_path();
6870         if (!path)
6871                 return -ENOMEM;
6872
6873         ret = btrfs_lookup_file_extent(NULL, root, path,
6874                         btrfs_ino(BTRFS_I(inode)), offset, 0);
6875         if (ret < 0)
6876                 goto out;
6877
6878         slot = path->slots[0];
6879         if (ret == 1) {
6880                 if (slot == 0) {
6881                         /* can't find the item, must cow */
6882                         ret = 0;
6883                         goto out;
6884                 }
6885                 slot--;
6886         }
6887         ret = 0;
6888         leaf = path->nodes[0];
6889         btrfs_item_key_to_cpu(leaf, &key, slot);
6890         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
6891             key.type != BTRFS_EXTENT_DATA_KEY) {
6892                 /* not our file or wrong item type, must cow */
6893                 goto out;
6894         }
6895
6896         if (key.offset > offset) {
6897                 /* Wrong offset, must cow */
6898                 goto out;
6899         }
6900
6901         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
6902         found_type = btrfs_file_extent_type(leaf, fi);
6903         if (found_type != BTRFS_FILE_EXTENT_REG &&
6904             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
6905                 /* not a regular extent, must cow */
6906                 goto out;
6907         }
6908
6909         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
6910                 goto out;
6911
6912         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
6913         if (extent_end <= offset)
6914                 goto out;
6915
6916         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
6917         if (disk_bytenr == 0)
6918                 goto out;
6919
6920         if (btrfs_file_extent_compression(leaf, fi) ||
6921             btrfs_file_extent_encryption(leaf, fi) ||
6922             btrfs_file_extent_other_encoding(leaf, fi))
6923                 goto out;
6924
6925         /*
6926          * Do the same check as in btrfs_cross_ref_exist but without the
6927          * unnecessary search.
6928          */
6929         if (btrfs_file_extent_generation(leaf, fi) <=
6930             btrfs_root_last_snapshot(&root->root_item))
6931                 goto out;
6932
6933         backref_offset = btrfs_file_extent_offset(leaf, fi);
6934
6935         if (orig_start) {
6936                 *orig_start = key.offset - backref_offset;
6937                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
6938                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
6939         }
6940
6941         if (btrfs_extent_readonly(fs_info, disk_bytenr))
6942                 goto out;
6943
6944         num_bytes = min(offset + *len, extent_end) - offset;
6945         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6946                 u64 range_end;
6947
6948                 range_end = round_up(offset + num_bytes,
6949                                      root->fs_info->sectorsize) - 1;
6950                 ret = test_range_bit(io_tree, offset, range_end,
6951                                      EXTENT_DELALLOC, 0, NULL);
6952                 if (ret) {
6953                         ret = -EAGAIN;
6954                         goto out;
6955                 }
6956         }
6957
6958         btrfs_release_path(path);
6959
6960         /*
6961          * look for other files referencing this extent, if we
6962          * find any we must cow
6963          */
6964
6965         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
6966                                     key.offset - backref_offset, disk_bytenr);
6967         if (ret) {
6968                 ret = 0;
6969                 goto out;
6970         }
6971
6972         /*
6973          * adjust disk_bytenr and num_bytes to cover just the bytes
6974          * in this extent we are about to write.  If there
6975          * are any csums in that range we have to cow in order
6976          * to keep the csums correct
6977          */
6978         disk_bytenr += backref_offset;
6979         disk_bytenr += offset - key.offset;
6980         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
6981                 goto out;
6982         /*
6983          * all of the above have passed, it is safe to overwrite this extent
6984          * without cow
6985          */
6986         *len = num_bytes;
6987         ret = 1;
6988 out:
6989         btrfs_free_path(path);
6990         return ret;
6991 }
6992
6993 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
6994                               struct extent_state **cached_state, int writing)
6995 {
6996         struct btrfs_ordered_extent *ordered;
6997         int ret = 0;
6998
6999         while (1) {
7000                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7001                                  cached_state);
7002                 /*
7003                  * We're concerned with the entire range that we're going to be
7004                  * doing DIO to, so we need to make sure there's no ordered
7005                  * extents in this range.
7006                  */
7007                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7008                                                      lockend - lockstart + 1);
7009
7010                 /*
7011                  * We need to make sure there are no buffered pages in this
7012                  * range either, we could have raced between the invalidate in
7013                  * generic_file_direct_write and locking the extent.  The
7014                  * invalidate needs to happen so that reads after a write do not
7015                  * get stale data.
7016                  */
7017                 if (!ordered &&
7018                     (!writing || !filemap_range_has_page(inode->i_mapping,
7019                                                          lockstart, lockend)))
7020                         break;
7021
7022                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7023                                      cached_state);
7024
7025                 if (ordered) {
7026                         /*
7027                          * If we are doing a DIO read and the ordered extent we
7028                          * found is for a buffered write, we can not wait for it
7029                          * to complete and retry, because if we do so we can
7030                          * deadlock with concurrent buffered writes on page
7031                          * locks. This happens only if our DIO read covers more
7032                          * than one extent map, if at this point has already
7033                          * created an ordered extent for a previous extent map
7034                          * and locked its range in the inode's io tree, and a
7035                          * concurrent write against that previous extent map's
7036                          * range and this range started (we unlock the ranges
7037                          * in the io tree only when the bios complete and
7038                          * buffered writes always lock pages before attempting
7039                          * to lock range in the io tree).
7040                          */
7041                         if (writing ||
7042                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7043                                 btrfs_start_ordered_extent(inode, ordered, 1);
7044                         else
7045                                 ret = -ENOTBLK;
7046                         btrfs_put_ordered_extent(ordered);
7047                 } else {
7048                         /*
7049                          * We could trigger writeback for this range (and wait
7050                          * for it to complete) and then invalidate the pages for
7051                          * this range (through invalidate_inode_pages2_range()),
7052                          * but that can lead us to a deadlock with a concurrent
7053                          * call to readpages() (a buffered read or a defrag call
7054                          * triggered a readahead) on a page lock due to an
7055                          * ordered dio extent we created before but did not have
7056                          * yet a corresponding bio submitted (whence it can not
7057                          * complete), which makes readpages() wait for that
7058                          * ordered extent to complete while holding a lock on
7059                          * that page.
7060                          */
7061                         ret = -ENOTBLK;
7062                 }
7063
7064                 if (ret)
7065                         break;
7066
7067                 cond_resched();
7068         }
7069
7070         return ret;
7071 }
7072
7073 /* The callers of this must take lock_extent() */
7074 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7075                                        u64 orig_start, u64 block_start,
7076                                        u64 block_len, u64 orig_block_len,
7077                                        u64 ram_bytes, int compress_type,
7078                                        int type)
7079 {
7080         struct extent_map_tree *em_tree;
7081         struct extent_map *em;
7082         int ret;
7083
7084         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7085                type == BTRFS_ORDERED_COMPRESSED ||
7086                type == BTRFS_ORDERED_NOCOW ||
7087                type == BTRFS_ORDERED_REGULAR);
7088
7089         em_tree = &BTRFS_I(inode)->extent_tree;
7090         em = alloc_extent_map();
7091         if (!em)
7092                 return ERR_PTR(-ENOMEM);
7093
7094         em->start = start;
7095         em->orig_start = orig_start;
7096         em->len = len;
7097         em->block_len = block_len;
7098         em->block_start = block_start;
7099         em->orig_block_len = orig_block_len;
7100         em->ram_bytes = ram_bytes;
7101         em->generation = -1;
7102         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7103         if (type == BTRFS_ORDERED_PREALLOC) {
7104                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7105         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7106                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7107                 em->compress_type = compress_type;
7108         }
7109
7110         do {
7111                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7112                                 em->start + em->len - 1, 0);
7113                 write_lock(&em_tree->lock);
7114                 ret = add_extent_mapping(em_tree, em, 1);
7115                 write_unlock(&em_tree->lock);
7116                 /*
7117                  * The caller has taken lock_extent(), who could race with us
7118                  * to add em?
7119                  */
7120         } while (ret == -EEXIST);
7121
7122         if (ret) {
7123                 free_extent_map(em);
7124                 return ERR_PTR(ret);
7125         }
7126
7127         /* em got 2 refs now, callers needs to do free_extent_map once. */
7128         return em;
7129 }
7130
7131
7132 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7133                                         struct buffer_head *bh_result,
7134                                         struct inode *inode,
7135                                         u64 start, u64 len)
7136 {
7137         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7138
7139         if (em->block_start == EXTENT_MAP_HOLE ||
7140                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7141                 return -ENOENT;
7142
7143         len = min(len, em->len - (start - em->start));
7144
7145         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7146                 inode->i_blkbits;
7147         bh_result->b_size = len;
7148         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7149         set_buffer_mapped(bh_result);
7150
7151         return 0;
7152 }
7153
7154 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7155                                          struct buffer_head *bh_result,
7156                                          struct inode *inode,
7157                                          struct btrfs_dio_data *dio_data,
7158                                          u64 start, u64 len)
7159 {
7160         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7161         struct extent_map *em = *map;
7162         int ret = 0;
7163
7164         /*
7165          * We don't allocate a new extent in the following cases
7166          *
7167          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7168          * existing extent.
7169          * 2) The extent is marked as PREALLOC. We're good to go here and can
7170          * just use the extent.
7171          *
7172          */
7173         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7174             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7175              em->block_start != EXTENT_MAP_HOLE)) {
7176                 int type;
7177                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7178
7179                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7180                         type = BTRFS_ORDERED_PREALLOC;
7181                 else
7182                         type = BTRFS_ORDERED_NOCOW;
7183                 len = min(len, em->len - (start - em->start));
7184                 block_start = em->block_start + (start - em->start);
7185
7186                 if (can_nocow_extent(inode, start, &len, &orig_start,
7187                                      &orig_block_len, &ram_bytes) == 1 &&
7188                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7189                         struct extent_map *em2;
7190
7191                         em2 = btrfs_create_dio_extent(inode, start, len,
7192                                                       orig_start, block_start,
7193                                                       len, orig_block_len,
7194                                                       ram_bytes, type);
7195                         btrfs_dec_nocow_writers(fs_info, block_start);
7196                         if (type == BTRFS_ORDERED_PREALLOC) {
7197                                 free_extent_map(em);
7198                                 *map = em = em2;
7199                         }
7200
7201                         if (em2 && IS_ERR(em2)) {
7202                                 ret = PTR_ERR(em2);
7203                                 goto out;
7204                         }
7205                         /*
7206                          * For inode marked NODATACOW or extent marked PREALLOC,
7207                          * use the existing or preallocated extent, so does not
7208                          * need to adjust btrfs_space_info's bytes_may_use.
7209                          */
7210                         btrfs_free_reserved_data_space_noquota(inode, start,
7211                                                                len);
7212                         goto skip_cow;
7213                 }
7214         }
7215
7216         /* this will cow the extent */
7217         len = bh_result->b_size;
7218         free_extent_map(em);
7219         *map = em = btrfs_new_extent_direct(inode, start, len);
7220         if (IS_ERR(em)) {
7221                 ret = PTR_ERR(em);
7222                 goto out;
7223         }
7224
7225         len = min(len, em->len - (start - em->start));
7226
7227 skip_cow:
7228         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7229                 inode->i_blkbits;
7230         bh_result->b_size = len;
7231         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7232         set_buffer_mapped(bh_result);
7233
7234         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7235                 set_buffer_new(bh_result);
7236
7237         /*
7238          * Need to update the i_size under the extent lock so buffered
7239          * readers will get the updated i_size when we unlock.
7240          */
7241         if (!dio_data->overwrite && start + len > i_size_read(inode))
7242                 i_size_write(inode, start + len);
7243
7244         WARN_ON(dio_data->reserve < len);
7245         dio_data->reserve -= len;
7246         dio_data->unsubmitted_oe_range_end = start + len;
7247         current->journal_info = dio_data;
7248 out:
7249         return ret;
7250 }
7251
7252 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7253                                    struct buffer_head *bh_result, int create)
7254 {
7255         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7256         struct extent_map *em;
7257         struct extent_state *cached_state = NULL;
7258         struct btrfs_dio_data *dio_data = NULL;
7259         u64 start = iblock << inode->i_blkbits;
7260         u64 lockstart, lockend;
7261         u64 len = bh_result->b_size;
7262         int ret = 0;
7263
7264         if (!create)
7265                 len = min_t(u64, len, fs_info->sectorsize);
7266
7267         lockstart = start;
7268         lockend = start + len - 1;
7269
7270         if (current->journal_info) {
7271                 /*
7272                  * Need to pull our outstanding extents and set journal_info to NULL so
7273                  * that anything that needs to check if there's a transaction doesn't get
7274                  * confused.
7275                  */
7276                 dio_data = current->journal_info;
7277                 current->journal_info = NULL;
7278         }
7279
7280         /*
7281          * If this errors out it's because we couldn't invalidate pagecache for
7282          * this range and we need to fallback to buffered.
7283          */
7284         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7285                                create)) {
7286                 ret = -ENOTBLK;
7287                 goto err;
7288         }
7289
7290         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len);
7291         if (IS_ERR(em)) {
7292                 ret = PTR_ERR(em);
7293                 goto unlock_err;
7294         }
7295
7296         /*
7297          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7298          * io.  INLINE is special, and we could probably kludge it in here, but
7299          * it's still buffered so for safety lets just fall back to the generic
7300          * buffered path.
7301          *
7302          * For COMPRESSED we _have_ to read the entire extent in so we can
7303          * decompress it, so there will be buffering required no matter what we
7304          * do, so go ahead and fallback to buffered.
7305          *
7306          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7307          * to buffered IO.  Don't blame me, this is the price we pay for using
7308          * the generic code.
7309          */
7310         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7311             em->block_start == EXTENT_MAP_INLINE) {
7312                 free_extent_map(em);
7313                 ret = -ENOTBLK;
7314                 goto unlock_err;
7315         }
7316
7317         if (create) {
7318                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7319                                                     dio_data, start, len);
7320                 if (ret < 0)
7321                         goto unlock_err;
7322
7323                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
7324                                      lockend, &cached_state);
7325         } else {
7326                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7327                                                    start, len);
7328                 /* Can be negative only if we read from a hole */
7329                 if (ret < 0) {
7330                         ret = 0;
7331                         free_extent_map(em);
7332                         goto unlock_err;
7333                 }
7334                 /*
7335                  * We need to unlock only the end area that we aren't using.
7336                  * The rest is going to be unlocked by the endio routine.
7337                  */
7338                 lockstart = start + bh_result->b_size;
7339                 if (lockstart < lockend) {
7340                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
7341                                              lockstart, lockend, &cached_state);
7342                 } else {
7343                         free_extent_state(cached_state);
7344                 }
7345         }
7346
7347         free_extent_map(em);
7348
7349         return 0;
7350
7351 unlock_err:
7352         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7353                              &cached_state);
7354 err:
7355         if (dio_data)
7356                 current->journal_info = dio_data;
7357         return ret;
7358 }
7359
7360 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7361                                                  struct bio *bio,
7362                                                  int mirror_num)
7363 {
7364         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7365         blk_status_t ret;
7366
7367         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7368
7369         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7370         if (ret)
7371                 return ret;
7372
7373         ret = btrfs_map_bio(fs_info, bio, mirror_num);
7374
7375         return ret;
7376 }
7377
7378 static int btrfs_check_dio_repairable(struct inode *inode,
7379                                       struct bio *failed_bio,
7380                                       struct io_failure_record *failrec,
7381                                       int failed_mirror)
7382 {
7383         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7384         int num_copies;
7385
7386         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7387         if (num_copies == 1) {
7388                 /*
7389                  * we only have a single copy of the data, so don't bother with
7390                  * all the retry and error correction code that follows. no
7391                  * matter what the error is, it is very likely to persist.
7392                  */
7393                 btrfs_debug(fs_info,
7394                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7395                         num_copies, failrec->this_mirror, failed_mirror);
7396                 return 0;
7397         }
7398
7399         failrec->failed_mirror = failed_mirror;
7400         failrec->this_mirror++;
7401         if (failrec->this_mirror == failed_mirror)
7402                 failrec->this_mirror++;
7403
7404         if (failrec->this_mirror > num_copies) {
7405                 btrfs_debug(fs_info,
7406                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7407                         num_copies, failrec->this_mirror, failed_mirror);
7408                 return 0;
7409         }
7410
7411         return 1;
7412 }
7413
7414 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7415                                    struct page *page, unsigned int pgoff,
7416                                    u64 start, u64 end, int failed_mirror,
7417                                    bio_end_io_t *repair_endio, void *repair_arg)
7418 {
7419         struct io_failure_record *failrec;
7420         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7421         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7422         struct bio *bio;
7423         int isector;
7424         unsigned int read_mode = 0;
7425         int segs;
7426         int ret;
7427         blk_status_t status;
7428         struct bio_vec bvec;
7429
7430         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7431
7432         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7433         if (ret)
7434                 return errno_to_blk_status(ret);
7435
7436         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7437                                          failed_mirror);
7438         if (!ret) {
7439                 free_io_failure(failure_tree, io_tree, failrec);
7440                 return BLK_STS_IOERR;
7441         }
7442
7443         segs = bio_segments(failed_bio);
7444         bio_get_first_bvec(failed_bio, &bvec);
7445         if (segs > 1 ||
7446             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7447                 read_mode |= REQ_FAILFAST_DEV;
7448
7449         isector = start - btrfs_io_bio(failed_bio)->logical;
7450         isector >>= inode->i_sb->s_blocksize_bits;
7451         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7452                                 pgoff, isector, repair_endio, repair_arg);
7453         bio->bi_opf = REQ_OP_READ | read_mode;
7454
7455         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7456                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7457                     read_mode, failrec->this_mirror, failrec->in_validation);
7458
7459         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7460         if (status) {
7461                 free_io_failure(failure_tree, io_tree, failrec);
7462                 bio_put(bio);
7463         }
7464
7465         return status;
7466 }
7467
7468 struct btrfs_retry_complete {
7469         struct completion done;
7470         struct inode *inode;
7471         u64 start;
7472         int uptodate;
7473 };
7474
7475 static void btrfs_retry_endio_nocsum(struct bio *bio)
7476 {
7477         struct btrfs_retry_complete *done = bio->bi_private;
7478         struct inode *inode = done->inode;
7479         struct bio_vec *bvec;
7480         struct extent_io_tree *io_tree, *failure_tree;
7481         struct bvec_iter_all iter_all;
7482
7483         if (bio->bi_status)
7484                 goto end;
7485
7486         ASSERT(bio->bi_vcnt == 1);
7487         io_tree = &BTRFS_I(inode)->io_tree;
7488         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7489         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7490
7491         done->uptodate = 1;
7492         ASSERT(!bio_flagged(bio, BIO_CLONED));
7493         bio_for_each_segment_all(bvec, bio, iter_all)
7494                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7495                                  io_tree, done->start, bvec->bv_page,
7496                                  btrfs_ino(BTRFS_I(inode)), 0);
7497 end:
7498         complete(&done->done);
7499         bio_put(bio);
7500 }
7501
7502 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7503                                                 struct btrfs_io_bio *io_bio)
7504 {
7505         struct btrfs_fs_info *fs_info;
7506         struct bio_vec bvec;
7507         struct bvec_iter iter;
7508         struct btrfs_retry_complete done;
7509         u64 start;
7510         unsigned int pgoff;
7511         u32 sectorsize;
7512         int nr_sectors;
7513         blk_status_t ret;
7514         blk_status_t err = BLK_STS_OK;
7515
7516         fs_info = BTRFS_I(inode)->root->fs_info;
7517         sectorsize = fs_info->sectorsize;
7518
7519         start = io_bio->logical;
7520         done.inode = inode;
7521         io_bio->bio.bi_iter = io_bio->iter;
7522
7523         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7524                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7525                 pgoff = bvec.bv_offset;
7526
7527 next_block_or_try_again:
7528                 done.uptodate = 0;
7529                 done.start = start;
7530                 init_completion(&done.done);
7531
7532                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7533                                 pgoff, start, start + sectorsize - 1,
7534                                 io_bio->mirror_num,
7535                                 btrfs_retry_endio_nocsum, &done);
7536                 if (ret) {
7537                         err = ret;
7538                         goto next;
7539                 }
7540
7541                 wait_for_completion_io(&done.done);
7542
7543                 if (!done.uptodate) {
7544                         /* We might have another mirror, so try again */
7545                         goto next_block_or_try_again;
7546                 }
7547
7548 next:
7549                 start += sectorsize;
7550
7551                 nr_sectors--;
7552                 if (nr_sectors) {
7553                         pgoff += sectorsize;
7554                         ASSERT(pgoff < PAGE_SIZE);
7555                         goto next_block_or_try_again;
7556                 }
7557         }
7558
7559         return err;
7560 }
7561
7562 static void btrfs_retry_endio(struct bio *bio)
7563 {
7564         struct btrfs_retry_complete *done = bio->bi_private;
7565         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7566         struct extent_io_tree *io_tree, *failure_tree;
7567         struct inode *inode = done->inode;
7568         struct bio_vec *bvec;
7569         int uptodate;
7570         int ret;
7571         int i = 0;
7572         struct bvec_iter_all iter_all;
7573
7574         if (bio->bi_status)
7575                 goto end;
7576
7577         uptodate = 1;
7578
7579         ASSERT(bio->bi_vcnt == 1);
7580         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
7581
7582         io_tree = &BTRFS_I(inode)->io_tree;
7583         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7584
7585         ASSERT(!bio_flagged(bio, BIO_CLONED));
7586         bio_for_each_segment_all(bvec, bio, iter_all) {
7587                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7588                                              bvec->bv_offset, done->start,
7589                                              bvec->bv_len);
7590                 if (!ret)
7591                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
7592                                          failure_tree, io_tree, done->start,
7593                                          bvec->bv_page,
7594                                          btrfs_ino(BTRFS_I(inode)),
7595                                          bvec->bv_offset);
7596                 else
7597                         uptodate = 0;
7598                 i++;
7599         }
7600
7601         done->uptodate = uptodate;
7602 end:
7603         complete(&done->done);
7604         bio_put(bio);
7605 }
7606
7607 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
7608                 struct btrfs_io_bio *io_bio, blk_status_t err)
7609 {
7610         struct btrfs_fs_info *fs_info;
7611         struct bio_vec bvec;
7612         struct bvec_iter iter;
7613         struct btrfs_retry_complete done;
7614         u64 start;
7615         u64 offset = 0;
7616         u32 sectorsize;
7617         int nr_sectors;
7618         unsigned int pgoff;
7619         int csum_pos;
7620         bool uptodate = (err == 0);
7621         int ret;
7622         blk_status_t status;
7623
7624         fs_info = BTRFS_I(inode)->root->fs_info;
7625         sectorsize = fs_info->sectorsize;
7626
7627         err = BLK_STS_OK;
7628         start = io_bio->logical;
7629         done.inode = inode;
7630         io_bio->bio.bi_iter = io_bio->iter;
7631
7632         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7633                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7634
7635                 pgoff = bvec.bv_offset;
7636 next_block:
7637                 if (uptodate) {
7638                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
7639                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
7640                                         bvec.bv_page, pgoff, start, sectorsize);
7641                         if (likely(!ret))
7642                                 goto next;
7643                 }
7644 try_again:
7645                 done.uptodate = 0;
7646                 done.start = start;
7647                 init_completion(&done.done);
7648
7649                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7650                                         pgoff, start, start + sectorsize - 1,
7651                                         io_bio->mirror_num, btrfs_retry_endio,
7652                                         &done);
7653                 if (status) {
7654                         err = status;
7655                         goto next;
7656                 }
7657
7658                 wait_for_completion_io(&done.done);
7659
7660                 if (!done.uptodate) {
7661                         /* We might have another mirror, so try again */
7662                         goto try_again;
7663                 }
7664 next:
7665                 offset += sectorsize;
7666                 start += sectorsize;
7667
7668                 ASSERT(nr_sectors);
7669
7670                 nr_sectors--;
7671                 if (nr_sectors) {
7672                         pgoff += sectorsize;
7673                         ASSERT(pgoff < PAGE_SIZE);
7674                         goto next_block;
7675                 }
7676         }
7677
7678         return err;
7679 }
7680
7681 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
7682                 struct btrfs_io_bio *io_bio, blk_status_t err)
7683 {
7684         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7685
7686         if (skip_csum) {
7687                 if (unlikely(err))
7688                         return __btrfs_correct_data_nocsum(inode, io_bio);
7689                 else
7690                         return BLK_STS_OK;
7691         } else {
7692                 return __btrfs_subio_endio_read(inode, io_bio, err);
7693         }
7694 }
7695
7696 static void btrfs_endio_direct_read(struct bio *bio)
7697 {
7698         struct btrfs_dio_private *dip = bio->bi_private;
7699         struct inode *inode = dip->inode;
7700         struct bio *dio_bio;
7701         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7702         blk_status_t err = bio->bi_status;
7703
7704         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
7705                 err = btrfs_subio_endio_read(inode, io_bio, err);
7706
7707         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
7708                       dip->logical_offset + dip->bytes - 1);
7709         dio_bio = dip->dio_bio;
7710
7711         kfree(dip);
7712
7713         dio_bio->bi_status = err;
7714         dio_end_io(dio_bio);
7715         btrfs_io_bio_free_csum(io_bio);
7716         bio_put(bio);
7717 }
7718
7719 static void __endio_write_update_ordered(struct inode *inode,
7720                                          const u64 offset, const u64 bytes,
7721                                          const bool uptodate)
7722 {
7723         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7724         struct btrfs_ordered_extent *ordered = NULL;
7725         struct btrfs_workqueue *wq;
7726         u64 ordered_offset = offset;
7727         u64 ordered_bytes = bytes;
7728         u64 last_offset;
7729
7730         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
7731                 wq = fs_info->endio_freespace_worker;
7732         else
7733                 wq = fs_info->endio_write_workers;
7734
7735         while (ordered_offset < offset + bytes) {
7736                 last_offset = ordered_offset;
7737                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
7738                                                            &ordered_offset,
7739                                                            ordered_bytes,
7740                                                            uptodate)) {
7741                         btrfs_init_work(&ordered->work, finish_ordered_fn, NULL,
7742                                         NULL);
7743                         btrfs_queue_work(wq, &ordered->work);
7744                 }
7745                 /*
7746                  * If btrfs_dec_test_ordered_pending does not find any ordered
7747                  * extent in the range, we can exit.
7748                  */
7749                 if (ordered_offset == last_offset)
7750                         return;
7751                 /*
7752                  * Our bio might span multiple ordered extents. In this case
7753                  * we keep going until we have accounted the whole dio.
7754                  */
7755                 if (ordered_offset < offset + bytes) {
7756                         ordered_bytes = offset + bytes - ordered_offset;
7757                         ordered = NULL;
7758                 }
7759         }
7760 }
7761
7762 static void btrfs_endio_direct_write(struct bio *bio)
7763 {
7764         struct btrfs_dio_private *dip = bio->bi_private;
7765         struct bio *dio_bio = dip->dio_bio;
7766
7767         __endio_write_update_ordered(dip->inode, dip->logical_offset,
7768                                      dip->bytes, !bio->bi_status);
7769
7770         kfree(dip);
7771
7772         dio_bio->bi_status = bio->bi_status;
7773         dio_end_io(dio_bio);
7774         bio_put(bio);
7775 }
7776
7777 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
7778                                     struct bio *bio, u64 offset)
7779 {
7780         struct inode *inode = private_data;
7781         blk_status_t ret;
7782         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
7783         BUG_ON(ret); /* -ENOMEM */
7784         return 0;
7785 }
7786
7787 static void btrfs_end_dio_bio(struct bio *bio)
7788 {
7789         struct btrfs_dio_private *dip = bio->bi_private;
7790         blk_status_t err = bio->bi_status;
7791
7792         if (err)
7793                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
7794                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
7795                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
7796                            bio->bi_opf,
7797                            (unsigned long long)bio->bi_iter.bi_sector,
7798                            bio->bi_iter.bi_size, err);
7799
7800         if (dip->subio_endio)
7801                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
7802
7803         if (err) {
7804                 /*
7805                  * We want to perceive the errors flag being set before
7806                  * decrementing the reference count. We don't need a barrier
7807                  * since atomic operations with a return value are fully
7808                  * ordered as per atomic_t.txt
7809                  */
7810                 dip->errors = 1;
7811         }
7812
7813         /* if there are more bios still pending for this dio, just exit */
7814         if (!atomic_dec_and_test(&dip->pending_bios))
7815                 goto out;
7816
7817         if (dip->errors) {
7818                 bio_io_error(dip->orig_bio);
7819         } else {
7820                 dip->dio_bio->bi_status = BLK_STS_OK;
7821                 bio_endio(dip->orig_bio);
7822         }
7823 out:
7824         bio_put(bio);
7825 }
7826
7827 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
7828                                                  struct btrfs_dio_private *dip,
7829                                                  struct bio *bio,
7830                                                  u64 file_offset)
7831 {
7832         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7833         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
7834         u16 csum_size;
7835         blk_status_t ret;
7836
7837         /*
7838          * We load all the csum data we need when we submit
7839          * the first bio to reduce the csum tree search and
7840          * contention.
7841          */
7842         if (dip->logical_offset == file_offset) {
7843                 ret = btrfs_lookup_bio_sums(inode, dip->orig_bio, file_offset,
7844                                             NULL);
7845                 if (ret)
7846                         return ret;
7847         }
7848
7849         if (bio == dip->orig_bio)
7850                 return 0;
7851
7852         file_offset -= dip->logical_offset;
7853         file_offset >>= inode->i_sb->s_blocksize_bits;
7854         csum_size = btrfs_super_csum_size(btrfs_sb(inode->i_sb)->super_copy);
7855         io_bio->csum = orig_io_bio->csum + csum_size * file_offset;
7856
7857         return 0;
7858 }
7859
7860 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
7861                 struct inode *inode, u64 file_offset, int async_submit)
7862 {
7863         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7864         struct btrfs_dio_private *dip = bio->bi_private;
7865         bool write = bio_op(bio) == REQ_OP_WRITE;
7866         blk_status_t ret;
7867
7868         /* Check btrfs_submit_bio_hook() for rules about async submit. */
7869         if (async_submit)
7870                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
7871
7872         if (!write) {
7873                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
7874                 if (ret)
7875                         goto err;
7876         }
7877
7878         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
7879                 goto map;
7880
7881         if (write && async_submit) {
7882                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
7883                                           file_offset, inode,
7884                                           btrfs_submit_bio_start_direct_io);
7885                 goto err;
7886         } else if (write) {
7887                 /*
7888                  * If we aren't doing async submit, calculate the csum of the
7889                  * bio now.
7890                  */
7891                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
7892                 if (ret)
7893                         goto err;
7894         } else {
7895                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
7896                                                      file_offset);
7897                 if (ret)
7898                         goto err;
7899         }
7900 map:
7901         ret = btrfs_map_bio(fs_info, bio, 0);
7902 err:
7903         return ret;
7904 }
7905
7906 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
7907 {
7908         struct inode *inode = dip->inode;
7909         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7910         struct bio *bio;
7911         struct bio *orig_bio = dip->orig_bio;
7912         u64 start_sector = orig_bio->bi_iter.bi_sector;
7913         u64 file_offset = dip->logical_offset;
7914         int async_submit = 0;
7915         u64 submit_len;
7916         int clone_offset = 0;
7917         int clone_len;
7918         int ret;
7919         blk_status_t status;
7920         struct btrfs_io_geometry geom;
7921
7922         submit_len = orig_bio->bi_iter.bi_size;
7923         ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
7924                                     start_sector << 9, submit_len, &geom);
7925         if (ret)
7926                 return -EIO;
7927
7928         if (geom.len >= submit_len) {
7929                 bio = orig_bio;
7930                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
7931                 goto submit;
7932         }
7933
7934         /* async crcs make it difficult to collect full stripe writes. */
7935         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
7936                 async_submit = 0;
7937         else
7938                 async_submit = 1;
7939
7940         /* bio split */
7941         ASSERT(geom.len <= INT_MAX);
7942         do {
7943                 clone_len = min_t(int, submit_len, geom.len);
7944
7945                 /*
7946                  * This will never fail as it's passing GPF_NOFS and
7947                  * the allocation is backed by btrfs_bioset.
7948                  */
7949                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
7950                                               clone_len);
7951                 bio->bi_private = dip;
7952                 bio->bi_end_io = btrfs_end_dio_bio;
7953                 btrfs_io_bio(bio)->logical = file_offset;
7954
7955                 ASSERT(submit_len >= clone_len);
7956                 submit_len -= clone_len;
7957                 if (submit_len == 0)
7958                         break;
7959
7960                 /*
7961                  * Increase the count before we submit the bio so we know
7962                  * the end IO handler won't happen before we increase the
7963                  * count. Otherwise, the dip might get freed before we're
7964                  * done setting it up.
7965                  */
7966                 atomic_inc(&dip->pending_bios);
7967
7968                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
7969                                                 async_submit);
7970                 if (status) {
7971                         bio_put(bio);
7972                         atomic_dec(&dip->pending_bios);
7973                         goto out_err;
7974                 }
7975
7976                 clone_offset += clone_len;
7977                 start_sector += clone_len >> 9;
7978                 file_offset += clone_len;
7979
7980                 ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
7981                                       start_sector << 9, submit_len, &geom);
7982                 if (ret)
7983                         goto out_err;
7984         } while (submit_len > 0);
7985
7986 submit:
7987         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
7988         if (!status)
7989                 return 0;
7990
7991         if (bio != orig_bio)
7992                 bio_put(bio);
7993 out_err:
7994         dip->errors = 1;
7995         /*
7996          * Before atomic variable goto zero, we must  make sure dip->errors is
7997          * perceived to be set. This ordering is ensured by the fact that an
7998          * atomic operations with a return value are fully ordered as per
7999          * atomic_t.txt
8000          */
8001         if (atomic_dec_and_test(&dip->pending_bios))
8002                 bio_io_error(dip->orig_bio);
8003
8004         /* bio_end_io() will handle error, so we needn't return it */
8005         return 0;
8006 }
8007
8008 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8009                                 loff_t file_offset)
8010 {
8011         struct btrfs_dio_private *dip = NULL;
8012         struct bio *bio = NULL;
8013         struct btrfs_io_bio *io_bio;
8014         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8015         int ret = 0;
8016
8017         bio = btrfs_bio_clone(dio_bio);
8018
8019         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8020         if (!dip) {
8021                 ret = -ENOMEM;
8022                 goto free_ordered;
8023         }
8024
8025         dip->private = dio_bio->bi_private;
8026         dip->inode = inode;
8027         dip->logical_offset = file_offset;
8028         dip->bytes = dio_bio->bi_iter.bi_size;
8029         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8030         bio->bi_private = dip;
8031         dip->orig_bio = bio;
8032         dip->dio_bio = dio_bio;
8033         atomic_set(&dip->pending_bios, 1);
8034         io_bio = btrfs_io_bio(bio);
8035         io_bio->logical = file_offset;
8036
8037         if (write) {
8038                 bio->bi_end_io = btrfs_endio_direct_write;
8039         } else {
8040                 bio->bi_end_io = btrfs_endio_direct_read;
8041                 dip->subio_endio = btrfs_subio_endio_read;
8042         }
8043
8044         /*
8045          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8046          * even if we fail to submit a bio, because in such case we do the
8047          * corresponding error handling below and it must not be done a second
8048          * time by btrfs_direct_IO().
8049          */
8050         if (write) {
8051                 struct btrfs_dio_data *dio_data = current->journal_info;
8052
8053                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8054                         dip->bytes;
8055                 dio_data->unsubmitted_oe_range_start =
8056                         dio_data->unsubmitted_oe_range_end;
8057         }
8058
8059         ret = btrfs_submit_direct_hook(dip);
8060         if (!ret)
8061                 return;
8062
8063         btrfs_io_bio_free_csum(io_bio);
8064
8065 free_ordered:
8066         /*
8067          * If we arrived here it means either we failed to submit the dip
8068          * or we either failed to clone the dio_bio or failed to allocate the
8069          * dip. If we cloned the dio_bio and allocated the dip, we can just
8070          * call bio_endio against our io_bio so that we get proper resource
8071          * cleanup if we fail to submit the dip, otherwise, we must do the
8072          * same as btrfs_endio_direct_[write|read] because we can't call these
8073          * callbacks - they require an allocated dip and a clone of dio_bio.
8074          */
8075         if (bio && dip) {
8076                 bio_io_error(bio);
8077                 /*
8078                  * The end io callbacks free our dip, do the final put on bio
8079                  * and all the cleanup and final put for dio_bio (through
8080                  * dio_end_io()).
8081                  */
8082                 dip = NULL;
8083                 bio = NULL;
8084         } else {
8085                 if (write)
8086                         __endio_write_update_ordered(inode,
8087                                                 file_offset,
8088                                                 dio_bio->bi_iter.bi_size,
8089                                                 false);
8090                 else
8091                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8092                               file_offset + dio_bio->bi_iter.bi_size - 1);
8093
8094                 dio_bio->bi_status = BLK_STS_IOERR;
8095                 /*
8096                  * Releases and cleans up our dio_bio, no need to bio_put()
8097                  * nor bio_endio()/bio_io_error() against dio_bio.
8098                  */
8099                 dio_end_io(dio_bio);
8100         }
8101         if (bio)
8102                 bio_put(bio);
8103         kfree(dip);
8104 }
8105
8106 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8107                                const struct iov_iter *iter, loff_t offset)
8108 {
8109         int seg;
8110         int i;
8111         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8112         ssize_t retval = -EINVAL;
8113
8114         if (offset & blocksize_mask)
8115                 goto out;
8116
8117         if (iov_iter_alignment(iter) & blocksize_mask)
8118                 goto out;
8119
8120         /* If this is a write we don't need to check anymore */
8121         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8122                 return 0;
8123         /*
8124          * Check to make sure we don't have duplicate iov_base's in this
8125          * iovec, if so return EINVAL, otherwise we'll get csum errors
8126          * when reading back.
8127          */
8128         for (seg = 0; seg < iter->nr_segs; seg++) {
8129                 for (i = seg + 1; i < iter->nr_segs; i++) {
8130                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8131                                 goto out;
8132                 }
8133         }
8134         retval = 0;
8135 out:
8136         return retval;
8137 }
8138
8139 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8140 {
8141         struct file *file = iocb->ki_filp;
8142         struct inode *inode = file->f_mapping->host;
8143         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8144         struct btrfs_dio_data dio_data = { 0 };
8145         struct extent_changeset *data_reserved = NULL;
8146         loff_t offset = iocb->ki_pos;
8147         size_t count = 0;
8148         int flags = 0;
8149         bool wakeup = true;
8150         bool relock = false;
8151         ssize_t ret;
8152
8153         if (check_direct_IO(fs_info, iter, offset))
8154                 return 0;
8155
8156         inode_dio_begin(inode);
8157
8158         /*
8159          * The generic stuff only does filemap_write_and_wait_range, which
8160          * isn't enough if we've written compressed pages to this area, so
8161          * we need to flush the dirty pages again to make absolutely sure
8162          * that any outstanding dirty pages are on disk.
8163          */
8164         count = iov_iter_count(iter);
8165         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8166                      &BTRFS_I(inode)->runtime_flags))
8167                 filemap_fdatawrite_range(inode->i_mapping, offset,
8168                                          offset + count - 1);
8169
8170         if (iov_iter_rw(iter) == WRITE) {
8171                 /*
8172                  * If the write DIO is beyond the EOF, we need update
8173                  * the isize, but it is protected by i_mutex. So we can
8174                  * not unlock the i_mutex at this case.
8175                  */
8176                 if (offset + count <= inode->i_size) {
8177                         dio_data.overwrite = 1;
8178                         inode_unlock(inode);
8179                         relock = true;
8180                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8181                         ret = -EAGAIN;
8182                         goto out;
8183                 }
8184                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8185                                                    offset, count);
8186                 if (ret)
8187                         goto out;
8188
8189                 /*
8190                  * We need to know how many extents we reserved so that we can
8191                  * do the accounting properly if we go over the number we
8192                  * originally calculated.  Abuse current->journal_info for this.
8193                  */
8194                 dio_data.reserve = round_up(count,
8195                                             fs_info->sectorsize);
8196                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8197                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8198                 current->journal_info = &dio_data;
8199                 down_read(&BTRFS_I(inode)->dio_sem);
8200         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8201                                      &BTRFS_I(inode)->runtime_flags)) {
8202                 inode_dio_end(inode);
8203                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8204                 wakeup = false;
8205         }
8206
8207         ret = __blockdev_direct_IO(iocb, inode,
8208                                    fs_info->fs_devices->latest_bdev,
8209                                    iter, btrfs_get_blocks_direct, NULL,
8210                                    btrfs_submit_direct, flags);
8211         if (iov_iter_rw(iter) == WRITE) {
8212                 up_read(&BTRFS_I(inode)->dio_sem);
8213                 current->journal_info = NULL;
8214                 if (ret < 0 && ret != -EIOCBQUEUED) {
8215                         if (dio_data.reserve)
8216                                 btrfs_delalloc_release_space(inode, data_reserved,
8217                                         offset, dio_data.reserve, true);
8218                         /*
8219                          * On error we might have left some ordered extents
8220                          * without submitting corresponding bios for them, so
8221                          * cleanup them up to avoid other tasks getting them
8222                          * and waiting for them to complete forever.
8223                          */
8224                         if (dio_data.unsubmitted_oe_range_start <
8225                             dio_data.unsubmitted_oe_range_end)
8226                                 __endio_write_update_ordered(inode,
8227                                         dio_data.unsubmitted_oe_range_start,
8228                                         dio_data.unsubmitted_oe_range_end -
8229                                         dio_data.unsubmitted_oe_range_start,
8230                                         false);
8231                 } else if (ret >= 0 && (size_t)ret < count)
8232                         btrfs_delalloc_release_space(inode, data_reserved,
8233                                         offset, count - (size_t)ret, true);
8234                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8235         }
8236 out:
8237         if (wakeup)
8238                 inode_dio_end(inode);
8239         if (relock)
8240                 inode_lock(inode);
8241
8242         extent_changeset_free(data_reserved);
8243         return ret;
8244 }
8245
8246 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8247
8248 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8249                 __u64 start, __u64 len)
8250 {
8251         int     ret;
8252
8253         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8254         if (ret)
8255                 return ret;
8256
8257         return extent_fiemap(inode, fieinfo, start, len);
8258 }
8259
8260 int btrfs_readpage(struct file *file, struct page *page)
8261 {
8262         return extent_read_full_page(page, btrfs_get_extent, 0);
8263 }
8264
8265 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8266 {
8267         struct inode *inode = page->mapping->host;
8268         int ret;
8269
8270         if (current->flags & PF_MEMALLOC) {
8271                 redirty_page_for_writepage(wbc, page);
8272                 unlock_page(page);
8273                 return 0;
8274         }
8275
8276         /*
8277          * If we are under memory pressure we will call this directly from the
8278          * VM, we need to make sure we have the inode referenced for the ordered
8279          * extent.  If not just return like we didn't do anything.
8280          */
8281         if (!igrab(inode)) {
8282                 redirty_page_for_writepage(wbc, page);
8283                 return AOP_WRITEPAGE_ACTIVATE;
8284         }
8285         ret = extent_write_full_page(page, wbc);
8286         btrfs_add_delayed_iput(inode);
8287         return ret;
8288 }
8289
8290 static int btrfs_writepages(struct address_space *mapping,
8291                             struct writeback_control *wbc)
8292 {
8293         return extent_writepages(mapping, wbc);
8294 }
8295
8296 static int
8297 btrfs_readpages(struct file *file, struct address_space *mapping,
8298                 struct list_head *pages, unsigned nr_pages)
8299 {
8300         return extent_readpages(mapping, pages, nr_pages);
8301 }
8302
8303 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8304 {
8305         int ret = try_release_extent_mapping(page, gfp_flags);
8306         if (ret == 1) {
8307                 ClearPagePrivate(page);
8308                 set_page_private(page, 0);
8309                 put_page(page);
8310         }
8311         return ret;
8312 }
8313
8314 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8315 {
8316         if (PageWriteback(page) || PageDirty(page))
8317                 return 0;
8318         return __btrfs_releasepage(page, gfp_flags);
8319 }
8320
8321 #ifdef CONFIG_MIGRATION
8322 static int btrfs_migratepage(struct address_space *mapping,
8323                              struct page *newpage, struct page *page,
8324                              enum migrate_mode mode)
8325 {
8326         int ret;
8327
8328         ret = migrate_page_move_mapping(mapping, newpage, page, 0);
8329         if (ret != MIGRATEPAGE_SUCCESS)
8330                 return ret;
8331
8332         if (page_has_private(page)) {
8333                 ClearPagePrivate(page);
8334                 get_page(newpage);
8335                 set_page_private(newpage, page_private(page));
8336                 set_page_private(page, 0);
8337                 put_page(page);
8338                 SetPagePrivate(newpage);
8339         }
8340
8341         if (PagePrivate2(page)) {
8342                 ClearPagePrivate2(page);
8343                 SetPagePrivate2(newpage);
8344         }
8345
8346         if (mode != MIGRATE_SYNC_NO_COPY)
8347                 migrate_page_copy(newpage, page);
8348         else
8349                 migrate_page_states(newpage, page);
8350         return MIGRATEPAGE_SUCCESS;
8351 }
8352 #endif
8353
8354 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8355                                  unsigned int length)
8356 {
8357         struct inode *inode = page->mapping->host;
8358         struct extent_io_tree *tree;
8359         struct btrfs_ordered_extent *ordered;
8360         struct extent_state *cached_state = NULL;
8361         u64 page_start = page_offset(page);
8362         u64 page_end = page_start + PAGE_SIZE - 1;
8363         u64 start;
8364         u64 end;
8365         int inode_evicting = inode->i_state & I_FREEING;
8366
8367         /*
8368          * we have the page locked, so new writeback can't start,
8369          * and the dirty bit won't be cleared while we are here.
8370          *
8371          * Wait for IO on this page so that we can safely clear
8372          * the PagePrivate2 bit and do ordered accounting
8373          */
8374         wait_on_page_writeback(page);
8375
8376         tree = &BTRFS_I(inode)->io_tree;
8377         if (offset) {
8378                 btrfs_releasepage(page, GFP_NOFS);
8379                 return;
8380         }
8381
8382         if (!inode_evicting)
8383                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8384 again:
8385         start = page_start;
8386         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8387                                         page_end - start + 1);
8388         if (ordered) {
8389                 end = min(page_end,
8390                           ordered->file_offset + ordered->num_bytes - 1);
8391                 /*
8392                  * IO on this page will never be started, so we need
8393                  * to account for any ordered extents now
8394                  */
8395                 if (!inode_evicting)
8396                         clear_extent_bit(tree, start, end,
8397                                          EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8398                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8399                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8400                 /*
8401                  * whoever cleared the private bit is responsible
8402                  * for the finish_ordered_io
8403                  */
8404                 if (TestClearPagePrivate2(page)) {
8405                         struct btrfs_ordered_inode_tree *tree;
8406                         u64 new_len;
8407
8408                         tree = &BTRFS_I(inode)->ordered_tree;
8409
8410                         spin_lock_irq(&tree->lock);
8411                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8412                         new_len = start - ordered->file_offset;
8413                         if (new_len < ordered->truncated_len)
8414                                 ordered->truncated_len = new_len;
8415                         spin_unlock_irq(&tree->lock);
8416
8417                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8418                                                            start,
8419                                                            end - start + 1, 1))
8420                                 btrfs_finish_ordered_io(ordered);
8421                 }
8422                 btrfs_put_ordered_extent(ordered);
8423                 if (!inode_evicting) {
8424                         cached_state = NULL;
8425                         lock_extent_bits(tree, start, end,
8426                                          &cached_state);
8427                 }
8428
8429                 start = end + 1;
8430                 if (start < page_end)
8431                         goto again;
8432         }
8433
8434         /*
8435          * Qgroup reserved space handler
8436          * Page here will be either
8437          * 1) Already written to disk
8438          *    In this case, its reserved space is released from data rsv map
8439          *    and will be freed by delayed_ref handler finally.
8440          *    So even we call qgroup_free_data(), it won't decrease reserved
8441          *    space.
8442          * 2) Not written to disk
8443          *    This means the reserved space should be freed here. However,
8444          *    if a truncate invalidates the page (by clearing PageDirty)
8445          *    and the page is accounted for while allocating extent
8446          *    in btrfs_check_data_free_space() we let delayed_ref to
8447          *    free the entire extent.
8448          */
8449         if (PageDirty(page))
8450                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8451         if (!inode_evicting) {
8452                 clear_extent_bit(tree, page_start, page_end, EXTENT_LOCKED |
8453                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8454                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8455                                  &cached_state);
8456
8457                 __btrfs_releasepage(page, GFP_NOFS);
8458         }
8459
8460         ClearPageChecked(page);
8461         if (PagePrivate(page)) {
8462                 ClearPagePrivate(page);
8463                 set_page_private(page, 0);
8464                 put_page(page);
8465         }
8466 }
8467
8468 /*
8469  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8470  * called from a page fault handler when a page is first dirtied. Hence we must
8471  * be careful to check for EOF conditions here. We set the page up correctly
8472  * for a written page which means we get ENOSPC checking when writing into
8473  * holes and correct delalloc and unwritten extent mapping on filesystems that
8474  * support these features.
8475  *
8476  * We are not allowed to take the i_mutex here so we have to play games to
8477  * protect against truncate races as the page could now be beyond EOF.  Because
8478  * truncate_setsize() writes the inode size before removing pages, once we have
8479  * the page lock we can determine safely if the page is beyond EOF. If it is not
8480  * beyond EOF, then the page is guaranteed safe against truncation until we
8481  * unlock the page.
8482  */
8483 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8484 {
8485         struct page *page = vmf->page;
8486         struct inode *inode = file_inode(vmf->vma->vm_file);
8487         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8488         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8489         struct btrfs_ordered_extent *ordered;
8490         struct extent_state *cached_state = NULL;
8491         struct extent_changeset *data_reserved = NULL;
8492         char *kaddr;
8493         unsigned long zero_start;
8494         loff_t size;
8495         vm_fault_t ret;
8496         int ret2;
8497         int reserved = 0;
8498         u64 reserved_space;
8499         u64 page_start;
8500         u64 page_end;
8501         u64 end;
8502
8503         reserved_space = PAGE_SIZE;
8504
8505         sb_start_pagefault(inode->i_sb);
8506         page_start = page_offset(page);
8507         page_end = page_start + PAGE_SIZE - 1;
8508         end = page_end;
8509
8510         /*
8511          * Reserving delalloc space after obtaining the page lock can lead to
8512          * deadlock. For example, if a dirty page is locked by this function
8513          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8514          * dirty page write out, then the btrfs_writepage() function could
8515          * end up waiting indefinitely to get a lock on the page currently
8516          * being processed by btrfs_page_mkwrite() function.
8517          */
8518         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8519                                            reserved_space);
8520         if (!ret2) {
8521                 ret2 = file_update_time(vmf->vma->vm_file);
8522                 reserved = 1;
8523         }
8524         if (ret2) {
8525                 ret = vmf_error(ret2);
8526                 if (reserved)
8527                         goto out;
8528                 goto out_noreserve;
8529         }
8530
8531         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8532 again:
8533         lock_page(page);
8534         size = i_size_read(inode);
8535
8536         if ((page->mapping != inode->i_mapping) ||
8537             (page_start >= size)) {
8538                 /* page got truncated out from underneath us */
8539                 goto out_unlock;
8540         }
8541         wait_on_page_writeback(page);
8542
8543         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8544         set_page_extent_mapped(page);
8545
8546         /*
8547          * we can't set the delalloc bits if there are pending ordered
8548          * extents.  Drop our locks and wait for them to finish
8549          */
8550         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8551                         PAGE_SIZE);
8552         if (ordered) {
8553                 unlock_extent_cached(io_tree, page_start, page_end,
8554                                      &cached_state);
8555                 unlock_page(page);
8556                 btrfs_start_ordered_extent(inode, ordered, 1);
8557                 btrfs_put_ordered_extent(ordered);
8558                 goto again;
8559         }
8560
8561         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8562                 reserved_space = round_up(size - page_start,
8563                                           fs_info->sectorsize);
8564                 if (reserved_space < PAGE_SIZE) {
8565                         end = page_start + reserved_space - 1;
8566                         btrfs_delalloc_release_space(inode, data_reserved,
8567                                         page_start, PAGE_SIZE - reserved_space,
8568                                         true);
8569                 }
8570         }
8571
8572         /*
8573          * page_mkwrite gets called when the page is firstly dirtied after it's
8574          * faulted in, but write(2) could also dirty a page and set delalloc
8575          * bits, thus in this case for space account reason, we still need to
8576          * clear any delalloc bits within this page range since we have to
8577          * reserve data&meta space before lock_page() (see above comments).
8578          */
8579         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8580                           EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8581                           EXTENT_DEFRAG, 0, 0, &cached_state);
8582
8583         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
8584                                         &cached_state);
8585         if (ret2) {
8586                 unlock_extent_cached(io_tree, page_start, page_end,
8587                                      &cached_state);
8588                 ret = VM_FAULT_SIGBUS;
8589                 goto out_unlock;
8590         }
8591
8592         /* page is wholly or partially inside EOF */
8593         if (page_start + PAGE_SIZE > size)
8594                 zero_start = offset_in_page(size);
8595         else
8596                 zero_start = PAGE_SIZE;
8597
8598         if (zero_start != PAGE_SIZE) {
8599                 kaddr = kmap(page);
8600                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
8601                 flush_dcache_page(page);
8602                 kunmap(page);
8603         }
8604         ClearPageChecked(page);
8605         set_page_dirty(page);
8606         SetPageUptodate(page);
8607
8608         BTRFS_I(inode)->last_trans = fs_info->generation;
8609         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8610         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8611
8612         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
8613
8614         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
8615         sb_end_pagefault(inode->i_sb);
8616         extent_changeset_free(data_reserved);
8617         return VM_FAULT_LOCKED;
8618
8619 out_unlock:
8620         unlock_page(page);
8621 out:
8622         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
8623         btrfs_delalloc_release_space(inode, data_reserved, page_start,
8624                                      reserved_space, (ret != 0));
8625 out_noreserve:
8626         sb_end_pagefault(inode->i_sb);
8627         extent_changeset_free(data_reserved);
8628         return ret;
8629 }
8630
8631 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
8632 {
8633         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8634         struct btrfs_root *root = BTRFS_I(inode)->root;
8635         struct btrfs_block_rsv *rsv;
8636         int ret;
8637         struct btrfs_trans_handle *trans;
8638         u64 mask = fs_info->sectorsize - 1;
8639         u64 min_size = btrfs_calc_metadata_size(fs_info, 1);
8640
8641         if (!skip_writeback) {
8642                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8643                                                (u64)-1);
8644                 if (ret)
8645                         return ret;
8646         }
8647
8648         /*
8649          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
8650          * things going on here:
8651          *
8652          * 1) We need to reserve space to update our inode.
8653          *
8654          * 2) We need to have something to cache all the space that is going to
8655          * be free'd up by the truncate operation, but also have some slack
8656          * space reserved in case it uses space during the truncate (thank you
8657          * very much snapshotting).
8658          *
8659          * And we need these to be separate.  The fact is we can use a lot of
8660          * space doing the truncate, and we have no earthly idea how much space
8661          * we will use, so we need the truncate reservation to be separate so it
8662          * doesn't end up using space reserved for updating the inode.  We also
8663          * need to be able to stop the transaction and start a new one, which
8664          * means we need to be able to update the inode several times, and we
8665          * have no idea of knowing how many times that will be, so we can't just
8666          * reserve 1 item for the entirety of the operation, so that has to be
8667          * done separately as well.
8668          *
8669          * So that leaves us with
8670          *
8671          * 1) rsv - for the truncate reservation, which we will steal from the
8672          * transaction reservation.
8673          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
8674          * updating the inode.
8675          */
8676         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
8677         if (!rsv)
8678                 return -ENOMEM;
8679         rsv->size = min_size;
8680         rsv->failfast = 1;
8681
8682         /*
8683          * 1 for the truncate slack space
8684          * 1 for updating the inode.
8685          */
8686         trans = btrfs_start_transaction(root, 2);
8687         if (IS_ERR(trans)) {
8688                 ret = PTR_ERR(trans);
8689                 goto out;
8690         }
8691
8692         /* Migrate the slack space for the truncate to our reserve */
8693         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
8694                                       min_size, false);
8695         BUG_ON(ret);
8696
8697         /*
8698          * So if we truncate and then write and fsync we normally would just
8699          * write the extents that changed, which is a problem if we need to
8700          * first truncate that entire inode.  So set this flag so we write out
8701          * all of the extents in the inode to the sync log so we're completely
8702          * safe.
8703          */
8704         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8705         trans->block_rsv = rsv;
8706
8707         while (1) {
8708                 ret = btrfs_truncate_inode_items(trans, root, inode,
8709                                                  inode->i_size,
8710                                                  BTRFS_EXTENT_DATA_KEY);
8711                 trans->block_rsv = &fs_info->trans_block_rsv;
8712                 if (ret != -ENOSPC && ret != -EAGAIN)
8713                         break;
8714
8715                 ret = btrfs_update_inode(trans, root, inode);
8716                 if (ret)
8717                         break;
8718
8719                 btrfs_end_transaction(trans);
8720                 btrfs_btree_balance_dirty(fs_info);
8721
8722                 trans = btrfs_start_transaction(root, 2);
8723                 if (IS_ERR(trans)) {
8724                         ret = PTR_ERR(trans);
8725                         trans = NULL;
8726                         break;
8727                 }
8728
8729                 btrfs_block_rsv_release(fs_info, rsv, -1, NULL);
8730                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
8731                                               rsv, min_size, false);
8732                 BUG_ON(ret);    /* shouldn't happen */
8733                 trans->block_rsv = rsv;
8734         }
8735
8736         /*
8737          * We can't call btrfs_truncate_block inside a trans handle as we could
8738          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
8739          * we've truncated everything except the last little bit, and can do
8740          * btrfs_truncate_block and then update the disk_i_size.
8741          */
8742         if (ret == NEED_TRUNCATE_BLOCK) {
8743                 btrfs_end_transaction(trans);
8744                 btrfs_btree_balance_dirty(fs_info);
8745
8746                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
8747                 if (ret)
8748                         goto out;
8749                 trans = btrfs_start_transaction(root, 1);
8750                 if (IS_ERR(trans)) {
8751                         ret = PTR_ERR(trans);
8752                         goto out;
8753                 }
8754                 btrfs_inode_safe_disk_i_size_write(inode, 0);
8755         }
8756
8757         if (trans) {
8758                 int ret2;
8759
8760                 trans->block_rsv = &fs_info->trans_block_rsv;
8761                 ret2 = btrfs_update_inode(trans, root, inode);
8762                 if (ret2 && !ret)
8763                         ret = ret2;
8764
8765                 ret2 = btrfs_end_transaction(trans);
8766                 if (ret2 && !ret)
8767                         ret = ret2;
8768                 btrfs_btree_balance_dirty(fs_info);
8769         }
8770 out:
8771         btrfs_free_block_rsv(fs_info, rsv);
8772
8773         return ret;
8774 }
8775
8776 /*
8777  * create a new subvolume directory/inode (helper for the ioctl).
8778  */
8779 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
8780                              struct btrfs_root *new_root,
8781                              struct btrfs_root *parent_root,
8782                              u64 new_dirid)
8783 {
8784         struct inode *inode;
8785         int err;
8786         u64 index = 0;
8787
8788         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
8789                                 new_dirid, new_dirid,
8790                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
8791                                 &index);
8792         if (IS_ERR(inode))
8793                 return PTR_ERR(inode);
8794         inode->i_op = &btrfs_dir_inode_operations;
8795         inode->i_fop = &btrfs_dir_file_operations;
8796
8797         set_nlink(inode, 1);
8798         btrfs_i_size_write(BTRFS_I(inode), 0);
8799         unlock_new_inode(inode);
8800
8801         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
8802         if (err)
8803                 btrfs_err(new_root->fs_info,
8804                           "error inheriting subvolume %llu properties: %d",
8805                           new_root->root_key.objectid, err);
8806
8807         err = btrfs_update_inode(trans, new_root, inode);
8808
8809         iput(inode);
8810         return err;
8811 }
8812
8813 struct inode *btrfs_alloc_inode(struct super_block *sb)
8814 {
8815         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
8816         struct btrfs_inode *ei;
8817         struct inode *inode;
8818
8819         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
8820         if (!ei)
8821                 return NULL;
8822
8823         ei->root = NULL;
8824         ei->generation = 0;
8825         ei->last_trans = 0;
8826         ei->last_sub_trans = 0;
8827         ei->logged_trans = 0;
8828         ei->delalloc_bytes = 0;
8829         ei->new_delalloc_bytes = 0;
8830         ei->defrag_bytes = 0;
8831         ei->disk_i_size = 0;
8832         ei->flags = 0;
8833         ei->csum_bytes = 0;
8834         ei->index_cnt = (u64)-1;
8835         ei->dir_index = 0;
8836         ei->last_unlink_trans = 0;
8837         ei->last_log_commit = 0;
8838
8839         spin_lock_init(&ei->lock);
8840         ei->outstanding_extents = 0;
8841         if (sb->s_magic != BTRFS_TEST_MAGIC)
8842                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
8843                                               BTRFS_BLOCK_RSV_DELALLOC);
8844         ei->runtime_flags = 0;
8845         ei->prop_compress = BTRFS_COMPRESS_NONE;
8846         ei->defrag_compress = BTRFS_COMPRESS_NONE;
8847
8848         ei->delayed_node = NULL;
8849
8850         ei->i_otime.tv_sec = 0;
8851         ei->i_otime.tv_nsec = 0;
8852
8853         inode = &ei->vfs_inode;
8854         extent_map_tree_init(&ei->extent_tree);
8855         extent_io_tree_init(fs_info, &ei->io_tree, IO_TREE_INODE_IO, inode);
8856         extent_io_tree_init(fs_info, &ei->io_failure_tree,
8857                             IO_TREE_INODE_IO_FAILURE, inode);
8858         extent_io_tree_init(fs_info, &ei->file_extent_tree,
8859                             IO_TREE_INODE_FILE_EXTENT, inode);
8860         ei->io_tree.track_uptodate = true;
8861         ei->io_failure_tree.track_uptodate = true;
8862         atomic_set(&ei->sync_writers, 0);
8863         mutex_init(&ei->log_mutex);
8864         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
8865         INIT_LIST_HEAD(&ei->delalloc_inodes);
8866         INIT_LIST_HEAD(&ei->delayed_iput);
8867         RB_CLEAR_NODE(&ei->rb_node);
8868         init_rwsem(&ei->dio_sem);
8869
8870         return inode;
8871 }
8872
8873 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8874 void btrfs_test_destroy_inode(struct inode *inode)
8875 {
8876         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
8877         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
8878 }
8879 #endif
8880
8881 void btrfs_free_inode(struct inode *inode)
8882 {
8883         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
8884 }
8885
8886 void btrfs_destroy_inode(struct inode *inode)
8887 {
8888         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8889         struct btrfs_ordered_extent *ordered;
8890         struct btrfs_root *root = BTRFS_I(inode)->root;
8891
8892         WARN_ON(!hlist_empty(&inode->i_dentry));
8893         WARN_ON(inode->i_data.nrpages);
8894         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
8895         WARN_ON(BTRFS_I(inode)->block_rsv.size);
8896         WARN_ON(BTRFS_I(inode)->outstanding_extents);
8897         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
8898         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
8899         WARN_ON(BTRFS_I(inode)->csum_bytes);
8900         WARN_ON(BTRFS_I(inode)->defrag_bytes);
8901
8902         /*
8903          * This can happen where we create an inode, but somebody else also
8904          * created the same inode and we need to destroy the one we already
8905          * created.
8906          */
8907         if (!root)
8908                 return;
8909
8910         while (1) {
8911                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
8912                 if (!ordered)
8913                         break;
8914                 else {
8915                         btrfs_err(fs_info,
8916                                   "found ordered extent %llu %llu on inode cleanup",
8917                                   ordered->file_offset, ordered->num_bytes);
8918                         btrfs_remove_ordered_extent(inode, ordered);
8919                         btrfs_put_ordered_extent(ordered);
8920                         btrfs_put_ordered_extent(ordered);
8921                 }
8922         }
8923         btrfs_qgroup_check_reserved_leak(inode);
8924         inode_tree_del(inode);
8925         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
8926         btrfs_inode_clear_file_extent_range(BTRFS_I(inode), 0, (u64)-1);
8927         btrfs_put_root(BTRFS_I(inode)->root);
8928 }
8929
8930 int btrfs_drop_inode(struct inode *inode)
8931 {
8932         struct btrfs_root *root = BTRFS_I(inode)->root;
8933
8934         if (root == NULL)
8935                 return 1;
8936
8937         /* the snap/subvol tree is on deleting */
8938         if (btrfs_root_refs(&root->root_item) == 0)
8939                 return 1;
8940         else
8941                 return generic_drop_inode(inode);
8942 }
8943
8944 static void init_once(void *foo)
8945 {
8946         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
8947
8948         inode_init_once(&ei->vfs_inode);
8949 }
8950
8951 void __cold btrfs_destroy_cachep(void)
8952 {
8953         /*
8954          * Make sure all delayed rcu free inodes are flushed before we
8955          * destroy cache.
8956          */
8957         rcu_barrier();
8958         kmem_cache_destroy(btrfs_inode_cachep);
8959         kmem_cache_destroy(btrfs_trans_handle_cachep);
8960         kmem_cache_destroy(btrfs_path_cachep);
8961         kmem_cache_destroy(btrfs_free_space_cachep);
8962         kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
8963 }
8964
8965 int __init btrfs_init_cachep(void)
8966 {
8967         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
8968                         sizeof(struct btrfs_inode), 0,
8969                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
8970                         init_once);
8971         if (!btrfs_inode_cachep)
8972                 goto fail;
8973
8974         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
8975                         sizeof(struct btrfs_trans_handle), 0,
8976                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
8977         if (!btrfs_trans_handle_cachep)
8978                 goto fail;
8979
8980         btrfs_path_cachep = kmem_cache_create("btrfs_path",
8981                         sizeof(struct btrfs_path), 0,
8982                         SLAB_MEM_SPREAD, NULL);
8983         if (!btrfs_path_cachep)
8984                 goto fail;
8985
8986         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
8987                         sizeof(struct btrfs_free_space), 0,
8988                         SLAB_MEM_SPREAD, NULL);
8989         if (!btrfs_free_space_cachep)
8990                 goto fail;
8991
8992         btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
8993                                                         PAGE_SIZE, PAGE_SIZE,
8994                                                         SLAB_RED_ZONE, NULL);
8995         if (!btrfs_free_space_bitmap_cachep)
8996                 goto fail;
8997
8998         return 0;
8999 fail:
9000         btrfs_destroy_cachep();
9001         return -ENOMEM;
9002 }
9003
9004 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9005                          u32 request_mask, unsigned int flags)
9006 {
9007         u64 delalloc_bytes;
9008         struct inode *inode = d_inode(path->dentry);
9009         u32 blocksize = inode->i_sb->s_blocksize;
9010         u32 bi_flags = BTRFS_I(inode)->flags;
9011
9012         stat->result_mask |= STATX_BTIME;
9013         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9014         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9015         if (bi_flags & BTRFS_INODE_APPEND)
9016                 stat->attributes |= STATX_ATTR_APPEND;
9017         if (bi_flags & BTRFS_INODE_COMPRESS)
9018                 stat->attributes |= STATX_ATTR_COMPRESSED;
9019         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9020                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9021         if (bi_flags & BTRFS_INODE_NODUMP)
9022                 stat->attributes |= STATX_ATTR_NODUMP;
9023
9024         stat->attributes_mask |= (STATX_ATTR_APPEND |
9025                                   STATX_ATTR_COMPRESSED |
9026                                   STATX_ATTR_IMMUTABLE |
9027                                   STATX_ATTR_NODUMP);
9028
9029         generic_fillattr(inode, stat);
9030         stat->dev = BTRFS_I(inode)->root->anon_dev;
9031
9032         spin_lock(&BTRFS_I(inode)->lock);
9033         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9034         spin_unlock(&BTRFS_I(inode)->lock);
9035         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9036                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9037         return 0;
9038 }
9039
9040 static int btrfs_rename_exchange(struct inode *old_dir,
9041                               struct dentry *old_dentry,
9042                               struct inode *new_dir,
9043                               struct dentry *new_dentry)
9044 {
9045         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9046         struct btrfs_trans_handle *trans;
9047         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9048         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9049         struct inode *new_inode = new_dentry->d_inode;
9050         struct inode *old_inode = old_dentry->d_inode;
9051         struct timespec64 ctime = current_time(old_inode);
9052         struct dentry *parent;
9053         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9054         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9055         u64 old_idx = 0;
9056         u64 new_idx = 0;
9057         int ret;
9058         bool root_log_pinned = false;
9059         bool dest_log_pinned = false;
9060         struct btrfs_log_ctx ctx_root;
9061         struct btrfs_log_ctx ctx_dest;
9062         bool sync_log_root = false;
9063         bool sync_log_dest = false;
9064         bool commit_transaction = false;
9065
9066         /* we only allow rename subvolume link between subvolumes */
9067         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9068                 return -EXDEV;
9069
9070         btrfs_init_log_ctx(&ctx_root, old_inode);
9071         btrfs_init_log_ctx(&ctx_dest, new_inode);
9072
9073         /* close the race window with snapshot create/destroy ioctl */
9074         if (old_ino == BTRFS_FIRST_FREE_OBJECTID ||
9075             new_ino == BTRFS_FIRST_FREE_OBJECTID)
9076                 down_read(&fs_info->subvol_sem);
9077
9078         /*
9079          * We want to reserve the absolute worst case amount of items.  So if
9080          * both inodes are subvols and we need to unlink them then that would
9081          * require 4 item modifications, but if they are both normal inodes it
9082          * would require 5 item modifications, so we'll assume their normal
9083          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9084          * should cover the worst case number of items we'll modify.
9085          */
9086         trans = btrfs_start_transaction(root, 12);
9087         if (IS_ERR(trans)) {
9088                 ret = PTR_ERR(trans);
9089                 goto out_notrans;
9090         }
9091
9092         if (dest != root)
9093                 btrfs_record_root_in_trans(trans, dest);
9094
9095         /*
9096          * We need to find a free sequence number both in the source and
9097          * in the destination directory for the exchange.
9098          */
9099         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9100         if (ret)
9101                 goto out_fail;
9102         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9103         if (ret)
9104                 goto out_fail;
9105
9106         BTRFS_I(old_inode)->dir_index = 0ULL;
9107         BTRFS_I(new_inode)->dir_index = 0ULL;
9108
9109         /* Reference for the source. */
9110         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9111                 /* force full log commit if subvolume involved. */
9112                 btrfs_set_log_full_commit(trans);
9113         } else {
9114                 btrfs_pin_log_trans(root);
9115                 root_log_pinned = true;
9116                 ret = btrfs_insert_inode_ref(trans, dest,
9117                                              new_dentry->d_name.name,
9118                                              new_dentry->d_name.len,
9119                                              old_ino,
9120                                              btrfs_ino(BTRFS_I(new_dir)),
9121                                              old_idx);
9122                 if (ret)
9123                         goto out_fail;
9124         }
9125
9126         /* And now for the dest. */
9127         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9128                 /* force full log commit if subvolume involved. */
9129                 btrfs_set_log_full_commit(trans);
9130         } else {
9131                 btrfs_pin_log_trans(dest);
9132                 dest_log_pinned = true;
9133                 ret = btrfs_insert_inode_ref(trans, root,
9134                                              old_dentry->d_name.name,
9135                                              old_dentry->d_name.len,
9136                                              new_ino,
9137                                              btrfs_ino(BTRFS_I(old_dir)),
9138                                              new_idx);
9139                 if (ret)
9140                         goto out_fail;
9141         }
9142
9143         /* Update inode version and ctime/mtime. */
9144         inode_inc_iversion(old_dir);
9145         inode_inc_iversion(new_dir);
9146         inode_inc_iversion(old_inode);
9147         inode_inc_iversion(new_inode);
9148         old_dir->i_ctime = old_dir->i_mtime = ctime;
9149         new_dir->i_ctime = new_dir->i_mtime = ctime;
9150         old_inode->i_ctime = ctime;
9151         new_inode->i_ctime = ctime;
9152
9153         if (old_dentry->d_parent != new_dentry->d_parent) {
9154                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9155                                 BTRFS_I(old_inode), 1);
9156                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9157                                 BTRFS_I(new_inode), 1);
9158         }
9159
9160         /* src is a subvolume */
9161         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9162                 ret = btrfs_unlink_subvol(trans, old_dir, old_dentry);
9163         } else { /* src is an inode */
9164                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9165                                            BTRFS_I(old_dentry->d_inode),
9166                                            old_dentry->d_name.name,
9167                                            old_dentry->d_name.len);
9168                 if (!ret)
9169                         ret = btrfs_update_inode(trans, root, old_inode);
9170         }
9171         if (ret) {
9172                 btrfs_abort_transaction(trans, ret);
9173                 goto out_fail;
9174         }
9175
9176         /* dest is a subvolume */
9177         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9178                 ret = btrfs_unlink_subvol(trans, new_dir, new_dentry);
9179         } else { /* dest is an inode */
9180                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9181                                            BTRFS_I(new_dentry->d_inode),
9182                                            new_dentry->d_name.name,
9183                                            new_dentry->d_name.len);
9184                 if (!ret)
9185                         ret = btrfs_update_inode(trans, dest, new_inode);
9186         }
9187         if (ret) {
9188                 btrfs_abort_transaction(trans, ret);
9189                 goto out_fail;
9190         }
9191
9192         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9193                              new_dentry->d_name.name,
9194                              new_dentry->d_name.len, 0, old_idx);
9195         if (ret) {
9196                 btrfs_abort_transaction(trans, ret);
9197                 goto out_fail;
9198         }
9199
9200         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9201                              old_dentry->d_name.name,
9202                              old_dentry->d_name.len, 0, new_idx);
9203         if (ret) {
9204                 btrfs_abort_transaction(trans, ret);
9205                 goto out_fail;
9206         }
9207
9208         if (old_inode->i_nlink == 1)
9209                 BTRFS_I(old_inode)->dir_index = old_idx;
9210         if (new_inode->i_nlink == 1)
9211                 BTRFS_I(new_inode)->dir_index = new_idx;
9212
9213         if (root_log_pinned) {
9214                 parent = new_dentry->d_parent;
9215                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9216                                          BTRFS_I(old_dir), parent,
9217                                          false, &ctx_root);
9218                 if (ret == BTRFS_NEED_LOG_SYNC)
9219                         sync_log_root = true;
9220                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9221                         commit_transaction = true;
9222                 ret = 0;
9223                 btrfs_end_log_trans(root);
9224                 root_log_pinned = false;
9225         }
9226         if (dest_log_pinned) {
9227                 if (!commit_transaction) {
9228                         parent = old_dentry->d_parent;
9229                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
9230                                                  BTRFS_I(new_dir), parent,
9231                                                  false, &ctx_dest);
9232                         if (ret == BTRFS_NEED_LOG_SYNC)
9233                                 sync_log_dest = true;
9234                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
9235                                 commit_transaction = true;
9236                         ret = 0;
9237                 }
9238                 btrfs_end_log_trans(dest);
9239                 dest_log_pinned = false;
9240         }
9241 out_fail:
9242         /*
9243          * If we have pinned a log and an error happened, we unpin tasks
9244          * trying to sync the log and force them to fallback to a transaction
9245          * commit if the log currently contains any of the inodes involved in
9246          * this rename operation (to ensure we do not persist a log with an
9247          * inconsistent state for any of these inodes or leading to any
9248          * inconsistencies when replayed). If the transaction was aborted, the
9249          * abortion reason is propagated to userspace when attempting to commit
9250          * the transaction. If the log does not contain any of these inodes, we
9251          * allow the tasks to sync it.
9252          */
9253         if (ret && (root_log_pinned || dest_log_pinned)) {
9254                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9255                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9256                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9257                     (new_inode &&
9258                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9259                         btrfs_set_log_full_commit(trans);
9260
9261                 if (root_log_pinned) {
9262                         btrfs_end_log_trans(root);
9263                         root_log_pinned = false;
9264                 }
9265                 if (dest_log_pinned) {
9266                         btrfs_end_log_trans(dest);
9267                         dest_log_pinned = false;
9268                 }
9269         }
9270         if (!ret && sync_log_root && !commit_transaction) {
9271                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
9272                                      &ctx_root);
9273                 if (ret)
9274                         commit_transaction = true;
9275         }
9276         if (!ret && sync_log_dest && !commit_transaction) {
9277                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
9278                                      &ctx_dest);
9279                 if (ret)
9280                         commit_transaction = true;
9281         }
9282         if (commit_transaction) {
9283                 /*
9284                  * We may have set commit_transaction when logging the new name
9285                  * in the destination root, in which case we left the source
9286                  * root context in the list of log contextes. So make sure we
9287                  * remove it to avoid invalid memory accesses, since the context
9288                  * was allocated in our stack frame.
9289                  */
9290                 if (sync_log_root) {
9291                         mutex_lock(&root->log_mutex);
9292                         list_del_init(&ctx_root.list);
9293                         mutex_unlock(&root->log_mutex);
9294                 }
9295                 ret = btrfs_commit_transaction(trans);
9296         } else {
9297                 int ret2;
9298
9299                 ret2 = btrfs_end_transaction(trans);
9300                 ret = ret ? ret : ret2;
9301         }
9302 out_notrans:
9303         if (new_ino == BTRFS_FIRST_FREE_OBJECTID ||
9304             old_ino == BTRFS_FIRST_FREE_OBJECTID)
9305                 up_read(&fs_info->subvol_sem);
9306
9307         ASSERT(list_empty(&ctx_root.list));
9308         ASSERT(list_empty(&ctx_dest.list));
9309
9310         return ret;
9311 }
9312
9313 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9314                                      struct btrfs_root *root,
9315                                      struct inode *dir,
9316                                      struct dentry *dentry)
9317 {
9318         int ret;
9319         struct inode *inode;
9320         u64 objectid;
9321         u64 index;
9322
9323         ret = btrfs_find_free_ino(root, &objectid);
9324         if (ret)
9325                 return ret;
9326
9327         inode = btrfs_new_inode(trans, root, dir,
9328                                 dentry->d_name.name,
9329                                 dentry->d_name.len,
9330                                 btrfs_ino(BTRFS_I(dir)),
9331                                 objectid,
9332                                 S_IFCHR | WHITEOUT_MODE,
9333                                 &index);
9334
9335         if (IS_ERR(inode)) {
9336                 ret = PTR_ERR(inode);
9337                 return ret;
9338         }
9339
9340         inode->i_op = &btrfs_special_inode_operations;
9341         init_special_inode(inode, inode->i_mode,
9342                 WHITEOUT_DEV);
9343
9344         ret = btrfs_init_inode_security(trans, inode, dir,
9345                                 &dentry->d_name);
9346         if (ret)
9347                 goto out;
9348
9349         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9350                                 BTRFS_I(inode), 0, index);
9351         if (ret)
9352                 goto out;
9353
9354         ret = btrfs_update_inode(trans, root, inode);
9355 out:
9356         unlock_new_inode(inode);
9357         if (ret)
9358                 inode_dec_link_count(inode);
9359         iput(inode);
9360
9361         return ret;
9362 }
9363
9364 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9365                            struct inode *new_dir, struct dentry *new_dentry,
9366                            unsigned int flags)
9367 {
9368         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9369         struct btrfs_trans_handle *trans;
9370         unsigned int trans_num_items;
9371         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9372         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9373         struct inode *new_inode = d_inode(new_dentry);
9374         struct inode *old_inode = d_inode(old_dentry);
9375         u64 index = 0;
9376         int ret;
9377         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9378         bool log_pinned = false;
9379         struct btrfs_log_ctx ctx;
9380         bool sync_log = false;
9381         bool commit_transaction = false;
9382
9383         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9384                 return -EPERM;
9385
9386         /* we only allow rename subvolume link between subvolumes */
9387         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9388                 return -EXDEV;
9389
9390         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9391             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9392                 return -ENOTEMPTY;
9393
9394         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9395             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9396                 return -ENOTEMPTY;
9397
9398
9399         /* check for collisions, even if the  name isn't there */
9400         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9401                              new_dentry->d_name.name,
9402                              new_dentry->d_name.len);
9403
9404         if (ret) {
9405                 if (ret == -EEXIST) {
9406                         /* we shouldn't get
9407                          * eexist without a new_inode */
9408                         if (WARN_ON(!new_inode)) {
9409                                 return ret;
9410                         }
9411                 } else {
9412                         /* maybe -EOVERFLOW */
9413                         return ret;
9414                 }
9415         }
9416         ret = 0;
9417
9418         /*
9419          * we're using rename to replace one file with another.  Start IO on it
9420          * now so  we don't add too much work to the end of the transaction
9421          */
9422         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9423                 filemap_flush(old_inode->i_mapping);
9424
9425         /* close the racy window with snapshot create/destroy ioctl */
9426         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9427                 down_read(&fs_info->subvol_sem);
9428         /*
9429          * We want to reserve the absolute worst case amount of items.  So if
9430          * both inodes are subvols and we need to unlink them then that would
9431          * require 4 item modifications, but if they are both normal inodes it
9432          * would require 5 item modifications, so we'll assume they are normal
9433          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9434          * should cover the worst case number of items we'll modify.
9435          * If our rename has the whiteout flag, we need more 5 units for the
9436          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9437          * when selinux is enabled).
9438          */
9439         trans_num_items = 11;
9440         if (flags & RENAME_WHITEOUT)
9441                 trans_num_items += 5;
9442         trans = btrfs_start_transaction(root, trans_num_items);
9443         if (IS_ERR(trans)) {
9444                 ret = PTR_ERR(trans);
9445                 goto out_notrans;
9446         }
9447
9448         if (dest != root)
9449                 btrfs_record_root_in_trans(trans, dest);
9450
9451         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9452         if (ret)
9453                 goto out_fail;
9454
9455         BTRFS_I(old_inode)->dir_index = 0ULL;
9456         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9457                 /* force full log commit if subvolume involved. */
9458                 btrfs_set_log_full_commit(trans);
9459         } else {
9460                 btrfs_pin_log_trans(root);
9461                 log_pinned = true;
9462                 ret = btrfs_insert_inode_ref(trans, dest,
9463                                              new_dentry->d_name.name,
9464                                              new_dentry->d_name.len,
9465                                              old_ino,
9466                                              btrfs_ino(BTRFS_I(new_dir)), index);
9467                 if (ret)
9468                         goto out_fail;
9469         }
9470
9471         inode_inc_iversion(old_dir);
9472         inode_inc_iversion(new_dir);
9473         inode_inc_iversion(old_inode);
9474         old_dir->i_ctime = old_dir->i_mtime =
9475         new_dir->i_ctime = new_dir->i_mtime =
9476         old_inode->i_ctime = current_time(old_dir);
9477
9478         if (old_dentry->d_parent != new_dentry->d_parent)
9479                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9480                                 BTRFS_I(old_inode), 1);
9481
9482         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9483                 ret = btrfs_unlink_subvol(trans, old_dir, old_dentry);
9484         } else {
9485                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9486                                         BTRFS_I(d_inode(old_dentry)),
9487                                         old_dentry->d_name.name,
9488                                         old_dentry->d_name.len);
9489                 if (!ret)
9490                         ret = btrfs_update_inode(trans, root, old_inode);
9491         }
9492         if (ret) {
9493                 btrfs_abort_transaction(trans, ret);
9494                 goto out_fail;
9495         }
9496
9497         if (new_inode) {
9498                 inode_inc_iversion(new_inode);
9499                 new_inode->i_ctime = current_time(new_inode);
9500                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9501                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9502                         ret = btrfs_unlink_subvol(trans, new_dir, new_dentry);
9503                         BUG_ON(new_inode->i_nlink == 0);
9504                 } else {
9505                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9506                                                  BTRFS_I(d_inode(new_dentry)),
9507                                                  new_dentry->d_name.name,
9508                                                  new_dentry->d_name.len);
9509                 }
9510                 if (!ret && new_inode->i_nlink == 0)
9511                         ret = btrfs_orphan_add(trans,
9512                                         BTRFS_I(d_inode(new_dentry)));
9513                 if (ret) {
9514                         btrfs_abort_transaction(trans, ret);
9515                         goto out_fail;
9516                 }
9517         }
9518
9519         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9520                              new_dentry->d_name.name,
9521                              new_dentry->d_name.len, 0, index);
9522         if (ret) {
9523                 btrfs_abort_transaction(trans, ret);
9524                 goto out_fail;
9525         }
9526
9527         if (old_inode->i_nlink == 1)
9528                 BTRFS_I(old_inode)->dir_index = index;
9529
9530         if (log_pinned) {
9531                 struct dentry *parent = new_dentry->d_parent;
9532
9533                 btrfs_init_log_ctx(&ctx, old_inode);
9534                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9535                                          BTRFS_I(old_dir), parent,
9536                                          false, &ctx);
9537                 if (ret == BTRFS_NEED_LOG_SYNC)
9538                         sync_log = true;
9539                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9540                         commit_transaction = true;
9541                 ret = 0;
9542                 btrfs_end_log_trans(root);
9543                 log_pinned = false;
9544         }
9545
9546         if (flags & RENAME_WHITEOUT) {
9547                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9548                                                 old_dentry);
9549
9550                 if (ret) {
9551                         btrfs_abort_transaction(trans, ret);
9552                         goto out_fail;
9553                 }
9554         }
9555 out_fail:
9556         /*
9557          * If we have pinned the log and an error happened, we unpin tasks
9558          * trying to sync the log and force them to fallback to a transaction
9559          * commit if the log currently contains any of the inodes involved in
9560          * this rename operation (to ensure we do not persist a log with an
9561          * inconsistent state for any of these inodes or leading to any
9562          * inconsistencies when replayed). If the transaction was aborted, the
9563          * abortion reason is propagated to userspace when attempting to commit
9564          * the transaction. If the log does not contain any of these inodes, we
9565          * allow the tasks to sync it.
9566          */
9567         if (ret && log_pinned) {
9568                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9569                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9570                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9571                     (new_inode &&
9572                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9573                         btrfs_set_log_full_commit(trans);
9574
9575                 btrfs_end_log_trans(root);
9576                 log_pinned = false;
9577         }
9578         if (!ret && sync_log) {
9579                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
9580                 if (ret)
9581                         commit_transaction = true;
9582         } else if (sync_log) {
9583                 mutex_lock(&root->log_mutex);
9584                 list_del(&ctx.list);
9585                 mutex_unlock(&root->log_mutex);
9586         }
9587         if (commit_transaction) {
9588                 ret = btrfs_commit_transaction(trans);
9589         } else {
9590                 int ret2;
9591
9592                 ret2 = btrfs_end_transaction(trans);
9593                 ret = ret ? ret : ret2;
9594         }
9595 out_notrans:
9596         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9597                 up_read(&fs_info->subvol_sem);
9598
9599         return ret;
9600 }
9601
9602 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9603                          struct inode *new_dir, struct dentry *new_dentry,
9604                          unsigned int flags)
9605 {
9606         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9607                 return -EINVAL;
9608
9609         if (flags & RENAME_EXCHANGE)
9610                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9611                                           new_dentry);
9612
9613         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9614 }
9615
9616 struct btrfs_delalloc_work {
9617         struct inode *inode;
9618         struct completion completion;
9619         struct list_head list;
9620         struct btrfs_work work;
9621 };
9622
9623 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9624 {
9625         struct btrfs_delalloc_work *delalloc_work;
9626         struct inode *inode;
9627
9628         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9629                                      work);
9630         inode = delalloc_work->inode;
9631         filemap_flush(inode->i_mapping);
9632         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9633                                 &BTRFS_I(inode)->runtime_flags))
9634                 filemap_flush(inode->i_mapping);
9635
9636         iput(inode);
9637         complete(&delalloc_work->completion);
9638 }
9639
9640 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
9641 {
9642         struct btrfs_delalloc_work *work;
9643
9644         work = kmalloc(sizeof(*work), GFP_NOFS);
9645         if (!work)
9646                 return NULL;
9647
9648         init_completion(&work->completion);
9649         INIT_LIST_HEAD(&work->list);
9650         work->inode = inode;
9651         btrfs_init_work(&work->work, btrfs_run_delalloc_work, NULL, NULL);
9652
9653         return work;
9654 }
9655
9656 /*
9657  * some fairly slow code that needs optimization. This walks the list
9658  * of all the inodes with pending delalloc and forces them to disk.
9659  */
9660 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
9661 {
9662         struct btrfs_inode *binode;
9663         struct inode *inode;
9664         struct btrfs_delalloc_work *work, *next;
9665         struct list_head works;
9666         struct list_head splice;
9667         int ret = 0;
9668
9669         INIT_LIST_HEAD(&works);
9670         INIT_LIST_HEAD(&splice);
9671
9672         mutex_lock(&root->delalloc_mutex);
9673         spin_lock(&root->delalloc_lock);
9674         list_splice_init(&root->delalloc_inodes, &splice);
9675         while (!list_empty(&splice)) {
9676                 binode = list_entry(splice.next, struct btrfs_inode,
9677                                     delalloc_inodes);
9678
9679                 list_move_tail(&binode->delalloc_inodes,
9680                                &root->delalloc_inodes);
9681                 inode = igrab(&binode->vfs_inode);
9682                 if (!inode) {
9683                         cond_resched_lock(&root->delalloc_lock);
9684                         continue;
9685                 }
9686                 spin_unlock(&root->delalloc_lock);
9687
9688                 if (snapshot)
9689                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
9690                                 &binode->runtime_flags);
9691                 work = btrfs_alloc_delalloc_work(inode);
9692                 if (!work) {
9693                         iput(inode);
9694                         ret = -ENOMEM;
9695                         goto out;
9696                 }
9697                 list_add_tail(&work->list, &works);
9698                 btrfs_queue_work(root->fs_info->flush_workers,
9699                                  &work->work);
9700                 ret++;
9701                 if (nr != -1 && ret >= nr)
9702                         goto out;
9703                 cond_resched();
9704                 spin_lock(&root->delalloc_lock);
9705         }
9706         spin_unlock(&root->delalloc_lock);
9707
9708 out:
9709         list_for_each_entry_safe(work, next, &works, list) {
9710                 list_del_init(&work->list);
9711                 wait_for_completion(&work->completion);
9712                 kfree(work);
9713         }
9714
9715         if (!list_empty(&splice)) {
9716                 spin_lock(&root->delalloc_lock);
9717                 list_splice_tail(&splice, &root->delalloc_inodes);
9718                 spin_unlock(&root->delalloc_lock);
9719         }
9720         mutex_unlock(&root->delalloc_mutex);
9721         return ret;
9722 }
9723
9724 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
9725 {
9726         struct btrfs_fs_info *fs_info = root->fs_info;
9727         int ret;
9728
9729         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9730                 return -EROFS;
9731
9732         ret = start_delalloc_inodes(root, -1, true);
9733         if (ret > 0)
9734                 ret = 0;
9735         return ret;
9736 }
9737
9738 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
9739 {
9740         struct btrfs_root *root;
9741         struct list_head splice;
9742         int ret;
9743
9744         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9745                 return -EROFS;
9746
9747         INIT_LIST_HEAD(&splice);
9748
9749         mutex_lock(&fs_info->delalloc_root_mutex);
9750         spin_lock(&fs_info->delalloc_root_lock);
9751         list_splice_init(&fs_info->delalloc_roots, &splice);
9752         while (!list_empty(&splice) && nr) {
9753                 root = list_first_entry(&splice, struct btrfs_root,
9754                                         delalloc_root);
9755                 root = btrfs_grab_root(root);
9756                 BUG_ON(!root);
9757                 list_move_tail(&root->delalloc_root,
9758                                &fs_info->delalloc_roots);
9759                 spin_unlock(&fs_info->delalloc_root_lock);
9760
9761                 ret = start_delalloc_inodes(root, nr, false);
9762                 btrfs_put_root(root);
9763                 if (ret < 0)
9764                         goto out;
9765
9766                 if (nr != -1) {
9767                         nr -= ret;
9768                         WARN_ON(nr < 0);
9769                 }
9770                 spin_lock(&fs_info->delalloc_root_lock);
9771         }
9772         spin_unlock(&fs_info->delalloc_root_lock);
9773
9774         ret = 0;
9775 out:
9776         if (!list_empty(&splice)) {
9777                 spin_lock(&fs_info->delalloc_root_lock);
9778                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9779                 spin_unlock(&fs_info->delalloc_root_lock);
9780         }
9781         mutex_unlock(&fs_info->delalloc_root_mutex);
9782         return ret;
9783 }
9784
9785 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9786                          const char *symname)
9787 {
9788         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
9789         struct btrfs_trans_handle *trans;
9790         struct btrfs_root *root = BTRFS_I(dir)->root;
9791         struct btrfs_path *path;
9792         struct btrfs_key key;
9793         struct inode *inode = NULL;
9794         int err;
9795         u64 objectid;
9796         u64 index = 0;
9797         int name_len;
9798         int datasize;
9799         unsigned long ptr;
9800         struct btrfs_file_extent_item *ei;
9801         struct extent_buffer *leaf;
9802
9803         name_len = strlen(symname);
9804         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
9805                 return -ENAMETOOLONG;
9806
9807         /*
9808          * 2 items for inode item and ref
9809          * 2 items for dir items
9810          * 1 item for updating parent inode item
9811          * 1 item for the inline extent item
9812          * 1 item for xattr if selinux is on
9813          */
9814         trans = btrfs_start_transaction(root, 7);
9815         if (IS_ERR(trans))
9816                 return PTR_ERR(trans);
9817
9818         err = btrfs_find_free_ino(root, &objectid);
9819         if (err)
9820                 goto out_unlock;
9821
9822         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9823                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
9824                                 objectid, S_IFLNK|S_IRWXUGO, &index);
9825         if (IS_ERR(inode)) {
9826                 err = PTR_ERR(inode);
9827                 inode = NULL;
9828                 goto out_unlock;
9829         }
9830
9831         /*
9832         * If the active LSM wants to access the inode during
9833         * d_instantiate it needs these. Smack checks to see
9834         * if the filesystem supports xattrs by looking at the
9835         * ops vector.
9836         */
9837         inode->i_fop = &btrfs_file_operations;
9838         inode->i_op = &btrfs_file_inode_operations;
9839         inode->i_mapping->a_ops = &btrfs_aops;
9840         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9841
9842         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9843         if (err)
9844                 goto out_unlock;
9845
9846         path = btrfs_alloc_path();
9847         if (!path) {
9848                 err = -ENOMEM;
9849                 goto out_unlock;
9850         }
9851         key.objectid = btrfs_ino(BTRFS_I(inode));
9852         key.offset = 0;
9853         key.type = BTRFS_EXTENT_DATA_KEY;
9854         datasize = btrfs_file_extent_calc_inline_size(name_len);
9855         err = btrfs_insert_empty_item(trans, root, path, &key,
9856                                       datasize);
9857         if (err) {
9858                 btrfs_free_path(path);
9859                 goto out_unlock;
9860         }
9861         leaf = path->nodes[0];
9862         ei = btrfs_item_ptr(leaf, path->slots[0],
9863                             struct btrfs_file_extent_item);
9864         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9865         btrfs_set_file_extent_type(leaf, ei,
9866                                    BTRFS_FILE_EXTENT_INLINE);
9867         btrfs_set_file_extent_encryption(leaf, ei, 0);
9868         btrfs_set_file_extent_compression(leaf, ei, 0);
9869         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9870         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9871
9872         ptr = btrfs_file_extent_inline_start(ei);
9873         write_extent_buffer(leaf, symname, ptr, name_len);
9874         btrfs_mark_buffer_dirty(leaf);
9875         btrfs_free_path(path);
9876
9877         inode->i_op = &btrfs_symlink_inode_operations;
9878         inode_nohighmem(inode);
9879         inode_set_bytes(inode, name_len);
9880         btrfs_i_size_write(BTRFS_I(inode), name_len);
9881         err = btrfs_update_inode(trans, root, inode);
9882         /*
9883          * Last step, add directory indexes for our symlink inode. This is the
9884          * last step to avoid extra cleanup of these indexes if an error happens
9885          * elsewhere above.
9886          */
9887         if (!err)
9888                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9889                                 BTRFS_I(inode), 0, index);
9890         if (err)
9891                 goto out_unlock;
9892
9893         d_instantiate_new(dentry, inode);
9894
9895 out_unlock:
9896         btrfs_end_transaction(trans);
9897         if (err && inode) {
9898                 inode_dec_link_count(inode);
9899                 discard_new_inode(inode);
9900         }
9901         btrfs_btree_balance_dirty(fs_info);
9902         return err;
9903 }
9904
9905 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9906                                        u64 start, u64 num_bytes, u64 min_size,
9907                                        loff_t actual_len, u64 *alloc_hint,
9908                                        struct btrfs_trans_handle *trans)
9909 {
9910         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9911         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9912         struct extent_map *em;
9913         struct btrfs_root *root = BTRFS_I(inode)->root;
9914         struct btrfs_key ins;
9915         u64 cur_offset = start;
9916         u64 clear_offset = start;
9917         u64 i_size;
9918         u64 cur_bytes;
9919         u64 last_alloc = (u64)-1;
9920         int ret = 0;
9921         bool own_trans = true;
9922         u64 end = start + num_bytes - 1;
9923
9924         if (trans)
9925                 own_trans = false;
9926         while (num_bytes > 0) {
9927                 if (own_trans) {
9928                         trans = btrfs_start_transaction(root, 3);
9929                         if (IS_ERR(trans)) {
9930                                 ret = PTR_ERR(trans);
9931                                 break;
9932                         }
9933                 }
9934
9935                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
9936                 cur_bytes = max(cur_bytes, min_size);
9937                 /*
9938                  * If we are severely fragmented we could end up with really
9939                  * small allocations, so if the allocator is returning small
9940                  * chunks lets make its job easier by only searching for those
9941                  * sized chunks.
9942                  */
9943                 cur_bytes = min(cur_bytes, last_alloc);
9944                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
9945                                 min_size, 0, *alloc_hint, &ins, 1, 0);
9946                 if (ret) {
9947                         if (own_trans)
9948                                 btrfs_end_transaction(trans);
9949                         break;
9950                 }
9951
9952                 /*
9953                  * We've reserved this space, and thus converted it from
9954                  * ->bytes_may_use to ->bytes_reserved.  Any error that happens
9955                  * from here on out we will only need to clear our reservation
9956                  * for the remaining unreserved area, so advance our
9957                  * clear_offset by our extent size.
9958                  */
9959                 clear_offset += ins.offset;
9960                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
9961
9962                 last_alloc = ins.offset;
9963                 ret = insert_reserved_file_extent(trans, inode,
9964                                                   cur_offset, ins.objectid,
9965                                                   ins.offset, ins.offset,
9966                                                   ins.offset, 0, 0, 0,
9967                                                   BTRFS_FILE_EXTENT_PREALLOC);
9968                 if (ret) {
9969                         btrfs_free_reserved_extent(fs_info, ins.objectid,
9970                                                    ins.offset, 0);
9971                         btrfs_abort_transaction(trans, ret);
9972                         if (own_trans)
9973                                 btrfs_end_transaction(trans);
9974                         break;
9975                 }
9976
9977                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
9978                                         cur_offset + ins.offset -1, 0);
9979
9980                 em = alloc_extent_map();
9981                 if (!em) {
9982                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9983                                 &BTRFS_I(inode)->runtime_flags);
9984                         goto next;
9985                 }
9986
9987                 em->start = cur_offset;
9988                 em->orig_start = cur_offset;
9989                 em->len = ins.offset;
9990                 em->block_start = ins.objectid;
9991                 em->block_len = ins.offset;
9992                 em->orig_block_len = ins.offset;
9993                 em->ram_bytes = ins.offset;
9994                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9995                 em->generation = trans->transid;
9996
9997                 while (1) {
9998                         write_lock(&em_tree->lock);
9999                         ret = add_extent_mapping(em_tree, em, 1);
10000                         write_unlock(&em_tree->lock);
10001                         if (ret != -EEXIST)
10002                                 break;
10003                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10004                                                 cur_offset + ins.offset - 1,
10005                                                 0);
10006                 }
10007                 free_extent_map(em);
10008 next:
10009                 num_bytes -= ins.offset;
10010                 cur_offset += ins.offset;
10011                 *alloc_hint = ins.objectid + ins.offset;
10012
10013                 inode_inc_iversion(inode);
10014                 inode->i_ctime = current_time(inode);
10015                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10016                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10017                     (actual_len > inode->i_size) &&
10018                     (cur_offset > inode->i_size)) {
10019                         if (cur_offset > actual_len)
10020                                 i_size = actual_len;
10021                         else
10022                                 i_size = cur_offset;
10023                         i_size_write(inode, i_size);
10024                         btrfs_inode_safe_disk_i_size_write(inode, 0);
10025                 }
10026
10027                 ret = btrfs_update_inode(trans, root, inode);
10028
10029                 if (ret) {
10030                         btrfs_abort_transaction(trans, ret);
10031                         if (own_trans)
10032                                 btrfs_end_transaction(trans);
10033                         break;
10034                 }
10035
10036                 if (own_trans)
10037                         btrfs_end_transaction(trans);
10038         }
10039         if (clear_offset < end)
10040                 btrfs_free_reserved_data_space(inode, NULL, clear_offset,
10041                         end - clear_offset + 1);
10042         return ret;
10043 }
10044
10045 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10046                               u64 start, u64 num_bytes, u64 min_size,
10047                               loff_t actual_len, u64 *alloc_hint)
10048 {
10049         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10050                                            min_size, actual_len, alloc_hint,
10051                                            NULL);
10052 }
10053
10054 int btrfs_prealloc_file_range_trans(struct inode *inode,
10055                                     struct btrfs_trans_handle *trans, int mode,
10056                                     u64 start, u64 num_bytes, u64 min_size,
10057                                     loff_t actual_len, u64 *alloc_hint)
10058 {
10059         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10060                                            min_size, actual_len, alloc_hint, trans);
10061 }
10062
10063 static int btrfs_set_page_dirty(struct page *page)
10064 {
10065         return __set_page_dirty_nobuffers(page);
10066 }
10067
10068 static int btrfs_permission(struct inode *inode, int mask)
10069 {
10070         struct btrfs_root *root = BTRFS_I(inode)->root;
10071         umode_t mode = inode->i_mode;
10072
10073         if (mask & MAY_WRITE &&
10074             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10075                 if (btrfs_root_readonly(root))
10076                         return -EROFS;
10077                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10078                         return -EACCES;
10079         }
10080         return generic_permission(inode, mask);
10081 }
10082
10083 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10084 {
10085         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10086         struct btrfs_trans_handle *trans;
10087         struct btrfs_root *root = BTRFS_I(dir)->root;
10088         struct inode *inode = NULL;
10089         u64 objectid;
10090         u64 index;
10091         int ret = 0;
10092
10093         /*
10094          * 5 units required for adding orphan entry
10095          */
10096         trans = btrfs_start_transaction(root, 5);
10097         if (IS_ERR(trans))
10098                 return PTR_ERR(trans);
10099
10100         ret = btrfs_find_free_ino(root, &objectid);
10101         if (ret)
10102                 goto out;
10103
10104         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10105                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10106         if (IS_ERR(inode)) {
10107                 ret = PTR_ERR(inode);
10108                 inode = NULL;
10109                 goto out;
10110         }
10111
10112         inode->i_fop = &btrfs_file_operations;
10113         inode->i_op = &btrfs_file_inode_operations;
10114
10115         inode->i_mapping->a_ops = &btrfs_aops;
10116         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10117
10118         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10119         if (ret)
10120                 goto out;
10121
10122         ret = btrfs_update_inode(trans, root, inode);
10123         if (ret)
10124                 goto out;
10125         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10126         if (ret)
10127                 goto out;
10128
10129         /*
10130          * We set number of links to 0 in btrfs_new_inode(), and here we set
10131          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10132          * through:
10133          *
10134          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10135          */
10136         set_nlink(inode, 1);
10137         d_tmpfile(dentry, inode);
10138         unlock_new_inode(inode);
10139         mark_inode_dirty(inode);
10140 out:
10141         btrfs_end_transaction(trans);
10142         if (ret && inode)
10143                 discard_new_inode(inode);
10144         btrfs_btree_balance_dirty(fs_info);
10145         return ret;
10146 }
10147
10148 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10149 {
10150         struct inode *inode = tree->private_data;
10151         unsigned long index = start >> PAGE_SHIFT;
10152         unsigned long end_index = end >> PAGE_SHIFT;
10153         struct page *page;
10154
10155         while (index <= end_index) {
10156                 page = find_get_page(inode->i_mapping, index);
10157                 ASSERT(page); /* Pages should be in the extent_io_tree */
10158                 set_page_writeback(page);
10159                 put_page(page);
10160                 index++;
10161         }
10162 }
10163
10164 #ifdef CONFIG_SWAP
10165 /*
10166  * Add an entry indicating a block group or device which is pinned by a
10167  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
10168  * negative errno on failure.
10169  */
10170 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
10171                                   bool is_block_group)
10172 {
10173         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10174         struct btrfs_swapfile_pin *sp, *entry;
10175         struct rb_node **p;
10176         struct rb_node *parent = NULL;
10177
10178         sp = kmalloc(sizeof(*sp), GFP_NOFS);
10179         if (!sp)
10180                 return -ENOMEM;
10181         sp->ptr = ptr;
10182         sp->inode = inode;
10183         sp->is_block_group = is_block_group;
10184
10185         spin_lock(&fs_info->swapfile_pins_lock);
10186         p = &fs_info->swapfile_pins.rb_node;
10187         while (*p) {
10188                 parent = *p;
10189                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
10190                 if (sp->ptr < entry->ptr ||
10191                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
10192                         p = &(*p)->rb_left;
10193                 } else if (sp->ptr > entry->ptr ||
10194                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
10195                         p = &(*p)->rb_right;
10196                 } else {
10197                         spin_unlock(&fs_info->swapfile_pins_lock);
10198                         kfree(sp);
10199                         return 1;
10200                 }
10201         }
10202         rb_link_node(&sp->node, parent, p);
10203         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
10204         spin_unlock(&fs_info->swapfile_pins_lock);
10205         return 0;
10206 }
10207
10208 /* Free all of the entries pinned by this swapfile. */
10209 static void btrfs_free_swapfile_pins(struct inode *inode)
10210 {
10211         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10212         struct btrfs_swapfile_pin *sp;
10213         struct rb_node *node, *next;
10214
10215         spin_lock(&fs_info->swapfile_pins_lock);
10216         node = rb_first(&fs_info->swapfile_pins);
10217         while (node) {
10218                 next = rb_next(node);
10219                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
10220                 if (sp->inode == inode) {
10221                         rb_erase(&sp->node, &fs_info->swapfile_pins);
10222                         if (sp->is_block_group)
10223                                 btrfs_put_block_group(sp->ptr);
10224                         kfree(sp);
10225                 }
10226                 node = next;
10227         }
10228         spin_unlock(&fs_info->swapfile_pins_lock);
10229 }
10230
10231 struct btrfs_swap_info {
10232         u64 start;
10233         u64 block_start;
10234         u64 block_len;
10235         u64 lowest_ppage;
10236         u64 highest_ppage;
10237         unsigned long nr_pages;
10238         int nr_extents;
10239 };
10240
10241 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
10242                                  struct btrfs_swap_info *bsi)
10243 {
10244         unsigned long nr_pages;
10245         u64 first_ppage, first_ppage_reported, next_ppage;
10246         int ret;
10247
10248         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
10249         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
10250                                 PAGE_SIZE) >> PAGE_SHIFT;
10251
10252         if (first_ppage >= next_ppage)
10253                 return 0;
10254         nr_pages = next_ppage - first_ppage;
10255
10256         first_ppage_reported = first_ppage;
10257         if (bsi->start == 0)
10258                 first_ppage_reported++;
10259         if (bsi->lowest_ppage > first_ppage_reported)
10260                 bsi->lowest_ppage = first_ppage_reported;
10261         if (bsi->highest_ppage < (next_ppage - 1))
10262                 bsi->highest_ppage = next_ppage - 1;
10263
10264         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
10265         if (ret < 0)
10266                 return ret;
10267         bsi->nr_extents += ret;
10268         bsi->nr_pages += nr_pages;
10269         return 0;
10270 }
10271
10272 static void btrfs_swap_deactivate(struct file *file)
10273 {
10274         struct inode *inode = file_inode(file);
10275
10276         btrfs_free_swapfile_pins(inode);
10277         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
10278 }
10279
10280 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10281                                sector_t *span)
10282 {
10283         struct inode *inode = file_inode(file);
10284         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10285         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
10286         struct extent_state *cached_state = NULL;
10287         struct extent_map *em = NULL;
10288         struct btrfs_device *device = NULL;
10289         struct btrfs_swap_info bsi = {
10290                 .lowest_ppage = (sector_t)-1ULL,
10291         };
10292         int ret = 0;
10293         u64 isize;
10294         u64 start;
10295
10296         /*
10297          * If the swap file was just created, make sure delalloc is done. If the
10298          * file changes again after this, the user is doing something stupid and
10299          * we don't really care.
10300          */
10301         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
10302         if (ret)
10303                 return ret;
10304
10305         /*
10306          * The inode is locked, so these flags won't change after we check them.
10307          */
10308         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
10309                 btrfs_warn(fs_info, "swapfile must not be compressed");
10310                 return -EINVAL;
10311         }
10312         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
10313                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
10314                 return -EINVAL;
10315         }
10316         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
10317                 btrfs_warn(fs_info, "swapfile must not be checksummed");
10318                 return -EINVAL;
10319         }
10320
10321         /*
10322          * Balance or device remove/replace/resize can move stuff around from
10323          * under us. The EXCL_OP flag makes sure they aren't running/won't run
10324          * concurrently while we are mapping the swap extents, and
10325          * fs_info->swapfile_pins prevents them from running while the swap file
10326          * is active and moving the extents. Note that this also prevents a
10327          * concurrent device add which isn't actually necessary, but it's not
10328          * really worth the trouble to allow it.
10329          */
10330         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
10331                 btrfs_warn(fs_info,
10332            "cannot activate swapfile while exclusive operation is running");
10333                 return -EBUSY;
10334         }
10335         /*
10336          * Snapshots can create extents which require COW even if NODATACOW is
10337          * set. We use this counter to prevent snapshots. We must increment it
10338          * before walking the extents because we don't want a concurrent
10339          * snapshot to run after we've already checked the extents.
10340          */
10341         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
10342
10343         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
10344
10345         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
10346         start = 0;
10347         while (start < isize) {
10348                 u64 logical_block_start, physical_block_start;
10349                 struct btrfs_block_group *bg;
10350                 u64 len = isize - start;
10351
10352                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len);
10353                 if (IS_ERR(em)) {
10354                         ret = PTR_ERR(em);
10355                         goto out;
10356                 }
10357
10358                 if (em->block_start == EXTENT_MAP_HOLE) {
10359                         btrfs_warn(fs_info, "swapfile must not have holes");
10360                         ret = -EINVAL;
10361                         goto out;
10362                 }
10363                 if (em->block_start == EXTENT_MAP_INLINE) {
10364                         /*
10365                          * It's unlikely we'll ever actually find ourselves
10366                          * here, as a file small enough to fit inline won't be
10367                          * big enough to store more than the swap header, but in
10368                          * case something changes in the future, let's catch it
10369                          * here rather than later.
10370                          */
10371                         btrfs_warn(fs_info, "swapfile must not be inline");
10372                         ret = -EINVAL;
10373                         goto out;
10374                 }
10375                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
10376                         btrfs_warn(fs_info, "swapfile must not be compressed");
10377                         ret = -EINVAL;
10378                         goto out;
10379                 }
10380
10381                 logical_block_start = em->block_start + (start - em->start);
10382                 len = min(len, em->len - (start - em->start));
10383                 free_extent_map(em);
10384                 em = NULL;
10385
10386                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
10387                 if (ret < 0) {
10388                         goto out;
10389                 } else if (ret) {
10390                         ret = 0;
10391                 } else {
10392                         btrfs_warn(fs_info,
10393                                    "swapfile must not be copy-on-write");
10394                         ret = -EINVAL;
10395                         goto out;
10396                 }
10397
10398                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
10399                 if (IS_ERR(em)) {
10400                         ret = PTR_ERR(em);
10401                         goto out;
10402                 }
10403
10404                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
10405                         btrfs_warn(fs_info,
10406                                    "swapfile must have single data profile");
10407                         ret = -EINVAL;
10408                         goto out;
10409                 }
10410
10411                 if (device == NULL) {
10412                         device = em->map_lookup->stripes[0].dev;
10413                         ret = btrfs_add_swapfile_pin(inode, device, false);
10414                         if (ret == 1)
10415                                 ret = 0;
10416                         else if (ret)
10417                                 goto out;
10418                 } else if (device != em->map_lookup->stripes[0].dev) {
10419                         btrfs_warn(fs_info, "swapfile must be on one device");
10420                         ret = -EINVAL;
10421                         goto out;
10422                 }
10423
10424                 physical_block_start = (em->map_lookup->stripes[0].physical +
10425                                         (logical_block_start - em->start));
10426                 len = min(len, em->len - (logical_block_start - em->start));
10427                 free_extent_map(em);
10428                 em = NULL;
10429
10430                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10431                 if (!bg) {
10432                         btrfs_warn(fs_info,
10433                            "could not find block group containing swapfile");
10434                         ret = -EINVAL;
10435                         goto out;
10436                 }
10437
10438                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10439                 if (ret) {
10440                         btrfs_put_block_group(bg);
10441                         if (ret == 1)
10442                                 ret = 0;
10443                         else
10444                                 goto out;
10445                 }
10446
10447                 if (bsi.block_len &&
10448                     bsi.block_start + bsi.block_len == physical_block_start) {
10449                         bsi.block_len += len;
10450                 } else {
10451                         if (bsi.block_len) {
10452                                 ret = btrfs_add_swap_extent(sis, &bsi);
10453                                 if (ret)
10454                                         goto out;
10455                         }
10456                         bsi.start = start;
10457                         bsi.block_start = physical_block_start;
10458                         bsi.block_len = len;
10459                 }
10460
10461                 start += len;
10462         }
10463
10464         if (bsi.block_len)
10465                 ret = btrfs_add_swap_extent(sis, &bsi);
10466
10467 out:
10468         if (!IS_ERR_OR_NULL(em))
10469                 free_extent_map(em);
10470
10471         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10472
10473         if (ret)
10474                 btrfs_swap_deactivate(file);
10475
10476         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10477
10478         if (ret)
10479                 return ret;
10480
10481         if (device)
10482                 sis->bdev = device->bdev;
10483         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10484         sis->max = bsi.nr_pages;
10485         sis->pages = bsi.nr_pages - 1;
10486         sis->highest_bit = bsi.nr_pages - 1;
10487         return bsi.nr_extents;
10488 }
10489 #else
10490 static void btrfs_swap_deactivate(struct file *file)
10491 {
10492 }
10493
10494 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10495                                sector_t *span)
10496 {
10497         return -EOPNOTSUPP;
10498 }
10499 #endif
10500
10501 static const struct inode_operations btrfs_dir_inode_operations = {
10502         .getattr        = btrfs_getattr,
10503         .lookup         = btrfs_lookup,
10504         .create         = btrfs_create,
10505         .unlink         = btrfs_unlink,
10506         .link           = btrfs_link,
10507         .mkdir          = btrfs_mkdir,
10508         .rmdir          = btrfs_rmdir,
10509         .rename         = btrfs_rename2,
10510         .symlink        = btrfs_symlink,
10511         .setattr        = btrfs_setattr,
10512         .mknod          = btrfs_mknod,
10513         .listxattr      = btrfs_listxattr,
10514         .permission     = btrfs_permission,
10515         .get_acl        = btrfs_get_acl,
10516         .set_acl        = btrfs_set_acl,
10517         .update_time    = btrfs_update_time,
10518         .tmpfile        = btrfs_tmpfile,
10519 };
10520
10521 static const struct file_operations btrfs_dir_file_operations = {
10522         .llseek         = generic_file_llseek,
10523         .read           = generic_read_dir,
10524         .iterate_shared = btrfs_real_readdir,
10525         .open           = btrfs_opendir,
10526         .unlocked_ioctl = btrfs_ioctl,
10527 #ifdef CONFIG_COMPAT
10528         .compat_ioctl   = btrfs_compat_ioctl,
10529 #endif
10530         .release        = btrfs_release_file,
10531         .fsync          = btrfs_sync_file,
10532 };
10533
10534 static const struct extent_io_ops btrfs_extent_io_ops = {
10535         /* mandatory callbacks */
10536         .submit_bio_hook = btrfs_submit_bio_hook,
10537         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10538 };
10539
10540 /*
10541  * btrfs doesn't support the bmap operation because swapfiles
10542  * use bmap to make a mapping of extents in the file.  They assume
10543  * these extents won't change over the life of the file and they
10544  * use the bmap result to do IO directly to the drive.
10545  *
10546  * the btrfs bmap call would return logical addresses that aren't
10547  * suitable for IO and they also will change frequently as COW
10548  * operations happen.  So, swapfile + btrfs == corruption.
10549  *
10550  * For now we're avoiding this by dropping bmap.
10551  */
10552 static const struct address_space_operations btrfs_aops = {
10553         .readpage       = btrfs_readpage,
10554         .writepage      = btrfs_writepage,
10555         .writepages     = btrfs_writepages,
10556         .readpages      = btrfs_readpages,
10557         .direct_IO      = btrfs_direct_IO,
10558         .invalidatepage = btrfs_invalidatepage,
10559         .releasepage    = btrfs_releasepage,
10560 #ifdef CONFIG_MIGRATION
10561         .migratepage    = btrfs_migratepage,
10562 #endif
10563         .set_page_dirty = btrfs_set_page_dirty,
10564         .error_remove_page = generic_error_remove_page,
10565         .swap_activate  = btrfs_swap_activate,
10566         .swap_deactivate = btrfs_swap_deactivate,
10567 };
10568
10569 static const struct inode_operations btrfs_file_inode_operations = {
10570         .getattr        = btrfs_getattr,
10571         .setattr        = btrfs_setattr,
10572         .listxattr      = btrfs_listxattr,
10573         .permission     = btrfs_permission,
10574         .fiemap         = btrfs_fiemap,
10575         .get_acl        = btrfs_get_acl,
10576         .set_acl        = btrfs_set_acl,
10577         .update_time    = btrfs_update_time,
10578 };
10579 static const struct inode_operations btrfs_special_inode_operations = {
10580         .getattr        = btrfs_getattr,
10581         .setattr        = btrfs_setattr,
10582         .permission     = btrfs_permission,
10583         .listxattr      = btrfs_listxattr,
10584         .get_acl        = btrfs_get_acl,
10585         .set_acl        = btrfs_set_acl,
10586         .update_time    = btrfs_update_time,
10587 };
10588 static const struct inode_operations btrfs_symlink_inode_operations = {
10589         .get_link       = page_get_link,
10590         .getattr        = btrfs_getattr,
10591         .setattr        = btrfs_setattr,
10592         .permission     = btrfs_permission,
10593         .listxattr      = btrfs_listxattr,
10594         .update_time    = btrfs_update_time,
10595 };
10596
10597 const struct dentry_operations btrfs_dentry_operations = {
10598         .d_delete       = btrfs_dentry_delete,
10599 };