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