fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include <linux/magic.h>
  46 #include <linux/iversion.h>
  47 #include "ctree.h"
  48 #include "disk-io.h"
  49 #include "transaction.h"
  50 #include "btrfs_inode.h"
  51 #include "print-tree.h"
  52 #include "ordered-data.h"
  53 #include "xattr.h"
  54 #include "tree-log.h"
  55 #include "volumes.h"
  56 #include "compression.h"
  57 #include "locking.h"
  58 #include "free-space-cache.h"
  59 #include "inode-map.h"
  60 #include "backref.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63 #include "dedupe.h"
  64
  65 struct btrfs_iget_args {
  66         struct btrfs_key *location;
  67         struct btrfs_root *root;
  68 };
  69
  70 struct btrfs_dio_data {
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74         int overwrite;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, u64 delalloc_end,
 109                                    int *page_started, unsigned long *nr_written,
 110                                    int unlock, struct btrfs_dedupe_hash *hash);
 111 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 112                                        u64 orig_start, u64 block_start,
 113                                        u64 block_len, u64 orig_block_len,
 114                                        u64 ram_bytes, int compress_type,
 115                                        int type);
 116
 117 static void __endio_write_update_ordered(struct inode *inode,
 118                                          const u64 offset, const u64 bytes,
 119                                          const bool uptodate);
 120
 121 /*
 122  * Cleanup all submitted ordered extents in specified range to handle errors
 123  * from the fill_dellaloc() callback.
 124  *
 125  * NOTE: caller must ensure that when an error happens, it can not call
 126  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 127  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 128  * to be released, which we want to happen only when finishing the ordered
 129  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 130  * fill_delalloc() callback already does proper cleanup for the first page of
 131  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 132  * range of the first page.
 133  */
 134 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 135                                                  const u64 offset,
 136                                                  const u64 bytes)
 137 {
 138         unsigned long index = offset >> PAGE_SHIFT;
 139         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 140         struct page *page;
 141
 142         while (index <= end_index) {
 143                 page = find_get_page(inode->i_mapping, index);
 144                 index++;
 145                 if (!page)
 146                         continue;
 147                 ClearPagePrivate2(page);
 148                 put_page(page);
 149         }
 150         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 151                                             bytes - PAGE_SIZE, false);
 152 }
 153
 154 static int btrfs_dirty_inode(struct inode *inode);
 155
 156 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 157 void btrfs_test_inode_set_ops(struct inode *inode)
 158 {
 159         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 160 }
 161 #endif
 162
 163 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 164                                      struct inode *inode,  struct inode *dir,
 165                                      const struct qstr *qstr)
 166 {
 167         int err;
 168
 169         err = btrfs_init_acl(trans, inode, dir);
 170         if (!err)
 171                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 172         return err;
 173 }
 174
 175 /*
 176  * this does all the hard work for inserting an inline extent into
 177  * the btree.  The caller should have done a btrfs_drop_extents so that
 178  * no overlapping inline items exist in the btree
 179  */
 180 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 181                                 struct btrfs_path *path, int extent_inserted,
 182                                 struct btrfs_root *root, struct inode *inode,
 183                                 u64 start, size_t size, size_t compressed_size,
 184                                 int compress_type,
 185                                 struct page **compressed_pages)
 186 {
 187         struct extent_buffer *leaf;
 188         struct page *page = NULL;
 189         char *kaddr;
 190         unsigned long ptr;
 191         struct btrfs_file_extent_item *ei;
 192         int ret;
 193         size_t cur_size = size;
 194         unsigned long offset;
 195
 196         if (compressed_size && compressed_pages)
 197                 cur_size = compressed_size;
 198
 199         inode_add_bytes(inode, size);
 200
 201         if (!extent_inserted) {
 202                 struct btrfs_key key;
 203                 size_t datasize;
 204
 205                 key.objectid = btrfs_ino(BTRFS_I(inode));
 206                 key.offset = start;
 207                 key.type = BTRFS_EXTENT_DATA_KEY;
 208
 209                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 210                 path->leave_spinning = 1;
 211                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 212                                               datasize);
 213                 if (ret)
 214                         goto fail;
 215         }
 216         leaf = path->nodes[0];
 217         ei = btrfs_item_ptr(leaf, path->slots[0],
 218                             struct btrfs_file_extent_item);
 219         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 220         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 221         btrfs_set_file_extent_encryption(leaf, ei, 0);
 222         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 223         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 224         ptr = btrfs_file_extent_inline_start(ei);
 225
 226         if (compress_type != BTRFS_COMPRESS_NONE) {
 227                 struct page *cpage;
 228                 int i = 0;
 229                 while (compressed_size > 0) {
 230                         cpage = compressed_pages[i];
 231                         cur_size = min_t(unsigned long, compressed_size,
 232                                        PAGE_SIZE);
 233
 234                         kaddr = kmap_atomic(cpage);
 235                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 236                         kunmap_atomic(kaddr);
 237
 238                         i++;
 239                         ptr += cur_size;
 240                         compressed_size -= cur_size;
 241                 }
 242                 btrfs_set_file_extent_compression(leaf, ei,
 243                                                   compress_type);
 244         } else {
 245                 page = find_get_page(inode->i_mapping,
 246                                      start >> PAGE_SHIFT);
 247                 btrfs_set_file_extent_compression(leaf, ei, 0);
 248                 kaddr = kmap_atomic(page);
 249                 offset = start & (PAGE_SIZE - 1);
 250                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 251                 kunmap_atomic(kaddr);
 252                 put_page(page);
 253         }
 254         btrfs_mark_buffer_dirty(leaf);
 255         btrfs_release_path(path);
 256
 257         /*
 258          * we're an inline extent, so nobody can
 259          * extend the file past i_size without locking
 260          * a page we already have locked.
 261          *
 262          * We must do any isize and inode updates
 263          * before we unlock the pages.  Otherwise we
 264          * could end up racing with unlink.
 265          */
 266         BTRFS_I(inode)->disk_i_size = inode->i_size;
 267         ret = btrfs_update_inode(trans, root, inode);
 268
 269 fail:
 270         return ret;
 271 }
 272
 273
 274 /*
 275  * conditionally insert an inline extent into the file.  This
 276  * does the checks required to make sure the data is small enough
 277  * to fit as an inline extent.
 278  */
 279 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 280                                           u64 end, size_t compressed_size,
 281                                           int compress_type,
 282                                           struct page **compressed_pages)
 283 {
 284         struct btrfs_root *root = BTRFS_I(inode)->root;
 285         struct btrfs_fs_info *fs_info = root->fs_info;
 286         struct btrfs_trans_handle *trans;
 287         u64 isize = i_size_read(inode);
 288         u64 actual_end = min(end + 1, isize);
 289         u64 inline_len = actual_end - start;
 290         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 291         u64 data_len = inline_len;
 292         int ret;
 293         struct btrfs_path *path;
 294         int extent_inserted = 0;
 295         u32 extent_item_size;
 296
 297         if (compressed_size)
 298                 data_len = compressed_size;
 299
 300         if (start > 0 ||
 301             actual_end > fs_info->sectorsize ||
 302             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 303             (!compressed_size &&
 304             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 305             end + 1 < isize ||
 306             data_len > fs_info->max_inline) {
 307                 return 1;
 308         }
 309
 310         path = btrfs_alloc_path();
 311         if (!path)
 312                 return -ENOMEM;
 313
 314         trans = btrfs_join_transaction(root);
 315         if (IS_ERR(trans)) {
 316                 btrfs_free_path(path);
 317                 return PTR_ERR(trans);
 318         }
 319         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 320
 321         if (compressed_size && compressed_pages)
 322                 extent_item_size = btrfs_file_extent_calc_inline_size(
 323                    compressed_size);
 324         else
 325                 extent_item_size = btrfs_file_extent_calc_inline_size(
 326                     inline_len);
 327
 328         ret = __btrfs_drop_extents(trans, root, inode, path,
 329                                    start, aligned_end, NULL,
 330                                    1, 1, extent_item_size, &extent_inserted);
 331         if (ret) {
 332                 btrfs_abort_transaction(trans, ret);
 333                 goto out;
 334         }
 335
 336         if (isize > actual_end)
 337                 inline_len = min_t(u64, isize, actual_end);
 338         ret = insert_inline_extent(trans, path, extent_inserted,
 339                                    root, inode, start,
 340                                    inline_len, compressed_size,
 341                                    compress_type, compressed_pages);
 342         if (ret && ret != -ENOSPC) {
 343                 btrfs_abort_transaction(trans, ret);
 344                 goto out;
 345         } else if (ret == -ENOSPC) {
 346                 ret = 1;
 347                 goto out;
 348         }
 349
 350         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 351         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 352 out:
 353         /*
 354          * Don't forget to free the reserved space, as for inlined extent
 355          * it won't count as data extent, free them directly here.
 356          * And at reserve time, it's always aligned to page size, so
 357          * just free one page here.
 358          */
 359         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 360         btrfs_free_path(path);
 361         btrfs_end_transaction(trans);
 362         return ret;
 363 }
 364
 365 struct async_extent {
 366         u64 start;
 367         u64 ram_size;
 368         u64 compressed_size;
 369         struct page **pages;
 370         unsigned long nr_pages;
 371         int compress_type;
 372         struct list_head list;
 373 };
 374
 375 struct async_cow {
 376         struct inode *inode;
 377         struct btrfs_root *root;
 378         struct page *locked_page;
 379         u64 start;
 380         u64 end;
 381         unsigned int write_flags;
 382         struct list_head extents;
 383         struct btrfs_work work;
 384 };
 385
 386 static noinline int add_async_extent(struct async_cow *cow,
 387                                      u64 start, u64 ram_size,
 388                                      u64 compressed_size,
 389                                      struct page **pages,
 390                                      unsigned long nr_pages,
 391                                      int compress_type)
 392 {
 393         struct async_extent *async_extent;
 394
 395         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 396         BUG_ON(!async_extent); /* -ENOMEM */
 397         async_extent->start = start;
 398         async_extent->ram_size = ram_size;
 399         async_extent->compressed_size = compressed_size;
 400         async_extent->pages = pages;
 401         async_extent->nr_pages = nr_pages;
 402         async_extent->compress_type = compress_type;
 403         list_add_tail(&async_extent->list, &cow->extents);
 404         return 0;
 405 }
 406
 407 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 408 {
 409         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 410
 411         /* force compress */
 412         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 413                 return 1;
 414         /* defrag ioctl */
 415         if (BTRFS_I(inode)->defrag_compress)
 416                 return 1;
 417         /* bad compression ratios */
 418         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 419                 return 0;
 420         if (btrfs_test_opt(fs_info, COMPRESS) ||
 421             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 422             BTRFS_I(inode)->prop_compress)
 423                 return btrfs_compress_heuristic(inode, start, end);
 424         return 0;
 425 }
 426
 427 static inline void inode_should_defrag(struct btrfs_inode *inode,
 428                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 429 {
 430         /* If this is a small write inside eof, kick off a defrag */
 431         if (num_bytes < small_write &&
 432             (start > 0 || end + 1 < inode->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434 }
 435
 436 /*
 437  * we create compressed extents in two phases.  The first
 438  * phase compresses a range of pages that have already been
 439  * locked (both pages and state bits are locked).
 440  *
 441  * This is done inside an ordered work queue, and the compression
 442  * is spread across many cpus.  The actual IO submission is step
 443  * two, and the ordered work queue takes care of making sure that
 444  * happens in the same order things were put onto the queue by
 445  * writepages and friends.
 446  *
 447  * If this code finds it can't get good compression, it puts an
 448  * entry onto the work queue to write the uncompressed bytes.  This
 449  * makes sure that both compressed inodes and uncompressed inodes
 450  * are written in the same order that the flusher thread sent them
 451  * down.
 452  */
 453 static noinline void compress_file_range(struct inode *inode,
 454                                         struct page *locked_page,
 455                                         u64 start, u64 end,
 456                                         struct async_cow *async_cow,
 457                                         int *num_added)
 458 {
 459         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 460         u64 blocksize = fs_info->sectorsize;
 461         u64 actual_end;
 462         u64 isize = i_size_read(inode);
 463         int ret = 0;
 464         struct page **pages = NULL;
 465         unsigned long nr_pages;
 466         unsigned long total_compressed = 0;
 467         unsigned long total_in = 0;
 468         int i;
 469         int will_compress;
 470         int compress_type = fs_info->compress_type;
 471         int redirty = 0;
 472
 473         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 474                         SZ_16K);
 475
 476         actual_end = min_t(u64, isize, end + 1);
 477 again:
 478         will_compress = 0;
 479         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 480         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 481         nr_pages = min_t(unsigned long, nr_pages,
 482                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 483
 484         /*
 485          * we don't want to send crud past the end of i_size through
 486          * compression, that's just a waste of CPU time.  So, if the
 487          * end of the file is before the start of our current
 488          * requested range of bytes, we bail out to the uncompressed
 489          * cleanup code that can deal with all of this.
 490          *
 491          * It isn't really the fastest way to fix things, but this is a
 492          * very uncommon corner.
 493          */
 494         if (actual_end <= start)
 495                 goto cleanup_and_bail_uncompressed;
 496
 497         total_compressed = actual_end - start;
 498
 499         /*
 500          * skip compression for a small file range(<=blocksize) that
 501          * isn't an inline extent, since it doesn't save disk space at all.
 502          */
 503         if (total_compressed <= blocksize &&
 504            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 505                 goto cleanup_and_bail_uncompressed;
 506
 507         total_compressed = min_t(unsigned long, total_compressed,
 508                         BTRFS_MAX_UNCOMPRESSED);
 509         total_in = 0;
 510         ret = 0;
 511
 512         /*
 513          * we do compression for mount -o compress and when the
 514          * inode has not been flagged as nocompress.  This flag can
 515          * change at any time if we discover bad compression ratios.
 516          */
 517         if (inode_need_compress(inode, start, end)) {
 518                 WARN_ON(pages);
 519                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 520                 if (!pages) {
 521                         /* just bail out to the uncompressed code */
 522                         goto cont;
 523                 }
 524
 525                 if (BTRFS_I(inode)->defrag_compress)
 526                         compress_type = BTRFS_I(inode)->defrag_compress;
 527                 else if (BTRFS_I(inode)->prop_compress)
 528                         compress_type = BTRFS_I(inode)->prop_compress;
 529
 530                 /*
 531                  * we need to call clear_page_dirty_for_io on each
 532                  * page in the range.  Otherwise applications with the file
 533                  * mmap'd can wander in and change the page contents while
 534                  * we are compressing them.
 535                  *
 536                  * If the compression fails for any reason, we set the pages
 537                  * dirty again later on.
 538                  *
 539                  * Note that the remaining part is redirtied, the start pointer
 540                  * has moved, the end is the original one.
 541                  */
 542                 if (!redirty) {
 543                         extent_range_clear_dirty_for_io(inode, start, end);
 544                         redirty = 1;
 545                 }
 546
 547                 /* Compression level is applied here and only here */
 548                 ret = btrfs_compress_pages(
 549                         compress_type | (fs_info->compress_level << 4),
 550                                            inode->i_mapping, start,
 551                                            pages,
 552                                            &nr_pages,
 553                                            &total_in,
 554                                            &total_compressed);
 555
 556                 if (!ret) {
 557                         unsigned long offset = total_compressed &
 558                                 (PAGE_SIZE - 1);
 559                         struct page *page = pages[nr_pages - 1];
 560                         char *kaddr;
 561
 562                         /* zero the tail end of the last page, we might be
 563                          * sending it down to disk
 564                          */
 565                         if (offset) {
 566                                 kaddr = kmap_atomic(page);
 567                                 memset(kaddr + offset, 0,
 568                                        PAGE_SIZE - offset);
 569                                 kunmap_atomic(kaddr);
 570                         }
 571                         will_compress = 1;
 572                 }
 573         }
 574 cont:
 575         if (start == 0) {
 576                 /* lets try to make an inline extent */
 577                 if (ret || total_in < actual_end) {
 578                         /* we didn't compress the entire range, try
 579                          * to make an uncompressed inline extent.
 580                          */
 581                         ret = cow_file_range_inline(inode, start, end, 0,
 582                                                     BTRFS_COMPRESS_NONE, NULL);
 583                 } else {
 584                         /* try making a compressed inline extent */
 585                         ret = cow_file_range_inline(inode, start, end,
 586                                                     total_compressed,
 587                                                     compress_type, pages);
 588                 }
 589                 if (ret <= 0) {
 590                         unsigned long clear_flags = EXTENT_DELALLOC |
 591                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 592                                 EXTENT_DO_ACCOUNTING;
 593                         unsigned long page_error_op;
 594
 595                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 596
 597                         /*
 598                          * inline extent creation worked or returned error,
 599                          * we don't need to create any more async work items.
 600                          * Unlock and free up our temp pages.
 601                          *
 602                          * We use DO_ACCOUNTING here because we need the
 603                          * delalloc_release_metadata to be done _after_ we drop
 604                          * our outstanding extent for clearing delalloc for this
 605                          * range.
 606                          */
 607                         extent_clear_unlock_delalloc(inode, start, end, end,
 608                                                      NULL, clear_flags,
 609                                                      PAGE_UNLOCK |
 610                                                      PAGE_CLEAR_DIRTY |
 611                                                      PAGE_SET_WRITEBACK |
 612                                                      page_error_op |
 613                                                      PAGE_END_WRITEBACK);
 614                         goto free_pages_out;
 615                 }
 616         }
 617
 618         if (will_compress) {
 619                 /*
 620                  * we aren't doing an inline extent round the compressed size
 621                  * up to a block size boundary so the allocator does sane
 622                  * things
 623                  */
 624                 total_compressed = ALIGN(total_compressed, blocksize);
 625
 626                 /*
 627                  * one last check to make sure the compression is really a
 628                  * win, compare the page count read with the blocks on disk,
 629                  * compression must free at least one sector size
 630                  */
 631                 total_in = ALIGN(total_in, PAGE_SIZE);
 632                 if (total_compressed + blocksize <= total_in) {
 633                         *num_added += 1;
 634
 635                         /*
 636                          * The async work queues will take care of doing actual
 637                          * allocation on disk for these compressed pages, and
 638                          * will submit them to the elevator.
 639                          */
 640                         add_async_extent(async_cow, start, total_in,
 641                                         total_compressed, pages, nr_pages,
 642                                         compress_type);
 643
 644                         if (start + total_in < end) {
 645                                 start += total_in;
 646                                 pages = NULL;
 647                                 cond_resched();
 648                                 goto again;
 649                         }
 650                         return;
 651                 }
 652         }
 653         if (pages) {
 654                 /*
 655                  * the compression code ran but failed to make things smaller,
 656                  * free any pages it allocated and our page pointer array
 657                  */
 658                 for (i = 0; i < nr_pages; i++) {
 659                         WARN_ON(pages[i]->mapping);
 660                         put_page(pages[i]);
 661                 }
 662                 kfree(pages);
 663                 pages = NULL;
 664                 total_compressed = 0;
 665                 nr_pages = 0;
 666
 667                 /* flag the file so we don't compress in the future */
 668                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 669                     !(BTRFS_I(inode)->prop_compress)) {
 670                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 671                 }
 672         }
 673 cleanup_and_bail_uncompressed:
 674         /*
 675          * No compression, but we still need to write the pages in the file
 676          * we've been given so far.  redirty the locked page if it corresponds
 677          * to our extent and set things up for the async work queue to run
 678          * cow_file_range to do the normal delalloc dance.
 679          */
 680         if (page_offset(locked_page) >= start &&
 681             page_offset(locked_page) <= end)
 682                 __set_page_dirty_nobuffers(locked_page);
 683                 /* unlocked later on in the async handlers */
 684
 685         if (redirty)
 686                 extent_range_redirty_for_io(inode, start, end);
 687         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 688                          BTRFS_COMPRESS_NONE);
 689         *num_added += 1;
 690
 691         return;
 692
 693 free_pages_out:
 694         for (i = 0; i < nr_pages; i++) {
 695                 WARN_ON(pages[i]->mapping);
 696                 put_page(pages[i]);
 697         }
 698         kfree(pages);
 699 }
 700
 701 static void free_async_extent_pages(struct async_extent *async_extent)
 702 {
 703         int i;
 704
 705         if (!async_extent->pages)
 706                 return;
 707
 708         for (i = 0; i < async_extent->nr_pages; i++) {
 709                 WARN_ON(async_extent->pages[i]->mapping);
 710                 put_page(async_extent->pages[i]);
 711         }
 712         kfree(async_extent->pages);
 713         async_extent->nr_pages = 0;
 714         async_extent->pages = NULL;
 715 }
 716
 717 /*
 718  * phase two of compressed writeback.  This is the ordered portion
 719  * of the code, which only gets called in the order the work was
 720  * queued.  We walk all the async extents created by compress_file_range
 721  * and send them down to the disk.
 722  */
 723 static noinline void submit_compressed_extents(struct inode *inode,
 724                                               struct async_cow *async_cow)
 725 {
 726         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 727         struct async_extent *async_extent;
 728         u64 alloc_hint = 0;
 729         struct btrfs_key ins;
 730         struct extent_map *em;
 731         struct btrfs_root *root = BTRFS_I(inode)->root;
 732         struct extent_io_tree *io_tree;
 733         int ret = 0;
 734
 735 again:
 736         while (!list_empty(&async_cow->extents)) {
 737                 async_extent = list_entry(async_cow->extents.next,
 738                                           struct async_extent, list);
 739                 list_del(&async_extent->list);
 740
 741                 io_tree = &BTRFS_I(inode)->io_tree;
 742
 743 retry:
 744                 /* did the compression code fall back to uncompressed IO? */
 745                 if (!async_extent->pages) {
 746                         int page_started = 0;
 747                         unsigned long nr_written = 0;
 748
 749                         lock_extent(io_tree, async_extent->start,
 750                                          async_extent->start +
 751                                          async_extent->ram_size - 1);
 752
 753                         /* allocate blocks */
 754                         ret = cow_file_range(inode, async_cow->locked_page,
 755                                              async_extent->start,
 756                                              async_extent->start +
 757                                              async_extent->ram_size - 1,
 758                                              async_extent->start +
 759                                              async_extent->ram_size - 1,
 760                                              &page_started, &nr_written, 0,
 761                                              NULL);
 762
 763                         /* JDM XXX */
 764
 765                         /*
 766                          * if page_started, cow_file_range inserted an
 767                          * inline extent and took care of all the unlocking
 768                          * and IO for us.  Otherwise, we need to submit
 769                          * all those pages down to the drive.
 770                          */
 771                         if (!page_started && !ret)
 772                                 extent_write_locked_range(inode,
 773                                                   async_extent->start,
 774                                                   async_extent->start +
 775                                                   async_extent->ram_size - 1,
 776                                                   WB_SYNC_ALL);
 777                         else if (ret)
 778                                 unlock_page(async_cow->locked_page);
 779                         kfree(async_extent);
 780                         cond_resched();
 781                         continue;
 782                 }
 783
 784                 lock_extent(io_tree, async_extent->start,
 785                             async_extent->start + async_extent->ram_size - 1);
 786
 787                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 788                                            async_extent->compressed_size,
 789                                            async_extent->compressed_size,
 790                                            0, alloc_hint, &ins, 1, 1);
 791                 if (ret) {
 792                         free_async_extent_pages(async_extent);
 793
 794                         if (ret == -ENOSPC) {
 795                                 unlock_extent(io_tree, async_extent->start,
 796                                               async_extent->start +
 797                                               async_extent->ram_size - 1);
 798
 799                                 /*
 800                                  * we need to redirty the pages if we decide to
 801                                  * fallback to uncompressed IO, otherwise we
 802                                  * will not submit these pages down to lower
 803                                  * layers.
 804                                  */
 805                                 extent_range_redirty_for_io(inode,
 806                                                 async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1);
 809
 810                                 goto retry;
 811                         }
 812                         goto out_free;
 813                 }
 814                 /*
 815                  * here we're doing allocation and writeback of the
 816                  * compressed pages
 817                  */
 818                 em = create_io_em(inode, async_extent->start,
 819                                   async_extent->ram_size, /* len */
 820                                   async_extent->start, /* orig_start */
 821                                   ins.objectid, /* block_start */
 822                                   ins.offset, /* block_len */
 823                                   ins.offset, /* orig_block_len */
 824                                   async_extent->ram_size, /* ram_bytes */
 825                                   async_extent->compress_type,
 826                                   BTRFS_ORDERED_COMPRESSED);
 827                 if (IS_ERR(em))
 828                         /* ret value is not necessary due to void function */
 829                         goto out_free_reserve;
 830                 free_extent_map(em);
 831
 832                 ret = btrfs_add_ordered_extent_compress(inode,
 833                                                 async_extent->start,
 834                                                 ins.objectid,
 835                                                 async_extent->ram_size,
 836                                                 ins.offset,
 837                                                 BTRFS_ORDERED_COMPRESSED,
 838                                                 async_extent->compress_type);
 839                 if (ret) {
 840                         btrfs_drop_extent_cache(BTRFS_I(inode),
 841                                                 async_extent->start,
 842                                                 async_extent->start +
 843                                                 async_extent->ram_size - 1, 0);
 844                         goto out_free_reserve;
 845                 }
 846                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 847
 848                 /*
 849                  * clear dirty, set writeback and unlock the pages.
 850                  */
 851                 extent_clear_unlock_delalloc(inode, async_extent->start,
 852                                 async_extent->start +
 853                                 async_extent->ram_size - 1,
 854                                 async_extent->start +
 855                                 async_extent->ram_size - 1,
 856                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 857                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 858                                 PAGE_SET_WRITEBACK);
 859                 if (btrfs_submit_compressed_write(inode,
 860                                     async_extent->start,
 861                                     async_extent->ram_size,
 862                                     ins.objectid,
 863                                     ins.offset, async_extent->pages,
 864                                     async_extent->nr_pages,
 865                                     async_cow->write_flags)) {
 866                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 867                         struct page *p = async_extent->pages[0];
 868                         const u64 start = async_extent->start;
 869                         const u64 end = start + async_extent->ram_size - 1;
 870
 871                         p->mapping = inode->i_mapping;
 872                         tree->ops->writepage_end_io_hook(p, start, end,
 873                                                          NULL, 0);
 874                         p->mapping = NULL;
 875                         extent_clear_unlock_delalloc(inode, start, end, end,
 876                                                      NULL, 0,
 877                                                      PAGE_END_WRITEBACK |
 878                                                      PAGE_SET_ERROR);
 879                         free_async_extent_pages(async_extent);
 880                 }
 881                 alloc_hint = ins.objectid + ins.offset;
 882                 kfree(async_extent);
 883                 cond_resched();
 884         }
 885         return;
 886 out_free_reserve:
 887         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 888         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 889 out_free:
 890         extent_clear_unlock_delalloc(inode, async_extent->start,
 891                                      async_extent->start +
 892                                      async_extent->ram_size - 1,
 893                                      async_extent->start +
 894                                      async_extent->ram_size - 1,
 895                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 896                                      EXTENT_DELALLOC_NEW |
 897                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 898                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 899                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 900                                      PAGE_SET_ERROR);
 901         free_async_extent_pages(async_extent);
 902         kfree(async_extent);
 903         goto again;
 904 }
 905
 906 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 907                                       u64 num_bytes)
 908 {
 909         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 910         struct extent_map *em;
 911         u64 alloc_hint = 0;
 912
 913         read_lock(&em_tree->lock);
 914         em = search_extent_mapping(em_tree, start, num_bytes);
 915         if (em) {
 916                 /*
 917                  * if block start isn't an actual block number then find the
 918                  * first block in this inode and use that as a hint.  If that
 919                  * block is also bogus then just don't worry about it.
 920                  */
 921                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 922                         free_extent_map(em);
 923                         em = search_extent_mapping(em_tree, 0, 0);
 924                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 925                                 alloc_hint = em->block_start;
 926                         if (em)
 927                                 free_extent_map(em);
 928                 } else {
 929                         alloc_hint = em->block_start;
 930                         free_extent_map(em);
 931                 }
 932         }
 933         read_unlock(&em_tree->lock);
 934
 935         return alloc_hint;
 936 }
 937
 938 /*
 939  * when extent_io.c finds a delayed allocation range in the file,
 940  * the call backs end up in this code.  The basic idea is to
 941  * allocate extents on disk for the range, and create ordered data structs
 942  * in ram to track those extents.
 943  *
 944  * locked_page is the page that writepage had locked already.  We use
 945  * it to make sure we don't do extra locks or unlocks.
 946  *
 947  * *page_started is set to one if we unlock locked_page and do everything
 948  * required to start IO on it.  It may be clean and already done with
 949  * IO when we return.
 950  */
 951 static noinline int cow_file_range(struct inode *inode,
 952                                    struct page *locked_page,
 953                                    u64 start, u64 end, u64 delalloc_end,
 954                                    int *page_started, unsigned long *nr_written,
 955                                    int unlock, struct btrfs_dedupe_hash *hash)
 956 {
 957         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 958         struct btrfs_root *root = BTRFS_I(inode)->root;
 959         u64 alloc_hint = 0;
 960         u64 num_bytes;
 961         unsigned long ram_size;
 962         u64 cur_alloc_size = 0;
 963         u64 blocksize = fs_info->sectorsize;
 964         struct btrfs_key ins;
 965         struct extent_map *em;
 966         unsigned clear_bits;
 967         unsigned long page_ops;
 968         bool extent_reserved = false;
 969         int ret = 0;
 970
 971         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 972                 WARN_ON_ONCE(1);
 973                 ret = -EINVAL;
 974                 goto out_unlock;
 975         }
 976
 977         num_bytes = ALIGN(end - start + 1, blocksize);
 978         num_bytes = max(blocksize,  num_bytes);
 979         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 980
 981         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 982
 983         if (start == 0) {
 984                 /* lets try to make an inline extent */
 985                 ret = cow_file_range_inline(inode, start, end, 0,
 986                                             BTRFS_COMPRESS_NONE, NULL);
 987                 if (ret == 0) {
 988                         /*
 989                          * We use DO_ACCOUNTING here because we need the
 990                          * delalloc_release_metadata to be run _after_ we drop
 991                          * our outstanding extent for clearing delalloc for this
 992                          * range.
 993                          */
 994                         extent_clear_unlock_delalloc(inode, start, end,
 995                                      delalloc_end, NULL,
 996                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 997                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 998                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 999                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1000                                      PAGE_END_WRITEBACK);
1001                         *nr_written = *nr_written +
1002                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1003                         *page_started = 1;
1004                         goto out;
1005                 } else if (ret < 0) {
1006                         goto out_unlock;
1007                 }
1008         }
1009
1010         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1011         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1012                         start + num_bytes - 1, 0);
1013
1014         while (num_bytes > 0) {
1015                 cur_alloc_size = num_bytes;
1016                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1017                                            fs_info->sectorsize, 0, alloc_hint,
1018                                            &ins, 1, 1);
1019                 if (ret < 0)
1020                         goto out_unlock;
1021                 cur_alloc_size = ins.offset;
1022                 extent_reserved = true;
1023
1024                 ram_size = ins.offset;
1025                 em = create_io_em(inode, start, ins.offset, /* len */
1026                                   start, /* orig_start */
1027                                   ins.objectid, /* block_start */
1028                                   ins.offset, /* block_len */
1029                                   ins.offset, /* orig_block_len */
1030                                   ram_size, /* ram_bytes */
1031                                   BTRFS_COMPRESS_NONE, /* compress_type */
1032                                   BTRFS_ORDERED_REGULAR /* type */);
1033                 if (IS_ERR(em))
1034                         goto out_reserve;
1035                 free_extent_map(em);
1036
1037                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1038                                                ram_size, cur_alloc_size, 0);
1039                 if (ret)
1040                         goto out_drop_extent_cache;
1041
1042                 if (root->root_key.objectid ==
1043                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1044                         ret = btrfs_reloc_clone_csums(inode, start,
1045                                                       cur_alloc_size);
1046                         /*
1047                          * Only drop cache here, and process as normal.
1048                          *
1049                          * We must not allow extent_clear_unlock_delalloc()
1050                          * at out_unlock label to free meta of this ordered
1051                          * extent, as its meta should be freed by
1052                          * btrfs_finish_ordered_io().
1053                          *
1054                          * So we must continue until @start is increased to
1055                          * skip current ordered extent.
1056                          */
1057                         if (ret)
1058                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1059                                                 start + ram_size - 1, 0);
1060                 }
1061
1062                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1063
1064                 /* we're not doing compressed IO, don't unlock the first
1065                  * page (which the caller expects to stay locked), don't
1066                  * clear any dirty bits and don't set any writeback bits
1067                  *
1068                  * Do set the Private2 bit so we know this page was properly
1069                  * setup for writepage
1070                  */
1071                 page_ops = unlock ? PAGE_UNLOCK : 0;
1072                 page_ops |= PAGE_SET_PRIVATE2;
1073
1074                 extent_clear_unlock_delalloc(inode, start,
1075                                              start + ram_size - 1,
1076                                              delalloc_end, locked_page,
1077                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1078                                              page_ops);
1079                 if (num_bytes < cur_alloc_size)
1080                         num_bytes = 0;
1081                 else
1082                         num_bytes -= cur_alloc_size;
1083                 alloc_hint = ins.objectid + ins.offset;
1084                 start += cur_alloc_size;
1085                 extent_reserved = false;
1086
1087                 /*
1088                  * btrfs_reloc_clone_csums() error, since start is increased
1089                  * extent_clear_unlock_delalloc() at out_unlock label won't
1090                  * free metadata of current ordered extent, we're OK to exit.
1091                  */
1092                 if (ret)
1093                         goto out_unlock;
1094         }
1095 out:
1096         return ret;
1097
1098 out_drop_extent_cache:
1099         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1100 out_reserve:
1101         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1102         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1103 out_unlock:
1104         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1105                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1106         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1107                 PAGE_END_WRITEBACK;
1108         /*
1109          * If we reserved an extent for our delalloc range (or a subrange) and
1110          * failed to create the respective ordered extent, then it means that
1111          * when we reserved the extent we decremented the extent's size from
1112          * the data space_info's bytes_may_use counter and incremented the
1113          * space_info's bytes_reserved counter by the same amount. We must make
1114          * sure extent_clear_unlock_delalloc() does not try to decrement again
1115          * the data space_info's bytes_may_use counter, therefore we do not pass
1116          * it the flag EXTENT_CLEAR_DATA_RESV.
1117          */
1118         if (extent_reserved) {
1119                 extent_clear_unlock_delalloc(inode, start,
1120                                              start + cur_alloc_size,
1121                                              start + cur_alloc_size,
1122                                              locked_page,
1123                                              clear_bits,
1124                                              page_ops);
1125                 start += cur_alloc_size;
1126                 if (start >= end)
1127                         goto out;
1128         }
1129         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1130                                      locked_page,
1131                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1132                                      page_ops);
1133         goto out;
1134 }
1135
1136 /*
1137  * work queue call back to started compression on a file and pages
1138  */
1139 static noinline void async_cow_start(struct btrfs_work *work)
1140 {
1141         struct async_cow *async_cow;
1142         int num_added = 0;
1143         async_cow = container_of(work, struct async_cow, work);
1144
1145         compress_file_range(async_cow->inode, async_cow->locked_page,
1146                             async_cow->start, async_cow->end, async_cow,
1147                             &num_added);
1148         if (num_added == 0) {
1149                 btrfs_add_delayed_iput(async_cow->inode);
1150                 async_cow->inode = NULL;
1151         }
1152 }
1153
1154 /*
1155  * work queue call back to submit previously compressed pages
1156  */
1157 static noinline void async_cow_submit(struct btrfs_work *work)
1158 {
1159         struct btrfs_fs_info *fs_info;
1160         struct async_cow *async_cow;
1161         struct btrfs_root *root;
1162         unsigned long nr_pages;
1163
1164         async_cow = container_of(work, struct async_cow, work);
1165
1166         root = async_cow->root;
1167         fs_info = root->fs_info;
1168         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1169                 PAGE_SHIFT;
1170
1171         /*
1172          * atomic_sub_return implies a barrier for waitqueue_active
1173          */
1174         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1175             5 * SZ_1M &&
1176             waitqueue_active(&fs_info->async_submit_wait))
1177                 wake_up(&fs_info->async_submit_wait);
1178
1179         if (async_cow->inode)
1180                 submit_compressed_extents(async_cow->inode, async_cow);
1181 }
1182
1183 static noinline void async_cow_free(struct btrfs_work *work)
1184 {
1185         struct async_cow *async_cow;
1186         async_cow = container_of(work, struct async_cow, work);
1187         if (async_cow->inode)
1188                 btrfs_add_delayed_iput(async_cow->inode);
1189         kfree(async_cow);
1190 }
1191
1192 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1193                                 u64 start, u64 end, int *page_started,
1194                                 unsigned long *nr_written,
1195                                 unsigned int write_flags)
1196 {
1197         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1198         struct async_cow *async_cow;
1199         struct btrfs_root *root = BTRFS_I(inode)->root;
1200         unsigned long nr_pages;
1201         u64 cur_end;
1202
1203         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1204                          1, 0, NULL);
1205         while (start < end) {
1206                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1207                 BUG_ON(!async_cow); /* -ENOMEM */
1208                 async_cow->inode = igrab(inode);
1209                 async_cow->root = root;
1210                 async_cow->locked_page = locked_page;
1211                 async_cow->start = start;
1212                 async_cow->write_flags = write_flags;
1213
1214                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1215                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1216                         cur_end = end;
1217                 else
1218                         cur_end = min(end, start + SZ_512K - 1);
1219
1220                 async_cow->end = cur_end;
1221                 INIT_LIST_HEAD(&async_cow->extents);
1222
1223                 btrfs_init_work(&async_cow->work,
1224                                 btrfs_delalloc_helper,
1225                                 async_cow_start, async_cow_submit,
1226                                 async_cow_free);
1227
1228                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1229                         PAGE_SHIFT;
1230                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1231
1232                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1233
1234                 *nr_written += nr_pages;
1235                 start = cur_end + 1;
1236         }
1237         *page_started = 1;
1238         return 0;
1239 }
1240
1241 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1242                                         u64 bytenr, u64 num_bytes)
1243 {
1244         int ret;
1245         struct btrfs_ordered_sum *sums;
1246         LIST_HEAD(list);
1247
1248         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1249                                        bytenr + num_bytes - 1, &list, 0);
1250         if (ret == 0 && list_empty(&list))
1251                 return 0;
1252
1253         while (!list_empty(&list)) {
1254                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1255                 list_del(&sums->list);
1256                 kfree(sums);
1257         }
1258         if (ret < 0)
1259                 return ret;
1260         return 1;
1261 }
1262
1263 /*
1264  * when nowcow writeback call back.  This checks for snapshots or COW copies
1265  * of the extents that exist in the file, and COWs the file as required.
1266  *
1267  * If no cow copies or snapshots exist, we write directly to the existing
1268  * blocks on disk
1269  */
1270 static noinline int run_delalloc_nocow(struct inode *inode,
1271                                        struct page *locked_page,
1272                               u64 start, u64 end, int *page_started, int force,
1273                               unsigned long *nr_written)
1274 {
1275         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1276         struct btrfs_root *root = BTRFS_I(inode)->root;
1277         struct extent_buffer *leaf;
1278         struct btrfs_path *path;
1279         struct btrfs_file_extent_item *fi;
1280         struct btrfs_key found_key;
1281         struct extent_map *em;
1282         u64 cow_start;
1283         u64 cur_offset;
1284         u64 extent_end;
1285         u64 extent_offset;
1286         u64 disk_bytenr;
1287         u64 num_bytes;
1288         u64 disk_num_bytes;
1289         u64 ram_bytes;
1290         int extent_type;
1291         int ret, err;
1292         int type;
1293         int nocow;
1294         int check_prev = 1;
1295         bool nolock;
1296         u64 ino = btrfs_ino(BTRFS_I(inode));
1297
1298         path = btrfs_alloc_path();
1299         if (!path) {
1300                 extent_clear_unlock_delalloc(inode, start, end, end,
1301                                              locked_page,
1302                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1303                                              EXTENT_DO_ACCOUNTING |
1304                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1305                                              PAGE_CLEAR_DIRTY |
1306                                              PAGE_SET_WRITEBACK |
1307                                              PAGE_END_WRITEBACK);
1308                 return -ENOMEM;
1309         }
1310
1311         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1312
1313         cow_start = (u64)-1;
1314         cur_offset = start;
1315         while (1) {
1316                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1317                                                cur_offset, 0);
1318                 if (ret < 0)
1319                         goto error;
1320                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1321                         leaf = path->nodes[0];
1322                         btrfs_item_key_to_cpu(leaf, &found_key,
1323                                               path->slots[0] - 1);
1324                         if (found_key.objectid == ino &&
1325                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1326                                 path->slots[0]--;
1327                 }
1328                 check_prev = 0;
1329 next_slot:
1330                 leaf = path->nodes[0];
1331                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1332                         ret = btrfs_next_leaf(root, path);
1333                         if (ret < 0) {
1334                                 if (cow_start != (u64)-1)
1335                                         cur_offset = cow_start;
1336                                 goto error;
1337                         }
1338                         if (ret > 0)
1339                                 break;
1340                         leaf = path->nodes[0];
1341                 }
1342
1343                 nocow = 0;
1344                 disk_bytenr = 0;
1345                 num_bytes = 0;
1346                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1347
1348                 if (found_key.objectid > ino)
1349                         break;
1350                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1351                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1352                         path->slots[0]++;
1353                         goto next_slot;
1354                 }
1355                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1356                     found_key.offset > end)
1357                         break;
1358
1359                 if (found_key.offset > cur_offset) {
1360                         extent_end = found_key.offset;
1361                         extent_type = 0;
1362                         goto out_check;
1363                 }
1364
1365                 fi = btrfs_item_ptr(leaf, path->slots[0],
1366                                     struct btrfs_file_extent_item);
1367                 extent_type = btrfs_file_extent_type(leaf, fi);
1368
1369                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1370                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1371                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1372                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1373                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1374                         extent_end = found_key.offset +
1375                                 btrfs_file_extent_num_bytes(leaf, fi);
1376                         disk_num_bytes =
1377                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1378                         if (extent_end <= start) {
1379                                 path->slots[0]++;
1380                                 goto next_slot;
1381                         }
1382                         if (disk_bytenr == 0)
1383                                 goto out_check;
1384                         if (btrfs_file_extent_compression(leaf, fi) ||
1385                             btrfs_file_extent_encryption(leaf, fi) ||
1386                             btrfs_file_extent_other_encoding(leaf, fi))
1387                                 goto out_check;
1388                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1389                                 goto out_check;
1390                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1391                                 goto out_check;
1392                         ret = btrfs_cross_ref_exist(root, ino,
1393                                                     found_key.offset -
1394                                                     extent_offset, disk_bytenr);
1395                         if (ret) {
1396                                 /*
1397                                  * ret could be -EIO if the above fails to read
1398                                  * metadata.
1399                                  */
1400                                 if (ret < 0) {
1401                                         if (cow_start != (u64)-1)
1402                                                 cur_offset = cow_start;
1403                                         goto error;
1404                                 }
1405
1406                                 WARN_ON_ONCE(nolock);
1407                                 goto out_check;
1408                         }
1409                         disk_bytenr += extent_offset;
1410                         disk_bytenr += cur_offset - found_key.offset;
1411                         num_bytes = min(end + 1, extent_end) - cur_offset;
1412                         /*
1413                          * if there are pending snapshots for this root,
1414                          * we fall into common COW way.
1415                          */
1416                         if (!nolock) {
1417                                 err = btrfs_start_write_no_snapshotting(root);
1418                                 if (!err)
1419                                         goto out_check;
1420                         }
1421                         /*
1422                          * force cow if csum exists in the range.
1423                          * this ensure that csum for a given extent are
1424                          * either valid or do not exist.
1425                          */
1426                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1427                                                   num_bytes);
1428                         if (ret) {
1429                                 if (!nolock)
1430                                         btrfs_end_write_no_snapshotting(root);
1431
1432                                 /*
1433                                  * ret could be -EIO if the above fails to read
1434                                  * metadata.
1435                                  */
1436                                 if (ret < 0) {
1437                                         if (cow_start != (u64)-1)
1438                                                 cur_offset = cow_start;
1439                                         goto error;
1440                                 }
1441                                 WARN_ON_ONCE(nolock);
1442                                 goto out_check;
1443                         }
1444                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1445                                 if (!nolock)
1446                                         btrfs_end_write_no_snapshotting(root);
1447                                 goto out_check;
1448                         }
1449                         nocow = 1;
1450                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1451                         extent_end = found_key.offset +
1452                                 btrfs_file_extent_inline_len(leaf,
1453                                                      path->slots[0], fi);
1454                         extent_end = ALIGN(extent_end,
1455                                            fs_info->sectorsize);
1456                 } else {
1457                         BUG_ON(1);
1458                 }
1459 out_check:
1460                 if (extent_end <= start) {
1461                         path->slots[0]++;
1462                         if (!nolock && nocow)
1463                                 btrfs_end_write_no_snapshotting(root);
1464                         if (nocow)
1465                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1466                         goto next_slot;
1467                 }
1468                 if (!nocow) {
1469                         if (cow_start == (u64)-1)
1470                                 cow_start = cur_offset;
1471                         cur_offset = extent_end;
1472                         if (cur_offset > end)
1473                                 break;
1474                         path->slots[0]++;
1475                         goto next_slot;
1476                 }
1477
1478                 btrfs_release_path(path);
1479                 if (cow_start != (u64)-1) {
1480                         ret = cow_file_range(inode, locked_page,
1481                                              cow_start, found_key.offset - 1,
1482                                              end, page_started, nr_written, 1,
1483                                              NULL);
1484                         if (ret) {
1485                                 if (!nolock && nocow)
1486                                         btrfs_end_write_no_snapshotting(root);
1487                                 if (nocow)
1488                                         btrfs_dec_nocow_writers(fs_info,
1489                                                                 disk_bytenr);
1490                                 goto error;
1491                         }
1492                         cow_start = (u64)-1;
1493                 }
1494
1495                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1496                         u64 orig_start = found_key.offset - extent_offset;
1497
1498                         em = create_io_em(inode, cur_offset, num_bytes,
1499                                           orig_start,
1500                                           disk_bytenr, /* block_start */
1501                                           num_bytes, /* block_len */
1502                                           disk_num_bytes, /* orig_block_len */
1503                                           ram_bytes, BTRFS_COMPRESS_NONE,
1504                                           BTRFS_ORDERED_PREALLOC);
1505                         if (IS_ERR(em)) {
1506                                 if (!nolock && nocow)
1507                                         btrfs_end_write_no_snapshotting(root);
1508                                 if (nocow)
1509                                         btrfs_dec_nocow_writers(fs_info,
1510                                                                 disk_bytenr);
1511                                 ret = PTR_ERR(em);
1512                                 goto error;
1513                         }
1514                         free_extent_map(em);
1515                 }
1516
1517                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1518                         type = BTRFS_ORDERED_PREALLOC;
1519                 } else {
1520                         type = BTRFS_ORDERED_NOCOW;
1521                 }
1522
1523                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1524                                                num_bytes, num_bytes, type);
1525                 if (nocow)
1526                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1527                 BUG_ON(ret); /* -ENOMEM */
1528
1529                 if (root->root_key.objectid ==
1530                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1531                         /*
1532                          * Error handled later, as we must prevent
1533                          * extent_clear_unlock_delalloc() in error handler
1534                          * from freeing metadata of created ordered extent.
1535                          */
1536                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1537                                                       num_bytes);
1538
1539                 extent_clear_unlock_delalloc(inode, cur_offset,
1540                                              cur_offset + num_bytes - 1, end,
1541                                              locked_page, EXTENT_LOCKED |
1542                                              EXTENT_DELALLOC |
1543                                              EXTENT_CLEAR_DATA_RESV,
1544                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1545
1546                 if (!nolock && nocow)
1547                         btrfs_end_write_no_snapshotting(root);
1548                 cur_offset = extent_end;
1549
1550                 /*
1551                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1552                  * handler, as metadata for created ordered extent will only
1553                  * be freed by btrfs_finish_ordered_io().
1554                  */
1555                 if (ret)
1556                         goto error;
1557                 if (cur_offset > end)
1558                         break;
1559         }
1560         btrfs_release_path(path);
1561
1562         if (cur_offset <= end && cow_start == (u64)-1) {
1563                 cow_start = cur_offset;
1564                 cur_offset = end;
1565         }
1566
1567         if (cow_start != (u64)-1) {
1568                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1569                                      page_started, nr_written, 1, NULL);
1570                 if (ret)
1571                         goto error;
1572         }
1573
1574 error:
1575         if (ret && cur_offset < end)
1576                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1577                                              locked_page, EXTENT_LOCKED |
1578                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1579                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1580                                              PAGE_CLEAR_DIRTY |
1581                                              PAGE_SET_WRITEBACK |
1582                                              PAGE_END_WRITEBACK);
1583         btrfs_free_path(path);
1584         return ret;
1585 }
1586
1587 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1588 {
1589
1590         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1591             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1592                 return 0;
1593
1594         /*
1595          * @defrag_bytes is a hint value, no spinlock held here,
1596          * if is not zero, it means the file is defragging.
1597          * Force cow if given extent needs to be defragged.
1598          */
1599         if (BTRFS_I(inode)->defrag_bytes &&
1600             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1601                            EXTENT_DEFRAG, 0, NULL))
1602                 return 1;
1603
1604         return 0;
1605 }
1606
1607 /*
1608  * extent_io.c call back to do delayed allocation processing
1609  */
1610 static int run_delalloc_range(void *private_data, struct page *locked_page,
1611                               u64 start, u64 end, int *page_started,
1612                               unsigned long *nr_written,
1613                               struct writeback_control *wbc)
1614 {
1615         struct inode *inode = private_data;
1616         int ret;
1617         int force_cow = need_force_cow(inode, start, end);
1618         unsigned int write_flags = wbc_to_write_flags(wbc);
1619
1620         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1621                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1622                                          page_started, 1, nr_written);
1623         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1624                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1625                                          page_started, 0, nr_written);
1626         } else if (!inode_need_compress(inode, start, end)) {
1627                 ret = cow_file_range(inode, locked_page, start, end, end,
1628                                       page_started, nr_written, 1, NULL);
1629         } else {
1630                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1631                         &BTRFS_I(inode)->runtime_flags);
1632                 ret = cow_file_range_async(inode, locked_page, start, end,
1633                                            page_started, nr_written,
1634                                            write_flags);
1635         }
1636         if (ret)
1637                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1638         return ret;
1639 }
1640
1641 static void btrfs_split_extent_hook(void *private_data,
1642                                     struct extent_state *orig, u64 split)
1643 {
1644         struct inode *inode = private_data;
1645         u64 size;
1646
1647         /* not delalloc, ignore it */
1648         if (!(orig->state & EXTENT_DELALLOC))
1649                 return;
1650
1651         size = orig->end - orig->start + 1;
1652         if (size > BTRFS_MAX_EXTENT_SIZE) {
1653                 u32 num_extents;
1654                 u64 new_size;
1655
1656                 /*
1657                  * See the explanation in btrfs_merge_extent_hook, the same
1658                  * applies here, just in reverse.
1659                  */
1660                 new_size = orig->end - split + 1;
1661                 num_extents = count_max_extents(new_size);
1662                 new_size = split - orig->start;
1663                 num_extents += count_max_extents(new_size);
1664                 if (count_max_extents(size) >= num_extents)
1665                         return;
1666         }
1667
1668         spin_lock(&BTRFS_I(inode)->lock);
1669         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1670         spin_unlock(&BTRFS_I(inode)->lock);
1671 }
1672
1673 /*
1674  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1675  * extents so we can keep track of new extents that are just merged onto old
1676  * extents, such as when we are doing sequential writes, so we can properly
1677  * account for the metadata space we'll need.
1678  */
1679 static void btrfs_merge_extent_hook(void *private_data,
1680                                     struct extent_state *new,
1681                                     struct extent_state *other)
1682 {
1683         struct inode *inode = private_data;
1684         u64 new_size, old_size;
1685         u32 num_extents;
1686
1687         /* not delalloc, ignore it */
1688         if (!(other->state & EXTENT_DELALLOC))
1689                 return;
1690
1691         if (new->start > other->start)
1692                 new_size = new->end - other->start + 1;
1693         else
1694                 new_size = other->end - new->start + 1;
1695
1696         /* we're not bigger than the max, unreserve the space and go */
1697         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1698                 spin_lock(&BTRFS_I(inode)->lock);
1699                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1700                 spin_unlock(&BTRFS_I(inode)->lock);
1701                 return;
1702         }
1703
1704         /*
1705          * We have to add up either side to figure out how many extents were
1706          * accounted for before we merged into one big extent.  If the number of
1707          * extents we accounted for is <= the amount we need for the new range
1708          * then we can return, otherwise drop.  Think of it like this
1709          *
1710          * [ 4k][MAX_SIZE]
1711          *
1712          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1713          * need 2 outstanding extents, on one side we have 1 and the other side
1714          * we have 1 so they are == and we can return.  But in this case
1715          *
1716          * [MAX_SIZE+4k][MAX_SIZE+4k]
1717          *
1718          * Each range on their own accounts for 2 extents, but merged together
1719          * they are only 3 extents worth of accounting, so we need to drop in
1720          * this case.
1721          */
1722         old_size = other->end - other->start + 1;
1723         num_extents = count_max_extents(old_size);
1724         old_size = new->end - new->start + 1;
1725         num_extents += count_max_extents(old_size);
1726         if (count_max_extents(new_size) >= num_extents)
1727                 return;
1728
1729         spin_lock(&BTRFS_I(inode)->lock);
1730         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1731         spin_unlock(&BTRFS_I(inode)->lock);
1732 }
1733
1734 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1735                                       struct inode *inode)
1736 {
1737         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1738
1739         spin_lock(&root->delalloc_lock);
1740         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1741                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1742                               &root->delalloc_inodes);
1743                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1744                         &BTRFS_I(inode)->runtime_flags);
1745                 root->nr_delalloc_inodes++;
1746                 if (root->nr_delalloc_inodes == 1) {
1747                         spin_lock(&fs_info->delalloc_root_lock);
1748                         BUG_ON(!list_empty(&root->delalloc_root));
1749                         list_add_tail(&root->delalloc_root,
1750                                       &fs_info->delalloc_roots);
1751                         spin_unlock(&fs_info->delalloc_root_lock);
1752                 }
1753         }
1754         spin_unlock(&root->delalloc_lock);
1755 }
1756
1757 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1758                                      struct btrfs_inode *inode)
1759 {
1760         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1761
1762         spin_lock(&root->delalloc_lock);
1763         if (!list_empty(&inode->delalloc_inodes)) {
1764                 list_del_init(&inode->delalloc_inodes);
1765                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1766                           &inode->runtime_flags);
1767                 root->nr_delalloc_inodes--;
1768                 if (!root->nr_delalloc_inodes) {
1769                         spin_lock(&fs_info->delalloc_root_lock);
1770                         BUG_ON(list_empty(&root->delalloc_root));
1771                         list_del_init(&root->delalloc_root);
1772                         spin_unlock(&fs_info->delalloc_root_lock);
1773                 }
1774         }
1775         spin_unlock(&root->delalloc_lock);
1776 }
1777
1778 /*
1779  * extent_io.c set_bit_hook, used to track delayed allocation
1780  * bytes in this file, and to maintain the list of inodes that
1781  * have pending delalloc work to be done.
1782  */
1783 static void btrfs_set_bit_hook(void *private_data,
1784                                struct extent_state *state, unsigned *bits)
1785 {
1786         struct inode *inode = private_data;
1787
1788         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1789
1790         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1791                 WARN_ON(1);
1792         /*
1793          * set_bit and clear bit hooks normally require _irqsave/restore
1794          * but in this case, we are only testing for the DELALLOC
1795          * bit, which is only set or cleared with irqs on
1796          */
1797         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1798                 struct btrfs_root *root = BTRFS_I(inode)->root;
1799                 u64 len = state->end + 1 - state->start;
1800                 u32 num_extents = count_max_extents(len);
1801                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1802
1803                 spin_lock(&BTRFS_I(inode)->lock);
1804                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1805                 spin_unlock(&BTRFS_I(inode)->lock);
1806
1807                 /* For sanity tests */
1808                 if (btrfs_is_testing(fs_info))
1809                         return;
1810
1811                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1812                                          fs_info->delalloc_batch);
1813                 spin_lock(&BTRFS_I(inode)->lock);
1814                 BTRFS_I(inode)->delalloc_bytes += len;
1815                 if (*bits & EXTENT_DEFRAG)
1816                         BTRFS_I(inode)->defrag_bytes += len;
1817                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1818                                          &BTRFS_I(inode)->runtime_flags))
1819                         btrfs_add_delalloc_inodes(root, inode);
1820                 spin_unlock(&BTRFS_I(inode)->lock);
1821         }
1822
1823         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1824             (*bits & EXTENT_DELALLOC_NEW)) {
1825                 spin_lock(&BTRFS_I(inode)->lock);
1826                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1827                         state->start;
1828                 spin_unlock(&BTRFS_I(inode)->lock);
1829         }
1830 }
1831
1832 /*
1833  * extent_io.c clear_bit_hook, see set_bit_hook for why
1834  */
1835 static void btrfs_clear_bit_hook(void *private_data,
1836                                  struct extent_state *state,
1837                                  unsigned *bits)
1838 {
1839         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1840         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1841         u64 len = state->end + 1 - state->start;
1842         u32 num_extents = count_max_extents(len);
1843
1844         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1845                 spin_lock(&inode->lock);
1846                 inode->defrag_bytes -= len;
1847                 spin_unlock(&inode->lock);
1848         }
1849
1850         /*
1851          * set_bit and clear bit hooks normally require _irqsave/restore
1852          * but in this case, we are only testing for the DELALLOC
1853          * bit, which is only set or cleared with irqs on
1854          */
1855         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1856                 struct btrfs_root *root = inode->root;
1857                 bool do_list = !btrfs_is_free_space_inode(inode);
1858
1859                 spin_lock(&inode->lock);
1860                 btrfs_mod_outstanding_extents(inode, -num_extents);
1861                 spin_unlock(&inode->lock);
1862
1863                 /*
1864                  * We don't reserve metadata space for space cache inodes so we
1865                  * don't need to call dellalloc_release_metadata if there is an
1866                  * error.
1867                  */
1868                 if (*bits & EXTENT_CLEAR_META_RESV &&
1869                     root != fs_info->tree_root)
1870                         btrfs_delalloc_release_metadata(inode, len);
1871
1872                 /* For sanity tests. */
1873                 if (btrfs_is_testing(fs_info))
1874                         return;
1875
1876                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1877                     do_list && !(state->state & EXTENT_NORESERVE) &&
1878                     (*bits & EXTENT_CLEAR_DATA_RESV))
1879                         btrfs_free_reserved_data_space_noquota(
1880                                         &inode->vfs_inode,
1881                                         state->start, len);
1882
1883                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1884                                          fs_info->delalloc_batch);
1885                 spin_lock(&inode->lock);
1886                 inode->delalloc_bytes -= len;
1887                 if (do_list && inode->delalloc_bytes == 0 &&
1888                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1889                                         &inode->runtime_flags))
1890                         btrfs_del_delalloc_inode(root, inode);
1891                 spin_unlock(&inode->lock);
1892         }
1893
1894         if ((state->state & EXTENT_DELALLOC_NEW) &&
1895             (*bits & EXTENT_DELALLOC_NEW)) {
1896                 spin_lock(&inode->lock);
1897                 ASSERT(inode->new_delalloc_bytes >= len);
1898                 inode->new_delalloc_bytes -= len;
1899                 spin_unlock(&inode->lock);
1900         }
1901 }
1902
1903 /*
1904  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1905  * we don't create bios that span stripes or chunks
1906  *
1907  * return 1 if page cannot be merged to bio
1908  * return 0 if page can be merged to bio
1909  * return error otherwise
1910  */
1911 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1912                          size_t size, struct bio *bio,
1913                          unsigned long bio_flags)
1914 {
1915         struct inode *inode = page->mapping->host;
1916         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1917         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1918         u64 length = 0;
1919         u64 map_length;
1920         int ret;
1921
1922         if (bio_flags & EXTENT_BIO_COMPRESSED)
1923                 return 0;
1924
1925         length = bio->bi_iter.bi_size;
1926         map_length = length;
1927         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1928                               NULL, 0);
1929         if (ret < 0)
1930                 return ret;
1931         if (map_length < length + size)
1932                 return 1;
1933         return 0;
1934 }
1935
1936 /*
1937  * in order to insert checksums into the metadata in large chunks,
1938  * we wait until bio submission time.   All the pages in the bio are
1939  * checksummed and sums are attached onto the ordered extent record.
1940  *
1941  * At IO completion time the cums attached on the ordered extent record
1942  * are inserted into the btree
1943  */
1944 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
1945                                     u64 bio_offset)
1946 {
1947         struct inode *inode = private_data;
1948         blk_status_t ret = 0;
1949
1950         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1951         BUG_ON(ret); /* -ENOMEM */
1952         return 0;
1953 }
1954
1955 /*
1956  * in order to insert checksums into the metadata in large chunks,
1957  * we wait until bio submission time.   All the pages in the bio are
1958  * checksummed and sums are attached onto the ordered extent record.
1959  *
1960  * At IO completion time the cums attached on the ordered extent record
1961  * are inserted into the btree
1962  */
1963 static blk_status_t btrfs_submit_bio_done(void *private_data, struct bio *bio,
1964                           int mirror_num)
1965 {
1966         struct inode *inode = private_data;
1967         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1968         blk_status_t ret;
1969
1970         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1971         if (ret) {
1972                 bio->bi_status = ret;
1973                 bio_endio(bio);
1974         }
1975         return ret;
1976 }
1977
1978 /*
1979  * extent_io.c submission hook. This does the right thing for csum calculation
1980  * on write, or reading the csums from the tree before a read.
1981  *
1982  * Rules about async/sync submit,
1983  * a) read:                             sync submit
1984  *
1985  * b) write without checksum:           sync submit
1986  *
1987  * c) write with checksum:
1988  *    c-1) if bio is issued by fsync:   sync submit
1989  *         (sync_writers != 0)
1990  *
1991  *    c-2) if root is reloc root:       sync submit
1992  *         (only in case of buffered IO)
1993  *
1994  *    c-3) otherwise:                   async submit
1995  */
1996 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1997                                  int mirror_num, unsigned long bio_flags,
1998                                  u64 bio_offset)
1999 {
2000         struct inode *inode = private_data;
2001         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2002         struct btrfs_root *root = BTRFS_I(inode)->root;
2003         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2004         blk_status_t ret = 0;
2005         int skip_sum;
2006         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2007
2008         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2009
2010         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2011                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2012
2013         if (bio_op(bio) != REQ_OP_WRITE) {
2014                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2015                 if (ret)
2016                         goto out;
2017
2018                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2019                         ret = btrfs_submit_compressed_read(inode, bio,
2020                                                            mirror_num,
2021                                                            bio_flags);
2022                         goto out;
2023                 } else if (!skip_sum) {
2024                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2025                         if (ret)
2026                                 goto out;
2027                 }
2028                 goto mapit;
2029         } else if (async && !skip_sum) {
2030                 /* csum items have already been cloned */
2031                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2032                         goto mapit;
2033                 /* we're doing a write, do the async checksumming */
2034                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2035                                           bio_offset, inode,
2036                                           btrfs_submit_bio_start,
2037                                           btrfs_submit_bio_done);
2038                 goto out;
2039         } else if (!skip_sum) {
2040                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2041                 if (ret)
2042                         goto out;
2043         }
2044
2045 mapit:
2046         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2047
2048 out:
2049         if (ret) {
2050                 bio->bi_status = ret;
2051                 bio_endio(bio);
2052         }
2053         return ret;
2054 }
2055
2056 /*
2057  * given a list of ordered sums record them in the inode.  This happens
2058  * at IO completion time based on sums calculated at bio submission time.
2059  */
2060 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2061                              struct inode *inode, struct list_head *list)
2062 {
2063         struct btrfs_ordered_sum *sum;
2064         int ret;
2065
2066         list_for_each_entry(sum, list, list) {
2067                 trans->adding_csums = true;
2068                 ret = btrfs_csum_file_blocks(trans,
2069                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2070                 trans->adding_csums = false;
2071                 if (ret)
2072                         return ret;
2073         }
2074         return 0;
2075 }
2076
2077 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2078                               unsigned int extra_bits,
2079                               struct extent_state **cached_state, int dedupe)
2080 {
2081         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2082         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2083                                    extra_bits, cached_state);
2084 }
2085
2086 /* see btrfs_writepage_start_hook for details on why this is required */
2087 struct btrfs_writepage_fixup {
2088         struct page *page;
2089         struct btrfs_work work;
2090 };
2091
2092 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2093 {
2094         struct btrfs_writepage_fixup *fixup;
2095         struct btrfs_ordered_extent *ordered;
2096         struct extent_state *cached_state = NULL;
2097         struct extent_changeset *data_reserved = NULL;
2098         struct page *page;
2099         struct inode *inode;
2100         u64 page_start;
2101         u64 page_end;
2102         int ret;
2103
2104         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2105         page = fixup->page;
2106 again:
2107         lock_page(page);
2108         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2109                 ClearPageChecked(page);
2110                 goto out_page;
2111         }
2112
2113         inode = page->mapping->host;
2114         page_start = page_offset(page);
2115         page_end = page_offset(page) + PAGE_SIZE - 1;
2116
2117         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2118                          &cached_state);
2119
2120         /* already ordered? We're done */
2121         if (PagePrivate2(page))
2122                 goto out;
2123
2124         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2125                                         PAGE_SIZE);
2126         if (ordered) {
2127                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2128                                      page_end, &cached_state);
2129                 unlock_page(page);
2130                 btrfs_start_ordered_extent(inode, ordered, 1);
2131                 btrfs_put_ordered_extent(ordered);
2132                 goto again;
2133         }
2134
2135         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2136                                            PAGE_SIZE);
2137         if (ret) {
2138                 mapping_set_error(page->mapping, ret);
2139                 end_extent_writepage(page, ret, page_start, page_end);
2140                 ClearPageChecked(page);
2141                 goto out;
2142          }
2143
2144         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2145                                         &cached_state, 0);
2146         if (ret) {
2147                 mapping_set_error(page->mapping, ret);
2148                 end_extent_writepage(page, ret, page_start, page_end);
2149                 ClearPageChecked(page);
2150                 goto out;
2151         }
2152
2153         ClearPageChecked(page);
2154         set_page_dirty(page);
2155         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2156 out:
2157         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2158                              &cached_state);
2159 out_page:
2160         unlock_page(page);
2161         put_page(page);
2162         kfree(fixup);
2163         extent_changeset_free(data_reserved);
2164 }
2165
2166 /*
2167  * There are a few paths in the higher layers of the kernel that directly
2168  * set the page dirty bit without asking the filesystem if it is a
2169  * good idea.  This causes problems because we want to make sure COW
2170  * properly happens and the data=ordered rules are followed.
2171  *
2172  * In our case any range that doesn't have the ORDERED bit set
2173  * hasn't been properly setup for IO.  We kick off an async process
2174  * to fix it up.  The async helper will wait for ordered extents, set
2175  * the delalloc bit and make it safe to write the page.
2176  */
2177 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2178 {
2179         struct inode *inode = page->mapping->host;
2180         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2181         struct btrfs_writepage_fixup *fixup;
2182
2183         /* this page is properly in the ordered list */
2184         if (TestClearPagePrivate2(page))
2185                 return 0;
2186
2187         if (PageChecked(page))
2188                 return -EAGAIN;
2189
2190         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2191         if (!fixup)
2192                 return -EAGAIN;
2193
2194         SetPageChecked(page);
2195         get_page(page);
2196         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2197                         btrfs_writepage_fixup_worker, NULL, NULL);
2198         fixup->page = page;
2199         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2200         return -EBUSY;
2201 }
2202
2203 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2204                                        struct inode *inode, u64 file_pos,
2205                                        u64 disk_bytenr, u64 disk_num_bytes,
2206                                        u64 num_bytes, u64 ram_bytes,
2207                                        u8 compression, u8 encryption,
2208                                        u16 other_encoding, int extent_type)
2209 {
2210         struct btrfs_root *root = BTRFS_I(inode)->root;
2211         struct btrfs_file_extent_item *fi;
2212         struct btrfs_path *path;
2213         struct extent_buffer *leaf;
2214         struct btrfs_key ins;
2215         u64 qg_released;
2216         int extent_inserted = 0;
2217         int ret;
2218
2219         path = btrfs_alloc_path();
2220         if (!path)
2221                 return -ENOMEM;
2222
2223         /*
2224          * we may be replacing one extent in the tree with another.
2225          * The new extent is pinned in the extent map, and we don't want
2226          * to drop it from the cache until it is completely in the btree.
2227          *
2228          * So, tell btrfs_drop_extents to leave this extent in the cache.
2229          * the caller is expected to unpin it and allow it to be merged
2230          * with the others.
2231          */
2232         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2233                                    file_pos + num_bytes, NULL, 0,
2234                                    1, sizeof(*fi), &extent_inserted);
2235         if (ret)
2236                 goto out;
2237
2238         if (!extent_inserted) {
2239                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2240                 ins.offset = file_pos;
2241                 ins.type = BTRFS_EXTENT_DATA_KEY;
2242
2243                 path->leave_spinning = 1;
2244                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2245                                               sizeof(*fi));
2246                 if (ret)
2247                         goto out;
2248         }
2249         leaf = path->nodes[0];
2250         fi = btrfs_item_ptr(leaf, path->slots[0],
2251                             struct btrfs_file_extent_item);
2252         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2253         btrfs_set_file_extent_type(leaf, fi, extent_type);
2254         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2255         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2256         btrfs_set_file_extent_offset(leaf, fi, 0);
2257         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2258         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2259         btrfs_set_file_extent_compression(leaf, fi, compression);
2260         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2261         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2262
2263         btrfs_mark_buffer_dirty(leaf);
2264         btrfs_release_path(path);
2265
2266         inode_add_bytes(inode, num_bytes);
2267
2268         ins.objectid = disk_bytenr;
2269         ins.offset = disk_num_bytes;
2270         ins.type = BTRFS_EXTENT_ITEM_KEY;
2271
2272         /*
2273          * Release the reserved range from inode dirty range map, as it is
2274          * already moved into delayed_ref_head
2275          */
2276         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2277         if (ret < 0)
2278                 goto out;
2279         qg_released = ret;
2280         ret = btrfs_alloc_reserved_file_extent(trans, root,
2281                                                btrfs_ino(BTRFS_I(inode)),
2282                                                file_pos, qg_released, &ins);
2283 out:
2284         btrfs_free_path(path);
2285
2286         return ret;
2287 }
2288
2289 /* snapshot-aware defrag */
2290 struct sa_defrag_extent_backref {
2291         struct rb_node node;
2292         struct old_sa_defrag_extent *old;
2293         u64 root_id;
2294         u64 inum;
2295         u64 file_pos;
2296         u64 extent_offset;
2297         u64 num_bytes;
2298         u64 generation;
2299 };
2300
2301 struct old_sa_defrag_extent {
2302         struct list_head list;
2303         struct new_sa_defrag_extent *new;
2304
2305         u64 extent_offset;
2306         u64 bytenr;
2307         u64 offset;
2308         u64 len;
2309         int count;
2310 };
2311
2312 struct new_sa_defrag_extent {
2313         struct rb_root root;
2314         struct list_head head;
2315         struct btrfs_path *path;
2316         struct inode *inode;
2317         u64 file_pos;
2318         u64 len;
2319         u64 bytenr;
2320         u64 disk_len;
2321         u8 compress_type;
2322 };
2323
2324 static int backref_comp(struct sa_defrag_extent_backref *b1,
2325                         struct sa_defrag_extent_backref *b2)
2326 {
2327         if (b1->root_id < b2->root_id)
2328                 return -1;
2329         else if (b1->root_id > b2->root_id)
2330                 return 1;
2331
2332         if (b1->inum < b2->inum)
2333                 return -1;
2334         else if (b1->inum > b2->inum)
2335                 return 1;
2336
2337         if (b1->file_pos < b2->file_pos)
2338                 return -1;
2339         else if (b1->file_pos > b2->file_pos)
2340                 return 1;
2341
2342         /*
2343          * [------------------------------] ===> (a range of space)
2344          *     |<--->|   |<---->| =============> (fs/file tree A)
2345          * |<---------------------------->| ===> (fs/file tree B)
2346          *
2347          * A range of space can refer to two file extents in one tree while
2348          * refer to only one file extent in another tree.
2349          *
2350          * So we may process a disk offset more than one time(two extents in A)
2351          * and locate at the same extent(one extent in B), then insert two same
2352          * backrefs(both refer to the extent in B).
2353          */
2354         return 0;
2355 }
2356
2357 static void backref_insert(struct rb_root *root,
2358                            struct sa_defrag_extent_backref *backref)
2359 {
2360         struct rb_node **p = &root->rb_node;
2361         struct rb_node *parent = NULL;
2362         struct sa_defrag_extent_backref *entry;
2363         int ret;
2364
2365         while (*p) {
2366                 parent = *p;
2367                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2368
2369                 ret = backref_comp(backref, entry);
2370                 if (ret < 0)
2371                         p = &(*p)->rb_left;
2372                 else
2373                         p = &(*p)->rb_right;
2374         }
2375
2376         rb_link_node(&backref->node, parent, p);
2377         rb_insert_color(&backref->node, root);
2378 }
2379
2380 /*
2381  * Note the backref might has changed, and in this case we just return 0.
2382  */
2383 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2384                                        void *ctx)
2385 {
2386         struct btrfs_file_extent_item *extent;
2387         struct old_sa_defrag_extent *old = ctx;
2388         struct new_sa_defrag_extent *new = old->new;
2389         struct btrfs_path *path = new->path;
2390         struct btrfs_key key;
2391         struct btrfs_root *root;
2392         struct sa_defrag_extent_backref *backref;
2393         struct extent_buffer *leaf;
2394         struct inode *inode = new->inode;
2395         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2396         int slot;
2397         int ret;
2398         u64 extent_offset;
2399         u64 num_bytes;
2400
2401         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2402             inum == btrfs_ino(BTRFS_I(inode)))
2403                 return 0;
2404
2405         key.objectid = root_id;
2406         key.type = BTRFS_ROOT_ITEM_KEY;
2407         key.offset = (u64)-1;
2408
2409         root = btrfs_read_fs_root_no_name(fs_info, &key);
2410         if (IS_ERR(root)) {
2411                 if (PTR_ERR(root) == -ENOENT)
2412                         return 0;
2413                 WARN_ON(1);
2414                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2415                          inum, offset, root_id);
2416                 return PTR_ERR(root);
2417         }
2418
2419         key.objectid = inum;
2420         key.type = BTRFS_EXTENT_DATA_KEY;
2421         if (offset > (u64)-1 << 32)
2422                 key.offset = 0;
2423         else
2424                 key.offset = offset;
2425
2426         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2427         if (WARN_ON(ret < 0))
2428                 return ret;
2429         ret = 0;
2430
2431         while (1) {
2432                 cond_resched();
2433
2434                 leaf = path->nodes[0];
2435                 slot = path->slots[0];
2436
2437                 if (slot >= btrfs_header_nritems(leaf)) {
2438                         ret = btrfs_next_leaf(root, path);
2439                         if (ret < 0) {
2440                                 goto out;
2441                         } else if (ret > 0) {
2442                                 ret = 0;
2443                                 goto out;
2444                         }
2445                         continue;
2446                 }
2447
2448                 path->slots[0]++;
2449
2450                 btrfs_item_key_to_cpu(leaf, &key, slot);
2451
2452                 if (key.objectid > inum)
2453                         goto out;
2454
2455                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2456                         continue;
2457
2458                 extent = btrfs_item_ptr(leaf, slot,
2459                                         struct btrfs_file_extent_item);
2460
2461                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2462                         continue;
2463
2464                 /*
2465                  * 'offset' refers to the exact key.offset,
2466                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2467                  * (key.offset - extent_offset).
2468                  */
2469                 if (key.offset != offset)
2470                         continue;
2471
2472                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2473                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2474
2475                 if (extent_offset >= old->extent_offset + old->offset +
2476                     old->len || extent_offset + num_bytes <=
2477                     old->extent_offset + old->offset)
2478                         continue;
2479                 break;
2480         }
2481
2482         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2483         if (!backref) {
2484                 ret = -ENOENT;
2485                 goto out;
2486         }
2487
2488         backref->root_id = root_id;
2489         backref->inum = inum;
2490         backref->file_pos = offset;
2491         backref->num_bytes = num_bytes;
2492         backref->extent_offset = extent_offset;
2493         backref->generation = btrfs_file_extent_generation(leaf, extent);
2494         backref->old = old;
2495         backref_insert(&new->root, backref);
2496         old->count++;
2497 out:
2498         btrfs_release_path(path);
2499         WARN_ON(ret);
2500         return ret;
2501 }
2502
2503 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2504                                    struct new_sa_defrag_extent *new)
2505 {
2506         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2507         struct old_sa_defrag_extent *old, *tmp;
2508         int ret;
2509
2510         new->path = path;
2511
2512         list_for_each_entry_safe(old, tmp, &new->head, list) {
2513                 ret = iterate_inodes_from_logical(old->bytenr +
2514                                                   old->extent_offset, fs_info,
2515                                                   path, record_one_backref,
2516                                                   old, false);
2517                 if (ret < 0 && ret != -ENOENT)
2518                         return false;
2519
2520                 /* no backref to be processed for this extent */
2521                 if (!old->count) {
2522                         list_del(&old->list);
2523                         kfree(old);
2524                 }
2525         }
2526
2527         if (list_empty(&new->head))
2528                 return false;
2529
2530         return true;
2531 }
2532
2533 static int relink_is_mergable(struct extent_buffer *leaf,
2534                               struct btrfs_file_extent_item *fi,
2535                               struct new_sa_defrag_extent *new)
2536 {
2537         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2538                 return 0;
2539
2540         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2541                 return 0;
2542
2543         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2544                 return 0;
2545
2546         if (btrfs_file_extent_encryption(leaf, fi) ||
2547             btrfs_file_extent_other_encoding(leaf, fi))
2548                 return 0;
2549
2550         return 1;
2551 }
2552
2553 /*
2554  * Note the backref might has changed, and in this case we just return 0.
2555  */
2556 static noinline int relink_extent_backref(struct btrfs_path *path,
2557                                  struct sa_defrag_extent_backref *prev,
2558                                  struct sa_defrag_extent_backref *backref)
2559 {
2560         struct btrfs_file_extent_item *extent;
2561         struct btrfs_file_extent_item *item;
2562         struct btrfs_ordered_extent *ordered;
2563         struct btrfs_trans_handle *trans;
2564         struct btrfs_root *root;
2565         struct btrfs_key key;
2566         struct extent_buffer *leaf;
2567         struct old_sa_defrag_extent *old = backref->old;
2568         struct new_sa_defrag_extent *new = old->new;
2569         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2570         struct inode *inode;
2571         struct extent_state *cached = NULL;
2572         int ret = 0;
2573         u64 start;
2574         u64 len;
2575         u64 lock_start;
2576         u64 lock_end;
2577         bool merge = false;
2578         int index;
2579
2580         if (prev && prev->root_id == backref->root_id &&
2581             prev->inum == backref->inum &&
2582             prev->file_pos + prev->num_bytes == backref->file_pos)
2583                 merge = true;
2584
2585         /* step 1: get root */
2586         key.objectid = backref->root_id;
2587         key.type = BTRFS_ROOT_ITEM_KEY;
2588         key.offset = (u64)-1;
2589
2590         index = srcu_read_lock(&fs_info->subvol_srcu);
2591
2592         root = btrfs_read_fs_root_no_name(fs_info, &key);
2593         if (IS_ERR(root)) {
2594                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2595                 if (PTR_ERR(root) == -ENOENT)
2596                         return 0;
2597                 return PTR_ERR(root);
2598         }
2599
2600         if (btrfs_root_readonly(root)) {
2601                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2602                 return 0;
2603         }
2604
2605         /* step 2: get inode */
2606         key.objectid = backref->inum;
2607         key.type = BTRFS_INODE_ITEM_KEY;
2608         key.offset = 0;
2609
2610         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2611         if (IS_ERR(inode)) {
2612                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2613                 return 0;
2614         }
2615
2616         srcu_read_unlock(&fs_info->subvol_srcu, index);
2617
2618         /* step 3: relink backref */
2619         lock_start = backref->file_pos;
2620         lock_end = backref->file_pos + backref->num_bytes - 1;
2621         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2622                          &cached);
2623
2624         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2625         if (ordered) {
2626                 btrfs_put_ordered_extent(ordered);
2627                 goto out_unlock;
2628         }
2629
2630         trans = btrfs_join_transaction(root);
2631         if (IS_ERR(trans)) {
2632                 ret = PTR_ERR(trans);
2633                 goto out_unlock;
2634         }
2635
2636         key.objectid = backref->inum;
2637         key.type = BTRFS_EXTENT_DATA_KEY;
2638         key.offset = backref->file_pos;
2639
2640         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2641         if (ret < 0) {
2642                 goto out_free_path;
2643         } else if (ret > 0) {
2644                 ret = 0;
2645                 goto out_free_path;
2646         }
2647
2648         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2649                                 struct btrfs_file_extent_item);
2650
2651         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2652             backref->generation)
2653                 goto out_free_path;
2654
2655         btrfs_release_path(path);
2656
2657         start = backref->file_pos;
2658         if (backref->extent_offset < old->extent_offset + old->offset)
2659                 start += old->extent_offset + old->offset -
2660                          backref->extent_offset;
2661
2662         len = min(backref->extent_offset + backref->num_bytes,
2663                   old->extent_offset + old->offset + old->len);
2664         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2665
2666         ret = btrfs_drop_extents(trans, root, inode, start,
2667                                  start + len, 1);
2668         if (ret)
2669                 goto out_free_path;
2670 again:
2671         key.objectid = btrfs_ino(BTRFS_I(inode));
2672         key.type = BTRFS_EXTENT_DATA_KEY;
2673         key.offset = start;
2674
2675         path->leave_spinning = 1;
2676         if (merge) {
2677                 struct btrfs_file_extent_item *fi;
2678                 u64 extent_len;
2679                 struct btrfs_key found_key;
2680
2681                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2682                 if (ret < 0)
2683                         goto out_free_path;
2684
2685                 path->slots[0]--;
2686                 leaf = path->nodes[0];
2687                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2688
2689                 fi = btrfs_item_ptr(leaf, path->slots[0],
2690                                     struct btrfs_file_extent_item);
2691                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2692
2693                 if (extent_len + found_key.offset == start &&
2694                     relink_is_mergable(leaf, fi, new)) {
2695                         btrfs_set_file_extent_num_bytes(leaf, fi,
2696                                                         extent_len + len);
2697                         btrfs_mark_buffer_dirty(leaf);
2698                         inode_add_bytes(inode, len);
2699
2700                         ret = 1;
2701                         goto out_free_path;
2702                 } else {
2703                         merge = false;
2704                         btrfs_release_path(path);
2705                         goto again;
2706                 }
2707         }
2708
2709         ret = btrfs_insert_empty_item(trans, root, path, &key,
2710                                         sizeof(*extent));
2711         if (ret) {
2712                 btrfs_abort_transaction(trans, ret);
2713                 goto out_free_path;
2714         }
2715
2716         leaf = path->nodes[0];
2717         item = btrfs_item_ptr(leaf, path->slots[0],
2718                                 struct btrfs_file_extent_item);
2719         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2720         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2721         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2722         btrfs_set_file_extent_num_bytes(leaf, item, len);
2723         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2724         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2725         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2726         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2727         btrfs_set_file_extent_encryption(leaf, item, 0);
2728         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2729
2730         btrfs_mark_buffer_dirty(leaf);
2731         inode_add_bytes(inode, len);
2732         btrfs_release_path(path);
2733
2734         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2735                         new->disk_len, 0,
2736                         backref->root_id, backref->inum,
2737                         new->file_pos); /* start - extent_offset */
2738         if (ret) {
2739                 btrfs_abort_transaction(trans, ret);
2740                 goto out_free_path;
2741         }
2742
2743         ret = 1;
2744 out_free_path:
2745         btrfs_release_path(path);
2746         path->leave_spinning = 0;
2747         btrfs_end_transaction(trans);
2748 out_unlock:
2749         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2750                              &cached);
2751         iput(inode);
2752         return ret;
2753 }
2754
2755 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2756 {
2757         struct old_sa_defrag_extent *old, *tmp;
2758
2759         if (!new)
2760                 return;
2761
2762         list_for_each_entry_safe(old, tmp, &new->head, list) {
2763                 kfree(old);
2764         }
2765         kfree(new);
2766 }
2767
2768 static void relink_file_extents(struct new_sa_defrag_extent *new)
2769 {
2770         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2771         struct btrfs_path *path;
2772         struct sa_defrag_extent_backref *backref;
2773         struct sa_defrag_extent_backref *prev = NULL;
2774         struct inode *inode;
2775         struct rb_node *node;
2776         int ret;
2777
2778         inode = new->inode;
2779
2780         path = btrfs_alloc_path();
2781         if (!path)
2782                 return;
2783
2784         if (!record_extent_backrefs(path, new)) {
2785                 btrfs_free_path(path);
2786                 goto out;
2787         }
2788         btrfs_release_path(path);
2789
2790         while (1) {
2791                 node = rb_first(&new->root);
2792                 if (!node)
2793                         break;
2794                 rb_erase(node, &new->root);
2795
2796                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2797
2798                 ret = relink_extent_backref(path, prev, backref);
2799                 WARN_ON(ret < 0);
2800
2801                 kfree(prev);
2802
2803                 if (ret == 1)
2804                         prev = backref;
2805                 else
2806                         prev = NULL;
2807                 cond_resched();
2808         }
2809         kfree(prev);
2810
2811         btrfs_free_path(path);
2812 out:
2813         free_sa_defrag_extent(new);
2814
2815         atomic_dec(&fs_info->defrag_running);
2816         wake_up(&fs_info->transaction_wait);
2817 }
2818
2819 static struct new_sa_defrag_extent *
2820 record_old_file_extents(struct inode *inode,
2821                         struct btrfs_ordered_extent *ordered)
2822 {
2823         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2824         struct btrfs_root *root = BTRFS_I(inode)->root;
2825         struct btrfs_path *path;
2826         struct btrfs_key key;
2827         struct old_sa_defrag_extent *old;
2828         struct new_sa_defrag_extent *new;
2829         int ret;
2830
2831         new = kmalloc(sizeof(*new), GFP_NOFS);
2832         if (!new)
2833                 return NULL;
2834
2835         new->inode = inode;
2836         new->file_pos = ordered->file_offset;
2837         new->len = ordered->len;
2838         new->bytenr = ordered->start;
2839         new->disk_len = ordered->disk_len;
2840         new->compress_type = ordered->compress_type;
2841         new->root = RB_ROOT;
2842         INIT_LIST_HEAD(&new->head);
2843
2844         path = btrfs_alloc_path();
2845         if (!path)
2846                 goto out_kfree;
2847
2848         key.objectid = btrfs_ino(BTRFS_I(inode));
2849         key.type = BTRFS_EXTENT_DATA_KEY;
2850         key.offset = new->file_pos;
2851
2852         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2853         if (ret < 0)
2854                 goto out_free_path;
2855         if (ret > 0 && path->slots[0] > 0)
2856                 path->slots[0]--;
2857
2858         /* find out all the old extents for the file range */
2859         while (1) {
2860                 struct btrfs_file_extent_item *extent;
2861                 struct extent_buffer *l;
2862                 int slot;
2863                 u64 num_bytes;
2864                 u64 offset;
2865                 u64 end;
2866                 u64 disk_bytenr;
2867                 u64 extent_offset;
2868
2869                 l = path->nodes[0];
2870                 slot = path->slots[0];
2871
2872                 if (slot >= btrfs_header_nritems(l)) {
2873                         ret = btrfs_next_leaf(root, path);
2874                         if (ret < 0)
2875                                 goto out_free_path;
2876                         else if (ret > 0)
2877                                 break;
2878                         continue;
2879                 }
2880
2881                 btrfs_item_key_to_cpu(l, &key, slot);
2882
2883                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2884                         break;
2885                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2886                         break;
2887                 if (key.offset >= new->file_pos + new->len)
2888                         break;
2889
2890                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2891
2892                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2893                 if (key.offset + num_bytes < new->file_pos)
2894                         goto next;
2895
2896                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2897                 if (!disk_bytenr)
2898                         goto next;
2899
2900                 extent_offset = btrfs_file_extent_offset(l, extent);
2901
2902                 old = kmalloc(sizeof(*old), GFP_NOFS);
2903                 if (!old)
2904                         goto out_free_path;
2905
2906                 offset = max(new->file_pos, key.offset);
2907                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2908
2909                 old->bytenr = disk_bytenr;
2910                 old->extent_offset = extent_offset;
2911                 old->offset = offset - key.offset;
2912                 old->len = end - offset;
2913                 old->new = new;
2914                 old->count = 0;
2915                 list_add_tail(&old->list, &new->head);
2916 next:
2917                 path->slots[0]++;
2918                 cond_resched();
2919         }
2920
2921         btrfs_free_path(path);
2922         atomic_inc(&fs_info->defrag_running);
2923
2924         return new;
2925
2926 out_free_path:
2927         btrfs_free_path(path);
2928 out_kfree:
2929         free_sa_defrag_extent(new);
2930         return NULL;
2931 }
2932
2933 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2934                                          u64 start, u64 len)
2935 {
2936         struct btrfs_block_group_cache *cache;
2937
2938         cache = btrfs_lookup_block_group(fs_info, start);
2939         ASSERT(cache);
2940
2941         spin_lock(&cache->lock);
2942         cache->delalloc_bytes -= len;
2943         spin_unlock(&cache->lock);
2944
2945         btrfs_put_block_group(cache);
2946 }
2947
2948 /* as ordered data IO finishes, this gets called so we can finish
2949  * an ordered extent if the range of bytes in the file it covers are
2950  * fully written.
2951  */
2952 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2953 {
2954         struct inode *inode = ordered_extent->inode;
2955         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2956         struct btrfs_root *root = BTRFS_I(inode)->root;
2957         struct btrfs_trans_handle *trans = NULL;
2958         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2959         struct extent_state *cached_state = NULL;
2960         struct new_sa_defrag_extent *new = NULL;
2961         int compress_type = 0;
2962         int ret = 0;
2963         u64 logical_len = ordered_extent->len;
2964         bool nolock;
2965         bool truncated = false;
2966         bool range_locked = false;
2967         bool clear_new_delalloc_bytes = false;
2968
2969         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2970             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2971             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2972                 clear_new_delalloc_bytes = true;
2973
2974         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2975
2976         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2977                 ret = -EIO;
2978                 goto out;
2979         }
2980
2981         btrfs_free_io_failure_record(BTRFS_I(inode),
2982                         ordered_extent->file_offset,
2983                         ordered_extent->file_offset +
2984                         ordered_extent->len - 1);
2985
2986         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2987                 truncated = true;
2988                 logical_len = ordered_extent->truncated_len;
2989                 /* Truncated the entire extent, don't bother adding */
2990                 if (!logical_len)
2991                         goto out;
2992         }
2993
2994         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2995                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2996
2997                 /*
2998                  * For mwrite(mmap + memset to write) case, we still reserve
2999                  * space for NOCOW range.
3000                  * As NOCOW won't cause a new delayed ref, just free the space
3001                  */
3002                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3003                                        ordered_extent->len);
3004                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3005                 if (nolock)
3006                         trans = btrfs_join_transaction_nolock(root);
3007                 else
3008                         trans = btrfs_join_transaction(root);
3009                 if (IS_ERR(trans)) {
3010                         ret = PTR_ERR(trans);
3011                         trans = NULL;
3012                         goto out;
3013                 }
3014                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3015                 ret = btrfs_update_inode_fallback(trans, root, inode);
3016                 if (ret) /* -ENOMEM or corruption */
3017                         btrfs_abort_transaction(trans, ret);
3018                 goto out;
3019         }
3020
3021         range_locked = true;
3022         lock_extent_bits(io_tree, ordered_extent->file_offset,
3023                          ordered_extent->file_offset + ordered_extent->len - 1,
3024                          &cached_state);
3025
3026         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3027                         ordered_extent->file_offset + ordered_extent->len - 1,
3028                         EXTENT_DEFRAG, 0, cached_state);
3029         if (ret) {
3030                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3031                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3032                         /* the inode is shared */
3033                         new = record_old_file_extents(inode, ordered_extent);
3034
3035                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3036                         ordered_extent->file_offset + ordered_extent->len - 1,
3037                         EXTENT_DEFRAG, 0, 0, &cached_state);
3038         }
3039
3040         if (nolock)
3041                 trans = btrfs_join_transaction_nolock(root);
3042         else
3043                 trans = btrfs_join_transaction(root);
3044         if (IS_ERR(trans)) {
3045                 ret = PTR_ERR(trans);
3046                 trans = NULL;
3047                 goto out;
3048         }
3049
3050         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3051
3052         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3053                 compress_type = ordered_extent->compress_type;
3054         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3055                 BUG_ON(compress_type);
3056                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3057                                        ordered_extent->len);
3058                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3059                                                 ordered_extent->file_offset,
3060                                                 ordered_extent->file_offset +
3061                                                 logical_len);
3062         } else {
3063                 BUG_ON(root == fs_info->tree_root);
3064                 ret = insert_reserved_file_extent(trans, inode,
3065                                                 ordered_extent->file_offset,
3066                                                 ordered_extent->start,
3067                                                 ordered_extent->disk_len,
3068                                                 logical_len, logical_len,
3069                                                 compress_type, 0, 0,
3070                                                 BTRFS_FILE_EXTENT_REG);
3071                 if (!ret)
3072                         btrfs_release_delalloc_bytes(fs_info,
3073                                                      ordered_extent->start,
3074                                                      ordered_extent->disk_len);
3075         }
3076         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3077                            ordered_extent->file_offset, ordered_extent->len,
3078                            trans->transid);
3079         if (ret < 0) {
3080                 btrfs_abort_transaction(trans, ret);
3081                 goto out;
3082         }
3083
3084         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3085         if (ret) {
3086                 btrfs_abort_transaction(trans, ret);
3087                 goto out;
3088         }
3089
3090         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3091         ret = btrfs_update_inode_fallback(trans, root, inode);
3092         if (ret) { /* -ENOMEM or corruption */
3093                 btrfs_abort_transaction(trans, ret);
3094                 goto out;
3095         }
3096         ret = 0;
3097 out:
3098         if (range_locked || clear_new_delalloc_bytes) {
3099                 unsigned int clear_bits = 0;
3100
3101                 if (range_locked)
3102                         clear_bits |= EXTENT_LOCKED;
3103                 if (clear_new_delalloc_bytes)
3104                         clear_bits |= EXTENT_DELALLOC_NEW;
3105                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3106                                  ordered_extent->file_offset,
3107                                  ordered_extent->file_offset +
3108                                  ordered_extent->len - 1,
3109                                  clear_bits,
3110                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3111                                  0, &cached_state);
3112         }
3113
3114         if (trans)
3115                 btrfs_end_transaction(trans);
3116
3117         if (ret || truncated) {
3118                 u64 start, end;
3119
3120                 if (truncated)
3121                         start = ordered_extent->file_offset + logical_len;
3122                 else
3123                         start = ordered_extent->file_offset;
3124                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3125                 clear_extent_uptodate(io_tree, start, end, NULL);
3126
3127                 /* Drop the cache for the part of the extent we didn't write. */
3128                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3129
3130                 /*
3131                  * If the ordered extent had an IOERR or something else went
3132                  * wrong we need to return the space for this ordered extent
3133                  * back to the allocator.  We only free the extent in the
3134                  * truncated case if we didn't write out the extent at all.
3135                  */
3136                 if ((ret || !logical_len) &&
3137                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3138                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3139                         btrfs_free_reserved_extent(fs_info,
3140                                                    ordered_extent->start,
3141                                                    ordered_extent->disk_len, 1);
3142         }
3143
3144
3145         /*
3146          * This needs to be done to make sure anybody waiting knows we are done
3147          * updating everything for this ordered extent.
3148          */
3149         btrfs_remove_ordered_extent(inode, ordered_extent);
3150
3151         /* for snapshot-aware defrag */
3152         if (new) {
3153                 if (ret) {
3154                         free_sa_defrag_extent(new);
3155                         atomic_dec(&fs_info->defrag_running);
3156                 } else {
3157                         relink_file_extents(new);
3158                 }
3159         }
3160
3161         /* once for us */
3162         btrfs_put_ordered_extent(ordered_extent);
3163         /* once for the tree */
3164         btrfs_put_ordered_extent(ordered_extent);
3165
3166         return ret;
3167 }
3168
3169 static void finish_ordered_fn(struct btrfs_work *work)
3170 {
3171         struct btrfs_ordered_extent *ordered_extent;
3172         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3173         btrfs_finish_ordered_io(ordered_extent);
3174 }
3175
3176 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3177                                 struct extent_state *state, int uptodate)
3178 {
3179         struct inode *inode = page->mapping->host;
3180         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3181         struct btrfs_ordered_extent *ordered_extent = NULL;
3182         struct btrfs_workqueue *wq;
3183         btrfs_work_func_t func;
3184
3185         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3186
3187         ClearPagePrivate2(page);
3188         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3189                                             end - start + 1, uptodate))
3190                 return;
3191
3192         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3193                 wq = fs_info->endio_freespace_worker;
3194                 func = btrfs_freespace_write_helper;
3195         } else {
3196                 wq = fs_info->endio_write_workers;
3197                 func = btrfs_endio_write_helper;
3198         }
3199
3200         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3201                         NULL);
3202         btrfs_queue_work(wq, &ordered_extent->work);
3203 }
3204
3205 static int __readpage_endio_check(struct inode *inode,
3206                                   struct btrfs_io_bio *io_bio,
3207                                   int icsum, struct page *page,
3208                                   int pgoff, u64 start, size_t len)
3209 {
3210         char *kaddr;
3211         u32 csum_expected;
3212         u32 csum = ~(u32)0;
3213
3214         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3215
3216         kaddr = kmap_atomic(page);
3217         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3218         btrfs_csum_final(csum, (u8 *)&csum);
3219         if (csum != csum_expected)
3220                 goto zeroit;
3221
3222         kunmap_atomic(kaddr);
3223         return 0;
3224 zeroit:
3225         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3226                                     io_bio->mirror_num);
3227         memset(kaddr + pgoff, 1, len);
3228         flush_dcache_page(page);
3229         kunmap_atomic(kaddr);
3230         return -EIO;
3231 }
3232
3233 /*
3234  * when reads are done, we need to check csums to verify the data is correct
3235  * if there's a match, we allow the bio to finish.  If not, the code in
3236  * extent_io.c will try to find good copies for us.
3237  */
3238 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3239                                       u64 phy_offset, struct page *page,
3240                                       u64 start, u64 end, int mirror)
3241 {
3242         size_t offset = start - page_offset(page);
3243         struct inode *inode = page->mapping->host;
3244         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3245         struct btrfs_root *root = BTRFS_I(inode)->root;
3246
3247         if (PageChecked(page)) {
3248                 ClearPageChecked(page);
3249                 return 0;
3250         }
3251
3252         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3253                 return 0;
3254
3255         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3256             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3257                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3258                 return 0;
3259         }
3260
3261         phy_offset >>= inode->i_sb->s_blocksize_bits;
3262         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3263                                       start, (size_t)(end - start + 1));
3264 }
3265
3266 /*
3267  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3268  *
3269  * @inode: The inode we want to perform iput on
3270  *
3271  * This function uses the generic vfs_inode::i_count to track whether we should
3272  * just decrement it (in case it's > 1) or if this is the last iput then link
3273  * the inode to the delayed iput machinery. Delayed iputs are processed at
3274  * transaction commit time/superblock commit/cleaner kthread.
3275  */
3276 void btrfs_add_delayed_iput(struct inode *inode)
3277 {
3278         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3279         struct btrfs_inode *binode = BTRFS_I(inode);
3280
3281         if (atomic_add_unless(&inode->i_count, -1, 1))
3282                 return;
3283
3284         spin_lock(&fs_info->delayed_iput_lock);
3285         ASSERT(list_empty(&binode->delayed_iput));
3286         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3287         spin_unlock(&fs_info->delayed_iput_lock);
3288 }
3289
3290 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3291 {
3292
3293         spin_lock(&fs_info->delayed_iput_lock);
3294         while (!list_empty(&fs_info->delayed_iputs)) {
3295                 struct btrfs_inode *inode;
3296
3297                 inode = list_first_entry(&fs_info->delayed_iputs,
3298                                 struct btrfs_inode, delayed_iput);
3299                 list_del_init(&inode->delayed_iput);
3300                 spin_unlock(&fs_info->delayed_iput_lock);
3301                 iput(&inode->vfs_inode);
3302                 spin_lock(&fs_info->delayed_iput_lock);
3303         }
3304         spin_unlock(&fs_info->delayed_iput_lock);
3305 }
3306
3307 /*
3308  * This is called in transaction commit time. If there are no orphan
3309  * files in the subvolume, it removes orphan item and frees block_rsv
3310  * structure.
3311  */
3312 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3313                               struct btrfs_root *root)
3314 {
3315         struct btrfs_fs_info *fs_info = root->fs_info;
3316         struct btrfs_block_rsv *block_rsv;
3317         int ret;
3318
3319         if (atomic_read(&root->orphan_inodes) ||
3320             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3321                 return;
3322
3323         spin_lock(&root->orphan_lock);
3324         if (atomic_read(&root->orphan_inodes)) {
3325                 spin_unlock(&root->orphan_lock);
3326                 return;
3327         }
3328
3329         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3330                 spin_unlock(&root->orphan_lock);
3331                 return;
3332         }
3333
3334         block_rsv = root->orphan_block_rsv;
3335         root->orphan_block_rsv = NULL;
3336         spin_unlock(&root->orphan_lock);
3337
3338         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3339             btrfs_root_refs(&root->root_item) > 0) {
3340                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3341                                             root->root_key.objectid);
3342                 if (ret)
3343                         btrfs_abort_transaction(trans, ret);
3344                 else
3345                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3346                                   &root->state);
3347         }
3348
3349         if (block_rsv) {
3350                 WARN_ON(block_rsv->size > 0);
3351                 btrfs_free_block_rsv(fs_info, block_rsv);
3352         }
3353 }
3354
3355 /*
3356  * This creates an orphan entry for the given inode in case something goes
3357  * wrong in the middle of an unlink/truncate.
3358  *
3359  * NOTE: caller of this function should reserve 5 units of metadata for
3360  *       this function.
3361  */
3362 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3363                 struct btrfs_inode *inode)
3364 {
3365         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3366         struct btrfs_root *root = inode->root;
3367         struct btrfs_block_rsv *block_rsv = NULL;
3368         int reserve = 0;
3369         int insert = 0;
3370         int ret;
3371
3372         if (!root->orphan_block_rsv) {
3373                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3374                                                   BTRFS_BLOCK_RSV_TEMP);
3375                 if (!block_rsv)
3376                         return -ENOMEM;
3377         }
3378
3379         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3380                               &inode->runtime_flags)) {
3381 #if 0
3382                 /*
3383                  * For proper ENOSPC handling, we should do orphan
3384                  * cleanup when mounting. But this introduces backward
3385                  * compatibility issue.
3386                  */
3387                 if (!xchg(&root->orphan_item_inserted, 1))
3388                         insert = 2;
3389                 else
3390                         insert = 1;
3391 #endif
3392                 insert = 1;
3393         }
3394
3395         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3396                               &inode->runtime_flags))
3397                 reserve = 1;
3398
3399         spin_lock(&root->orphan_lock);
3400         /* If someone has created ->orphan_block_rsv, be happy to use it. */
3401         if (!root->orphan_block_rsv) {
3402                 root->orphan_block_rsv = block_rsv;
3403         } else if (block_rsv) {
3404                 btrfs_free_block_rsv(fs_info, block_rsv);
3405                 block_rsv = NULL;
3406         }
3407
3408         if (insert)
3409                 atomic_inc(&root->orphan_inodes);
3410         spin_unlock(&root->orphan_lock);
3411
3412         /* grab metadata reservation from transaction handle */
3413         if (reserve) {
3414                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3415                 ASSERT(!ret);
3416                 if (ret) {
3417                         /*
3418                          * dec doesn't need spin_lock as ->orphan_block_rsv
3419                          * would be released only if ->orphan_inodes is
3420                          * zero.
3421                          */
3422                         atomic_dec(&root->orphan_inodes);
3423                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3424                                   &inode->runtime_flags);
3425                         if (insert)
3426                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3427                                           &inode->runtime_flags);
3428                         return ret;
3429                 }
3430         }
3431
3432         /* insert an orphan item to track this unlinked/truncated file */
3433         if (insert >= 1) {
3434                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3435                 if (ret) {
3436                         if (reserve) {
3437                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3438                                           &inode->runtime_flags);
3439                                 btrfs_orphan_release_metadata(inode);
3440                         }
3441                         /*
3442                          * btrfs_orphan_commit_root may race with us and set
3443                          * ->orphan_block_rsv to zero, in order to avoid that,
3444                          * decrease ->orphan_inodes after everything is done.
3445                          */
3446                         atomic_dec(&root->orphan_inodes);
3447                         if (ret != -EEXIST) {
3448                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3449                                           &inode->runtime_flags);
3450                                 btrfs_abort_transaction(trans, ret);
3451                                 return ret;
3452                         }
3453                 }
3454                 ret = 0;
3455         }
3456
3457         /* insert an orphan item to track subvolume contains orphan files */
3458         if (insert >= 2) {
3459                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3460                                                root->root_key.objectid);
3461                 if (ret && ret != -EEXIST) {
3462                         btrfs_abort_transaction(trans, ret);
3463                         return ret;
3464                 }
3465         }
3466         return 0;
3467 }
3468
3469 /*
3470  * We have done the truncate/delete so we can go ahead and remove the orphan
3471  * item for this particular inode.
3472  */
3473 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3474                             struct btrfs_inode *inode)
3475 {
3476         struct btrfs_root *root = inode->root;
3477         int delete_item = 0;
3478         int ret = 0;
3479
3480         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3481                                &inode->runtime_flags))
3482                 delete_item = 1;
3483
3484         if (delete_item && trans)
3485                 ret = btrfs_del_orphan_item(trans, root, btrfs_ino(inode));
3486
3487         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3488                                &inode->runtime_flags))
3489                 btrfs_orphan_release_metadata(inode);
3490
3491         /*
3492          * btrfs_orphan_commit_root may race with us and set ->orphan_block_rsv
3493          * to zero, in order to avoid that, decrease ->orphan_inodes after
3494          * everything is done.
3495          */
3496         if (delete_item)
3497                 atomic_dec(&root->orphan_inodes);
3498
3499         return ret;
3500 }
3501
3502 /*
3503  * this cleans up any orphans that may be left on the list from the last use
3504  * of this root.
3505  */
3506 int btrfs_orphan_cleanup(struct btrfs_root *root)
3507 {
3508         struct btrfs_fs_info *fs_info = root->fs_info;
3509         struct btrfs_path *path;
3510         struct extent_buffer *leaf;
3511         struct btrfs_key key, found_key;
3512         struct btrfs_trans_handle *trans;
3513         struct inode *inode;
3514         u64 last_objectid = 0;
3515         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3516
3517         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3518                 return 0;
3519
3520         path = btrfs_alloc_path();
3521         if (!path) {
3522                 ret = -ENOMEM;
3523                 goto out;
3524         }
3525         path->reada = READA_BACK;
3526
3527         key.objectid = BTRFS_ORPHAN_OBJECTID;
3528         key.type = BTRFS_ORPHAN_ITEM_KEY;
3529         key.offset = (u64)-1;
3530
3531         while (1) {
3532                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3533                 if (ret < 0)
3534                         goto out;
3535
3536                 /*
3537                  * if ret == 0 means we found what we were searching for, which
3538                  * is weird, but possible, so only screw with path if we didn't
3539                  * find the key and see if we have stuff that matches
3540                  */
3541                 if (ret > 0) {
3542                         ret = 0;
3543                         if (path->slots[0] == 0)
3544                                 break;
3545                         path->slots[0]--;
3546                 }
3547
3548                 /* pull out the item */
3549                 leaf = path->nodes[0];
3550                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3551
3552                 /* make sure the item matches what we want */
3553                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3554                         break;
3555                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3556                         break;
3557
3558                 /* release the path since we're done with it */
3559                 btrfs_release_path(path);
3560
3561                 /*
3562                  * this is where we are basically btrfs_lookup, without the
3563                  * crossing root thing.  we store the inode number in the
3564                  * offset of the orphan item.
3565                  */
3566
3567                 if (found_key.offset == last_objectid) {
3568                         btrfs_err(fs_info,
3569                                   "Error removing orphan entry, stopping orphan cleanup");
3570                         ret = -EINVAL;
3571                         goto out;
3572                 }
3573
3574                 last_objectid = found_key.offset;
3575
3576                 found_key.objectid = found_key.offset;
3577                 found_key.type = BTRFS_INODE_ITEM_KEY;
3578                 found_key.offset = 0;
3579                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3580                 ret = PTR_ERR_OR_ZERO(inode);
3581                 if (ret && ret != -ENOENT)
3582                         goto out;
3583
3584                 if (ret == -ENOENT && root == fs_info->tree_root) {
3585                         struct btrfs_root *dead_root;
3586                         struct btrfs_fs_info *fs_info = root->fs_info;
3587                         int is_dead_root = 0;
3588
3589                         /*
3590                          * this is an orphan in the tree root. Currently these
3591                          * could come from 2 sources:
3592                          *  a) a snapshot deletion in progress
3593                          *  b) a free space cache inode
3594                          * We need to distinguish those two, as the snapshot
3595                          * orphan must not get deleted.
3596                          * find_dead_roots already ran before us, so if this
3597                          * is a snapshot deletion, we should find the root
3598                          * in the dead_roots list
3599                          */
3600                         spin_lock(&fs_info->trans_lock);
3601                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3602                                             root_list) {
3603                                 if (dead_root->root_key.objectid ==
3604                                     found_key.objectid) {
3605                                         is_dead_root = 1;
3606                                         break;
3607                                 }
3608                         }
3609                         spin_unlock(&fs_info->trans_lock);
3610                         if (is_dead_root) {
3611                                 /* prevent this orphan from being found again */
3612                                 key.offset = found_key.objectid - 1;
3613                                 continue;
3614                         }
3615                 }
3616                 /*
3617                  * Inode is already gone but the orphan item is still there,
3618                  * kill the orphan item.
3619                  */
3620                 if (ret == -ENOENT) {
3621                         trans = btrfs_start_transaction(root, 1);
3622                         if (IS_ERR(trans)) {
3623                                 ret = PTR_ERR(trans);
3624                                 goto out;
3625                         }
3626                         btrfs_debug(fs_info, "auto deleting %Lu",
3627                                     found_key.objectid);
3628                         ret = btrfs_del_orphan_item(trans, root,
3629                                                     found_key.objectid);
3630                         btrfs_end_transaction(trans);
3631                         if (ret)
3632                                 goto out;
3633                         continue;
3634                 }
3635
3636                 /*
3637                  * add this inode to the orphan list so btrfs_orphan_del does
3638                  * the proper thing when we hit it
3639                  */
3640                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3641                         &BTRFS_I(inode)->runtime_flags);
3642                 atomic_inc(&root->orphan_inodes);
3643
3644                 /* if we have links, this was a truncate, lets do that */
3645                 if (inode->i_nlink) {
3646                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3647                                 iput(inode);
3648                                 continue;
3649                         }
3650                         nr_truncate++;
3651
3652                         /* 1 for the orphan item deletion. */
3653                         trans = btrfs_start_transaction(root, 1);
3654                         if (IS_ERR(trans)) {
3655                                 iput(inode);
3656                                 ret = PTR_ERR(trans);
3657                                 goto out;
3658                         }
3659                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3660                         btrfs_end_transaction(trans);
3661                         if (ret) {
3662                                 iput(inode);
3663                                 goto out;
3664                         }
3665
3666                         ret = btrfs_truncate(inode, false);
3667                         if (ret)
3668                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3669                 } else {
3670                         nr_unlink++;
3671                 }
3672
3673                 /* this will do delete_inode and everything for us */
3674                 iput(inode);
3675                 if (ret)
3676                         goto out;
3677         }
3678         /* release the path since we're done with it */
3679         btrfs_release_path(path);
3680
3681         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3682
3683         if (root->orphan_block_rsv)
3684                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3685                                         (u64)-1);
3686
3687         if (root->orphan_block_rsv ||
3688             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3689                 trans = btrfs_join_transaction(root);
3690                 if (!IS_ERR(trans))
3691                         btrfs_end_transaction(trans);
3692         }
3693
3694         if (nr_unlink)
3695                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3696         if (nr_truncate)
3697                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3698
3699 out:
3700         if (ret)
3701                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3702         btrfs_free_path(path);
3703         return ret;
3704 }
3705
3706 /*
3707  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3708  * don't find any xattrs, we know there can't be any acls.
3709  *
3710  * slot is the slot the inode is in, objectid is the objectid of the inode
3711  */
3712 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3713                                           int slot, u64 objectid,
3714                                           int *first_xattr_slot)
3715 {
3716         u32 nritems = btrfs_header_nritems(leaf);
3717         struct btrfs_key found_key;
3718         static u64 xattr_access = 0;
3719         static u64 xattr_default = 0;
3720         int scanned = 0;
3721
3722         if (!xattr_access) {
3723                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3724                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3725                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3726                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3727         }
3728
3729         slot++;
3730         *first_xattr_slot = -1;
3731         while (slot < nritems) {
3732                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3733
3734                 /* we found a different objectid, there must not be acls */
3735                 if (found_key.objectid != objectid)
3736                         return 0;
3737
3738                 /* we found an xattr, assume we've got an acl */
3739                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3740                         if (*first_xattr_slot == -1)
3741                                 *first_xattr_slot = slot;
3742                         if (found_key.offset == xattr_access ||
3743                             found_key.offset == xattr_default)
3744                                 return 1;
3745                 }
3746
3747                 /*
3748                  * we found a key greater than an xattr key, there can't
3749                  * be any acls later on
3750                  */
3751                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3752                         return 0;
3753
3754                 slot++;
3755                 scanned++;
3756
3757                 /*
3758                  * it goes inode, inode backrefs, xattrs, extents,
3759                  * so if there are a ton of hard links to an inode there can
3760                  * be a lot of backrefs.  Don't waste time searching too hard,
3761                  * this is just an optimization
3762                  */
3763                 if (scanned >= 8)
3764                         break;
3765         }
3766         /* we hit the end of the leaf before we found an xattr or
3767          * something larger than an xattr.  We have to assume the inode
3768          * has acls
3769          */
3770         if (*first_xattr_slot == -1)
3771                 *first_xattr_slot = slot;
3772         return 1;
3773 }
3774
3775 /*
3776  * read an inode from the btree into the in-memory inode
3777  */
3778 static int btrfs_read_locked_inode(struct inode *inode)
3779 {
3780         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3781         struct btrfs_path *path;
3782         struct extent_buffer *leaf;
3783         struct btrfs_inode_item *inode_item;
3784         struct btrfs_root *root = BTRFS_I(inode)->root;
3785         struct btrfs_key location;
3786         unsigned long ptr;
3787         int maybe_acls;
3788         u32 rdev;
3789         int ret;
3790         bool filled = false;
3791         int first_xattr_slot;
3792
3793         ret = btrfs_fill_inode(inode, &rdev);
3794         if (!ret)
3795                 filled = true;
3796
3797         path = btrfs_alloc_path();
3798         if (!path) {
3799                 ret = -ENOMEM;
3800                 goto make_bad;
3801         }
3802
3803         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3804
3805         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3806         if (ret) {
3807                 if (ret > 0)
3808                         ret = -ENOENT;
3809                 goto make_bad;
3810         }
3811
3812         leaf = path->nodes[0];
3813
3814         if (filled)
3815                 goto cache_index;
3816
3817         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3818                                     struct btrfs_inode_item);
3819         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3820         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3821         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3822         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3823         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3824
3825         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3826         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3827
3828         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3829         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3830
3831         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3832         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3833
3834         BTRFS_I(inode)->i_otime.tv_sec =
3835                 btrfs_timespec_sec(leaf, &inode_item->otime);
3836         BTRFS_I(inode)->i_otime.tv_nsec =
3837                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3838
3839         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3840         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3841         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3842
3843         inode_set_iversion_queried(inode,
3844                                    btrfs_inode_sequence(leaf, inode_item));
3845         inode->i_generation = BTRFS_I(inode)->generation;
3846         inode->i_rdev = 0;
3847         rdev = btrfs_inode_rdev(leaf, inode_item);
3848
3849         BTRFS_I(inode)->index_cnt = (u64)-1;
3850         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3851
3852 cache_index:
3853         /*
3854          * If we were modified in the current generation and evicted from memory
3855          * and then re-read we need to do a full sync since we don't have any
3856          * idea about which extents were modified before we were evicted from
3857          * cache.
3858          *
3859          * This is required for both inode re-read from disk and delayed inode
3860          * in delayed_nodes_tree.
3861          */
3862         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3863                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3864                         &BTRFS_I(inode)->runtime_flags);
3865
3866         /*
3867          * We don't persist the id of the transaction where an unlink operation
3868          * against the inode was last made. So here we assume the inode might
3869          * have been evicted, and therefore the exact value of last_unlink_trans
3870          * lost, and set it to last_trans to avoid metadata inconsistencies
3871          * between the inode and its parent if the inode is fsync'ed and the log
3872          * replayed. For example, in the scenario:
3873          *
3874          * touch mydir/foo
3875          * ln mydir/foo mydir/bar
3876          * sync
3877          * unlink mydir/bar
3878          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3879          * xfs_io -c fsync mydir/foo
3880          * <power failure>
3881          * mount fs, triggers fsync log replay
3882          *
3883          * We must make sure that when we fsync our inode foo we also log its
3884          * parent inode, otherwise after log replay the parent still has the
3885          * dentry with the "bar" name but our inode foo has a link count of 1
3886          * and doesn't have an inode ref with the name "bar" anymore.
3887          *
3888          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3889          * but it guarantees correctness at the expense of occasional full
3890          * transaction commits on fsync if our inode is a directory, or if our
3891          * inode is not a directory, logging its parent unnecessarily.
3892          */
3893         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3894
3895         path->slots[0]++;
3896         if (inode->i_nlink != 1 ||
3897             path->slots[0] >= btrfs_header_nritems(leaf))
3898                 goto cache_acl;
3899
3900         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3901         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3902                 goto cache_acl;
3903
3904         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3905         if (location.type == BTRFS_INODE_REF_KEY) {
3906                 struct btrfs_inode_ref *ref;
3907
3908                 ref = (struct btrfs_inode_ref *)ptr;
3909                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3910         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3911                 struct btrfs_inode_extref *extref;
3912
3913                 extref = (struct btrfs_inode_extref *)ptr;
3914                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3915                                                                      extref);
3916         }
3917 cache_acl:
3918         /*
3919          * try to precache a NULL acl entry for files that don't have
3920          * any xattrs or acls
3921          */
3922         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3923                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3924         if (first_xattr_slot != -1) {
3925                 path->slots[0] = first_xattr_slot;
3926                 ret = btrfs_load_inode_props(inode, path);
3927                 if (ret)
3928                         btrfs_err(fs_info,
3929                                   "error loading props for ino %llu (root %llu): %d",
3930                                   btrfs_ino(BTRFS_I(inode)),
3931                                   root->root_key.objectid, ret);
3932         }
3933         btrfs_free_path(path);
3934
3935         if (!maybe_acls)
3936                 cache_no_acl(inode);
3937
3938         switch (inode->i_mode & S_IFMT) {
3939         case S_IFREG:
3940                 inode->i_mapping->a_ops = &btrfs_aops;
3941                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3942                 inode->i_fop = &btrfs_file_operations;
3943                 inode->i_op = &btrfs_file_inode_operations;
3944                 break;
3945         case S_IFDIR:
3946                 inode->i_fop = &btrfs_dir_file_operations;
3947                 inode->i_op = &btrfs_dir_inode_operations;
3948                 break;
3949         case S_IFLNK:
3950                 inode->i_op = &btrfs_symlink_inode_operations;
3951                 inode_nohighmem(inode);
3952                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3953                 break;
3954         default:
3955                 inode->i_op = &btrfs_special_inode_operations;
3956                 init_special_inode(inode, inode->i_mode, rdev);
3957                 break;
3958         }
3959
3960         btrfs_update_iflags(inode);
3961         return 0;
3962
3963 make_bad:
3964         btrfs_free_path(path);
3965         make_bad_inode(inode);
3966         return ret;
3967 }
3968
3969 /*
3970  * given a leaf and an inode, copy the inode fields into the leaf
3971  */
3972 static void fill_inode_item(struct btrfs_trans_handle *trans,
3973                             struct extent_buffer *leaf,
3974                             struct btrfs_inode_item *item,
3975                             struct inode *inode)
3976 {
3977         struct btrfs_map_token token;
3978
3979         btrfs_init_map_token(&token);
3980
3981         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3982         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3983         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3984                                    &token);
3985         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3986         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3987
3988         btrfs_set_token_timespec_sec(leaf, &item->atime,
3989                                      inode->i_atime.tv_sec, &token);
3990         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3991                                       inode->i_atime.tv_nsec, &token);
3992
3993         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3994                                      inode->i_mtime.tv_sec, &token);
3995         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3996                                       inode->i_mtime.tv_nsec, &token);
3997
3998         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3999                                      inode->i_ctime.tv_sec, &token);
4000         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
4001                                       inode->i_ctime.tv_nsec, &token);
4002
4003         btrfs_set_token_timespec_sec(leaf, &item->otime,
4004                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
4005         btrfs_set_token_timespec_nsec(leaf, &item->otime,
4006                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
4007
4008         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
4009                                      &token);
4010         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
4011                                          &token);
4012         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
4013                                        &token);
4014         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
4015         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
4016         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
4017         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
4018 }
4019
4020 /*
4021  * copy everything in the in-memory inode into the btree.
4022  */
4023 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
4024                                 struct btrfs_root *root, struct inode *inode)
4025 {
4026         struct btrfs_inode_item *inode_item;
4027         struct btrfs_path *path;
4028         struct extent_buffer *leaf;
4029         int ret;
4030
4031         path = btrfs_alloc_path();
4032         if (!path)
4033                 return -ENOMEM;
4034
4035         path->leave_spinning = 1;
4036         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
4037                                  1);
4038         if (ret) {
4039                 if (ret > 0)
4040                         ret = -ENOENT;
4041                 goto failed;
4042         }
4043
4044         leaf = path->nodes[0];
4045         inode_item = btrfs_item_ptr(leaf, path->slots[0],
4046                                     struct btrfs_inode_item);
4047
4048         fill_inode_item(trans, leaf, inode_item, inode);
4049         btrfs_mark_buffer_dirty(leaf);
4050         btrfs_set_inode_last_trans(trans, inode);
4051         ret = 0;
4052 failed:
4053         btrfs_free_path(path);
4054         return ret;
4055 }
4056
4057 /*
4058  * copy everything in the in-memory inode into the btree.
4059  */
4060 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
4061                                 struct btrfs_root *root, struct inode *inode)
4062 {
4063         struct btrfs_fs_info *fs_info = root->fs_info;
4064         int ret;
4065
4066         /*
4067          * If the inode is a free space inode, we can deadlock during commit
4068          * if we put it into the delayed code.
4069          *
4070          * The data relocation inode should also be directly updated
4071          * without delay
4072          */
4073         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4074             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4075             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4076                 btrfs_update_root_times(trans, root);
4077
4078                 ret = btrfs_delayed_update_inode(trans, root, inode);
4079                 if (!ret)
4080                         btrfs_set_inode_last_trans(trans, inode);
4081                 return ret;
4082         }
4083
4084         return btrfs_update_inode_item(trans, root, inode);
4085 }
4086
4087 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4088                                          struct btrfs_root *root,
4089                                          struct inode *inode)
4090 {
4091         int ret;
4092
4093         ret = btrfs_update_inode(trans, root, inode);
4094         if (ret == -ENOSPC)
4095                 return btrfs_update_inode_item(trans, root, inode);
4096         return ret;
4097 }
4098
4099 /*
4100  * unlink helper that gets used here in inode.c and in the tree logging
4101  * recovery code.  It remove a link in a directory with a given name, and
4102  * also drops the back refs in the inode to the directory
4103  */
4104 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4105                                 struct btrfs_root *root,
4106                                 struct btrfs_inode *dir,
4107                                 struct btrfs_inode *inode,
4108                                 const char *name, int name_len)
4109 {
4110         struct btrfs_fs_info *fs_info = root->fs_info;
4111         struct btrfs_path *path;
4112         int ret = 0;
4113         struct extent_buffer *leaf;
4114         struct btrfs_dir_item *di;
4115         struct btrfs_key key;
4116         u64 index;
4117         u64 ino = btrfs_ino(inode);
4118         u64 dir_ino = btrfs_ino(dir);
4119
4120         path = btrfs_alloc_path();
4121         if (!path) {
4122                 ret = -ENOMEM;
4123                 goto out;
4124         }
4125
4126         path->leave_spinning = 1;
4127         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4128                                     name, name_len, -1);
4129         if (IS_ERR(di)) {
4130                 ret = PTR_ERR(di);
4131                 goto err;
4132         }
4133         if (!di) {
4134                 ret = -ENOENT;
4135                 goto err;
4136         }
4137         leaf = path->nodes[0];
4138         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4139         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4140         if (ret)
4141                 goto err;
4142         btrfs_release_path(path);
4143
4144         /*
4145          * If we don't have dir index, we have to get it by looking up
4146          * the inode ref, since we get the inode ref, remove it directly,
4147          * it is unnecessary to do delayed deletion.
4148          *
4149          * But if we have dir index, needn't search inode ref to get it.
4150          * Since the inode ref is close to the inode item, it is better
4151          * that we delay to delete it, and just do this deletion when
4152          * we update the inode item.
4153          */
4154         if (inode->dir_index) {
4155                 ret = btrfs_delayed_delete_inode_ref(inode);
4156                 if (!ret) {
4157                         index = inode->dir_index;
4158                         goto skip_backref;
4159                 }
4160         }
4161
4162         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4163                                   dir_ino, &index);
4164         if (ret) {
4165                 btrfs_info(fs_info,
4166                         "failed to delete reference to %.*s, inode %llu parent %llu",
4167                         name_len, name, ino, dir_ino);
4168                 btrfs_abort_transaction(trans, ret);
4169                 goto err;
4170         }
4171 skip_backref:
4172         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4173         if (ret) {
4174                 btrfs_abort_transaction(trans, ret);
4175                 goto err;
4176         }
4177
4178         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4179                         dir_ino);
4180         if (ret != 0 && ret != -ENOENT) {
4181                 btrfs_abort_transaction(trans, ret);
4182                 goto err;
4183         }
4184
4185         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4186                         index);
4187         if (ret == -ENOENT)
4188                 ret = 0;
4189         else if (ret)
4190                 btrfs_abort_transaction(trans, ret);
4191 err:
4192         btrfs_free_path(path);
4193         if (ret)
4194                 goto out;
4195
4196         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4197         inode_inc_iversion(&inode->vfs_inode);
4198         inode_inc_iversion(&dir->vfs_inode);
4199         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4200                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4201         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4202 out:
4203         return ret;
4204 }
4205
4206 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4207                        struct btrfs_root *root,
4208                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4209                        const char *name, int name_len)
4210 {
4211         int ret;
4212         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4213         if (!ret) {
4214                 drop_nlink(&inode->vfs_inode);
4215                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4216         }
4217         return ret;
4218 }
4219
4220 /*
4221  * helper to start transaction for unlink and rmdir.
4222  *
4223  * unlink and rmdir are special in btrfs, they do not always free space, so
4224  * if we cannot make our reservations the normal way try and see if there is
4225  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4226  * allow the unlink to occur.
4227  */
4228 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4229 {
4230         struct btrfs_root *root = BTRFS_I(dir)->root;
4231
4232         /*
4233          * 1 for the possible orphan item
4234          * 1 for the dir item
4235          * 1 for the dir index
4236          * 1 for the inode ref
4237          * 1 for the inode
4238          */
4239         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4240 }
4241
4242 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4243 {
4244         struct btrfs_root *root = BTRFS_I(dir)->root;
4245         struct btrfs_trans_handle *trans;
4246         struct inode *inode = d_inode(dentry);
4247         int ret;
4248
4249         trans = __unlink_start_trans(dir);
4250         if (IS_ERR(trans))
4251                 return PTR_ERR(trans);
4252
4253         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4254                         0);
4255
4256         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4257                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4258                         dentry->d_name.len);
4259         if (ret)
4260                 goto out;
4261
4262         if (inode->i_nlink == 0) {
4263                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4264                 if (ret)
4265                         goto out;
4266         }
4267
4268 out:
4269         btrfs_end_transaction(trans);
4270         btrfs_btree_balance_dirty(root->fs_info);
4271         return ret;
4272 }
4273
4274 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4275                         struct btrfs_root *root,
4276                         struct inode *dir, u64 objectid,
4277                         const char *name, int name_len)
4278 {
4279         struct btrfs_fs_info *fs_info = root->fs_info;
4280         struct btrfs_path *path;
4281         struct extent_buffer *leaf;
4282         struct btrfs_dir_item *di;
4283         struct btrfs_key key;
4284         u64 index;
4285         int ret;
4286         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4287
4288         path = btrfs_alloc_path();
4289         if (!path)
4290                 return -ENOMEM;
4291
4292         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4293                                    name, name_len, -1);
4294         if (IS_ERR_OR_NULL(di)) {
4295                 if (!di)
4296                         ret = -ENOENT;
4297                 else
4298                         ret = PTR_ERR(di);
4299                 goto out;
4300         }
4301
4302         leaf = path->nodes[0];
4303         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4304         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4305         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4306         if (ret) {
4307                 btrfs_abort_transaction(trans, ret);
4308                 goto out;
4309         }
4310         btrfs_release_path(path);
4311
4312         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4313                                  root->root_key.objectid, dir_ino,
4314                                  &index, name, name_len);
4315         if (ret < 0) {
4316                 if (ret != -ENOENT) {
4317                         btrfs_abort_transaction(trans, ret);
4318                         goto out;
4319                 }
4320                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4321                                                  name, name_len);
4322                 if (IS_ERR_OR_NULL(di)) {
4323                         if (!di)
4324                                 ret = -ENOENT;
4325                         else
4326                                 ret = PTR_ERR(di);
4327                         btrfs_abort_transaction(trans, ret);
4328                         goto out;
4329                 }
4330
4331                 leaf = path->nodes[0];
4332                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4333                 btrfs_release_path(path);
4334                 index = key.offset;
4335         }
4336         btrfs_release_path(path);
4337
4338         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4339         if (ret) {
4340                 btrfs_abort_transaction(trans, ret);
4341                 goto out;
4342         }
4343
4344         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4345         inode_inc_iversion(dir);
4346         dir->i_mtime = dir->i_ctime = current_time(dir);
4347         ret = btrfs_update_inode_fallback(trans, root, dir);
4348         if (ret)
4349                 btrfs_abort_transaction(trans, ret);
4350 out:
4351         btrfs_free_path(path);
4352         return ret;
4353 }
4354
4355 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4356 {
4357         struct inode *inode = d_inode(dentry);
4358         int err = 0;
4359         struct btrfs_root *root = BTRFS_I(dir)->root;
4360         struct btrfs_trans_handle *trans;
4361         u64 last_unlink_trans;
4362
4363         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4364                 return -ENOTEMPTY;
4365         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4366                 return -EPERM;
4367
4368         trans = __unlink_start_trans(dir);
4369         if (IS_ERR(trans))
4370                 return PTR_ERR(trans);
4371
4372         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4373                 err = btrfs_unlink_subvol(trans, root, dir,
4374                                           BTRFS_I(inode)->location.objectid,
4375                                           dentry->d_name.name,
4376                                           dentry->d_name.len);
4377                 goto out;
4378         }
4379
4380         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4381         if (err)
4382                 goto out;
4383
4384         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4385
4386         /* now the directory is empty */
4387         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4388                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4389                         dentry->d_name.len);
4390         if (!err) {
4391                 btrfs_i_size_write(BTRFS_I(inode), 0);
4392                 /*
4393                  * Propagate the last_unlink_trans value of the deleted dir to
4394                  * its parent directory. This is to prevent an unrecoverable
4395                  * log tree in the case we do something like this:
4396                  * 1) create dir foo
4397                  * 2) create snapshot under dir foo
4398                  * 3) delete the snapshot
4399                  * 4) rmdir foo
4400                  * 5) mkdir foo
4401                  * 6) fsync foo or some file inside foo
4402                  */
4403                 if (last_unlink_trans >= trans->transid)
4404                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4405         }
4406 out:
4407         btrfs_end_transaction(trans);
4408         btrfs_btree_balance_dirty(root->fs_info);
4409
4410         return err;
4411 }
4412
4413 static int truncate_space_check(struct btrfs_trans_handle *trans,
4414                                 struct btrfs_root *root,
4415                                 u64 bytes_deleted)
4416 {
4417         struct btrfs_fs_info *fs_info = root->fs_info;
4418         int ret;
4419
4420         /*
4421          * This is only used to apply pressure to the enospc system, we don't
4422          * intend to use this reservation at all.
4423          */
4424         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4425         bytes_deleted *= fs_info->nodesize;
4426         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4427                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4428         if (!ret) {
4429                 trace_btrfs_space_reservation(fs_info, "transaction",
4430                                               trans->transid,
4431                                               bytes_deleted, 1);
4432                 trans->bytes_reserved += bytes_deleted;
4433         }
4434         return ret;
4435
4436 }
4437
4438 /*
4439  * Return this if we need to call truncate_block for the last bit of the
4440  * truncate.
4441  */
4442 #define NEED_TRUNCATE_BLOCK 1
4443
4444 /*
4445  * this can truncate away extent items, csum items and directory items.
4446  * It starts at a high offset and removes keys until it can't find
4447  * any higher than new_size
4448  *
4449  * csum items that cross the new i_size are truncated to the new size
4450  * as well.
4451  *
4452  * min_type is the minimum key type to truncate down to.  If set to 0, this
4453  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4454  */
4455 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4456                                struct btrfs_root *root,
4457                                struct inode *inode,
4458                                u64 new_size, u32 min_type)
4459 {
4460         struct btrfs_fs_info *fs_info = root->fs_info;
4461         struct btrfs_path *path;
4462         struct extent_buffer *leaf;
4463         struct btrfs_file_extent_item *fi;
4464         struct btrfs_key key;
4465         struct btrfs_key found_key;
4466         u64 extent_start = 0;
4467         u64 extent_num_bytes = 0;
4468         u64 extent_offset = 0;
4469         u64 item_end = 0;
4470         u64 last_size = new_size;
4471         u32 found_type = (u8)-1;
4472         int found_extent;
4473         int del_item;
4474         int pending_del_nr = 0;
4475         int pending_del_slot = 0;
4476         int extent_type = -1;
4477         int ret;
4478         int err = 0;
4479         u64 ino = btrfs_ino(BTRFS_I(inode));
4480         u64 bytes_deleted = 0;
4481         bool be_nice = false;
4482         bool should_throttle = false;
4483         bool should_end = false;
4484
4485         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4486
4487         /*
4488          * for non-free space inodes and ref cows, we want to back off from
4489          * time to time
4490          */
4491         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4492             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4493                 be_nice = true;
4494
4495         path = btrfs_alloc_path();
4496         if (!path)
4497                 return -ENOMEM;
4498         path->reada = READA_BACK;
4499
4500         /*
4501          * We want to drop from the next block forward in case this new size is
4502          * not block aligned since we will be keeping the last block of the
4503          * extent just the way it is.
4504          */
4505         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4506             root == fs_info->tree_root)
4507                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4508                                         fs_info->sectorsize),
4509                                         (u64)-1, 0);
4510
4511         /*
4512          * This function is also used to drop the items in the log tree before
4513          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4514          * it is used to drop the loged items. So we shouldn't kill the delayed
4515          * items.
4516          */
4517         if (min_type == 0 && root == BTRFS_I(inode)->root)
4518                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4519
4520         key.objectid = ino;
4521         key.offset = (u64)-1;
4522         key.type = (u8)-1;
4523
4524 search_again:
4525         /*
4526          * with a 16K leaf size and 128MB extents, you can actually queue
4527          * up a huge file in a single leaf.  Most of the time that
4528          * bytes_deleted is > 0, it will be huge by the time we get here
4529          */
4530         if (be_nice && bytes_deleted > SZ_32M) {
4531                 if (btrfs_should_end_transaction(trans)) {
4532                         err = -EAGAIN;
4533                         goto error;
4534                 }
4535         }
4536
4537
4538         path->leave_spinning = 1;
4539         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4540         if (ret < 0) {
4541                 err = ret;
4542                 goto out;
4543         }
4544
4545         if (ret > 0) {
4546                 /* there are no items in the tree for us to truncate, we're
4547                  * done
4548                  */
4549                 if (path->slots[0] == 0)
4550                         goto out;
4551                 path->slots[0]--;
4552         }
4553
4554         while (1) {
4555                 fi = NULL;
4556                 leaf = path->nodes[0];
4557                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4558                 found_type = found_key.type;
4559
4560                 if (found_key.objectid != ino)
4561                         break;
4562
4563                 if (found_type < min_type)
4564                         break;
4565
4566                 item_end = found_key.offset;
4567                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4568                         fi = btrfs_item_ptr(leaf, path->slots[0],
4569                                             struct btrfs_file_extent_item);
4570                         extent_type = btrfs_file_extent_type(leaf, fi);
4571                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4572                                 item_end +=
4573                                     btrfs_file_extent_num_bytes(leaf, fi);
4574
4575                                 trace_btrfs_truncate_show_fi_regular(
4576                                         BTRFS_I(inode), leaf, fi,
4577                                         found_key.offset);
4578                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4579                                 item_end += btrfs_file_extent_inline_len(leaf,
4580                                                          path->slots[0], fi);
4581
4582                                 trace_btrfs_truncate_show_fi_inline(
4583                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4584                                         found_key.offset);
4585                         }
4586                         item_end--;
4587                 }
4588                 if (found_type > min_type) {
4589                         del_item = 1;
4590                 } else {
4591                         if (item_end < new_size)
4592                                 break;
4593                         if (found_key.offset >= new_size)
4594                                 del_item = 1;
4595                         else
4596                                 del_item = 0;
4597                 }
4598                 found_extent = 0;
4599                 /* FIXME, shrink the extent if the ref count is only 1 */
4600                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4601                         goto delete;
4602
4603                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4604                         u64 num_dec;
4605                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4606                         if (!del_item) {
4607                                 u64 orig_num_bytes =
4608                                         btrfs_file_extent_num_bytes(leaf, fi);
4609                                 extent_num_bytes = ALIGN(new_size -
4610                                                 found_key.offset,
4611                                                 fs_info->sectorsize);
4612                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4613                                                          extent_num_bytes);
4614                                 num_dec = (orig_num_bytes -
4615                                            extent_num_bytes);
4616                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4617                                              &root->state) &&
4618                                     extent_start != 0)
4619                                         inode_sub_bytes(inode, num_dec);
4620                                 btrfs_mark_buffer_dirty(leaf);
4621                         } else {
4622                                 extent_num_bytes =
4623                                         btrfs_file_extent_disk_num_bytes(leaf,
4624                                                                          fi);
4625                                 extent_offset = found_key.offset -
4626                                         btrfs_file_extent_offset(leaf, fi);
4627
4628                                 /* FIXME blocksize != 4096 */
4629                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4630                                 if (extent_start != 0) {
4631                                         found_extent = 1;
4632                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4633                                                      &root->state))
4634                                                 inode_sub_bytes(inode, num_dec);
4635                                 }
4636                         }
4637                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4638                         /*
4639                          * we can't truncate inline items that have had
4640                          * special encodings
4641                          */
4642                         if (!del_item &&
4643                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4644                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4645                             btrfs_file_extent_compression(leaf, fi) == 0) {
4646                                 u32 size = (u32)(new_size - found_key.offset);
4647
4648                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4649                                 size = btrfs_file_extent_calc_inline_size(size);
4650                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4651                         } else if (!del_item) {
4652                                 /*
4653                                  * We have to bail so the last_size is set to
4654                                  * just before this extent.
4655                                  */
4656                                 err = NEED_TRUNCATE_BLOCK;
4657                                 break;
4658                         }
4659
4660                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4661                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4662                 }
4663 delete:
4664                 if (del_item)
4665                         last_size = found_key.offset;
4666                 else
4667                         last_size = new_size;
4668                 if (del_item) {
4669                         if (!pending_del_nr) {
4670                                 /* no pending yet, add ourselves */
4671                                 pending_del_slot = path->slots[0];
4672                                 pending_del_nr = 1;
4673                         } else if (pending_del_nr &&
4674                                    path->slots[0] + 1 == pending_del_slot) {
4675                                 /* hop on the pending chunk */
4676                                 pending_del_nr++;
4677                                 pending_del_slot = path->slots[0];
4678                         } else {
4679                                 BUG();
4680                         }
4681                 } else {
4682                         break;
4683                 }
4684                 should_throttle = false;
4685
4686                 if (found_extent &&
4687                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4688                      root == fs_info->tree_root)) {
4689                         btrfs_set_path_blocking(path);
4690                         bytes_deleted += extent_num_bytes;
4691                         ret = btrfs_free_extent(trans, root, extent_start,
4692                                                 extent_num_bytes, 0,
4693                                                 btrfs_header_owner(leaf),
4694                                                 ino, extent_offset);
4695                         BUG_ON(ret);
4696                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4697                                 btrfs_async_run_delayed_refs(fs_info,
4698                                         trans->delayed_ref_updates * 2,
4699                                         trans->transid, 0);
4700                         if (be_nice) {
4701                                 if (truncate_space_check(trans, root,
4702                                                          extent_num_bytes)) {
4703                                         should_end = true;
4704                                 }
4705                                 if (btrfs_should_throttle_delayed_refs(trans,
4706                                                                        fs_info))
4707                                         should_throttle = true;
4708                         }
4709                 }
4710
4711                 if (found_type == BTRFS_INODE_ITEM_KEY)
4712                         break;
4713
4714                 if (path->slots[0] == 0 ||
4715                     path->slots[0] != pending_del_slot ||
4716                     should_throttle || should_end) {
4717                         if (pending_del_nr) {
4718                                 ret = btrfs_del_items(trans, root, path,
4719                                                 pending_del_slot,
4720                                                 pending_del_nr);
4721                                 if (ret) {
4722                                         btrfs_abort_transaction(trans, ret);
4723                                         goto error;
4724                                 }
4725                                 pending_del_nr = 0;
4726                         }
4727                         btrfs_release_path(path);
4728                         if (should_throttle) {
4729                                 unsigned long updates = trans->delayed_ref_updates;
4730                                 if (updates) {
4731                                         trans->delayed_ref_updates = 0;
4732                                         ret = btrfs_run_delayed_refs(trans,
4733                                                                    fs_info,
4734                                                                    updates * 2);
4735                                         if (ret && !err)
4736                                                 err = ret;
4737                                 }
4738                         }
4739                         /*
4740                          * if we failed to refill our space rsv, bail out
4741                          * and let the transaction restart
4742                          */
4743                         if (should_end) {
4744                                 err = -EAGAIN;
4745                                 goto error;
4746                         }
4747                         goto search_again;
4748                 } else {
4749                         path->slots[0]--;
4750                 }
4751         }
4752 out:
4753         if (pending_del_nr) {
4754                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4755                                       pending_del_nr);
4756                 if (ret)
4757                         btrfs_abort_transaction(trans, ret);
4758         }
4759 error:
4760         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4761                 ASSERT(last_size >= new_size);
4762                 if (!err && last_size > new_size)
4763                         last_size = new_size;
4764                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4765         }
4766
4767         btrfs_free_path(path);
4768
4769         if (be_nice && bytes_deleted > SZ_32M) {
4770                 unsigned long updates = trans->delayed_ref_updates;
4771                 if (updates) {
4772                         trans->delayed_ref_updates = 0;
4773                         ret = btrfs_run_delayed_refs(trans, fs_info,
4774                                                      updates * 2);
4775                         if (ret && !err)
4776                                 err = ret;
4777                 }
4778         }
4779         return err;
4780 }
4781
4782 /*
4783  * btrfs_truncate_block - read, zero a chunk and write a block
4784  * @inode - inode that we're zeroing
4785  * @from - the offset to start zeroing
4786  * @len - the length to zero, 0 to zero the entire range respective to the
4787  *      offset
4788  * @front - zero up to the offset instead of from the offset on
4789  *
4790  * This will find the block for the "from" offset and cow the block and zero the
4791  * part we want to zero.  This is used with truncate and hole punching.
4792  */
4793 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4794                         int front)
4795 {
4796         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4797         struct address_space *mapping = inode->i_mapping;
4798         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4799         struct btrfs_ordered_extent *ordered;
4800         struct extent_state *cached_state = NULL;
4801         struct extent_changeset *data_reserved = NULL;
4802         char *kaddr;
4803         u32 blocksize = fs_info->sectorsize;
4804         pgoff_t index = from >> PAGE_SHIFT;
4805         unsigned offset = from & (blocksize - 1);
4806         struct page *page;
4807         gfp_t mask = btrfs_alloc_write_mask(mapping);
4808         int ret = 0;
4809         u64 block_start;
4810         u64 block_end;
4811
4812         if (IS_ALIGNED(offset, blocksize) &&
4813             (!len || IS_ALIGNED(len, blocksize)))
4814                 goto out;
4815
4816         block_start = round_down(from, blocksize);
4817         block_end = block_start + blocksize - 1;
4818
4819         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4820                                            block_start, blocksize);
4821         if (ret)
4822                 goto out;
4823
4824 again:
4825         page = find_or_create_page(mapping, index, mask);
4826         if (!page) {
4827                 btrfs_delalloc_release_space(inode, data_reserved,
4828                                              block_start, blocksize);
4829                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4830                 ret = -ENOMEM;
4831                 goto out;
4832         }
4833
4834         if (!PageUptodate(page)) {
4835                 ret = btrfs_readpage(NULL, page);
4836                 lock_page(page);
4837                 if (page->mapping != mapping) {
4838                         unlock_page(page);
4839                         put_page(page);
4840                         goto again;
4841                 }
4842                 if (!PageUptodate(page)) {
4843                         ret = -EIO;
4844                         goto out_unlock;
4845                 }
4846         }
4847         wait_on_page_writeback(page);
4848
4849         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4850         set_page_extent_mapped(page);
4851
4852         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4853         if (ordered) {
4854                 unlock_extent_cached(io_tree, block_start, block_end,
4855                                      &cached_state);
4856                 unlock_page(page);
4857                 put_page(page);
4858                 btrfs_start_ordered_extent(inode, ordered, 1);
4859                 btrfs_put_ordered_extent(ordered);
4860                 goto again;
4861         }
4862
4863         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4864                           EXTENT_DIRTY | EXTENT_DELALLOC |
4865                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4866                           0, 0, &cached_state);
4867
4868         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4869                                         &cached_state, 0);
4870         if (ret) {
4871                 unlock_extent_cached(io_tree, block_start, block_end,
4872                                      &cached_state);
4873                 goto out_unlock;
4874         }
4875
4876         if (offset != blocksize) {
4877                 if (!len)
4878                         len = blocksize - offset;
4879                 kaddr = kmap(page);
4880                 if (front)
4881                         memset(kaddr + (block_start - page_offset(page)),
4882                                 0, offset);
4883                 else
4884                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4885                                 0, len);
4886                 flush_dcache_page(page);
4887                 kunmap(page);
4888         }
4889         ClearPageChecked(page);
4890         set_page_dirty(page);
4891         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
4892
4893 out_unlock:
4894         if (ret)
4895                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4896                                              blocksize);
4897         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4898         unlock_page(page);
4899         put_page(page);
4900 out:
4901         extent_changeset_free(data_reserved);
4902         return ret;
4903 }
4904
4905 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4906                              u64 offset, u64 len)
4907 {
4908         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4909         struct btrfs_trans_handle *trans;
4910         int ret;
4911
4912         /*
4913          * Still need to make sure the inode looks like it's been updated so
4914          * that any holes get logged if we fsync.
4915          */
4916         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4917                 BTRFS_I(inode)->last_trans = fs_info->generation;
4918                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4919                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4920                 return 0;
4921         }
4922
4923         /*
4924          * 1 - for the one we're dropping
4925          * 1 - for the one we're adding
4926          * 1 - for updating the inode.
4927          */
4928         trans = btrfs_start_transaction(root, 3);
4929         if (IS_ERR(trans))
4930                 return PTR_ERR(trans);
4931
4932         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4933         if (ret) {
4934                 btrfs_abort_transaction(trans, ret);
4935                 btrfs_end_transaction(trans);
4936                 return ret;
4937         }
4938
4939         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4940                         offset, 0, 0, len, 0, len, 0, 0, 0);
4941         if (ret)
4942                 btrfs_abort_transaction(trans, ret);
4943         else
4944                 btrfs_update_inode(trans, root, inode);
4945         btrfs_end_transaction(trans);
4946         return ret;
4947 }
4948
4949 /*
4950  * This function puts in dummy file extents for the area we're creating a hole
4951  * for.  So if we are truncating this file to a larger size we need to insert
4952  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4953  * the range between oldsize and size
4954  */
4955 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4956 {
4957         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4958         struct btrfs_root *root = BTRFS_I(inode)->root;
4959         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4960         struct extent_map *em = NULL;
4961         struct extent_state *cached_state = NULL;
4962         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4963         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4964         u64 block_end = ALIGN(size, fs_info->sectorsize);
4965         u64 last_byte;
4966         u64 cur_offset;
4967         u64 hole_size;
4968         int err = 0;
4969
4970         /*
4971          * If our size started in the middle of a block we need to zero out the
4972          * rest of the block before we expand the i_size, otherwise we could
4973          * expose stale data.
4974          */
4975         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4976         if (err)
4977                 return err;
4978
4979         if (size <= hole_start)
4980                 return 0;
4981
4982         while (1) {
4983                 struct btrfs_ordered_extent *ordered;
4984
4985                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4986                                  &cached_state);
4987                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4988                                                      block_end - hole_start);
4989                 if (!ordered)
4990                         break;
4991                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4992                                      &cached_state);
4993                 btrfs_start_ordered_extent(inode, ordered, 1);
4994                 btrfs_put_ordered_extent(ordered);
4995         }
4996
4997         cur_offset = hole_start;
4998         while (1) {
4999                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5000                                 block_end - cur_offset, 0);
5001                 if (IS_ERR(em)) {
5002                         err = PTR_ERR(em);
5003                         em = NULL;
5004                         break;
5005                 }
5006                 last_byte = min(extent_map_end(em), block_end);
5007                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5008                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5009                         struct extent_map *hole_em;
5010                         hole_size = last_byte - cur_offset;
5011
5012                         err = maybe_insert_hole(root, inode, cur_offset,
5013                                                 hole_size);
5014                         if (err)
5015                                 break;
5016                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5017                                                 cur_offset + hole_size - 1, 0);
5018                         hole_em = alloc_extent_map();
5019                         if (!hole_em) {
5020                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5021                                         &BTRFS_I(inode)->runtime_flags);
5022                                 goto next;
5023                         }
5024                         hole_em->start = cur_offset;
5025                         hole_em->len = hole_size;
5026                         hole_em->orig_start = cur_offset;
5027
5028                         hole_em->block_start = EXTENT_MAP_HOLE;
5029                         hole_em->block_len = 0;
5030                         hole_em->orig_block_len = 0;
5031                         hole_em->ram_bytes = hole_size;
5032                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5033                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5034                         hole_em->generation = fs_info->generation;
5035
5036                         while (1) {
5037                                 write_lock(&em_tree->lock);
5038                                 err = add_extent_mapping(em_tree, hole_em, 1);
5039                                 write_unlock(&em_tree->lock);
5040                                 if (err != -EEXIST)
5041                                         break;
5042                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5043                                                         cur_offset,
5044                                                         cur_offset +
5045                                                         hole_size - 1, 0);
5046                         }
5047                         free_extent_map(hole_em);
5048                 }
5049 next:
5050                 free_extent_map(em);
5051                 em = NULL;
5052                 cur_offset = last_byte;
5053                 if (cur_offset >= block_end)
5054                         break;
5055         }
5056         free_extent_map(em);
5057         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5058         return err;
5059 }
5060
5061 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5062 {
5063         struct btrfs_root *root = BTRFS_I(inode)->root;
5064         struct btrfs_trans_handle *trans;
5065         loff_t oldsize = i_size_read(inode);
5066         loff_t newsize = attr->ia_size;
5067         int mask = attr->ia_valid;
5068         int ret;
5069
5070         /*
5071          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5072          * special case where we need to update the times despite not having
5073          * these flags set.  For all other operations the VFS set these flags
5074          * explicitly if it wants a timestamp update.
5075          */
5076         if (newsize != oldsize) {
5077                 inode_inc_iversion(inode);
5078                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5079                         inode->i_ctime = inode->i_mtime =
5080                                 current_time(inode);
5081         }
5082
5083         if (newsize > oldsize) {
5084                 /*
5085                  * Don't do an expanding truncate while snapshotting is ongoing.
5086                  * This is to ensure the snapshot captures a fully consistent
5087                  * state of this file - if the snapshot captures this expanding
5088                  * truncation, it must capture all writes that happened before
5089                  * this truncation.
5090                  */
5091                 btrfs_wait_for_snapshot_creation(root);
5092                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5093                 if (ret) {
5094                         btrfs_end_write_no_snapshotting(root);
5095                         return ret;
5096                 }
5097
5098                 trans = btrfs_start_transaction(root, 1);
5099                 if (IS_ERR(trans)) {
5100                         btrfs_end_write_no_snapshotting(root);
5101                         return PTR_ERR(trans);
5102                 }
5103
5104                 i_size_write(inode, newsize);
5105                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5106                 pagecache_isize_extended(inode, oldsize, newsize);
5107                 ret = btrfs_update_inode(trans, root, inode);
5108                 btrfs_end_write_no_snapshotting(root);
5109                 btrfs_end_transaction(trans);
5110         } else {
5111
5112                 /*
5113                  * We're truncating a file that used to have good data down to
5114                  * zero. Make sure it gets into the ordered flush list so that
5115                  * any new writes get down to disk quickly.
5116                  */
5117                 if (newsize == 0)
5118                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5119                                 &BTRFS_I(inode)->runtime_flags);
5120
5121                 /*
5122                  * 1 for the orphan item we're going to add
5123                  * 1 for the orphan item deletion.
5124                  */
5125                 trans = btrfs_start_transaction(root, 2);
5126                 if (IS_ERR(trans))
5127                         return PTR_ERR(trans);
5128
5129                 /*
5130                  * We need to do this in case we fail at _any_ point during the
5131                  * actual truncate.  Once we do the truncate_setsize we could
5132                  * invalidate pages which forces any outstanding ordered io to
5133                  * be instantly completed which will give us extents that need
5134                  * to be truncated.  If we fail to get an orphan inode down we
5135                  * could have left over extents that were never meant to live,
5136                  * so we need to guarantee from this point on that everything
5137                  * will be consistent.
5138                  */
5139                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5140                 btrfs_end_transaction(trans);
5141                 if (ret)
5142                         return ret;
5143
5144                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5145                 truncate_setsize(inode, newsize);
5146
5147                 /* Disable nonlocked read DIO to avoid the end less truncate */
5148                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5149                 inode_dio_wait(inode);
5150                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5151
5152                 ret = btrfs_truncate(inode, newsize == oldsize);
5153                 if (ret && inode->i_nlink) {
5154                         int err;
5155
5156                         /* To get a stable disk_i_size */
5157                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5158                         if (err) {
5159                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5160                                 return err;
5161                         }
5162
5163                         /*
5164                          * failed to truncate, disk_i_size is only adjusted down
5165                          * as we remove extents, so it should represent the true
5166                          * size of the inode, so reset the in memory size and
5167                          * delete our orphan entry.
5168                          */
5169                         trans = btrfs_join_transaction(root);
5170                         if (IS_ERR(trans)) {
5171                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5172                                 return ret;
5173                         }
5174                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5175                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5176                         if (err)
5177                                 btrfs_abort_transaction(trans, err);
5178                         btrfs_end_transaction(trans);
5179                 }
5180         }
5181
5182         return ret;
5183 }
5184
5185 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5186 {
5187         struct inode *inode = d_inode(dentry);
5188         struct btrfs_root *root = BTRFS_I(inode)->root;
5189         int err;
5190
5191         if (btrfs_root_readonly(root))
5192                 return -EROFS;
5193
5194         err = setattr_prepare(dentry, attr);
5195         if (err)
5196                 return err;
5197
5198         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5199                 err = btrfs_setsize(inode, attr);
5200                 if (err)
5201                         return err;
5202         }
5203
5204         if (attr->ia_valid) {
5205                 setattr_copy(inode, attr);
5206                 inode_inc_iversion(inode);
5207                 err = btrfs_dirty_inode(inode);
5208
5209                 if (!err && attr->ia_valid & ATTR_MODE)
5210                         err = posix_acl_chmod(inode, inode->i_mode);
5211         }
5212
5213         return err;
5214 }
5215
5216 /*
5217  * While truncating the inode pages during eviction, we get the VFS calling
5218  * btrfs_invalidatepage() against each page of the inode. This is slow because
5219  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5220  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5221  * extent_state structures over and over, wasting lots of time.
5222  *
5223  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5224  * those expensive operations on a per page basis and do only the ordered io
5225  * finishing, while we release here the extent_map and extent_state structures,
5226  * without the excessive merging and splitting.
5227  */
5228 static void evict_inode_truncate_pages(struct inode *inode)
5229 {
5230         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5231         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5232         struct rb_node *node;
5233
5234         ASSERT(inode->i_state & I_FREEING);
5235         truncate_inode_pages_final(&inode->i_data);
5236
5237         write_lock(&map_tree->lock);
5238         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5239                 struct extent_map *em;
5240
5241                 node = rb_first(&map_tree->map);
5242                 em = rb_entry(node, struct extent_map, rb_node);
5243                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5244                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5245                 remove_extent_mapping(map_tree, em);
5246                 free_extent_map(em);
5247                 if (need_resched()) {
5248                         write_unlock(&map_tree->lock);
5249                         cond_resched();
5250                         write_lock(&map_tree->lock);
5251                 }
5252         }
5253         write_unlock(&map_tree->lock);
5254
5255         /*
5256          * Keep looping until we have no more ranges in the io tree.
5257          * We can have ongoing bios started by readpages (called from readahead)
5258          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5259          * still in progress (unlocked the pages in the bio but did not yet
5260          * unlocked the ranges in the io tree). Therefore this means some
5261          * ranges can still be locked and eviction started because before
5262          * submitting those bios, which are executed by a separate task (work
5263          * queue kthread), inode references (inode->i_count) were not taken
5264          * (which would be dropped in the end io callback of each bio).
5265          * Therefore here we effectively end up waiting for those bios and
5266          * anyone else holding locked ranges without having bumped the inode's
5267          * reference count - if we don't do it, when they access the inode's
5268          * io_tree to unlock a range it may be too late, leading to an
5269          * use-after-free issue.
5270          */
5271         spin_lock(&io_tree->lock);
5272         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5273                 struct extent_state *state;
5274                 struct extent_state *cached_state = NULL;
5275                 u64 start;
5276                 u64 end;
5277
5278                 node = rb_first(&io_tree->state);
5279                 state = rb_entry(node, struct extent_state, rb_node);
5280                 start = state->start;
5281                 end = state->end;
5282                 spin_unlock(&io_tree->lock);
5283
5284                 lock_extent_bits(io_tree, start, end, &cached_state);
5285
5286                 /*
5287                  * If still has DELALLOC flag, the extent didn't reach disk,
5288                  * and its reserved space won't be freed by delayed_ref.
5289                  * So we need to free its reserved space here.
5290                  * (Refer to comment in btrfs_invalidatepage, case 2)
5291                  *
5292                  * Note, end is the bytenr of last byte, so we need + 1 here.
5293                  */
5294                 if (state->state & EXTENT_DELALLOC)
5295                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5296
5297                 clear_extent_bit(io_tree, start, end,
5298                                  EXTENT_LOCKED | EXTENT_DIRTY |
5299                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5300                                  EXTENT_DEFRAG, 1, 1, &cached_state);
5301
5302                 cond_resched();
5303                 spin_lock(&io_tree->lock);
5304         }
5305         spin_unlock(&io_tree->lock);
5306 }
5307
5308 void btrfs_evict_inode(struct inode *inode)
5309 {
5310         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5311         struct btrfs_trans_handle *trans;
5312         struct btrfs_root *root = BTRFS_I(inode)->root;
5313         struct btrfs_block_rsv *rsv, *global_rsv;
5314         int steal_from_global = 0;
5315         u64 min_size;
5316         int ret;
5317
5318         trace_btrfs_inode_evict(inode);
5319
5320         if (!root) {
5321                 clear_inode(inode);
5322                 return;
5323         }
5324
5325         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5326
5327         evict_inode_truncate_pages(inode);
5328
5329         if (inode->i_nlink &&
5330             ((btrfs_root_refs(&root->root_item) != 0 &&
5331               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5332              btrfs_is_free_space_inode(BTRFS_I(inode))))
5333                 goto no_delete;
5334
5335         if (is_bad_inode(inode)) {
5336                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5337                 goto no_delete;
5338         }
5339         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5340         if (!special_file(inode->i_mode))
5341                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5342
5343         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5344
5345         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5346                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5347                                  &BTRFS_I(inode)->runtime_flags));
5348                 goto no_delete;
5349         }
5350
5351         if (inode->i_nlink > 0) {
5352                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5353                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5354                 goto no_delete;
5355         }
5356
5357         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5358         if (ret) {
5359                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5360                 goto no_delete;
5361         }
5362
5363         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5364         if (!rsv) {
5365                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5366                 goto no_delete;
5367         }
5368         rsv->size = min_size;
5369         rsv->failfast = 1;
5370         global_rsv = &fs_info->global_block_rsv;
5371
5372         btrfs_i_size_write(BTRFS_I(inode), 0);
5373
5374         /*
5375          * This is a bit simpler than btrfs_truncate since we've already
5376          * reserved our space for our orphan item in the unlink, so we just
5377          * need to reserve some slack space in case we add bytes and update
5378          * inode item when doing the truncate.
5379          */
5380         while (1) {
5381                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5382                                              BTRFS_RESERVE_FLUSH_LIMIT);
5383
5384                 /*
5385                  * Try and steal from the global reserve since we will
5386                  * likely not use this space anyway, we want to try as
5387                  * hard as possible to get this to work.
5388                  */
5389                 if (ret)
5390                         steal_from_global++;
5391                 else
5392                         steal_from_global = 0;
5393                 ret = 0;
5394
5395                 /*
5396                  * steal_from_global == 0: we reserved stuff, hooray!
5397                  * steal_from_global == 1: we didn't reserve stuff, boo!
5398                  * steal_from_global == 2: we've committed, still not a lot of
5399                  * room but maybe we'll have room in the global reserve this
5400                  * time.
5401                  * steal_from_global == 3: abandon all hope!
5402                  */
5403                 if (steal_from_global > 2) {
5404                         btrfs_warn(fs_info,
5405                                    "Could not get space for a delete, will truncate on mount %d",
5406                                    ret);
5407                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5408                         btrfs_free_block_rsv(fs_info, rsv);
5409                         goto no_delete;
5410                 }
5411
5412                 trans = btrfs_join_transaction(root);
5413                 if (IS_ERR(trans)) {
5414                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5415                         btrfs_free_block_rsv(fs_info, rsv);
5416                         goto no_delete;
5417                 }
5418
5419                 /*
5420                  * We can't just steal from the global reserve, we need to make
5421                  * sure there is room to do it, if not we need to commit and try
5422                  * again.
5423                  */
5424                 if (steal_from_global) {
5425                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5426                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5427                                                               min_size, 0);
5428                         else
5429                                 ret = -ENOSPC;
5430                 }
5431
5432                 /*
5433                  * Couldn't steal from the global reserve, we have too much
5434                  * pending stuff built up, commit the transaction and try it
5435                  * again.
5436                  */
5437                 if (ret) {
5438                         ret = btrfs_commit_transaction(trans);
5439                         if (ret) {
5440                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5441                                 btrfs_free_block_rsv(fs_info, rsv);
5442                                 goto no_delete;
5443                         }
5444                         continue;
5445                 } else {
5446                         steal_from_global = 0;
5447                 }
5448
5449                 trans->block_rsv = rsv;
5450
5451                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5452                 if (ret != -ENOSPC && ret != -EAGAIN)
5453                         break;
5454
5455                 trans->block_rsv = &fs_info->trans_block_rsv;
5456                 btrfs_end_transaction(trans);
5457                 trans = NULL;
5458                 btrfs_btree_balance_dirty(fs_info);
5459         }
5460
5461         btrfs_free_block_rsv(fs_info, rsv);
5462
5463         /*
5464          * Errors here aren't a big deal, it just means we leave orphan items
5465          * in the tree.  They will be cleaned up on the next mount.
5466          */
5467         if (ret == 0) {
5468                 trans->block_rsv = root->orphan_block_rsv;
5469                 btrfs_orphan_del(trans, BTRFS_I(inode));
5470         } else {
5471                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5472         }
5473
5474         trans->block_rsv = &fs_info->trans_block_rsv;
5475         if (!(root == fs_info->tree_root ||
5476               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5477                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5478
5479         btrfs_end_transaction(trans);
5480         btrfs_btree_balance_dirty(fs_info);
5481 no_delete:
5482         btrfs_remove_delayed_node(BTRFS_I(inode));
5483         clear_inode(inode);
5484 }
5485
5486 /*
5487  * this returns the key found in the dir entry in the location pointer.
5488  * If no dir entries were found, returns -ENOENT.
5489  * If found a corrupted location in dir entry, returns -EUCLEAN.
5490  */
5491 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5492                                struct btrfs_key *location)
5493 {
5494         const char *name = dentry->d_name.name;
5495         int namelen = dentry->d_name.len;
5496         struct btrfs_dir_item *di;
5497         struct btrfs_path *path;
5498         struct btrfs_root *root = BTRFS_I(dir)->root;
5499         int ret = 0;
5500
5501         path = btrfs_alloc_path();
5502         if (!path)
5503                 return -ENOMEM;
5504
5505         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5506                         name, namelen, 0);
5507         if (!di) {
5508                 ret = -ENOENT;
5509                 goto out;
5510         }
5511         if (IS_ERR(di)) {
5512                 ret = PTR_ERR(di);
5513                 goto out;
5514         }
5515
5516         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5517         if (location->type != BTRFS_INODE_ITEM_KEY &&
5518             location->type != BTRFS_ROOT_ITEM_KEY) {
5519                 ret = -EUCLEAN;
5520                 btrfs_warn(root->fs_info,
5521 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5522                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5523                            location->objectid, location->type, location->offset);
5524         }
5525 out:
5526         btrfs_free_path(path);
5527         return ret;
5528 }
5529
5530 /*
5531  * when we hit a tree root in a directory, the btrfs part of the inode
5532  * needs to be changed to reflect the root directory of the tree root.  This
5533  * is kind of like crossing a mount point.
5534  */
5535 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5536                                     struct inode *dir,
5537                                     struct dentry *dentry,
5538                                     struct btrfs_key *location,
5539                                     struct btrfs_root **sub_root)
5540 {
5541         struct btrfs_path *path;
5542         struct btrfs_root *new_root;
5543         struct btrfs_root_ref *ref;
5544         struct extent_buffer *leaf;
5545         struct btrfs_key key;
5546         int ret;
5547         int err = 0;
5548
5549         path = btrfs_alloc_path();
5550         if (!path) {
5551                 err = -ENOMEM;
5552                 goto out;
5553         }
5554
5555         err = -ENOENT;
5556         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5557         key.type = BTRFS_ROOT_REF_KEY;
5558         key.offset = location->objectid;
5559
5560         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5561         if (ret) {
5562                 if (ret < 0)
5563                         err = ret;
5564                 goto out;
5565         }
5566
5567         leaf = path->nodes[0];
5568         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5569         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5570             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5571                 goto out;
5572
5573         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5574                                    (unsigned long)(ref + 1),
5575                                    dentry->d_name.len);
5576         if (ret)
5577                 goto out;
5578
5579         btrfs_release_path(path);
5580
5581         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5582         if (IS_ERR(new_root)) {
5583                 err = PTR_ERR(new_root);
5584                 goto out;
5585         }
5586
5587         *sub_root = new_root;
5588         location->objectid = btrfs_root_dirid(&new_root->root_item);
5589         location->type = BTRFS_INODE_ITEM_KEY;
5590         location->offset = 0;
5591         err = 0;
5592 out:
5593         btrfs_free_path(path);
5594         return err;
5595 }
5596
5597 static void inode_tree_add(struct inode *inode)
5598 {
5599         struct btrfs_root *root = BTRFS_I(inode)->root;
5600         struct btrfs_inode *entry;
5601         struct rb_node **p;
5602         struct rb_node *parent;
5603         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5604         u64 ino = btrfs_ino(BTRFS_I(inode));
5605
5606         if (inode_unhashed(inode))
5607                 return;
5608         parent = NULL;
5609         spin_lock(&root->inode_lock);
5610         p = &root->inode_tree.rb_node;
5611         while (*p) {
5612                 parent = *p;
5613                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5614
5615                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5616                         p = &parent->rb_left;
5617                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5618                         p = &parent->rb_right;
5619                 else {
5620                         WARN_ON(!(entry->vfs_inode.i_state &
5621                                   (I_WILL_FREE | I_FREEING)));
5622                         rb_replace_node(parent, new, &root->inode_tree);
5623                         RB_CLEAR_NODE(parent);
5624                         spin_unlock(&root->inode_lock);
5625                         return;
5626                 }
5627         }
5628         rb_link_node(new, parent, p);
5629         rb_insert_color(new, &root->inode_tree);
5630         spin_unlock(&root->inode_lock);
5631 }
5632
5633 static void inode_tree_del(struct inode *inode)
5634 {
5635         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5636         struct btrfs_root *root = BTRFS_I(inode)->root;
5637         int empty = 0;
5638
5639         spin_lock(&root->inode_lock);
5640         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5641                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5642                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5643                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5644         }
5645         spin_unlock(&root->inode_lock);
5646
5647         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5648                 synchronize_srcu(&fs_info->subvol_srcu);
5649                 spin_lock(&root->inode_lock);
5650                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5651                 spin_unlock(&root->inode_lock);
5652                 if (empty)
5653                         btrfs_add_dead_root(root);
5654         }
5655 }
5656
5657 void btrfs_invalidate_inodes(struct btrfs_root *root)
5658 {
5659         struct btrfs_fs_info *fs_info = root->fs_info;
5660         struct rb_node *node;
5661         struct rb_node *prev;
5662         struct btrfs_inode *entry;
5663         struct inode *inode;
5664         u64 objectid = 0;
5665
5666         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5667                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5668
5669         spin_lock(&root->inode_lock);
5670 again:
5671         node = root->inode_tree.rb_node;
5672         prev = NULL;
5673         while (node) {
5674                 prev = node;
5675                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5676
5677                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5678                         node = node->rb_left;
5679                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5680                         node = node->rb_right;
5681                 else
5682                         break;
5683         }
5684         if (!node) {
5685                 while (prev) {
5686                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5687                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5688                                 node = prev;
5689                                 break;
5690                         }
5691                         prev = rb_next(prev);
5692                 }
5693         }
5694         while (node) {
5695                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5696                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5697                 inode = igrab(&entry->vfs_inode);
5698                 if (inode) {
5699                         spin_unlock(&root->inode_lock);
5700                         if (atomic_read(&inode->i_count) > 1)
5701                                 d_prune_aliases(inode);
5702                         /*
5703                          * btrfs_drop_inode will have it removed from
5704                          * the inode cache when its usage count
5705                          * hits zero.
5706                          */
5707                         iput(inode);
5708                         cond_resched();
5709                         spin_lock(&root->inode_lock);
5710                         goto again;
5711                 }
5712
5713                 if (cond_resched_lock(&root->inode_lock))
5714                         goto again;
5715
5716                 node = rb_next(node);
5717         }
5718         spin_unlock(&root->inode_lock);
5719 }
5720
5721 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5722 {
5723         struct btrfs_iget_args *args = p;
5724         inode->i_ino = args->location->objectid;
5725         memcpy(&BTRFS_I(inode)->location, args->location,
5726                sizeof(*args->location));
5727         BTRFS_I(inode)->root = args->root;
5728         return 0;
5729 }
5730
5731 static int btrfs_find_actor(struct inode *inode, void *opaque)
5732 {
5733         struct btrfs_iget_args *args = opaque;
5734         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5735                 args->root == BTRFS_I(inode)->root;
5736 }
5737
5738 static struct inode *btrfs_iget_locked(struct super_block *s,
5739                                        struct btrfs_key *location,
5740                                        struct btrfs_root *root)
5741 {
5742         struct inode *inode;
5743         struct btrfs_iget_args args;
5744         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5745
5746         args.location = location;
5747         args.root = root;
5748
5749         inode = iget5_locked(s, hashval, btrfs_find_actor,
5750                              btrfs_init_locked_inode,
5751                              (void *)&args);
5752         return inode;
5753 }
5754
5755 /* Get an inode object given its location and corresponding root.
5756  * Returns in *is_new if the inode was read from disk
5757  */
5758 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5759                          struct btrfs_root *root, int *new)
5760 {
5761         struct inode *inode;
5762
5763         inode = btrfs_iget_locked(s, location, root);
5764         if (!inode)
5765                 return ERR_PTR(-ENOMEM);
5766
5767         if (inode->i_state & I_NEW) {
5768                 int ret;
5769
5770                 ret = btrfs_read_locked_inode(inode);
5771                 if (!is_bad_inode(inode)) {
5772                         inode_tree_add(inode);
5773                         unlock_new_inode(inode);
5774                         if (new)
5775                                 *new = 1;
5776                 } else {
5777                         unlock_new_inode(inode);
5778                         iput(inode);
5779                         ASSERT(ret < 0);
5780                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5781                 }
5782         }
5783
5784         return inode;
5785 }
5786
5787 static struct inode *new_simple_dir(struct super_block *s,
5788                                     struct btrfs_key *key,
5789                                     struct btrfs_root *root)
5790 {
5791         struct inode *inode = new_inode(s);
5792
5793         if (!inode)
5794                 return ERR_PTR(-ENOMEM);
5795
5796         BTRFS_I(inode)->root = root;
5797         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5798         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5799
5800         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5801         inode->i_op = &btrfs_dir_ro_inode_operations;
5802         inode->i_opflags &= ~IOP_XATTR;
5803         inode->i_fop = &simple_dir_operations;
5804         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5805         inode->i_mtime = current_time(inode);
5806         inode->i_atime = inode->i_mtime;
5807         inode->i_ctime = inode->i_mtime;
5808         BTRFS_I(inode)->i_otime = inode->i_mtime;
5809
5810         return inode;
5811 }
5812
5813 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5814 {
5815         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5816         struct inode *inode;
5817         struct btrfs_root *root = BTRFS_I(dir)->root;
5818         struct btrfs_root *sub_root = root;
5819         struct btrfs_key location;
5820         int index;
5821         int ret = 0;
5822
5823         if (dentry->d_name.len > BTRFS_NAME_LEN)
5824                 return ERR_PTR(-ENAMETOOLONG);
5825
5826         ret = btrfs_inode_by_name(dir, dentry, &location);
5827         if (ret < 0)
5828                 return ERR_PTR(ret);
5829
5830         if (location.type == BTRFS_INODE_ITEM_KEY) {
5831                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5832                 return inode;
5833         }
5834
5835         index = srcu_read_lock(&fs_info->subvol_srcu);
5836         ret = fixup_tree_root_location(fs_info, dir, dentry,
5837                                        &location, &sub_root);
5838         if (ret < 0) {
5839                 if (ret != -ENOENT)
5840                         inode = ERR_PTR(ret);
5841                 else
5842                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5843         } else {
5844                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5845         }
5846         srcu_read_unlock(&fs_info->subvol_srcu, index);
5847
5848         if (!IS_ERR(inode) && root != sub_root) {
5849                 down_read(&fs_info->cleanup_work_sem);
5850                 if (!sb_rdonly(inode->i_sb))
5851                         ret = btrfs_orphan_cleanup(sub_root);
5852                 up_read(&fs_info->cleanup_work_sem);
5853                 if (ret) {
5854                         iput(inode);
5855                         inode = ERR_PTR(ret);
5856                 }
5857         }
5858
5859         return inode;
5860 }
5861
5862 static int btrfs_dentry_delete(const struct dentry *dentry)
5863 {
5864         struct btrfs_root *root;
5865         struct inode *inode = d_inode(dentry);
5866
5867         if (!inode && !IS_ROOT(dentry))
5868                 inode = d_inode(dentry->d_parent);
5869
5870         if (inode) {
5871                 root = BTRFS_I(inode)->root;
5872                 if (btrfs_root_refs(&root->root_item) == 0)
5873                         return 1;
5874
5875                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5876                         return 1;
5877         }
5878         return 0;
5879 }
5880
5881 static void btrfs_dentry_release(struct dentry *dentry)
5882 {
5883         kfree(dentry->d_fsdata);
5884 }
5885
5886 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5887                                    unsigned int flags)
5888 {
5889         struct inode *inode;
5890
5891         inode = btrfs_lookup_dentry(dir, dentry);
5892         if (IS_ERR(inode)) {
5893                 if (PTR_ERR(inode) == -ENOENT)
5894                         inode = NULL;
5895                 else
5896                         return ERR_CAST(inode);
5897         }
5898
5899         return d_splice_alias(inode, dentry);
5900 }
5901
5902 unsigned char btrfs_filetype_table[] = {
5903         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5904 };
5905
5906 /*
5907  * All this infrastructure exists because dir_emit can fault, and we are holding
5908  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5909  * our information into that, and then dir_emit from the buffer.  This is
5910  * similar to what NFS does, only we don't keep the buffer around in pagecache
5911  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5912  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5913  * tree lock.
5914  */
5915 static int btrfs_opendir(struct inode *inode, struct file *file)
5916 {
5917         struct btrfs_file_private *private;
5918
5919         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5920         if (!private)
5921                 return -ENOMEM;
5922         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5923         if (!private->filldir_buf) {
5924                 kfree(private);
5925                 return -ENOMEM;
5926         }
5927         file->private_data = private;
5928         return 0;
5929 }
5930
5931 struct dir_entry {
5932         u64 ino;
5933         u64 offset;
5934         unsigned type;
5935         int name_len;
5936 };
5937
5938 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5939 {
5940         while (entries--) {
5941                 struct dir_entry *entry = addr;
5942                 char *name = (char *)(entry + 1);
5943
5944                 ctx->pos = entry->offset;
5945                 if (!dir_emit(ctx, name, entry->name_len, entry->ino,
5946                               entry->type))
5947                         return 1;
5948                 addr += sizeof(struct dir_entry) + entry->name_len;
5949                 ctx->pos++;
5950         }
5951         return 0;
5952 }
5953
5954 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5955 {
5956         struct inode *inode = file_inode(file);
5957         struct btrfs_root *root = BTRFS_I(inode)->root;
5958         struct btrfs_file_private *private = file->private_data;
5959         struct btrfs_dir_item *di;
5960         struct btrfs_key key;
5961         struct btrfs_key found_key;
5962         struct btrfs_path *path;
5963         void *addr;
5964         struct list_head ins_list;
5965         struct list_head del_list;
5966         int ret;
5967         struct extent_buffer *leaf;
5968         int slot;
5969         char *name_ptr;
5970         int name_len;
5971         int entries = 0;
5972         int total_len = 0;
5973         bool put = false;
5974         struct btrfs_key location;
5975
5976         if (!dir_emit_dots(file, ctx))
5977                 return 0;
5978
5979         path = btrfs_alloc_path();
5980         if (!path)
5981                 return -ENOMEM;
5982
5983         addr = private->filldir_buf;
5984         path->reada = READA_FORWARD;
5985
5986         INIT_LIST_HEAD(&ins_list);
5987         INIT_LIST_HEAD(&del_list);
5988         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5989
5990 again:
5991         key.type = BTRFS_DIR_INDEX_KEY;
5992         key.offset = ctx->pos;
5993         key.objectid = btrfs_ino(BTRFS_I(inode));
5994
5995         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5996         if (ret < 0)
5997                 goto err;
5998
5999         while (1) {
6000                 struct dir_entry *entry;
6001
6002                 leaf = path->nodes[0];
6003                 slot = path->slots[0];
6004                 if (slot >= btrfs_header_nritems(leaf)) {
6005                         ret = btrfs_next_leaf(root, path);
6006                         if (ret < 0)
6007                                 goto err;
6008                         else if (ret > 0)
6009                                 break;
6010                         continue;
6011                 }
6012
6013                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6014
6015                 if (found_key.objectid != key.objectid)
6016                         break;
6017                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
6018                         break;
6019                 if (found_key.offset < ctx->pos)
6020                         goto next;
6021                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
6022                         goto next;
6023                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
6024                 name_len = btrfs_dir_name_len(leaf, di);
6025                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6026                     PAGE_SIZE) {
6027                         btrfs_release_path(path);
6028                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6029                         if (ret)
6030                                 goto nopos;
6031                         addr = private->filldir_buf;
6032                         entries = 0;
6033                         total_len = 0;
6034                         goto again;
6035                 }
6036
6037                 entry = addr;
6038                 entry->name_len = name_len;
6039                 name_ptr = (char *)(entry + 1);
6040                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6041                                    name_len);
6042                 entry->type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
6043                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6044                 entry->ino = location.objectid;
6045                 entry->offset = found_key.offset;
6046                 entries++;
6047                 addr += sizeof(struct dir_entry) + name_len;
6048                 total_len += sizeof(struct dir_entry) + name_len;
6049 next:
6050                 path->slots[0]++;
6051         }
6052         btrfs_release_path(path);
6053
6054         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6055         if (ret)
6056                 goto nopos;
6057
6058         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6059         if (ret)
6060                 goto nopos;
6061
6062         /*
6063          * Stop new entries from being returned after we return the last
6064          * entry.
6065          *
6066          * New directory entries are assigned a strictly increasing
6067          * offset.  This means that new entries created during readdir
6068          * are *guaranteed* to be seen in the future by that readdir.
6069          * This has broken buggy programs which operate on names as
6070          * they're returned by readdir.  Until we re-use freed offsets
6071          * we have this hack to stop new entries from being returned
6072          * under the assumption that they'll never reach this huge
6073          * offset.
6074          *
6075          * This is being careful not to overflow 32bit loff_t unless the
6076          * last entry requires it because doing so has broken 32bit apps
6077          * in the past.
6078          */
6079         if (ctx->pos >= INT_MAX)
6080                 ctx->pos = LLONG_MAX;
6081         else
6082                 ctx->pos = INT_MAX;
6083 nopos:
6084         ret = 0;
6085 err:
6086         if (put)
6087                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6088         btrfs_free_path(path);
6089         return ret;
6090 }
6091
6092 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
6093 {
6094         struct btrfs_root *root = BTRFS_I(inode)->root;
6095         struct btrfs_trans_handle *trans;
6096         int ret = 0;
6097         bool nolock = false;
6098
6099         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6100                 return 0;
6101
6102         if (btrfs_fs_closing(root->fs_info) &&
6103                         btrfs_is_free_space_inode(BTRFS_I(inode)))
6104                 nolock = true;
6105
6106         if (wbc->sync_mode == WB_SYNC_ALL) {
6107                 if (nolock)
6108                         trans = btrfs_join_transaction_nolock(root);
6109                 else
6110                         trans = btrfs_join_transaction(root);
6111                 if (IS_ERR(trans))
6112                         return PTR_ERR(trans);
6113                 ret = btrfs_commit_transaction(trans);
6114         }
6115         return ret;
6116 }
6117
6118 /*
6119  * This is somewhat expensive, updating the tree every time the
6120  * inode changes.  But, it is most likely to find the inode in cache.
6121  * FIXME, needs more benchmarking...there are no reasons other than performance
6122  * to keep or drop this code.
6123  */
6124 static int btrfs_dirty_inode(struct inode *inode)
6125 {
6126         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6127         struct btrfs_root *root = BTRFS_I(inode)->root;
6128         struct btrfs_trans_handle *trans;
6129         int ret;
6130
6131         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6132                 return 0;
6133
6134         trans = btrfs_join_transaction(root);
6135         if (IS_ERR(trans))
6136                 return PTR_ERR(trans);
6137
6138         ret = btrfs_update_inode(trans, root, inode);
6139         if (ret && ret == -ENOSPC) {
6140                 /* whoops, lets try again with the full transaction */
6141                 btrfs_end_transaction(trans);
6142                 trans = btrfs_start_transaction(root, 1);
6143                 if (IS_ERR(trans))
6144                         return PTR_ERR(trans);
6145
6146                 ret = btrfs_update_inode(trans, root, inode);
6147         }
6148         btrfs_end_transaction(trans);
6149         if (BTRFS_I(inode)->delayed_node)
6150                 btrfs_balance_delayed_items(fs_info);
6151
6152         return ret;
6153 }
6154
6155 /*
6156  * This is a copy of file_update_time.  We need this so we can return error on
6157  * ENOSPC for updating the inode in the case of file write and mmap writes.
6158  */
6159 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6160                              int flags)
6161 {
6162         struct btrfs_root *root = BTRFS_I(inode)->root;
6163         bool dirty = flags & ~S_VERSION;
6164
6165         if (btrfs_root_readonly(root))
6166                 return -EROFS;
6167
6168         if (flags & S_VERSION)
6169                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6170         if (flags & S_CTIME)
6171                 inode->i_ctime = *now;
6172         if (flags & S_MTIME)
6173                 inode->i_mtime = *now;
6174         if (flags & S_ATIME)
6175                 inode->i_atime = *now;
6176         return dirty ? btrfs_dirty_inode(inode) : 0;
6177 }
6178
6179 /*
6180  * find the highest existing sequence number in a directory
6181  * and then set the in-memory index_cnt variable to reflect
6182  * free sequence numbers
6183  */
6184 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6185 {
6186         struct btrfs_root *root = inode->root;
6187         struct btrfs_key key, found_key;
6188         struct btrfs_path *path;
6189         struct extent_buffer *leaf;
6190         int ret;
6191
6192         key.objectid = btrfs_ino(inode);
6193         key.type = BTRFS_DIR_INDEX_KEY;
6194         key.offset = (u64)-1;
6195
6196         path = btrfs_alloc_path();
6197         if (!path)
6198                 return -ENOMEM;
6199
6200         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6201         if (ret < 0)
6202                 goto out;
6203         /* FIXME: we should be able to handle this */
6204         if (ret == 0)
6205                 goto out;
6206         ret = 0;
6207
6208         /*
6209          * MAGIC NUMBER EXPLANATION:
6210          * since we search a directory based on f_pos we have to start at 2
6211          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6212          * else has to start at 2
6213          */
6214         if (path->slots[0] == 0) {
6215                 inode->index_cnt = 2;
6216                 goto out;
6217         }
6218
6219         path->slots[0]--;
6220
6221         leaf = path->nodes[0];
6222         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6223
6224         if (found_key.objectid != btrfs_ino(inode) ||
6225             found_key.type != BTRFS_DIR_INDEX_KEY) {
6226                 inode->index_cnt = 2;
6227                 goto out;
6228         }
6229
6230         inode->index_cnt = found_key.offset + 1;
6231 out:
6232         btrfs_free_path(path);
6233         return ret;
6234 }
6235
6236 /*
6237  * helper to find a free sequence number in a given directory.  This current
6238  * code is very simple, later versions will do smarter things in the btree
6239  */
6240 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6241 {
6242         int ret = 0;
6243
6244         if (dir->index_cnt == (u64)-1) {
6245                 ret = btrfs_inode_delayed_dir_index_count(dir);
6246                 if (ret) {
6247                         ret = btrfs_set_inode_index_count(dir);
6248                         if (ret)
6249                                 return ret;
6250                 }
6251         }
6252
6253         *index = dir->index_cnt;
6254         dir->index_cnt++;
6255
6256         return ret;
6257 }
6258
6259 static int btrfs_insert_inode_locked(struct inode *inode)
6260 {
6261         struct btrfs_iget_args args;
6262         args.location = &BTRFS_I(inode)->location;
6263         args.root = BTRFS_I(inode)->root;
6264
6265         return insert_inode_locked4(inode,
6266                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6267                    btrfs_find_actor, &args);
6268 }
6269
6270 /*
6271  * Inherit flags from the parent inode.
6272  *
6273  * Currently only the compression flags and the cow flags are inherited.
6274  */
6275 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6276 {
6277         unsigned int flags;
6278
6279         if (!dir)
6280                 return;
6281
6282         flags = BTRFS_I(dir)->flags;
6283
6284         if (flags & BTRFS_INODE_NOCOMPRESS) {
6285                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6286                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6287         } else if (flags & BTRFS_INODE_COMPRESS) {
6288                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6289                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6290         }
6291
6292         if (flags & BTRFS_INODE_NODATACOW) {
6293                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6294                 if (S_ISREG(inode->i_mode))
6295                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6296         }
6297
6298         btrfs_update_iflags(inode);
6299 }
6300
6301 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6302                                      struct btrfs_root *root,
6303                                      struct inode *dir,
6304                                      const char *name, int name_len,
6305                                      u64 ref_objectid, u64 objectid,
6306                                      umode_t mode, u64 *index)
6307 {
6308         struct btrfs_fs_info *fs_info = root->fs_info;
6309         struct inode *inode;
6310         struct btrfs_inode_item *inode_item;
6311         struct btrfs_key *location;
6312         struct btrfs_path *path;
6313         struct btrfs_inode_ref *ref;
6314         struct btrfs_key key[2];
6315         u32 sizes[2];
6316         int nitems = name ? 2 : 1;
6317         unsigned long ptr;
6318         int ret;
6319
6320         path = btrfs_alloc_path();
6321         if (!path)
6322                 return ERR_PTR(-ENOMEM);
6323
6324         inode = new_inode(fs_info->sb);
6325         if (!inode) {
6326                 btrfs_free_path(path);
6327                 return ERR_PTR(-ENOMEM);
6328         }
6329
6330         /*
6331          * O_TMPFILE, set link count to 0, so that after this point,
6332          * we fill in an inode item with the correct link count.
6333          */
6334         if (!name)
6335                 set_nlink(inode, 0);
6336
6337         /*
6338          * we have to initialize this early, so we can reclaim the inode
6339          * number if we fail afterwards in this function.
6340          */
6341         inode->i_ino = objectid;
6342
6343         if (dir && name) {
6344                 trace_btrfs_inode_request(dir);
6345
6346                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6347                 if (ret) {
6348                         btrfs_free_path(path);
6349                         iput(inode);
6350                         return ERR_PTR(ret);
6351                 }
6352         } else if (dir) {
6353                 *index = 0;
6354         }
6355         /*
6356          * index_cnt is ignored for everything but a dir,
6357          * btrfs_set_inode_index_count has an explanation for the magic
6358          * number
6359          */
6360         BTRFS_I(inode)->index_cnt = 2;
6361         BTRFS_I(inode)->dir_index = *index;
6362         BTRFS_I(inode)->root = root;
6363         BTRFS_I(inode)->generation = trans->transid;
6364         inode->i_generation = BTRFS_I(inode)->generation;
6365
6366         /*
6367          * We could have gotten an inode number from somebody who was fsynced
6368          * and then removed in this same transaction, so let's just set full
6369          * sync since it will be a full sync anyway and this will blow away the
6370          * old info in the log.
6371          */
6372         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6373
6374         key[0].objectid = objectid;
6375         key[0].type = BTRFS_INODE_ITEM_KEY;
6376         key[0].offset = 0;
6377
6378         sizes[0] = sizeof(struct btrfs_inode_item);
6379
6380         if (name) {
6381                 /*
6382                  * Start new inodes with an inode_ref. This is slightly more
6383                  * efficient for small numbers of hard links since they will
6384                  * be packed into one item. Extended refs will kick in if we
6385                  * add more hard links than can fit in the ref item.
6386                  */
6387                 key[1].objectid = objectid;
6388                 key[1].type = BTRFS_INODE_REF_KEY;
6389                 key[1].offset = ref_objectid;
6390
6391                 sizes[1] = name_len + sizeof(*ref);
6392         }
6393
6394         location = &BTRFS_I(inode)->location;
6395         location->objectid = objectid;
6396         location->offset = 0;
6397         location->type = BTRFS_INODE_ITEM_KEY;
6398
6399         ret = btrfs_insert_inode_locked(inode);
6400         if (ret < 0)
6401                 goto fail;
6402
6403         path->leave_spinning = 1;
6404         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6405         if (ret != 0)
6406                 goto fail_unlock;
6407
6408         inode_init_owner(inode, dir, mode);
6409         inode_set_bytes(inode, 0);
6410
6411         inode->i_mtime = current_time(inode);
6412         inode->i_atime = inode->i_mtime;
6413         inode->i_ctime = inode->i_mtime;
6414         BTRFS_I(inode)->i_otime = inode->i_mtime;
6415
6416         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6417                                   struct btrfs_inode_item);
6418         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6419                              sizeof(*inode_item));
6420         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6421
6422         if (name) {
6423                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6424                                      struct btrfs_inode_ref);
6425                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6426                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6427                 ptr = (unsigned long)(ref + 1);
6428                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6429         }
6430
6431         btrfs_mark_buffer_dirty(path->nodes[0]);
6432         btrfs_free_path(path);
6433
6434         btrfs_inherit_iflags(inode, dir);
6435
6436         if (S_ISREG(mode)) {
6437                 if (btrfs_test_opt(fs_info, NODATASUM))
6438                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6439                 if (btrfs_test_opt(fs_info, NODATACOW))
6440                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6441                                 BTRFS_INODE_NODATASUM;
6442         }
6443
6444         inode_tree_add(inode);
6445
6446         trace_btrfs_inode_new(inode);
6447         btrfs_set_inode_last_trans(trans, inode);
6448
6449         btrfs_update_root_times(trans, root);
6450
6451         ret = btrfs_inode_inherit_props(trans, inode, dir);
6452         if (ret)
6453                 btrfs_err(fs_info,
6454                           "error inheriting props for ino %llu (root %llu): %d",
6455                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6456
6457         return inode;
6458
6459 fail_unlock:
6460         unlock_new_inode(inode);
6461 fail:
6462         if (dir && name)
6463                 BTRFS_I(dir)->index_cnt--;
6464         btrfs_free_path(path);
6465         iput(inode);
6466         return ERR_PTR(ret);
6467 }
6468
6469 static inline u8 btrfs_inode_type(struct inode *inode)
6470 {
6471         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6472 }
6473
6474 /*
6475  * utility function to add 'inode' into 'parent_inode' with
6476  * a give name and a given sequence number.
6477  * if 'add_backref' is true, also insert a backref from the
6478  * inode to the parent directory.
6479  */
6480 int btrfs_add_link(struct btrfs_trans_handle *trans,
6481                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6482                    const char *name, int name_len, int add_backref, u64 index)
6483 {
6484         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6485         int ret = 0;
6486         struct btrfs_key key;
6487         struct btrfs_root *root = parent_inode->root;
6488         u64 ino = btrfs_ino(inode);
6489         u64 parent_ino = btrfs_ino(parent_inode);
6490
6491         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6492                 memcpy(&key, &inode->root->root_key, sizeof(key));
6493         } else {
6494                 key.objectid = ino;
6495                 key.type = BTRFS_INODE_ITEM_KEY;
6496                 key.offset = 0;
6497         }
6498
6499         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6500                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6501                                          root->root_key.objectid, parent_ino,
6502                                          index, name, name_len);
6503         } else if (add_backref) {
6504                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6505                                              parent_ino, index);
6506         }
6507
6508         /* Nothing to clean up yet */
6509         if (ret)
6510                 return ret;
6511
6512         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6513                                     parent_inode, &key,
6514                                     btrfs_inode_type(&inode->vfs_inode), index);
6515         if (ret == -EEXIST || ret == -EOVERFLOW)
6516                 goto fail_dir_item;
6517         else if (ret) {
6518                 btrfs_abort_transaction(trans, ret);
6519                 return ret;
6520         }
6521
6522         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6523                            name_len * 2);
6524         inode_inc_iversion(&parent_inode->vfs_inode);
6525         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6526                 current_time(&parent_inode->vfs_inode);
6527         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6528         if (ret)
6529                 btrfs_abort_transaction(trans, ret);
6530         return ret;
6531
6532 fail_dir_item:
6533         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6534                 u64 local_index;
6535                 int err;
6536                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6537                                          root->root_key.objectid, parent_ino,
6538                                          &local_index, name, name_len);
6539
6540         } else if (add_backref) {
6541                 u64 local_index;
6542                 int err;
6543
6544                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6545                                           ino, parent_ino, &local_index);
6546         }
6547         return ret;
6548 }
6549
6550 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6551                             struct btrfs_inode *dir, struct dentry *dentry,
6552                             struct btrfs_inode *inode, int backref, u64 index)
6553 {
6554         int err = btrfs_add_link(trans, dir, inode,
6555                                  dentry->d_name.name, dentry->d_name.len,
6556                                  backref, index);
6557         if (err > 0)
6558                 err = -EEXIST;
6559         return err;
6560 }
6561
6562 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6563                         umode_t mode, dev_t rdev)
6564 {
6565         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6566         struct btrfs_trans_handle *trans;
6567         struct btrfs_root *root = BTRFS_I(dir)->root;
6568         struct inode *inode = NULL;
6569         int err;
6570         int drop_inode = 0;
6571         u64 objectid;
6572         u64 index = 0;
6573
6574         /*
6575          * 2 for inode item and ref
6576          * 2 for dir items
6577          * 1 for xattr if selinux is on
6578          */
6579         trans = btrfs_start_transaction(root, 5);
6580         if (IS_ERR(trans))
6581                 return PTR_ERR(trans);
6582
6583         err = btrfs_find_free_ino(root, &objectid);
6584         if (err)
6585                 goto out_unlock;
6586
6587         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6588                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6589                         mode, &index);
6590         if (IS_ERR(inode)) {
6591                 err = PTR_ERR(inode);
6592                 goto out_unlock;
6593         }
6594
6595         /*
6596         * If the active LSM wants to access the inode during
6597         * d_instantiate it needs these. Smack checks to see
6598         * if the filesystem supports xattrs by looking at the
6599         * ops vector.
6600         */
6601         inode->i_op = &btrfs_special_inode_operations;
6602         init_special_inode(inode, inode->i_mode, rdev);
6603
6604         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6605         if (err)
6606                 goto out_unlock_inode;
6607
6608         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6609                         0, index);
6610         if (err) {
6611                 goto out_unlock_inode;
6612         } else {
6613                 btrfs_update_inode(trans, root, inode);
6614                 unlock_new_inode(inode);
6615                 d_instantiate(dentry, inode);
6616         }
6617
6618 out_unlock:
6619         btrfs_end_transaction(trans);
6620         btrfs_btree_balance_dirty(fs_info);
6621         if (drop_inode) {
6622                 inode_dec_link_count(inode);
6623                 iput(inode);
6624         }
6625         return err;
6626
6627 out_unlock_inode:
6628         drop_inode = 1;
6629         unlock_new_inode(inode);
6630         goto out_unlock;
6631
6632 }
6633
6634 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6635                         umode_t mode, bool excl)
6636 {
6637         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6638         struct btrfs_trans_handle *trans;
6639         struct btrfs_root *root = BTRFS_I(dir)->root;
6640         struct inode *inode = NULL;
6641         int drop_inode_on_err = 0;
6642         int err;
6643         u64 objectid;
6644         u64 index = 0;
6645
6646         /*
6647          * 2 for inode item and ref
6648          * 2 for dir items
6649          * 1 for xattr if selinux is on
6650          */
6651         trans = btrfs_start_transaction(root, 5);
6652         if (IS_ERR(trans))
6653                 return PTR_ERR(trans);
6654
6655         err = btrfs_find_free_ino(root, &objectid);
6656         if (err)
6657                 goto out_unlock;
6658
6659         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6660                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6661                         mode, &index);
6662         if (IS_ERR(inode)) {
6663                 err = PTR_ERR(inode);
6664                 goto out_unlock;
6665         }
6666         drop_inode_on_err = 1;
6667         /*
6668         * If the active LSM wants to access the inode during
6669         * d_instantiate it needs these. Smack checks to see
6670         * if the filesystem supports xattrs by looking at the
6671         * ops vector.
6672         */
6673         inode->i_fop = &btrfs_file_operations;
6674         inode->i_op = &btrfs_file_inode_operations;
6675         inode->i_mapping->a_ops = &btrfs_aops;
6676
6677         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6678         if (err)
6679                 goto out_unlock_inode;
6680
6681         err = btrfs_update_inode(trans, root, inode);
6682         if (err)
6683                 goto out_unlock_inode;
6684
6685         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6686                         0, index);
6687         if (err)
6688                 goto out_unlock_inode;
6689
6690         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6691         unlock_new_inode(inode);
6692         d_instantiate(dentry, inode);
6693
6694 out_unlock:
6695         btrfs_end_transaction(trans);
6696         if (err && drop_inode_on_err) {
6697                 inode_dec_link_count(inode);
6698                 iput(inode);
6699         }
6700         btrfs_btree_balance_dirty(fs_info);
6701         return err;
6702
6703 out_unlock_inode:
6704         unlock_new_inode(inode);
6705         goto out_unlock;
6706
6707 }
6708
6709 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6710                       struct dentry *dentry)
6711 {
6712         struct btrfs_trans_handle *trans = NULL;
6713         struct btrfs_root *root = BTRFS_I(dir)->root;
6714         struct inode *inode = d_inode(old_dentry);
6715         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6716         u64 index;
6717         int err;
6718         int drop_inode = 0;
6719
6720         /* do not allow sys_link's with other subvols of the same device */
6721         if (root->objectid != BTRFS_I(inode)->root->objectid)
6722                 return -EXDEV;
6723
6724         if (inode->i_nlink >= BTRFS_LINK_MAX)
6725                 return -EMLINK;
6726
6727         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6728         if (err)
6729                 goto fail;
6730
6731         /*
6732          * 2 items for inode and inode ref
6733          * 2 items for dir items
6734          * 1 item for parent inode
6735          */
6736         trans = btrfs_start_transaction(root, 5);
6737         if (IS_ERR(trans)) {
6738                 err = PTR_ERR(trans);
6739                 trans = NULL;
6740                 goto fail;
6741         }
6742
6743         /* There are several dir indexes for this inode, clear the cache. */
6744         BTRFS_I(inode)->dir_index = 0ULL;
6745         inc_nlink(inode);
6746         inode_inc_iversion(inode);
6747         inode->i_ctime = current_time(inode);
6748         ihold(inode);
6749         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6750
6751         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6752                         1, index);
6753
6754         if (err) {
6755                 drop_inode = 1;
6756         } else {
6757                 struct dentry *parent = dentry->d_parent;
6758                 err = btrfs_update_inode(trans, root, inode);
6759                 if (err)
6760                         goto fail;
6761                 if (inode->i_nlink == 1) {
6762                         /*
6763                          * If new hard link count is 1, it's a file created
6764                          * with open(2) O_TMPFILE flag.
6765                          */
6766                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6767                         if (err)
6768                                 goto fail;
6769                 }
6770                 d_instantiate(dentry, inode);
6771                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6772         }
6773
6774 fail:
6775         if (trans)
6776                 btrfs_end_transaction(trans);
6777         if (drop_inode) {
6778                 inode_dec_link_count(inode);
6779                 iput(inode);
6780         }
6781         btrfs_btree_balance_dirty(fs_info);
6782         return err;
6783 }
6784
6785 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6786 {
6787         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6788         struct inode *inode = NULL;
6789         struct btrfs_trans_handle *trans;
6790         struct btrfs_root *root = BTRFS_I(dir)->root;
6791         int err = 0;
6792         int drop_on_err = 0;
6793         u64 objectid = 0;
6794         u64 index = 0;
6795
6796         /*
6797          * 2 items for inode and ref
6798          * 2 items for dir items
6799          * 1 for xattr if selinux is on
6800          */
6801         trans = btrfs_start_transaction(root, 5);
6802         if (IS_ERR(trans))
6803                 return PTR_ERR(trans);
6804
6805         err = btrfs_find_free_ino(root, &objectid);
6806         if (err)
6807                 goto out_fail;
6808
6809         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6810                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6811                         S_IFDIR | mode, &index);
6812         if (IS_ERR(inode)) {
6813                 err = PTR_ERR(inode);
6814                 goto out_fail;
6815         }
6816
6817         drop_on_err = 1;
6818         /* these must be set before we unlock the inode */
6819         inode->i_op = &btrfs_dir_inode_operations;
6820         inode->i_fop = &btrfs_dir_file_operations;
6821
6822         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6823         if (err)
6824                 goto out_fail_inode;
6825
6826         btrfs_i_size_write(BTRFS_I(inode), 0);
6827         err = btrfs_update_inode(trans, root, inode);
6828         if (err)
6829                 goto out_fail_inode;
6830
6831         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6832                         dentry->d_name.name,
6833                         dentry->d_name.len, 0, index);
6834         if (err)
6835                 goto out_fail_inode;
6836
6837         d_instantiate(dentry, inode);
6838         /*
6839          * mkdir is special.  We're unlocking after we call d_instantiate
6840          * to avoid a race with nfsd calling d_instantiate.
6841          */
6842         unlock_new_inode(inode);
6843         drop_on_err = 0;
6844
6845 out_fail:
6846         btrfs_end_transaction(trans);
6847         if (drop_on_err) {
6848                 inode_dec_link_count(inode);
6849                 iput(inode);
6850         }
6851         btrfs_btree_balance_dirty(fs_info);
6852         return err;
6853
6854 out_fail_inode:
6855         unlock_new_inode(inode);
6856         goto out_fail;
6857 }
6858
6859 static noinline int uncompress_inline(struct btrfs_path *path,
6860                                       struct page *page,
6861                                       size_t pg_offset, u64 extent_offset,
6862                                       struct btrfs_file_extent_item *item)
6863 {
6864         int ret;
6865         struct extent_buffer *leaf = path->nodes[0];
6866         char *tmp;
6867         size_t max_size;
6868         unsigned long inline_size;
6869         unsigned long ptr;
6870         int compress_type;
6871
6872         WARN_ON(pg_offset != 0);
6873         compress_type = btrfs_file_extent_compression(leaf, item);
6874         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6875         inline_size = btrfs_file_extent_inline_item_len(leaf,
6876                                         btrfs_item_nr(path->slots[0]));
6877         tmp = kmalloc(inline_size, GFP_NOFS);
6878         if (!tmp)
6879                 return -ENOMEM;
6880         ptr = btrfs_file_extent_inline_start(item);
6881
6882         read_extent_buffer(leaf, tmp, ptr, inline_size);
6883
6884         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6885         ret = btrfs_decompress(compress_type, tmp, page,
6886                                extent_offset, inline_size, max_size);
6887
6888         /*
6889          * decompression code contains a memset to fill in any space between the end
6890          * of the uncompressed data and the end of max_size in case the decompressed
6891          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6892          * the end of an inline extent and the beginning of the next block, so we
6893          * cover that region here.
6894          */
6895
6896         if (max_size + pg_offset < PAGE_SIZE) {
6897                 char *map = kmap(page);
6898                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6899                 kunmap(page);
6900         }
6901         kfree(tmp);
6902         return ret;
6903 }
6904
6905 /*
6906  * a bit scary, this does extent mapping from logical file offset to the disk.
6907  * the ugly parts come from merging extents from the disk with the in-ram
6908  * representation.  This gets more complex because of the data=ordered code,
6909  * where the in-ram extents might be locked pending data=ordered completion.
6910  *
6911  * This also copies inline extents directly into the page.
6912  */
6913 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6914                 struct page *page,
6915             size_t pg_offset, u64 start, u64 len,
6916                 int create)
6917 {
6918         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6919         int ret;
6920         int err = 0;
6921         u64 extent_start = 0;
6922         u64 extent_end = 0;
6923         u64 objectid = btrfs_ino(inode);
6924         u32 found_type;
6925         struct btrfs_path *path = NULL;
6926         struct btrfs_root *root = inode->root;
6927         struct btrfs_file_extent_item *item;
6928         struct extent_buffer *leaf;
6929         struct btrfs_key found_key;
6930         struct extent_map *em = NULL;
6931         struct extent_map_tree *em_tree = &inode->extent_tree;
6932         struct extent_io_tree *io_tree = &inode->io_tree;
6933         const bool new_inline = !page || create;
6934
6935         read_lock(&em_tree->lock);
6936         em = lookup_extent_mapping(em_tree, start, len);
6937         if (em)
6938                 em->bdev = fs_info->fs_devices->latest_bdev;
6939         read_unlock(&em_tree->lock);
6940
6941         if (em) {
6942                 if (em->start > start || em->start + em->len <= start)
6943                         free_extent_map(em);
6944                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6945                         free_extent_map(em);
6946                 else
6947                         goto out;
6948         }
6949         em = alloc_extent_map();
6950         if (!em) {
6951                 err = -ENOMEM;
6952                 goto out;
6953         }
6954         em->bdev = fs_info->fs_devices->latest_bdev;
6955         em->start = EXTENT_MAP_HOLE;
6956         em->orig_start = EXTENT_MAP_HOLE;
6957         em->len = (u64)-1;
6958         em->block_len = (u64)-1;
6959
6960         if (!path) {
6961                 path = btrfs_alloc_path();
6962                 if (!path) {
6963                         err = -ENOMEM;
6964                         goto out;
6965                 }
6966                 /*
6967                  * Chances are we'll be called again, so go ahead and do
6968                  * readahead
6969                  */
6970                 path->reada = READA_FORWARD;
6971         }
6972
6973         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
6974         if (ret < 0) {
6975                 err = ret;
6976                 goto out;
6977         }
6978
6979         if (ret != 0) {
6980                 if (path->slots[0] == 0)
6981                         goto not_found;
6982                 path->slots[0]--;
6983         }
6984
6985         leaf = path->nodes[0];
6986         item = btrfs_item_ptr(leaf, path->slots[0],
6987                               struct btrfs_file_extent_item);
6988         /* are we inside the extent that was found? */
6989         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6990         found_type = found_key.type;
6991         if (found_key.objectid != objectid ||
6992             found_type != BTRFS_EXTENT_DATA_KEY) {
6993                 /*
6994                  * If we backup past the first extent we want to move forward
6995                  * and see if there is an extent in front of us, otherwise we'll
6996                  * say there is a hole for our whole search range which can
6997                  * cause problems.
6998                  */
6999                 extent_end = start;
7000                 goto next;
7001         }
7002
7003         found_type = btrfs_file_extent_type(leaf, item);
7004         extent_start = found_key.offset;
7005         if (found_type == BTRFS_FILE_EXTENT_REG ||
7006             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7007                 extent_end = extent_start +
7008                        btrfs_file_extent_num_bytes(leaf, item);
7009
7010                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7011                                                        extent_start);
7012         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7013                 size_t size;
7014                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7015                 extent_end = ALIGN(extent_start + size,
7016                                    fs_info->sectorsize);
7017
7018                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7019                                                       path->slots[0],
7020                                                       extent_start);
7021         }
7022 next:
7023         if (start >= extent_end) {
7024                 path->slots[0]++;
7025                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7026                         ret = btrfs_next_leaf(root, path);
7027                         if (ret < 0) {
7028                                 err = ret;
7029                                 goto out;
7030                         }
7031                         if (ret > 0)
7032                                 goto not_found;
7033                         leaf = path->nodes[0];
7034                 }
7035                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7036                 if (found_key.objectid != objectid ||
7037                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7038                         goto not_found;
7039                 if (start + len <= found_key.offset)
7040                         goto not_found;
7041                 if (start > found_key.offset)
7042                         goto next;
7043                 em->start = start;
7044                 em->orig_start = start;
7045                 em->len = found_key.offset - start;
7046                 goto not_found_em;
7047         }
7048
7049         btrfs_extent_item_to_extent_map(inode, path, item,
7050                         new_inline, em);
7051
7052         if (found_type == BTRFS_FILE_EXTENT_REG ||
7053             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7054                 goto insert;
7055         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7056                 unsigned long ptr;
7057                 char *map;
7058                 size_t size;
7059                 size_t extent_offset;
7060                 size_t copy_size;
7061
7062                 if (new_inline)
7063                         goto out;
7064
7065                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7066                 extent_offset = page_offset(page) + pg_offset - extent_start;
7067                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7068                                   size - extent_offset);
7069                 em->start = extent_start + extent_offset;
7070                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7071                 em->orig_block_len = em->len;
7072                 em->orig_start = em->start;
7073                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7074                 if (!PageUptodate(page)) {
7075                         if (btrfs_file_extent_compression(leaf, item) !=
7076                             BTRFS_COMPRESS_NONE) {
7077                                 ret = uncompress_inline(path, page, pg_offset,
7078                                                         extent_offset, item);
7079                                 if (ret) {
7080                                         err = ret;
7081                                         goto out;
7082                                 }
7083                         } else {
7084                                 map = kmap(page);
7085                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7086                                                    copy_size);
7087                                 if (pg_offset + copy_size < PAGE_SIZE) {
7088                                         memset(map + pg_offset + copy_size, 0,
7089                                                PAGE_SIZE - pg_offset -
7090                                                copy_size);
7091                                 }
7092                                 kunmap(page);
7093                         }
7094                         flush_dcache_page(page);
7095                 }
7096                 set_extent_uptodate(io_tree, em->start,
7097                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7098                 goto insert;
7099         }
7100 not_found:
7101         em->start = start;
7102         em->orig_start = start;
7103         em->len = len;
7104 not_found_em:
7105         em->block_start = EXTENT_MAP_HOLE;
7106 insert:
7107         btrfs_release_path(path);
7108         if (em->start > start || extent_map_end(em) <= start) {
7109                 btrfs_err(fs_info,
7110                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7111                           em->start, em->len, start, len);
7112                 err = -EIO;
7113                 goto out;
7114         }
7115
7116         err = 0;
7117         write_lock(&em_tree->lock);
7118         err = btrfs_add_extent_mapping(em_tree, &em, start, len);
7119         write_unlock(&em_tree->lock);
7120 out:
7121
7122         trace_btrfs_get_extent(root, inode, em);
7123
7124         btrfs_free_path(path);
7125         if (err) {
7126                 free_extent_map(em);
7127                 return ERR_PTR(err);
7128         }
7129         BUG_ON(!em); /* Error is always set */
7130         return em;
7131 }
7132
7133 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7134                 struct page *page,
7135                 size_t pg_offset, u64 start, u64 len,
7136                 int create)
7137 {
7138         struct extent_map *em;
7139         struct extent_map *hole_em = NULL;
7140         u64 range_start = start;
7141         u64 end;
7142         u64 found;
7143         u64 found_end;
7144         int err = 0;
7145
7146         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7147         if (IS_ERR(em))
7148                 return em;
7149         /*
7150          * If our em maps to:
7151          * - a hole or
7152          * - a pre-alloc extent,
7153          * there might actually be delalloc bytes behind it.
7154          */
7155         if (em->block_start != EXTENT_MAP_HOLE &&
7156             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7157                 return em;
7158         else
7159                 hole_em = em;
7160
7161         /* check to see if we've wrapped (len == -1 or similar) */
7162         end = start + len;
7163         if (end < start)
7164                 end = (u64)-1;
7165         else
7166                 end -= 1;
7167
7168         em = NULL;
7169
7170         /* ok, we didn't find anything, lets look for delalloc */
7171         found = count_range_bits(&inode->io_tree, &range_start,
7172                                  end, len, EXTENT_DELALLOC, 1);
7173         found_end = range_start + found;
7174         if (found_end < range_start)
7175                 found_end = (u64)-1;
7176
7177         /*
7178          * we didn't find anything useful, return
7179          * the original results from get_extent()
7180          */
7181         if (range_start > end || found_end <= start) {
7182                 em = hole_em;
7183                 hole_em = NULL;
7184                 goto out;
7185         }
7186
7187         /* adjust the range_start to make sure it doesn't
7188          * go backwards from the start they passed in
7189          */
7190         range_start = max(start, range_start);
7191         found = found_end - range_start;
7192
7193         if (found > 0) {
7194                 u64 hole_start = start;
7195                 u64 hole_len = len;
7196
7197                 em = alloc_extent_map();
7198                 if (!em) {
7199                         err = -ENOMEM;
7200                         goto out;
7201                 }
7202                 /*
7203                  * when btrfs_get_extent can't find anything it
7204                  * returns one huge hole
7205                  *
7206                  * make sure what it found really fits our range, and
7207                  * adjust to make sure it is based on the start from
7208                  * the caller
7209                  */
7210                 if (hole_em) {
7211                         u64 calc_end = extent_map_end(hole_em);
7212
7213                         if (calc_end <= start || (hole_em->start > end)) {
7214                                 free_extent_map(hole_em);
7215                                 hole_em = NULL;
7216                         } else {
7217                                 hole_start = max(hole_em->start, start);
7218                                 hole_len = calc_end - hole_start;
7219                         }
7220                 }
7221                 em->bdev = NULL;
7222                 if (hole_em && range_start > hole_start) {
7223                         /* our hole starts before our delalloc, so we
7224                          * have to return just the parts of the hole
7225                          * that go until  the delalloc starts
7226                          */
7227                         em->len = min(hole_len,
7228                                       range_start - hole_start);
7229                         em->start = hole_start;
7230                         em->orig_start = hole_start;
7231                         /*
7232                          * don't adjust block start at all,
7233                          * it is fixed at EXTENT_MAP_HOLE
7234                          */
7235                         em->block_start = hole_em->block_start;
7236                         em->block_len = hole_len;
7237                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7238                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7239                 } else {
7240                         em->start = range_start;
7241                         em->len = found;
7242                         em->orig_start = range_start;
7243                         em->block_start = EXTENT_MAP_DELALLOC;
7244                         em->block_len = found;
7245                 }
7246         } else {
7247                 return hole_em;
7248         }
7249 out:
7250
7251         free_extent_map(hole_em);
7252         if (err) {
7253                 free_extent_map(em);
7254                 return ERR_PTR(err);
7255         }
7256         return em;
7257 }
7258
7259 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7260                                                   const u64 start,
7261                                                   const u64 len,
7262                                                   const u64 orig_start,
7263                                                   const u64 block_start,
7264                                                   const u64 block_len,
7265                                                   const u64 orig_block_len,
7266                                                   const u64 ram_bytes,
7267                                                   const int type)
7268 {
7269         struct extent_map *em = NULL;
7270         int ret;
7271
7272         if (type != BTRFS_ORDERED_NOCOW) {
7273                 em = create_io_em(inode, start, len, orig_start,
7274                                   block_start, block_len, orig_block_len,
7275                                   ram_bytes,
7276                                   BTRFS_COMPRESS_NONE, /* compress_type */
7277                                   type);
7278                 if (IS_ERR(em))
7279                         goto out;
7280         }
7281         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7282                                            len, block_len, type);
7283         if (ret) {
7284                 if (em) {
7285                         free_extent_map(em);
7286                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7287                                                 start + len - 1, 0);
7288                 }
7289                 em = ERR_PTR(ret);
7290         }
7291  out:
7292
7293         return em;
7294 }
7295
7296 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7297                                                   u64 start, u64 len)
7298 {
7299         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7300         struct btrfs_root *root = BTRFS_I(inode)->root;
7301         struct extent_map *em;
7302         struct btrfs_key ins;
7303         u64 alloc_hint;
7304         int ret;
7305
7306         alloc_hint = get_extent_allocation_hint(inode, start, len);
7307         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7308                                    0, alloc_hint, &ins, 1, 1);
7309         if (ret)
7310                 return ERR_PTR(ret);
7311
7312         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7313                                      ins.objectid, ins.offset, ins.offset,
7314                                      ins.offset, BTRFS_ORDERED_REGULAR);
7315         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7316         if (IS_ERR(em))
7317                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7318                                            ins.offset, 1);
7319
7320         return em;
7321 }
7322
7323 /*
7324  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7325  * block must be cow'd
7326  */
7327 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7328                               u64 *orig_start, u64 *orig_block_len,
7329                               u64 *ram_bytes)
7330 {
7331         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7332         struct btrfs_path *path;
7333         int ret;
7334         struct extent_buffer *leaf;
7335         struct btrfs_root *root = BTRFS_I(inode)->root;
7336         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7337         struct btrfs_file_extent_item *fi;
7338         struct btrfs_key key;
7339         u64 disk_bytenr;
7340         u64 backref_offset;
7341         u64 extent_end;
7342         u64 num_bytes;
7343         int slot;
7344         int found_type;
7345         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7346
7347         path = btrfs_alloc_path();
7348         if (!path)
7349                 return -ENOMEM;
7350
7351         ret = btrfs_lookup_file_extent(NULL, root, path,
7352                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7353         if (ret < 0)
7354                 goto out;
7355
7356         slot = path->slots[0];
7357         if (ret == 1) {
7358                 if (slot == 0) {
7359                         /* can't find the item, must cow */
7360                         ret = 0;
7361                         goto out;
7362                 }
7363                 slot--;
7364         }
7365         ret = 0;
7366         leaf = path->nodes[0];
7367         btrfs_item_key_to_cpu(leaf, &key, slot);
7368         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7369             key.type != BTRFS_EXTENT_DATA_KEY) {
7370                 /* not our file or wrong item type, must cow */
7371                 goto out;
7372         }
7373
7374         if (key.offset > offset) {
7375                 /* Wrong offset, must cow */
7376                 goto out;
7377         }
7378
7379         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7380         found_type = btrfs_file_extent_type(leaf, fi);
7381         if (found_type != BTRFS_FILE_EXTENT_REG &&
7382             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7383                 /* not a regular extent, must cow */
7384                 goto out;
7385         }
7386
7387         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7388                 goto out;
7389
7390         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7391         if (extent_end <= offset)
7392                 goto out;
7393
7394         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7395         if (disk_bytenr == 0)
7396                 goto out;
7397
7398         if (btrfs_file_extent_compression(leaf, fi) ||
7399             btrfs_file_extent_encryption(leaf, fi) ||
7400             btrfs_file_extent_other_encoding(leaf, fi))
7401                 goto out;
7402
7403         backref_offset = btrfs_file_extent_offset(leaf, fi);
7404
7405         if (orig_start) {
7406                 *orig_start = key.offset - backref_offset;
7407                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7408                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7409         }
7410
7411         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7412                 goto out;
7413
7414         num_bytes = min(offset + *len, extent_end) - offset;
7415         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7416                 u64 range_end;
7417
7418                 range_end = round_up(offset + num_bytes,
7419                                      root->fs_info->sectorsize) - 1;
7420                 ret = test_range_bit(io_tree, offset, range_end,
7421                                      EXTENT_DELALLOC, 0, NULL);
7422                 if (ret) {
7423                         ret = -EAGAIN;
7424                         goto out;
7425                 }
7426         }
7427
7428         btrfs_release_path(path);
7429
7430         /*
7431          * look for other files referencing this extent, if we
7432          * find any we must cow
7433          */
7434
7435         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7436                                     key.offset - backref_offset, disk_bytenr);
7437         if (ret) {
7438                 ret = 0;
7439                 goto out;
7440         }
7441
7442         /*
7443          * adjust disk_bytenr and num_bytes to cover just the bytes
7444          * in this extent we are about to write.  If there
7445          * are any csums in that range we have to cow in order
7446          * to keep the csums correct
7447          */
7448         disk_bytenr += backref_offset;
7449         disk_bytenr += offset - key.offset;
7450         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7451                 goto out;
7452         /*
7453          * all of the above have passed, it is safe to overwrite this extent
7454          * without cow
7455          */
7456         *len = num_bytes;
7457         ret = 1;
7458 out:
7459         btrfs_free_path(path);
7460         return ret;
7461 }
7462
7463 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7464                               struct extent_state **cached_state, int writing)
7465 {
7466         struct btrfs_ordered_extent *ordered;
7467         int ret = 0;
7468
7469         while (1) {
7470                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7471                                  cached_state);
7472                 /*
7473                  * We're concerned with the entire range that we're going to be
7474                  * doing DIO to, so we need to make sure there's no ordered
7475                  * extents in this range.
7476                  */
7477                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7478                                                      lockend - lockstart + 1);
7479
7480                 /*
7481                  * We need to make sure there are no buffered pages in this
7482                  * range either, we could have raced between the invalidate in
7483                  * generic_file_direct_write and locking the extent.  The
7484                  * invalidate needs to happen so that reads after a write do not
7485                  * get stale data.
7486                  */
7487                 if (!ordered &&
7488                     (!writing || !filemap_range_has_page(inode->i_mapping,
7489                                                          lockstart, lockend)))
7490                         break;
7491
7492                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7493                                      cached_state);
7494
7495                 if (ordered) {
7496                         /*
7497                          * If we are doing a DIO read and the ordered extent we
7498                          * found is for a buffered write, we can not wait for it
7499                          * to complete and retry, because if we do so we can
7500                          * deadlock with concurrent buffered writes on page
7501                          * locks. This happens only if our DIO read covers more
7502                          * than one extent map, if at this point has already
7503                          * created an ordered extent for a previous extent map
7504                          * and locked its range in the inode's io tree, and a
7505                          * concurrent write against that previous extent map's
7506                          * range and this range started (we unlock the ranges
7507                          * in the io tree only when the bios complete and
7508                          * buffered writes always lock pages before attempting
7509                          * to lock range in the io tree).
7510                          */
7511                         if (writing ||
7512                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7513                                 btrfs_start_ordered_extent(inode, ordered, 1);
7514                         else
7515                                 ret = -ENOTBLK;
7516                         btrfs_put_ordered_extent(ordered);
7517                 } else {
7518                         /*
7519                          * We could trigger writeback for this range (and wait
7520                          * for it to complete) and then invalidate the pages for
7521                          * this range (through invalidate_inode_pages2_range()),
7522                          * but that can lead us to a deadlock with a concurrent
7523                          * call to readpages() (a buffered read or a defrag call
7524                          * triggered a readahead) on a page lock due to an
7525                          * ordered dio extent we created before but did not have
7526                          * yet a corresponding bio submitted (whence it can not
7527                          * complete), which makes readpages() wait for that
7528                          * ordered extent to complete while holding a lock on
7529                          * that page.
7530                          */
7531                         ret = -ENOTBLK;
7532                 }
7533
7534                 if (ret)
7535                         break;
7536
7537                 cond_resched();
7538         }
7539
7540         return ret;
7541 }
7542
7543 /* The callers of this must take lock_extent() */
7544 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7545                                        u64 orig_start, u64 block_start,
7546                                        u64 block_len, u64 orig_block_len,
7547                                        u64 ram_bytes, int compress_type,
7548                                        int type)
7549 {
7550         struct extent_map_tree *em_tree;
7551         struct extent_map *em;
7552         struct btrfs_root *root = BTRFS_I(inode)->root;
7553         int ret;
7554
7555         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7556                type == BTRFS_ORDERED_COMPRESSED ||
7557                type == BTRFS_ORDERED_NOCOW ||
7558                type == BTRFS_ORDERED_REGULAR);
7559
7560         em_tree = &BTRFS_I(inode)->extent_tree;
7561         em = alloc_extent_map();
7562         if (!em)
7563                 return ERR_PTR(-ENOMEM);
7564
7565         em->start = start;
7566         em->orig_start = orig_start;
7567         em->len = len;
7568         em->block_len = block_len;
7569         em->block_start = block_start;
7570         em->bdev = root->fs_info->fs_devices->latest_bdev;
7571         em->orig_block_len = orig_block_len;
7572         em->ram_bytes = ram_bytes;
7573         em->generation = -1;
7574         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7575         if (type == BTRFS_ORDERED_PREALLOC) {
7576                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7577         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7578                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7579                 em->compress_type = compress_type;
7580         }
7581
7582         do {
7583                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7584                                 em->start + em->len - 1, 0);
7585                 write_lock(&em_tree->lock);
7586                 ret = add_extent_mapping(em_tree, em, 1);
7587                 write_unlock(&em_tree->lock);
7588                 /*
7589                  * The caller has taken lock_extent(), who could race with us
7590                  * to add em?
7591                  */
7592         } while (ret == -EEXIST);
7593
7594         if (ret) {
7595                 free_extent_map(em);
7596                 return ERR_PTR(ret);
7597         }
7598
7599         /* em got 2 refs now, callers needs to do free_extent_map once. */
7600         return em;
7601 }
7602
7603 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7604                                    struct buffer_head *bh_result, int create)
7605 {
7606         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7607         struct extent_map *em;
7608         struct extent_state *cached_state = NULL;
7609         struct btrfs_dio_data *dio_data = NULL;
7610         u64 start = iblock << inode->i_blkbits;
7611         u64 lockstart, lockend;
7612         u64 len = bh_result->b_size;
7613         int unlock_bits = EXTENT_LOCKED;
7614         int ret = 0;
7615
7616         if (create)
7617                 unlock_bits |= EXTENT_DIRTY;
7618         else
7619                 len = min_t(u64, len, fs_info->sectorsize);
7620
7621         lockstart = start;
7622         lockend = start + len - 1;
7623
7624         if (current->journal_info) {
7625                 /*
7626                  * Need to pull our outstanding extents and set journal_info to NULL so
7627                  * that anything that needs to check if there's a transaction doesn't get
7628                  * confused.
7629                  */
7630                 dio_data = current->journal_info;
7631                 current->journal_info = NULL;
7632         }
7633
7634         /*
7635          * If this errors out it's because we couldn't invalidate pagecache for
7636          * this range and we need to fallback to buffered.
7637          */
7638         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7639                                create)) {
7640                 ret = -ENOTBLK;
7641                 goto err;
7642         }
7643
7644         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7645         if (IS_ERR(em)) {
7646                 ret = PTR_ERR(em);
7647                 goto unlock_err;
7648         }
7649
7650         /*
7651          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7652          * io.  INLINE is special, and we could probably kludge it in here, but
7653          * it's still buffered so for safety lets just fall back to the generic
7654          * buffered path.
7655          *
7656          * For COMPRESSED we _have_ to read the entire extent in so we can
7657          * decompress it, so there will be buffering required no matter what we
7658          * do, so go ahead and fallback to buffered.
7659          *
7660          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7661          * to buffered IO.  Don't blame me, this is the price we pay for using
7662          * the generic code.
7663          */
7664         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7665             em->block_start == EXTENT_MAP_INLINE) {
7666                 free_extent_map(em);
7667                 ret = -ENOTBLK;
7668                 goto unlock_err;
7669         }
7670
7671         /* Just a good old fashioned hole, return */
7672         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7673                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7674                 free_extent_map(em);
7675                 goto unlock_err;
7676         }
7677
7678         /*
7679          * We don't allocate a new extent in the following cases
7680          *
7681          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7682          * existing extent.
7683          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7684          * just use the extent.
7685          *
7686          */
7687         if (!create) {
7688                 len = min(len, em->len - (start - em->start));
7689                 lockstart = start + len;
7690                 goto unlock;
7691         }
7692
7693         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7694             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7695              em->block_start != EXTENT_MAP_HOLE)) {
7696                 int type;
7697                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7698
7699                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7700                         type = BTRFS_ORDERED_PREALLOC;
7701                 else
7702                         type = BTRFS_ORDERED_NOCOW;
7703                 len = min(len, em->len - (start - em->start));
7704                 block_start = em->block_start + (start - em->start);
7705
7706                 if (can_nocow_extent(inode, start, &len, &orig_start,
7707                                      &orig_block_len, &ram_bytes) == 1 &&
7708                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7709                         struct extent_map *em2;
7710
7711                         em2 = btrfs_create_dio_extent(inode, start, len,
7712                                                       orig_start, block_start,
7713                                                       len, orig_block_len,
7714                                                       ram_bytes, type);
7715                         btrfs_dec_nocow_writers(fs_info, block_start);
7716                         if (type == BTRFS_ORDERED_PREALLOC) {
7717                                 free_extent_map(em);
7718                                 em = em2;
7719                         }
7720                         if (em2 && IS_ERR(em2)) {
7721                                 ret = PTR_ERR(em2);
7722                                 goto unlock_err;
7723                         }
7724                         /*
7725                          * For inode marked NODATACOW or extent marked PREALLOC,
7726                          * use the existing or preallocated extent, so does not
7727                          * need to adjust btrfs_space_info's bytes_may_use.
7728                          */
7729                         btrfs_free_reserved_data_space_noquota(inode,
7730                                         start, len);
7731                         goto unlock;
7732                 }
7733         }
7734
7735         /*
7736          * this will cow the extent, reset the len in case we changed
7737          * it above
7738          */
7739         len = bh_result->b_size;
7740         free_extent_map(em);
7741         em = btrfs_new_extent_direct(inode, start, len);
7742         if (IS_ERR(em)) {
7743                 ret = PTR_ERR(em);
7744                 goto unlock_err;
7745         }
7746         len = min(len, em->len - (start - em->start));
7747 unlock:
7748         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7749                 inode->i_blkbits;
7750         bh_result->b_size = len;
7751         bh_result->b_bdev = em->bdev;
7752         set_buffer_mapped(bh_result);
7753         if (create) {
7754                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7755                         set_buffer_new(bh_result);
7756
7757                 /*
7758                  * Need to update the i_size under the extent lock so buffered
7759                  * readers will get the updated i_size when we unlock.
7760                  */
7761                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7762                         i_size_write(inode, start + len);
7763
7764                 WARN_ON(dio_data->reserve < len);
7765                 dio_data->reserve -= len;
7766                 dio_data->unsubmitted_oe_range_end = start + len;
7767                 current->journal_info = dio_data;
7768         }
7769
7770         /*
7771          * In the case of write we need to clear and unlock the entire range,
7772          * in the case of read we need to unlock only the end area that we
7773          * aren't using if there is any left over space.
7774          */
7775         if (lockstart < lockend) {
7776                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7777                                  lockend, unlock_bits, 1, 0,
7778                                  &cached_state);
7779         } else {
7780                 free_extent_state(cached_state);
7781         }
7782
7783         free_extent_map(em);
7784
7785         return 0;
7786
7787 unlock_err:
7788         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7789                          unlock_bits, 1, 0, &cached_state);
7790 err:
7791         if (dio_data)
7792                 current->journal_info = dio_data;
7793         return ret;
7794 }
7795
7796 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7797                                                  struct bio *bio,
7798                                                  int mirror_num)
7799 {
7800         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7801         blk_status_t ret;
7802
7803         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7804
7805         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7806         if (ret)
7807                 return ret;
7808
7809         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7810
7811         return ret;
7812 }
7813
7814 static int btrfs_check_dio_repairable(struct inode *inode,
7815                                       struct bio *failed_bio,
7816                                       struct io_failure_record *failrec,
7817                                       int failed_mirror)
7818 {
7819         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7820         int num_copies;
7821
7822         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7823         if (num_copies == 1) {
7824                 /*
7825                  * we only have a single copy of the data, so don't bother with
7826                  * all the retry and error correction code that follows. no
7827                  * matter what the error is, it is very likely to persist.
7828                  */
7829                 btrfs_debug(fs_info,
7830                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7831                         num_copies, failrec->this_mirror, failed_mirror);
7832                 return 0;
7833         }
7834
7835         failrec->failed_mirror = failed_mirror;
7836         failrec->this_mirror++;
7837         if (failrec->this_mirror == failed_mirror)
7838                 failrec->this_mirror++;
7839
7840         if (failrec->this_mirror > num_copies) {
7841                 btrfs_debug(fs_info,
7842                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7843                         num_copies, failrec->this_mirror, failed_mirror);
7844                 return 0;
7845         }
7846
7847         return 1;
7848 }
7849
7850 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7851                                    struct page *page, unsigned int pgoff,
7852                                    u64 start, u64 end, int failed_mirror,
7853                                    bio_end_io_t *repair_endio, void *repair_arg)
7854 {
7855         struct io_failure_record *failrec;
7856         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7857         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7858         struct bio *bio;
7859         int isector;
7860         unsigned int read_mode = 0;
7861         int segs;
7862         int ret;
7863         blk_status_t status;
7864         struct bio_vec bvec;
7865
7866         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7867
7868         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7869         if (ret)
7870                 return errno_to_blk_status(ret);
7871
7872         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7873                                          failed_mirror);
7874         if (!ret) {
7875                 free_io_failure(failure_tree, io_tree, failrec);
7876                 return BLK_STS_IOERR;
7877         }
7878
7879         segs = bio_segments(failed_bio);
7880         bio_get_first_bvec(failed_bio, &bvec);
7881         if (segs > 1 ||
7882             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7883                 read_mode |= REQ_FAILFAST_DEV;
7884
7885         isector = start - btrfs_io_bio(failed_bio)->logical;
7886         isector >>= inode->i_sb->s_blocksize_bits;
7887         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7888                                 pgoff, isector, repair_endio, repair_arg);
7889         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
7890
7891         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7892                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7893                     read_mode, failrec->this_mirror, failrec->in_validation);
7894
7895         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7896         if (status) {
7897                 free_io_failure(failure_tree, io_tree, failrec);
7898                 bio_put(bio);
7899         }
7900
7901         return status;
7902 }
7903
7904 struct btrfs_retry_complete {
7905         struct completion done;
7906         struct inode *inode;
7907         u64 start;
7908         int uptodate;
7909 };
7910
7911 static void btrfs_retry_endio_nocsum(struct bio *bio)
7912 {
7913         struct btrfs_retry_complete *done = bio->bi_private;
7914         struct inode *inode = done->inode;
7915         struct bio_vec *bvec;
7916         struct extent_io_tree *io_tree, *failure_tree;
7917         int i;
7918
7919         if (bio->bi_status)
7920                 goto end;
7921
7922         ASSERT(bio->bi_vcnt == 1);
7923         io_tree = &BTRFS_I(inode)->io_tree;
7924         failure_tree = &BTRFS_I(inode)->io_failure_tree;
7925         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
7926
7927         done->uptodate = 1;
7928         ASSERT(!bio_flagged(bio, BIO_CLONED));
7929         bio_for_each_segment_all(bvec, bio, i)
7930                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
7931                                  io_tree, done->start, bvec->bv_page,
7932                                  btrfs_ino(BTRFS_I(inode)), 0);
7933 end:
7934         complete(&done->done);
7935         bio_put(bio);
7936 }
7937
7938 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
7939                                                 struct btrfs_io_bio *io_bio)
7940 {
7941         struct btrfs_fs_info *fs_info;
7942         struct bio_vec bvec;
7943         struct bvec_iter iter;
7944         struct btrfs_retry_complete done;
7945         u64 start;
7946         unsigned int pgoff;
7947         u32 sectorsize;
7948         int nr_sectors;
7949         blk_status_t ret;
7950         blk_status_t err = BLK_STS_OK;
7951
7952         fs_info = BTRFS_I(inode)->root->fs_info;
7953         sectorsize = fs_info->sectorsize;
7954
7955         start = io_bio->logical;
7956         done.inode = inode;
7957         io_bio->bio.bi_iter = io_bio->iter;
7958
7959         bio_for_each_segment(bvec, &io_bio->bio, iter) {
7960                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
7961                 pgoff = bvec.bv_offset;
7962
7963 next_block_or_try_again:
7964                 done.uptodate = 0;
7965                 done.start = start;
7966                 init_completion(&done.done);
7967
7968                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
7969                                 pgoff, start, start + sectorsize - 1,
7970                                 io_bio->mirror_num,
7971                                 btrfs_retry_endio_nocsum, &done);
7972                 if (ret) {
7973                         err = ret;
7974                         goto next;
7975                 }
7976
7977                 wait_for_completion_io(&done.done);
7978
7979                 if (!done.uptodate) {
7980                         /* We might have another mirror, so try again */
7981                         goto next_block_or_try_again;
7982                 }
7983
7984 next:
7985                 start += sectorsize;
7986
7987                 nr_sectors--;
7988                 if (nr_sectors) {
7989                         pgoff += sectorsize;
7990                         ASSERT(pgoff < PAGE_SIZE);
7991                         goto next_block_or_try_again;
7992                 }
7993         }
7994
7995         return err;
7996 }
7997
7998 static void btrfs_retry_endio(struct bio *bio)
7999 {
8000         struct btrfs_retry_complete *done = bio->bi_private;
8001         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8002         struct extent_io_tree *io_tree, *failure_tree;
8003         struct inode *inode = done->inode;
8004         struct bio_vec *bvec;
8005         int uptodate;
8006         int ret;
8007         int i;
8008
8009         if (bio->bi_status)
8010                 goto end;
8011
8012         uptodate = 1;
8013
8014         ASSERT(bio->bi_vcnt == 1);
8015         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
8016
8017         io_tree = &BTRFS_I(inode)->io_tree;
8018         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8019
8020         ASSERT(!bio_flagged(bio, BIO_CLONED));
8021         bio_for_each_segment_all(bvec, bio, i) {
8022                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8023                                              bvec->bv_offset, done->start,
8024                                              bvec->bv_len);
8025                 if (!ret)
8026                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8027                                          failure_tree, io_tree, done->start,
8028                                          bvec->bv_page,
8029                                          btrfs_ino(BTRFS_I(inode)),
8030                                          bvec->bv_offset);
8031                 else
8032                         uptodate = 0;
8033         }
8034
8035         done->uptodate = uptodate;
8036 end:
8037         complete(&done->done);
8038         bio_put(bio);
8039 }
8040
8041 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8042                 struct btrfs_io_bio *io_bio, blk_status_t err)
8043 {
8044         struct btrfs_fs_info *fs_info;
8045         struct bio_vec bvec;
8046         struct bvec_iter iter;
8047         struct btrfs_retry_complete done;
8048         u64 start;
8049         u64 offset = 0;
8050         u32 sectorsize;
8051         int nr_sectors;
8052         unsigned int pgoff;
8053         int csum_pos;
8054         bool uptodate = (err == 0);
8055         int ret;
8056         blk_status_t status;
8057
8058         fs_info = BTRFS_I(inode)->root->fs_info;
8059         sectorsize = fs_info->sectorsize;
8060
8061         err = BLK_STS_OK;
8062         start = io_bio->logical;
8063         done.inode = inode;
8064         io_bio->bio.bi_iter = io_bio->iter;
8065
8066         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8067                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8068
8069                 pgoff = bvec.bv_offset;
8070 next_block:
8071                 if (uptodate) {
8072                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8073                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8074                                         bvec.bv_page, pgoff, start, sectorsize);
8075                         if (likely(!ret))
8076                                 goto next;
8077                 }
8078 try_again:
8079                 done.uptodate = 0;
8080                 done.start = start;
8081                 init_completion(&done.done);
8082
8083                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8084                                         pgoff, start, start + sectorsize - 1,
8085                                         io_bio->mirror_num, btrfs_retry_endio,
8086                                         &done);
8087                 if (status) {
8088                         err = status;
8089                         goto next;
8090                 }
8091
8092                 wait_for_completion_io(&done.done);
8093
8094                 if (!done.uptodate) {
8095                         /* We might have another mirror, so try again */
8096                         goto try_again;
8097                 }
8098 next:
8099                 offset += sectorsize;
8100                 start += sectorsize;
8101
8102                 ASSERT(nr_sectors);
8103
8104                 nr_sectors--;
8105                 if (nr_sectors) {
8106                         pgoff += sectorsize;
8107                         ASSERT(pgoff < PAGE_SIZE);
8108                         goto next_block;
8109                 }
8110         }
8111
8112         return err;
8113 }
8114
8115 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8116                 struct btrfs_io_bio *io_bio, blk_status_t err)
8117 {
8118         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8119
8120         if (skip_csum) {
8121                 if (unlikely(err))
8122                         return __btrfs_correct_data_nocsum(inode, io_bio);
8123                 else
8124                         return BLK_STS_OK;
8125         } else {
8126                 return __btrfs_subio_endio_read(inode, io_bio, err);
8127         }
8128 }
8129
8130 static void btrfs_endio_direct_read(struct bio *bio)
8131 {
8132         struct btrfs_dio_private *dip = bio->bi_private;
8133         struct inode *inode = dip->inode;
8134         struct bio *dio_bio;
8135         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8136         blk_status_t err = bio->bi_status;
8137
8138         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8139                 err = btrfs_subio_endio_read(inode, io_bio, err);
8140
8141         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8142                       dip->logical_offset + dip->bytes - 1);
8143         dio_bio = dip->dio_bio;
8144
8145         kfree(dip);
8146
8147         dio_bio->bi_status = err;
8148         dio_end_io(dio_bio);
8149
8150         if (io_bio->end_io)
8151                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8152         bio_put(bio);
8153 }
8154
8155 static void __endio_write_update_ordered(struct inode *inode,
8156                                          const u64 offset, const u64 bytes,
8157                                          const bool uptodate)
8158 {
8159         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8160         struct btrfs_ordered_extent *ordered = NULL;
8161         struct btrfs_workqueue *wq;
8162         btrfs_work_func_t func;
8163         u64 ordered_offset = offset;
8164         u64 ordered_bytes = bytes;
8165         u64 last_offset;
8166         int ret;
8167
8168         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8169                 wq = fs_info->endio_freespace_worker;
8170                 func = btrfs_freespace_write_helper;
8171         } else {
8172                 wq = fs_info->endio_write_workers;
8173                 func = btrfs_endio_write_helper;
8174         }
8175
8176 again:
8177         last_offset = ordered_offset;
8178         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8179                                                    &ordered_offset,
8180                                                    ordered_bytes,
8181                                                    uptodate);
8182         if (!ret)
8183                 goto out_test;
8184
8185         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8186         btrfs_queue_work(wq, &ordered->work);
8187 out_test:
8188         /*
8189          * If btrfs_dec_test_ordered_pending does not find any ordered extent
8190          * in the range, we can exit.
8191          */
8192         if (ordered_offset == last_offset)
8193                 return;
8194         /*
8195          * our bio might span multiple ordered extents.  If we haven't
8196          * completed the accounting for the whole dio, go back and try again
8197          */
8198         if (ordered_offset < offset + bytes) {
8199                 ordered_bytes = offset + bytes - ordered_offset;
8200                 ordered = NULL;
8201                 goto again;
8202         }
8203 }
8204
8205 static void btrfs_endio_direct_write(struct bio *bio)
8206 {
8207         struct btrfs_dio_private *dip = bio->bi_private;
8208         struct bio *dio_bio = dip->dio_bio;
8209
8210         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8211                                      dip->bytes, !bio->bi_status);
8212
8213         kfree(dip);
8214
8215         dio_bio->bi_status = bio->bi_status;
8216         dio_end_io(dio_bio);
8217         bio_put(bio);
8218 }
8219
8220 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8221                                     struct bio *bio, u64 offset)
8222 {
8223         struct inode *inode = private_data;
8224         blk_status_t ret;
8225         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8226         BUG_ON(ret); /* -ENOMEM */
8227         return 0;
8228 }
8229
8230 static void btrfs_end_dio_bio(struct bio *bio)
8231 {
8232         struct btrfs_dio_private *dip = bio->bi_private;
8233         blk_status_t err = bio->bi_status;
8234
8235         if (err)
8236                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8237                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8238                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8239                            bio->bi_opf,
8240                            (unsigned long long)bio->bi_iter.bi_sector,
8241                            bio->bi_iter.bi_size, err);
8242
8243         if (dip->subio_endio)
8244                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8245
8246         if (err) {
8247                 /*
8248                  * We want to perceive the errors flag being set before
8249                  * decrementing the reference count. We don't need a barrier
8250                  * since atomic operations with a return value are fully
8251                  * ordered as per atomic_t.txt
8252                  */
8253                 dip->errors = 1;
8254         }
8255
8256         /* if there are more bios still pending for this dio, just exit */
8257         if (!atomic_dec_and_test(&dip->pending_bios))
8258                 goto out;
8259
8260         if (dip->errors) {
8261                 bio_io_error(dip->orig_bio);
8262         } else {
8263                 dip->dio_bio->bi_status = BLK_STS_OK;
8264                 bio_endio(dip->orig_bio);
8265         }
8266 out:
8267         bio_put(bio);
8268 }
8269
8270 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8271                                                  struct btrfs_dio_private *dip,
8272                                                  struct bio *bio,
8273                                                  u64 file_offset)
8274 {
8275         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8276         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8277         blk_status_t ret;
8278
8279         /*
8280          * We load all the csum data we need when we submit
8281          * the first bio to reduce the csum tree search and
8282          * contention.
8283          */
8284         if (dip->logical_offset == file_offset) {
8285                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8286                                                 file_offset);
8287                 if (ret)
8288                         return ret;
8289         }
8290
8291         if (bio == dip->orig_bio)
8292                 return 0;
8293
8294         file_offset -= dip->logical_offset;
8295         file_offset >>= inode->i_sb->s_blocksize_bits;
8296         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8297
8298         return 0;
8299 }
8300
8301 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8302                 struct inode *inode, u64 file_offset, int async_submit)
8303 {
8304         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8305         struct btrfs_dio_private *dip = bio->bi_private;
8306         bool write = bio_op(bio) == REQ_OP_WRITE;
8307         blk_status_t ret;
8308
8309         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8310         if (async_submit)
8311                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8312
8313         if (!write) {
8314                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8315                 if (ret)
8316                         goto err;
8317         }
8318
8319         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8320                 goto map;
8321
8322         if (write && async_submit) {
8323                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8324                                           file_offset, inode,
8325                                           btrfs_submit_bio_start_direct_io,
8326                                           btrfs_submit_bio_done);
8327                 goto err;
8328         } else if (write) {
8329                 /*
8330                  * If we aren't doing async submit, calculate the csum of the
8331                  * bio now.
8332                  */
8333                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8334                 if (ret)
8335                         goto err;
8336         } else {
8337                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8338                                                      file_offset);
8339                 if (ret)
8340                         goto err;
8341         }
8342 map:
8343         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8344 err:
8345         return ret;
8346 }
8347
8348 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8349 {
8350         struct inode *inode = dip->inode;
8351         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8352         struct bio *bio;
8353         struct bio *orig_bio = dip->orig_bio;
8354         u64 start_sector = orig_bio->bi_iter.bi_sector;
8355         u64 file_offset = dip->logical_offset;
8356         u64 map_length;
8357         int async_submit = 0;
8358         u64 submit_len;
8359         int clone_offset = 0;
8360         int clone_len;
8361         int ret;
8362         blk_status_t status;
8363
8364         map_length = orig_bio->bi_iter.bi_size;
8365         submit_len = map_length;
8366         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8367                               &map_length, NULL, 0);
8368         if (ret)
8369                 return -EIO;
8370
8371         if (map_length >= submit_len) {
8372                 bio = orig_bio;
8373                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8374                 goto submit;
8375         }
8376
8377         /* async crcs make it difficult to collect full stripe writes. */
8378         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8379                 async_submit = 0;
8380         else
8381                 async_submit = 1;
8382
8383         /* bio split */
8384         ASSERT(map_length <= INT_MAX);
8385         atomic_inc(&dip->pending_bios);
8386         do {
8387                 clone_len = min_t(int, submit_len, map_length);
8388
8389                 /*
8390                  * This will never fail as it's passing GPF_NOFS and
8391                  * the allocation is backed by btrfs_bioset.
8392                  */
8393                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8394                                               clone_len);
8395                 bio->bi_private = dip;
8396                 bio->bi_end_io = btrfs_end_dio_bio;
8397                 btrfs_io_bio(bio)->logical = file_offset;
8398
8399                 ASSERT(submit_len >= clone_len);
8400                 submit_len -= clone_len;
8401                 if (submit_len == 0)
8402                         break;
8403
8404                 /*
8405                  * Increase the count before we submit the bio so we know
8406                  * the end IO handler won't happen before we increase the
8407                  * count. Otherwise, the dip might get freed before we're
8408                  * done setting it up.
8409                  */
8410                 atomic_inc(&dip->pending_bios);
8411
8412                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8413                                                 async_submit);
8414                 if (status) {
8415                         bio_put(bio);
8416                         atomic_dec(&dip->pending_bios);
8417                         goto out_err;
8418                 }
8419
8420                 clone_offset += clone_len;
8421                 start_sector += clone_len >> 9;
8422                 file_offset += clone_len;
8423
8424                 map_length = submit_len;
8425                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8426                                       start_sector << 9, &map_length, NULL, 0);
8427                 if (ret)
8428                         goto out_err;
8429         } while (submit_len > 0);
8430
8431 submit:
8432         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8433         if (!status)
8434                 return 0;
8435
8436         bio_put(bio);
8437 out_err:
8438         dip->errors = 1;
8439         /*
8440          * Before atomic variable goto zero, we must  make sure dip->errors is
8441          * perceived to be set. This ordering is ensured by the fact that an
8442          * atomic operations with a return value are fully ordered as per
8443          * atomic_t.txt
8444          */
8445         if (atomic_dec_and_test(&dip->pending_bios))
8446                 bio_io_error(dip->orig_bio);
8447
8448         /* bio_end_io() will handle error, so we needn't return it */
8449         return 0;
8450 }
8451
8452 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8453                                 loff_t file_offset)
8454 {
8455         struct btrfs_dio_private *dip = NULL;
8456         struct bio *bio = NULL;
8457         struct btrfs_io_bio *io_bio;
8458         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8459         int ret = 0;
8460
8461         bio = btrfs_bio_clone(dio_bio);
8462
8463         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8464         if (!dip) {
8465                 ret = -ENOMEM;
8466                 goto free_ordered;
8467         }
8468
8469         dip->private = dio_bio->bi_private;
8470         dip->inode = inode;
8471         dip->logical_offset = file_offset;
8472         dip->bytes = dio_bio->bi_iter.bi_size;
8473         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8474         bio->bi_private = dip;
8475         dip->orig_bio = bio;
8476         dip->dio_bio = dio_bio;
8477         atomic_set(&dip->pending_bios, 0);
8478         io_bio = btrfs_io_bio(bio);
8479         io_bio->logical = file_offset;
8480
8481         if (write) {
8482                 bio->bi_end_io = btrfs_endio_direct_write;
8483         } else {
8484                 bio->bi_end_io = btrfs_endio_direct_read;
8485                 dip->subio_endio = btrfs_subio_endio_read;
8486         }
8487
8488         /*
8489          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8490          * even if we fail to submit a bio, because in such case we do the
8491          * corresponding error handling below and it must not be done a second
8492          * time by btrfs_direct_IO().
8493          */
8494         if (write) {
8495                 struct btrfs_dio_data *dio_data = current->journal_info;
8496
8497                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8498                         dip->bytes;
8499                 dio_data->unsubmitted_oe_range_start =
8500                         dio_data->unsubmitted_oe_range_end;
8501         }
8502
8503         ret = btrfs_submit_direct_hook(dip);
8504         if (!ret)
8505                 return;
8506
8507         if (io_bio->end_io)
8508                 io_bio->end_io(io_bio, ret);
8509
8510 free_ordered:
8511         /*
8512          * If we arrived here it means either we failed to submit the dip
8513          * or we either failed to clone the dio_bio or failed to allocate the
8514          * dip. If we cloned the dio_bio and allocated the dip, we can just
8515          * call bio_endio against our io_bio so that we get proper resource
8516          * cleanup if we fail to submit the dip, otherwise, we must do the
8517          * same as btrfs_endio_direct_[write|read] because we can't call these
8518          * callbacks - they require an allocated dip and a clone of dio_bio.
8519          */
8520         if (bio && dip) {
8521                 bio_io_error(bio);
8522                 /*
8523                  * The end io callbacks free our dip, do the final put on bio
8524                  * and all the cleanup and final put for dio_bio (through
8525                  * dio_end_io()).
8526                  */
8527                 dip = NULL;
8528                 bio = NULL;
8529         } else {
8530                 if (write)
8531                         __endio_write_update_ordered(inode,
8532                                                 file_offset,
8533                                                 dio_bio->bi_iter.bi_size,
8534                                                 false);
8535                 else
8536                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8537                               file_offset + dio_bio->bi_iter.bi_size - 1);
8538
8539                 dio_bio->bi_status = BLK_STS_IOERR;
8540                 /*
8541                  * Releases and cleans up our dio_bio, no need to bio_put()
8542                  * nor bio_endio()/bio_io_error() against dio_bio.
8543                  */
8544                 dio_end_io(dio_bio);
8545         }
8546         if (bio)
8547                 bio_put(bio);
8548         kfree(dip);
8549 }
8550
8551 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8552                                const struct iov_iter *iter, loff_t offset)
8553 {
8554         int seg;
8555         int i;
8556         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8557         ssize_t retval = -EINVAL;
8558
8559         if (offset & blocksize_mask)
8560                 goto out;
8561
8562         if (iov_iter_alignment(iter) & blocksize_mask)
8563                 goto out;
8564
8565         /* If this is a write we don't need to check anymore */
8566         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8567                 return 0;
8568         /*
8569          * Check to make sure we don't have duplicate iov_base's in this
8570          * iovec, if so return EINVAL, otherwise we'll get csum errors
8571          * when reading back.
8572          */
8573         for (seg = 0; seg < iter->nr_segs; seg++) {
8574                 for (i = seg + 1; i < iter->nr_segs; i++) {
8575                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8576                                 goto out;
8577                 }
8578         }
8579         retval = 0;
8580 out:
8581         return retval;
8582 }
8583
8584 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8585 {
8586         struct file *file = iocb->ki_filp;
8587         struct inode *inode = file->f_mapping->host;
8588         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8589         struct btrfs_dio_data dio_data = { 0 };
8590         struct extent_changeset *data_reserved = NULL;
8591         loff_t offset = iocb->ki_pos;
8592         size_t count = 0;
8593         int flags = 0;
8594         bool wakeup = true;
8595         bool relock = false;
8596         ssize_t ret;
8597
8598         if (check_direct_IO(fs_info, iter, offset))
8599                 return 0;
8600
8601         inode_dio_begin(inode);
8602
8603         /*
8604          * The generic stuff only does filemap_write_and_wait_range, which
8605          * isn't enough if we've written compressed pages to this area, so
8606          * we need to flush the dirty pages again to make absolutely sure
8607          * that any outstanding dirty pages are on disk.
8608          */
8609         count = iov_iter_count(iter);
8610         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8611                      &BTRFS_I(inode)->runtime_flags))
8612                 filemap_fdatawrite_range(inode->i_mapping, offset,
8613                                          offset + count - 1);
8614
8615         if (iov_iter_rw(iter) == WRITE) {
8616                 /*
8617                  * If the write DIO is beyond the EOF, we need update
8618                  * the isize, but it is protected by i_mutex. So we can
8619                  * not unlock the i_mutex at this case.
8620                  */
8621                 if (offset + count <= inode->i_size) {
8622                         dio_data.overwrite = 1;
8623                         inode_unlock(inode);
8624                         relock = true;
8625                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8626                         ret = -EAGAIN;
8627                         goto out;
8628                 }
8629                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8630                                                    offset, count);
8631                 if (ret)
8632                         goto out;
8633
8634                 /*
8635                  * We need to know how many extents we reserved so that we can
8636                  * do the accounting properly if we go over the number we
8637                  * originally calculated.  Abuse current->journal_info for this.
8638                  */
8639                 dio_data.reserve = round_up(count,
8640                                             fs_info->sectorsize);
8641                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8642                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8643                 current->journal_info = &dio_data;
8644                 down_read(&BTRFS_I(inode)->dio_sem);
8645         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8646                                      &BTRFS_I(inode)->runtime_flags)) {
8647                 inode_dio_end(inode);
8648                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8649                 wakeup = false;
8650         }
8651
8652         ret = __blockdev_direct_IO(iocb, inode,
8653                                    fs_info->fs_devices->latest_bdev,
8654                                    iter, btrfs_get_blocks_direct, NULL,
8655                                    btrfs_submit_direct, flags);
8656         if (iov_iter_rw(iter) == WRITE) {
8657                 up_read(&BTRFS_I(inode)->dio_sem);
8658                 current->journal_info = NULL;
8659                 if (ret < 0 && ret != -EIOCBQUEUED) {
8660                         if (dio_data.reserve)
8661                                 btrfs_delalloc_release_space(inode, data_reserved,
8662                                         offset, dio_data.reserve);
8663                         /*
8664                          * On error we might have left some ordered extents
8665                          * without submitting corresponding bios for them, so
8666                          * cleanup them up to avoid other tasks getting them
8667                          * and waiting for them to complete forever.
8668                          */
8669                         if (dio_data.unsubmitted_oe_range_start <
8670                             dio_data.unsubmitted_oe_range_end)
8671                                 __endio_write_update_ordered(inode,
8672                                         dio_data.unsubmitted_oe_range_start,
8673                                         dio_data.unsubmitted_oe_range_end -
8674                                         dio_data.unsubmitted_oe_range_start,
8675                                         false);
8676                 } else if (ret >= 0 && (size_t)ret < count)
8677                         btrfs_delalloc_release_space(inode, data_reserved,
8678                                         offset, count - (size_t)ret);
8679                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8680         }
8681 out:
8682         if (wakeup)
8683                 inode_dio_end(inode);
8684         if (relock)
8685                 inode_lock(inode);
8686
8687         extent_changeset_free(data_reserved);
8688         return ret;
8689 }
8690
8691 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8692
8693 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8694                 __u64 start, __u64 len)
8695 {
8696         int     ret;
8697
8698         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8699         if (ret)
8700                 return ret;
8701
8702         return extent_fiemap(inode, fieinfo, start, len);
8703 }
8704
8705 int btrfs_readpage(struct file *file, struct page *page)
8706 {
8707         struct extent_io_tree *tree;
8708         tree = &BTRFS_I(page->mapping->host)->io_tree;
8709         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8710 }
8711
8712 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8713 {
8714         struct inode *inode = page->mapping->host;
8715         int ret;
8716
8717         if (current->flags & PF_MEMALLOC) {
8718                 redirty_page_for_writepage(wbc, page);
8719                 unlock_page(page);
8720                 return 0;
8721         }
8722
8723         /*
8724          * If we are under memory pressure we will call this directly from the
8725          * VM, we need to make sure we have the inode referenced for the ordered
8726          * extent.  If not just return like we didn't do anything.
8727          */
8728         if (!igrab(inode)) {
8729                 redirty_page_for_writepage(wbc, page);
8730                 return AOP_WRITEPAGE_ACTIVATE;
8731         }
8732         ret = extent_write_full_page(page, wbc);
8733         btrfs_add_delayed_iput(inode);
8734         return ret;
8735 }
8736
8737 static int btrfs_writepages(struct address_space *mapping,
8738                             struct writeback_control *wbc)
8739 {
8740         struct extent_io_tree *tree;
8741
8742         tree = &BTRFS_I(mapping->host)->io_tree;
8743         return extent_writepages(tree, mapping, wbc);
8744 }
8745
8746 static int
8747 btrfs_readpages(struct file *file, struct address_space *mapping,
8748                 struct list_head *pages, unsigned nr_pages)
8749 {
8750         struct extent_io_tree *tree;
8751         tree = &BTRFS_I(mapping->host)->io_tree;
8752         return extent_readpages(tree, mapping, pages, nr_pages);
8753 }
8754 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8755 {
8756         struct extent_io_tree *tree;
8757         struct extent_map_tree *map;
8758         int ret;
8759
8760         tree = &BTRFS_I(page->mapping->host)->io_tree;
8761         map = &BTRFS_I(page->mapping->host)->extent_tree;
8762         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8763         if (ret == 1) {
8764                 ClearPagePrivate(page);
8765                 set_page_private(page, 0);
8766                 put_page(page);
8767         }
8768         return ret;
8769 }
8770
8771 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8772 {
8773         if (PageWriteback(page) || PageDirty(page))
8774                 return 0;
8775         return __btrfs_releasepage(page, gfp_flags);
8776 }
8777
8778 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8779                                  unsigned int length)
8780 {
8781         struct inode *inode = page->mapping->host;
8782         struct extent_io_tree *tree;
8783         struct btrfs_ordered_extent *ordered;
8784         struct extent_state *cached_state = NULL;
8785         u64 page_start = page_offset(page);
8786         u64 page_end = page_start + PAGE_SIZE - 1;
8787         u64 start;
8788         u64 end;
8789         int inode_evicting = inode->i_state & I_FREEING;
8790
8791         /*
8792          * we have the page locked, so new writeback can't start,
8793          * and the dirty bit won't be cleared while we are here.
8794          *
8795          * Wait for IO on this page so that we can safely clear
8796          * the PagePrivate2 bit and do ordered accounting
8797          */
8798         wait_on_page_writeback(page);
8799
8800         tree = &BTRFS_I(inode)->io_tree;
8801         if (offset) {
8802                 btrfs_releasepage(page, GFP_NOFS);
8803                 return;
8804         }
8805
8806         if (!inode_evicting)
8807                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8808 again:
8809         start = page_start;
8810         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8811                                         page_end - start + 1);
8812         if (ordered) {
8813                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8814                 /*
8815                  * IO on this page will never be started, so we need
8816                  * to account for any ordered extents now
8817                  */
8818                 if (!inode_evicting)
8819                         clear_extent_bit(tree, start, end,
8820                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8821                                          EXTENT_DELALLOC_NEW |
8822                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8823                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8824                 /*
8825                  * whoever cleared the private bit is responsible
8826                  * for the finish_ordered_io
8827                  */
8828                 if (TestClearPagePrivate2(page)) {
8829                         struct btrfs_ordered_inode_tree *tree;
8830                         u64 new_len;
8831
8832                         tree = &BTRFS_I(inode)->ordered_tree;
8833
8834                         spin_lock_irq(&tree->lock);
8835                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8836                         new_len = start - ordered->file_offset;
8837                         if (new_len < ordered->truncated_len)
8838                                 ordered->truncated_len = new_len;
8839                         spin_unlock_irq(&tree->lock);
8840
8841                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8842                                                            start,
8843                                                            end - start + 1, 1))
8844                                 btrfs_finish_ordered_io(ordered);
8845                 }
8846                 btrfs_put_ordered_extent(ordered);
8847                 if (!inode_evicting) {
8848                         cached_state = NULL;
8849                         lock_extent_bits(tree, start, end,
8850                                          &cached_state);
8851                 }
8852
8853                 start = end + 1;
8854                 if (start < page_end)
8855                         goto again;
8856         }
8857
8858         /*
8859          * Qgroup reserved space handler
8860          * Page here will be either
8861          * 1) Already written to disk
8862          *    In this case, its reserved space is released from data rsv map
8863          *    and will be freed by delayed_ref handler finally.
8864          *    So even we call qgroup_free_data(), it won't decrease reserved
8865          *    space.
8866          * 2) Not written to disk
8867          *    This means the reserved space should be freed here. However,
8868          *    if a truncate invalidates the page (by clearing PageDirty)
8869          *    and the page is accounted for while allocating extent
8870          *    in btrfs_check_data_free_space() we let delayed_ref to
8871          *    free the entire extent.
8872          */
8873         if (PageDirty(page))
8874                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8875         if (!inode_evicting) {
8876                 clear_extent_bit(tree, page_start, page_end,
8877                                  EXTENT_LOCKED | EXTENT_DIRTY |
8878                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8879                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8880                                  &cached_state);
8881
8882                 __btrfs_releasepage(page, GFP_NOFS);
8883         }
8884
8885         ClearPageChecked(page);
8886         if (PagePrivate(page)) {
8887                 ClearPagePrivate(page);
8888                 set_page_private(page, 0);
8889                 put_page(page);
8890         }
8891 }
8892
8893 /*
8894  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8895  * called from a page fault handler when a page is first dirtied. Hence we must
8896  * be careful to check for EOF conditions here. We set the page up correctly
8897  * for a written page which means we get ENOSPC checking when writing into
8898  * holes and correct delalloc and unwritten extent mapping on filesystems that
8899  * support these features.
8900  *
8901  * We are not allowed to take the i_mutex here so we have to play games to
8902  * protect against truncate races as the page could now be beyond EOF.  Because
8903  * vmtruncate() writes the inode size before removing pages, once we have the
8904  * page lock we can determine safely if the page is beyond EOF. If it is not
8905  * beyond EOF, then the page is guaranteed safe against truncation until we
8906  * unlock the page.
8907  */
8908 int btrfs_page_mkwrite(struct vm_fault *vmf)
8909 {
8910         struct page *page = vmf->page;
8911         struct inode *inode = file_inode(vmf->vma->vm_file);
8912         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8913         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8914         struct btrfs_ordered_extent *ordered;
8915         struct extent_state *cached_state = NULL;
8916         struct extent_changeset *data_reserved = NULL;
8917         char *kaddr;
8918         unsigned long zero_start;
8919         loff_t size;
8920         int ret;
8921         int reserved = 0;
8922         u64 reserved_space;
8923         u64 page_start;
8924         u64 page_end;
8925         u64 end;
8926
8927         reserved_space = PAGE_SIZE;
8928
8929         sb_start_pagefault(inode->i_sb);
8930         page_start = page_offset(page);
8931         page_end = page_start + PAGE_SIZE - 1;
8932         end = page_end;
8933
8934         /*
8935          * Reserving delalloc space after obtaining the page lock can lead to
8936          * deadlock. For example, if a dirty page is locked by this function
8937          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8938          * dirty page write out, then the btrfs_writepage() function could
8939          * end up waiting indefinitely to get a lock on the page currently
8940          * being processed by btrfs_page_mkwrite() function.
8941          */
8942         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
8943                                            reserved_space);
8944         if (!ret) {
8945                 ret = file_update_time(vmf->vma->vm_file);
8946                 reserved = 1;
8947         }
8948         if (ret) {
8949                 if (ret == -ENOMEM)
8950                         ret = VM_FAULT_OOM;
8951                 else /* -ENOSPC, -EIO, etc */
8952                         ret = VM_FAULT_SIGBUS;
8953                 if (reserved)
8954                         goto out;
8955                 goto out_noreserve;
8956         }
8957
8958         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8959 again:
8960         lock_page(page);
8961         size = i_size_read(inode);
8962
8963         if ((page->mapping != inode->i_mapping) ||
8964             (page_start >= size)) {
8965                 /* page got truncated out from underneath us */
8966                 goto out_unlock;
8967         }
8968         wait_on_page_writeback(page);
8969
8970         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8971         set_page_extent_mapped(page);
8972
8973         /*
8974          * we can't set the delalloc bits if there are pending ordered
8975          * extents.  Drop our locks and wait for them to finish
8976          */
8977         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8978                         PAGE_SIZE);
8979         if (ordered) {
8980                 unlock_extent_cached(io_tree, page_start, page_end,
8981                                      &cached_state);
8982                 unlock_page(page);
8983                 btrfs_start_ordered_extent(inode, ordered, 1);
8984                 btrfs_put_ordered_extent(ordered);
8985                 goto again;
8986         }
8987
8988         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8989                 reserved_space = round_up(size - page_start,
8990                                           fs_info->sectorsize);
8991                 if (reserved_space < PAGE_SIZE) {
8992                         end = page_start + reserved_space - 1;
8993                         btrfs_delalloc_release_space(inode, data_reserved,
8994                                         page_start, PAGE_SIZE - reserved_space);
8995                 }
8996         }
8997
8998         /*
8999          * page_mkwrite gets called when the page is firstly dirtied after it's
9000          * faulted in, but write(2) could also dirty a page and set delalloc
9001          * bits, thus in this case for space account reason, we still need to
9002          * clear any delalloc bits within this page range since we have to
9003          * reserve data&meta space before lock_page() (see above comments).
9004          */
9005         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9006                           EXTENT_DIRTY | EXTENT_DELALLOC |
9007                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9008                           0, 0, &cached_state);
9009
9010         ret = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9011                                         &cached_state, 0);
9012         if (ret) {
9013                 unlock_extent_cached(io_tree, page_start, page_end,
9014                                      &cached_state);
9015                 ret = VM_FAULT_SIGBUS;
9016                 goto out_unlock;
9017         }
9018         ret = 0;
9019
9020         /* page is wholly or partially inside EOF */
9021         if (page_start + PAGE_SIZE > size)
9022                 zero_start = size & ~PAGE_MASK;
9023         else
9024                 zero_start = PAGE_SIZE;
9025
9026         if (zero_start != PAGE_SIZE) {
9027                 kaddr = kmap(page);
9028                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9029                 flush_dcache_page(page);
9030                 kunmap(page);
9031         }
9032         ClearPageChecked(page);
9033         set_page_dirty(page);
9034         SetPageUptodate(page);
9035
9036         BTRFS_I(inode)->last_trans = fs_info->generation;
9037         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9038         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9039
9040         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
9041
9042 out_unlock:
9043         if (!ret) {
9044                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9045                 sb_end_pagefault(inode->i_sb);
9046                 extent_changeset_free(data_reserved);
9047                 return VM_FAULT_LOCKED;
9048         }
9049         unlock_page(page);
9050 out:
9051         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9052         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9053                                      reserved_space);
9054 out_noreserve:
9055         sb_end_pagefault(inode->i_sb);
9056         extent_changeset_free(data_reserved);
9057         return ret;
9058 }
9059
9060 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
9061 {
9062         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9063         struct btrfs_root *root = BTRFS_I(inode)->root;
9064         struct btrfs_block_rsv *rsv;
9065         int ret = 0;
9066         int err = 0;
9067         struct btrfs_trans_handle *trans;
9068         u64 mask = fs_info->sectorsize - 1;
9069         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9070
9071         if (!skip_writeback) {
9072                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9073                                                (u64)-1);
9074                 if (ret)
9075                         return ret;
9076         }
9077
9078         /*
9079          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9080          * 3 things going on here
9081          *
9082          * 1) We need to reserve space for our orphan item and the space to
9083          * delete our orphan item.  Lord knows we don't want to have a dangling
9084          * orphan item because we didn't reserve space to remove it.
9085          *
9086          * 2) We need to reserve space to update our inode.
9087          *
9088          * 3) We need to have something to cache all the space that is going to
9089          * be free'd up by the truncate operation, but also have some slack
9090          * space reserved in case it uses space during the truncate (thank you
9091          * very much snapshotting).
9092          *
9093          * And we need these to all be separate.  The fact is we can use a lot of
9094          * space doing the truncate, and we have no earthly idea how much space
9095          * we will use, so we need the truncate reservation to be separate so it
9096          * doesn't end up using space reserved for updating the inode or
9097          * removing the orphan item.  We also need to be able to stop the
9098          * transaction and start a new one, which means we need to be able to
9099          * update the inode several times, and we have no idea of knowing how
9100          * many times that will be, so we can't just reserve 1 item for the
9101          * entirety of the operation, so that has to be done separately as well.
9102          * Then there is the orphan item, which does indeed need to be held on
9103          * to for the whole operation, and we need nobody to touch this reserved
9104          * space except the orphan code.
9105          *
9106          * So that leaves us with
9107          *
9108          * 1) root->orphan_block_rsv - for the orphan deletion.
9109          * 2) rsv - for the truncate reservation, which we will steal from the
9110          * transaction reservation.
9111          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9112          * updating the inode.
9113          */
9114         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9115         if (!rsv)
9116                 return -ENOMEM;
9117         rsv->size = min_size;
9118         rsv->failfast = 1;
9119
9120         /*
9121          * 1 for the truncate slack space
9122          * 1 for updating the inode.
9123          */
9124         trans = btrfs_start_transaction(root, 2);
9125         if (IS_ERR(trans)) {
9126                 err = PTR_ERR(trans);
9127                 goto out;
9128         }
9129
9130         /* Migrate the slack space for the truncate to our reserve */
9131         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9132                                       min_size, 0);
9133         BUG_ON(ret);
9134
9135         /*
9136          * So if we truncate and then write and fsync we normally would just
9137          * write the extents that changed, which is a problem if we need to
9138          * first truncate that entire inode.  So set this flag so we write out
9139          * all of the extents in the inode to the sync log so we're completely
9140          * safe.
9141          */
9142         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9143         trans->block_rsv = rsv;
9144
9145         while (1) {
9146                 ret = btrfs_truncate_inode_items(trans, root, inode,
9147                                                  inode->i_size,
9148                                                  BTRFS_EXTENT_DATA_KEY);
9149                 trans->block_rsv = &fs_info->trans_block_rsv;
9150                 if (ret != -ENOSPC && ret != -EAGAIN) {
9151                         err = ret;
9152                         break;
9153                 }
9154
9155                 ret = btrfs_update_inode(trans, root, inode);
9156                 if (ret) {
9157                         err = ret;
9158                         break;
9159                 }
9160
9161                 btrfs_end_transaction(trans);
9162                 btrfs_btree_balance_dirty(fs_info);
9163
9164                 trans = btrfs_start_transaction(root, 2);
9165                 if (IS_ERR(trans)) {
9166                         ret = err = PTR_ERR(trans);
9167                         trans = NULL;
9168                         break;
9169                 }
9170
9171                 btrfs_block_rsv_release(fs_info, rsv, -1);
9172                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9173                                               rsv, min_size, 0);
9174                 BUG_ON(ret);    /* shouldn't happen */
9175                 trans->block_rsv = rsv;
9176         }
9177
9178         /*
9179          * We can't call btrfs_truncate_block inside a trans handle as we could
9180          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9181          * we've truncated everything except the last little bit, and can do
9182          * btrfs_truncate_block and then update the disk_i_size.
9183          */
9184         if (ret == NEED_TRUNCATE_BLOCK) {
9185                 btrfs_end_transaction(trans);
9186                 btrfs_btree_balance_dirty(fs_info);
9187
9188                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9189                 if (ret)
9190                         goto out;
9191                 trans = btrfs_start_transaction(root, 1);
9192                 if (IS_ERR(trans)) {
9193                         ret = PTR_ERR(trans);
9194                         goto out;
9195                 }
9196                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9197         }
9198
9199         if (ret == 0 && inode->i_nlink > 0) {
9200                 trans->block_rsv = root->orphan_block_rsv;
9201                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9202                 if (ret)
9203                         err = ret;
9204         }
9205
9206         if (trans) {
9207                 trans->block_rsv = &fs_info->trans_block_rsv;
9208                 ret = btrfs_update_inode(trans, root, inode);
9209                 if (ret && !err)
9210                         err = ret;
9211
9212                 ret = btrfs_end_transaction(trans);
9213                 btrfs_btree_balance_dirty(fs_info);
9214         }
9215 out:
9216         btrfs_free_block_rsv(fs_info, rsv);
9217
9218         if (ret && !err)
9219                 err = ret;
9220
9221         return err;
9222 }
9223
9224 /*
9225  * create a new subvolume directory/inode (helper for the ioctl).
9226  */
9227 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9228                              struct btrfs_root *new_root,
9229                              struct btrfs_root *parent_root,
9230                              u64 new_dirid)
9231 {
9232         struct inode *inode;
9233         int err;
9234         u64 index = 0;
9235
9236         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9237                                 new_dirid, new_dirid,
9238                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9239                                 &index);
9240         if (IS_ERR(inode))
9241                 return PTR_ERR(inode);
9242         inode->i_op = &btrfs_dir_inode_operations;
9243         inode->i_fop = &btrfs_dir_file_operations;
9244
9245         set_nlink(inode, 1);
9246         btrfs_i_size_write(BTRFS_I(inode), 0);
9247         unlock_new_inode(inode);
9248
9249         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9250         if (err)
9251                 btrfs_err(new_root->fs_info,
9252                           "error inheriting subvolume %llu properties: %d",
9253                           new_root->root_key.objectid, err);
9254
9255         err = btrfs_update_inode(trans, new_root, inode);
9256
9257         iput(inode);
9258         return err;
9259 }
9260
9261 struct inode *btrfs_alloc_inode(struct super_block *sb)
9262 {
9263         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9264         struct btrfs_inode *ei;
9265         struct inode *inode;
9266
9267         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9268         if (!ei)
9269                 return NULL;
9270
9271         ei->root = NULL;
9272         ei->generation = 0;
9273         ei->last_trans = 0;
9274         ei->last_sub_trans = 0;
9275         ei->logged_trans = 0;
9276         ei->delalloc_bytes = 0;
9277         ei->new_delalloc_bytes = 0;
9278         ei->defrag_bytes = 0;
9279         ei->disk_i_size = 0;
9280         ei->flags = 0;
9281         ei->csum_bytes = 0;
9282         ei->index_cnt = (u64)-1;
9283         ei->dir_index = 0;
9284         ei->last_unlink_trans = 0;
9285         ei->last_log_commit = 0;
9286
9287         spin_lock_init(&ei->lock);
9288         ei->outstanding_extents = 0;
9289         if (sb->s_magic != BTRFS_TEST_MAGIC)
9290                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9291                                               BTRFS_BLOCK_RSV_DELALLOC);
9292         ei->runtime_flags = 0;
9293         ei->prop_compress = BTRFS_COMPRESS_NONE;
9294         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9295
9296         ei->delayed_node = NULL;
9297
9298         ei->i_otime.tv_sec = 0;
9299         ei->i_otime.tv_nsec = 0;
9300
9301         inode = &ei->vfs_inode;
9302         extent_map_tree_init(&ei->extent_tree);
9303         extent_io_tree_init(&ei->io_tree, inode);
9304         extent_io_tree_init(&ei->io_failure_tree, inode);
9305         ei->io_tree.track_uptodate = 1;
9306         ei->io_failure_tree.track_uptodate = 1;
9307         atomic_set(&ei->sync_writers, 0);
9308         mutex_init(&ei->log_mutex);
9309         mutex_init(&ei->delalloc_mutex);
9310         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9311         INIT_LIST_HEAD(&ei->delalloc_inodes);
9312         INIT_LIST_HEAD(&ei->delayed_iput);
9313         RB_CLEAR_NODE(&ei->rb_node);
9314         init_rwsem(&ei->dio_sem);
9315
9316         return inode;
9317 }
9318
9319 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9320 void btrfs_test_destroy_inode(struct inode *inode)
9321 {
9322         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9323         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9324 }
9325 #endif
9326
9327 static void btrfs_i_callback(struct rcu_head *head)
9328 {
9329         struct inode *inode = container_of(head, struct inode, i_rcu);
9330         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9331 }
9332
9333 void btrfs_destroy_inode(struct inode *inode)
9334 {
9335         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9336         struct btrfs_ordered_extent *ordered;
9337         struct btrfs_root *root = BTRFS_I(inode)->root;
9338
9339         WARN_ON(!hlist_empty(&inode->i_dentry));
9340         WARN_ON(inode->i_data.nrpages);
9341         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9342         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9343         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9344         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9345         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9346         WARN_ON(BTRFS_I(inode)->csum_bytes);
9347         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9348
9349         /*
9350          * This can happen where we create an inode, but somebody else also
9351          * created the same inode and we need to destroy the one we already
9352          * created.
9353          */
9354         if (!root)
9355                 goto free;
9356
9357         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9358                      &BTRFS_I(inode)->runtime_flags)) {
9359                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9360                            btrfs_ino(BTRFS_I(inode)));
9361                 atomic_dec(&root->orphan_inodes);
9362         }
9363
9364         while (1) {
9365                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9366                 if (!ordered)
9367                         break;
9368                 else {
9369                         btrfs_err(fs_info,
9370                                   "found ordered extent %llu %llu on inode cleanup",
9371                                   ordered->file_offset, ordered->len);
9372                         btrfs_remove_ordered_extent(inode, ordered);
9373                         btrfs_put_ordered_extent(ordered);
9374                         btrfs_put_ordered_extent(ordered);
9375                 }
9376         }
9377         btrfs_qgroup_check_reserved_leak(inode);
9378         inode_tree_del(inode);
9379         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9380 free:
9381         call_rcu(&inode->i_rcu, btrfs_i_callback);
9382 }
9383
9384 int btrfs_drop_inode(struct inode *inode)
9385 {
9386         struct btrfs_root *root = BTRFS_I(inode)->root;
9387
9388         if (root == NULL)
9389                 return 1;
9390
9391         /* the snap/subvol tree is on deleting */
9392         if (btrfs_root_refs(&root->root_item) == 0)
9393                 return 1;
9394         else
9395                 return generic_drop_inode(inode);
9396 }
9397
9398 static void init_once(void *foo)
9399 {
9400         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9401
9402         inode_init_once(&ei->vfs_inode);
9403 }
9404
9405 void __cold btrfs_destroy_cachep(void)
9406 {
9407         /*
9408          * Make sure all delayed rcu free inodes are flushed before we
9409          * destroy cache.
9410          */
9411         rcu_barrier();
9412         kmem_cache_destroy(btrfs_inode_cachep);
9413         kmem_cache_destroy(btrfs_trans_handle_cachep);
9414         kmem_cache_destroy(btrfs_path_cachep);
9415         kmem_cache_destroy(btrfs_free_space_cachep);
9416 }
9417
9418 int __init btrfs_init_cachep(void)
9419 {
9420         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9421                         sizeof(struct btrfs_inode), 0,
9422                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9423                         init_once);
9424         if (!btrfs_inode_cachep)
9425                 goto fail;
9426
9427         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9428                         sizeof(struct btrfs_trans_handle), 0,
9429                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9430         if (!btrfs_trans_handle_cachep)
9431                 goto fail;
9432
9433         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9434                         sizeof(struct btrfs_path), 0,
9435                         SLAB_MEM_SPREAD, NULL);
9436         if (!btrfs_path_cachep)
9437                 goto fail;
9438
9439         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9440                         sizeof(struct btrfs_free_space), 0,
9441                         SLAB_MEM_SPREAD, NULL);
9442         if (!btrfs_free_space_cachep)
9443                 goto fail;
9444
9445         return 0;
9446 fail:
9447         btrfs_destroy_cachep();
9448         return -ENOMEM;
9449 }
9450
9451 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9452                          u32 request_mask, unsigned int flags)
9453 {
9454         u64 delalloc_bytes;
9455         struct inode *inode = d_inode(path->dentry);
9456         u32 blocksize = inode->i_sb->s_blocksize;
9457         u32 bi_flags = BTRFS_I(inode)->flags;
9458
9459         stat->result_mask |= STATX_BTIME;
9460         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9461         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9462         if (bi_flags & BTRFS_INODE_APPEND)
9463                 stat->attributes |= STATX_ATTR_APPEND;
9464         if (bi_flags & BTRFS_INODE_COMPRESS)
9465                 stat->attributes |= STATX_ATTR_COMPRESSED;
9466         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9467                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9468         if (bi_flags & BTRFS_INODE_NODUMP)
9469                 stat->attributes |= STATX_ATTR_NODUMP;
9470
9471         stat->attributes_mask |= (STATX_ATTR_APPEND |
9472                                   STATX_ATTR_COMPRESSED |
9473                                   STATX_ATTR_IMMUTABLE |
9474                                   STATX_ATTR_NODUMP);
9475
9476         generic_fillattr(inode, stat);
9477         stat->dev = BTRFS_I(inode)->root->anon_dev;
9478
9479         spin_lock(&BTRFS_I(inode)->lock);
9480         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9481         spin_unlock(&BTRFS_I(inode)->lock);
9482         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9483                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9484         return 0;
9485 }
9486
9487 static int btrfs_rename_exchange(struct inode *old_dir,
9488                               struct dentry *old_dentry,
9489                               struct inode *new_dir,
9490                               struct dentry *new_dentry)
9491 {
9492         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9493         struct btrfs_trans_handle *trans;
9494         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9495         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9496         struct inode *new_inode = new_dentry->d_inode;
9497         struct inode *old_inode = old_dentry->d_inode;
9498         struct timespec ctime = current_time(old_inode);
9499         struct dentry *parent;
9500         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9501         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9502         u64 old_idx = 0;
9503         u64 new_idx = 0;
9504         u64 root_objectid;
9505         int ret;
9506         bool root_log_pinned = false;
9507         bool dest_log_pinned = false;
9508
9509         /* we only allow rename subvolume link between subvolumes */
9510         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9511                 return -EXDEV;
9512
9513         /* close the race window with snapshot create/destroy ioctl */
9514         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9515                 down_read(&fs_info->subvol_sem);
9516         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9517                 down_read(&fs_info->subvol_sem);
9518
9519         /*
9520          * We want to reserve the absolute worst case amount of items.  So if
9521          * both inodes are subvols and we need to unlink them then that would
9522          * require 4 item modifications, but if they are both normal inodes it
9523          * would require 5 item modifications, so we'll assume their normal
9524          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9525          * should cover the worst case number of items we'll modify.
9526          */
9527         trans = btrfs_start_transaction(root, 12);
9528         if (IS_ERR(trans)) {
9529                 ret = PTR_ERR(trans);
9530                 goto out_notrans;
9531         }
9532
9533         /*
9534          * We need to find a free sequence number both in the source and
9535          * in the destination directory for the exchange.
9536          */
9537         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9538         if (ret)
9539                 goto out_fail;
9540         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9541         if (ret)
9542                 goto out_fail;
9543
9544         BTRFS_I(old_inode)->dir_index = 0ULL;
9545         BTRFS_I(new_inode)->dir_index = 0ULL;
9546
9547         /* Reference for the source. */
9548         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9549                 /* force full log commit if subvolume involved. */
9550                 btrfs_set_log_full_commit(fs_info, trans);
9551         } else {
9552                 btrfs_pin_log_trans(root);
9553                 root_log_pinned = true;
9554                 ret = btrfs_insert_inode_ref(trans, dest,
9555                                              new_dentry->d_name.name,
9556                                              new_dentry->d_name.len,
9557                                              old_ino,
9558                                              btrfs_ino(BTRFS_I(new_dir)),
9559                                              old_idx);
9560                 if (ret)
9561                         goto out_fail;
9562         }
9563
9564         /* And now for the dest. */
9565         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9566                 /* force full log commit if subvolume involved. */
9567                 btrfs_set_log_full_commit(fs_info, trans);
9568         } else {
9569                 btrfs_pin_log_trans(dest);
9570                 dest_log_pinned = true;
9571                 ret = btrfs_insert_inode_ref(trans, root,
9572                                              old_dentry->d_name.name,
9573                                              old_dentry->d_name.len,
9574                                              new_ino,
9575                                              btrfs_ino(BTRFS_I(old_dir)),
9576                                              new_idx);
9577                 if (ret)
9578                         goto out_fail;
9579         }
9580
9581         /* Update inode version and ctime/mtime. */
9582         inode_inc_iversion(old_dir);
9583         inode_inc_iversion(new_dir);
9584         inode_inc_iversion(old_inode);
9585         inode_inc_iversion(new_inode);
9586         old_dir->i_ctime = old_dir->i_mtime = ctime;
9587         new_dir->i_ctime = new_dir->i_mtime = ctime;
9588         old_inode->i_ctime = ctime;
9589         new_inode->i_ctime = ctime;
9590
9591         if (old_dentry->d_parent != new_dentry->d_parent) {
9592                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9593                                 BTRFS_I(old_inode), 1);
9594                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9595                                 BTRFS_I(new_inode), 1);
9596         }
9597
9598         /* src is a subvolume */
9599         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9600                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9601                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9602                                           root_objectid,
9603                                           old_dentry->d_name.name,
9604                                           old_dentry->d_name.len);
9605         } else { /* src is an inode */
9606                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9607                                            BTRFS_I(old_dentry->d_inode),
9608                                            old_dentry->d_name.name,
9609                                            old_dentry->d_name.len);
9610                 if (!ret)
9611                         ret = btrfs_update_inode(trans, root, old_inode);
9612         }
9613         if (ret) {
9614                 btrfs_abort_transaction(trans, ret);
9615                 goto out_fail;
9616         }
9617
9618         /* dest is a subvolume */
9619         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9620                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9621                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9622                                           root_objectid,
9623                                           new_dentry->d_name.name,
9624                                           new_dentry->d_name.len);
9625         } else { /* dest is an inode */
9626                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9627                                            BTRFS_I(new_dentry->d_inode),
9628                                            new_dentry->d_name.name,
9629                                            new_dentry->d_name.len);
9630                 if (!ret)
9631                         ret = btrfs_update_inode(trans, dest, new_inode);
9632         }
9633         if (ret) {
9634                 btrfs_abort_transaction(trans, ret);
9635                 goto out_fail;
9636         }
9637
9638         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9639                              new_dentry->d_name.name,
9640                              new_dentry->d_name.len, 0, old_idx);
9641         if (ret) {
9642                 btrfs_abort_transaction(trans, ret);
9643                 goto out_fail;
9644         }
9645
9646         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9647                              old_dentry->d_name.name,
9648                              old_dentry->d_name.len, 0, new_idx);
9649         if (ret) {
9650                 btrfs_abort_transaction(trans, ret);
9651                 goto out_fail;
9652         }
9653
9654         if (old_inode->i_nlink == 1)
9655                 BTRFS_I(old_inode)->dir_index = old_idx;
9656         if (new_inode->i_nlink == 1)
9657                 BTRFS_I(new_inode)->dir_index = new_idx;
9658
9659         if (root_log_pinned) {
9660                 parent = new_dentry->d_parent;
9661                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9662                                 parent);
9663                 btrfs_end_log_trans(root);
9664                 root_log_pinned = false;
9665         }
9666         if (dest_log_pinned) {
9667                 parent = old_dentry->d_parent;
9668                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9669                                 parent);
9670                 btrfs_end_log_trans(dest);
9671                 dest_log_pinned = false;
9672         }
9673 out_fail:
9674         /*
9675          * If we have pinned a log and an error happened, we unpin tasks
9676          * trying to sync the log and force them to fallback to a transaction
9677          * commit if the log currently contains any of the inodes involved in
9678          * this rename operation (to ensure we do not persist a log with an
9679          * inconsistent state for any of these inodes or leading to any
9680          * inconsistencies when replayed). If the transaction was aborted, the
9681          * abortion reason is propagated to userspace when attempting to commit
9682          * the transaction. If the log does not contain any of these inodes, we
9683          * allow the tasks to sync it.
9684          */
9685         if (ret && (root_log_pinned || dest_log_pinned)) {
9686                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9687                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9688                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9689                     (new_inode &&
9690                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9691                         btrfs_set_log_full_commit(fs_info, trans);
9692
9693                 if (root_log_pinned) {
9694                         btrfs_end_log_trans(root);
9695                         root_log_pinned = false;
9696                 }
9697                 if (dest_log_pinned) {
9698                         btrfs_end_log_trans(dest);
9699                         dest_log_pinned = false;
9700                 }
9701         }
9702         ret = btrfs_end_transaction(trans);
9703 out_notrans:
9704         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9705                 up_read(&fs_info->subvol_sem);
9706         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9707                 up_read(&fs_info->subvol_sem);
9708
9709         return ret;
9710 }
9711
9712 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9713                                      struct btrfs_root *root,
9714                                      struct inode *dir,
9715                                      struct dentry *dentry)
9716 {
9717         int ret;
9718         struct inode *inode;
9719         u64 objectid;
9720         u64 index;
9721
9722         ret = btrfs_find_free_ino(root, &objectid);
9723         if (ret)
9724                 return ret;
9725
9726         inode = btrfs_new_inode(trans, root, dir,
9727                                 dentry->d_name.name,
9728                                 dentry->d_name.len,
9729                                 btrfs_ino(BTRFS_I(dir)),
9730                                 objectid,
9731                                 S_IFCHR | WHITEOUT_MODE,
9732                                 &index);
9733
9734         if (IS_ERR(inode)) {
9735                 ret = PTR_ERR(inode);
9736                 return ret;
9737         }
9738
9739         inode->i_op = &btrfs_special_inode_operations;
9740         init_special_inode(inode, inode->i_mode,
9741                 WHITEOUT_DEV);
9742
9743         ret = btrfs_init_inode_security(trans, inode, dir,
9744                                 &dentry->d_name);
9745         if (ret)
9746                 goto out;
9747
9748         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9749                                 BTRFS_I(inode), 0, index);
9750         if (ret)
9751                 goto out;
9752
9753         ret = btrfs_update_inode(trans, root, inode);
9754 out:
9755         unlock_new_inode(inode);
9756         if (ret)
9757                 inode_dec_link_count(inode);
9758         iput(inode);
9759
9760         return ret;
9761 }
9762
9763 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9764                            struct inode *new_dir, struct dentry *new_dentry,
9765                            unsigned int flags)
9766 {
9767         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9768         struct btrfs_trans_handle *trans;
9769         unsigned int trans_num_items;
9770         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9771         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9772         struct inode *new_inode = d_inode(new_dentry);
9773         struct inode *old_inode = d_inode(old_dentry);
9774         u64 index = 0;
9775         u64 root_objectid;
9776         int ret;
9777         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9778         bool log_pinned = false;
9779
9780         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9781                 return -EPERM;
9782
9783         /* we only allow rename subvolume link between subvolumes */
9784         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9785                 return -EXDEV;
9786
9787         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9788             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9789                 return -ENOTEMPTY;
9790
9791         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9792             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9793                 return -ENOTEMPTY;
9794
9795
9796         /* check for collisions, even if the  name isn't there */
9797         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9798                              new_dentry->d_name.name,
9799                              new_dentry->d_name.len);
9800
9801         if (ret) {
9802                 if (ret == -EEXIST) {
9803                         /* we shouldn't get
9804                          * eexist without a new_inode */
9805                         if (WARN_ON(!new_inode)) {
9806                                 return ret;
9807                         }
9808                 } else {
9809                         /* maybe -EOVERFLOW */
9810                         return ret;
9811                 }
9812         }
9813         ret = 0;
9814
9815         /*
9816          * we're using rename to replace one file with another.  Start IO on it
9817          * now so  we don't add too much work to the end of the transaction
9818          */
9819         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9820                 filemap_flush(old_inode->i_mapping);
9821
9822         /* close the racy window with snapshot create/destroy ioctl */
9823         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9824                 down_read(&fs_info->subvol_sem);
9825         /*
9826          * We want to reserve the absolute worst case amount of items.  So if
9827          * both inodes are subvols and we need to unlink them then that would
9828          * require 4 item modifications, but if they are both normal inodes it
9829          * would require 5 item modifications, so we'll assume they are normal
9830          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9831          * should cover the worst case number of items we'll modify.
9832          * If our rename has the whiteout flag, we need more 5 units for the
9833          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9834          * when selinux is enabled).
9835          */
9836         trans_num_items = 11;
9837         if (flags & RENAME_WHITEOUT)
9838                 trans_num_items += 5;
9839         trans = btrfs_start_transaction(root, trans_num_items);
9840         if (IS_ERR(trans)) {
9841                 ret = PTR_ERR(trans);
9842                 goto out_notrans;
9843         }
9844
9845         if (dest != root)
9846                 btrfs_record_root_in_trans(trans, dest);
9847
9848         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9849         if (ret)
9850                 goto out_fail;
9851
9852         BTRFS_I(old_inode)->dir_index = 0ULL;
9853         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9854                 /* force full log commit if subvolume involved. */
9855                 btrfs_set_log_full_commit(fs_info, trans);
9856         } else {
9857                 btrfs_pin_log_trans(root);
9858                 log_pinned = true;
9859                 ret = btrfs_insert_inode_ref(trans, dest,
9860                                              new_dentry->d_name.name,
9861                                              new_dentry->d_name.len,
9862                                              old_ino,
9863                                              btrfs_ino(BTRFS_I(new_dir)), index);
9864                 if (ret)
9865                         goto out_fail;
9866         }
9867
9868         inode_inc_iversion(old_dir);
9869         inode_inc_iversion(new_dir);
9870         inode_inc_iversion(old_inode);
9871         old_dir->i_ctime = old_dir->i_mtime =
9872         new_dir->i_ctime = new_dir->i_mtime =
9873         old_inode->i_ctime = current_time(old_dir);
9874
9875         if (old_dentry->d_parent != new_dentry->d_parent)
9876                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9877                                 BTRFS_I(old_inode), 1);
9878
9879         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9880                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9881                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9882                                         old_dentry->d_name.name,
9883                                         old_dentry->d_name.len);
9884         } else {
9885                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9886                                         BTRFS_I(d_inode(old_dentry)),
9887                                         old_dentry->d_name.name,
9888                                         old_dentry->d_name.len);
9889                 if (!ret)
9890                         ret = btrfs_update_inode(trans, root, old_inode);
9891         }
9892         if (ret) {
9893                 btrfs_abort_transaction(trans, ret);
9894                 goto out_fail;
9895         }
9896
9897         if (new_inode) {
9898                 inode_inc_iversion(new_inode);
9899                 new_inode->i_ctime = current_time(new_inode);
9900                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9901                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9902                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9903                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9904                                                 root_objectid,
9905                                                 new_dentry->d_name.name,
9906                                                 new_dentry->d_name.len);
9907                         BUG_ON(new_inode->i_nlink == 0);
9908                 } else {
9909                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9910                                                  BTRFS_I(d_inode(new_dentry)),
9911                                                  new_dentry->d_name.name,
9912                                                  new_dentry->d_name.len);
9913                 }
9914                 if (!ret && new_inode->i_nlink == 0)
9915                         ret = btrfs_orphan_add(trans,
9916                                         BTRFS_I(d_inode(new_dentry)));
9917                 if (ret) {
9918                         btrfs_abort_transaction(trans, ret);
9919                         goto out_fail;
9920                 }
9921         }
9922
9923         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9924                              new_dentry->d_name.name,
9925                              new_dentry->d_name.len, 0, index);
9926         if (ret) {
9927                 btrfs_abort_transaction(trans, ret);
9928                 goto out_fail;
9929         }
9930
9931         if (old_inode->i_nlink == 1)
9932                 BTRFS_I(old_inode)->dir_index = index;
9933
9934         if (log_pinned) {
9935                 struct dentry *parent = new_dentry->d_parent;
9936
9937                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9938                                 parent);
9939                 btrfs_end_log_trans(root);
9940                 log_pinned = false;
9941         }
9942
9943         if (flags & RENAME_WHITEOUT) {
9944                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9945                                                 old_dentry);
9946
9947                 if (ret) {
9948                         btrfs_abort_transaction(trans, ret);
9949                         goto out_fail;
9950                 }
9951         }
9952 out_fail:
9953         /*
9954          * If we have pinned the log and an error happened, we unpin tasks
9955          * trying to sync the log and force them to fallback to a transaction
9956          * commit if the log currently contains any of the inodes involved in
9957          * this rename operation (to ensure we do not persist a log with an
9958          * inconsistent state for any of these inodes or leading to any
9959          * inconsistencies when replayed). If the transaction was aborted, the
9960          * abortion reason is propagated to userspace when attempting to commit
9961          * the transaction. If the log does not contain any of these inodes, we
9962          * allow the tasks to sync it.
9963          */
9964         if (ret && log_pinned) {
9965                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9966                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9967                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9968                     (new_inode &&
9969                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9970                         btrfs_set_log_full_commit(fs_info, trans);
9971
9972                 btrfs_end_log_trans(root);
9973                 log_pinned = false;
9974         }
9975         btrfs_end_transaction(trans);
9976 out_notrans:
9977         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9978                 up_read(&fs_info->subvol_sem);
9979
9980         return ret;
9981 }
9982
9983 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9984                          struct inode *new_dir, struct dentry *new_dentry,
9985                          unsigned int flags)
9986 {
9987         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9988                 return -EINVAL;
9989
9990         if (flags & RENAME_EXCHANGE)
9991                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9992                                           new_dentry);
9993
9994         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9995 }
9996
9997 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9998 {
9999         struct btrfs_delalloc_work *delalloc_work;
10000         struct inode *inode;
10001
10002         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10003                                      work);
10004         inode = delalloc_work->inode;
10005         filemap_flush(inode->i_mapping);
10006         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10007                                 &BTRFS_I(inode)->runtime_flags))
10008                 filemap_flush(inode->i_mapping);
10009
10010         if (delalloc_work->delay_iput)
10011                 btrfs_add_delayed_iput(inode);
10012         else
10013                 iput(inode);
10014         complete(&delalloc_work->completion);
10015 }
10016
10017 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10018                                                     int delay_iput)
10019 {
10020         struct btrfs_delalloc_work *work;
10021
10022         work = kmalloc(sizeof(*work), GFP_NOFS);
10023         if (!work)
10024                 return NULL;
10025
10026         init_completion(&work->completion);
10027         INIT_LIST_HEAD(&work->list);
10028         work->inode = inode;
10029         work->delay_iput = delay_iput;
10030         WARN_ON_ONCE(!inode);
10031         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10032                         btrfs_run_delalloc_work, NULL, NULL);
10033
10034         return work;
10035 }
10036
10037 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10038 {
10039         wait_for_completion(&work->completion);
10040         kfree(work);
10041 }
10042
10043 /*
10044  * some fairly slow code that needs optimization. This walks the list
10045  * of all the inodes with pending delalloc and forces them to disk.
10046  */
10047 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10048                                    int nr)
10049 {
10050         struct btrfs_inode *binode;
10051         struct inode *inode;
10052         struct btrfs_delalloc_work *work, *next;
10053         struct list_head works;
10054         struct list_head splice;
10055         int ret = 0;
10056
10057         INIT_LIST_HEAD(&works);
10058         INIT_LIST_HEAD(&splice);
10059
10060         mutex_lock(&root->delalloc_mutex);
10061         spin_lock(&root->delalloc_lock);
10062         list_splice_init(&root->delalloc_inodes, &splice);
10063         while (!list_empty(&splice)) {
10064                 binode = list_entry(splice.next, struct btrfs_inode,
10065                                     delalloc_inodes);
10066
10067                 list_move_tail(&binode->delalloc_inodes,
10068                                &root->delalloc_inodes);
10069                 inode = igrab(&binode->vfs_inode);
10070                 if (!inode) {
10071                         cond_resched_lock(&root->delalloc_lock);
10072                         continue;
10073                 }
10074                 spin_unlock(&root->delalloc_lock);
10075
10076                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10077                 if (!work) {
10078                         if (delay_iput)
10079                                 btrfs_add_delayed_iput(inode);
10080                         else
10081                                 iput(inode);
10082                         ret = -ENOMEM;
10083                         goto out;
10084                 }
10085                 list_add_tail(&work->list, &works);
10086                 btrfs_queue_work(root->fs_info->flush_workers,
10087                                  &work->work);
10088                 ret++;
10089                 if (nr != -1 && ret >= nr)
10090                         goto out;
10091                 cond_resched();
10092                 spin_lock(&root->delalloc_lock);
10093         }
10094         spin_unlock(&root->delalloc_lock);
10095
10096 out:
10097         list_for_each_entry_safe(work, next, &works, list) {
10098                 list_del_init(&work->list);
10099                 btrfs_wait_and_free_delalloc_work(work);
10100         }
10101
10102         if (!list_empty_careful(&splice)) {
10103                 spin_lock(&root->delalloc_lock);
10104                 list_splice_tail(&splice, &root->delalloc_inodes);
10105                 spin_unlock(&root->delalloc_lock);
10106         }
10107         mutex_unlock(&root->delalloc_mutex);
10108         return ret;
10109 }
10110
10111 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10112 {
10113         struct btrfs_fs_info *fs_info = root->fs_info;
10114         int ret;
10115
10116         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10117                 return -EROFS;
10118
10119         ret = __start_delalloc_inodes(root, delay_iput, -1);
10120         if (ret > 0)
10121                 ret = 0;
10122         return ret;
10123 }
10124
10125 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10126                                int nr)
10127 {
10128         struct btrfs_root *root;
10129         struct list_head splice;
10130         int ret;
10131
10132         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10133                 return -EROFS;
10134
10135         INIT_LIST_HEAD(&splice);
10136
10137         mutex_lock(&fs_info->delalloc_root_mutex);
10138         spin_lock(&fs_info->delalloc_root_lock);
10139         list_splice_init(&fs_info->delalloc_roots, &splice);
10140         while (!list_empty(&splice) && nr) {
10141                 root = list_first_entry(&splice, struct btrfs_root,
10142                                         delalloc_root);
10143                 root = btrfs_grab_fs_root(root);
10144                 BUG_ON(!root);
10145                 list_move_tail(&root->delalloc_root,
10146                                &fs_info->delalloc_roots);
10147                 spin_unlock(&fs_info->delalloc_root_lock);
10148
10149                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10150                 btrfs_put_fs_root(root);
10151                 if (ret < 0)
10152                         goto out;
10153
10154                 if (nr != -1) {
10155                         nr -= ret;
10156                         WARN_ON(nr < 0);
10157                 }
10158                 spin_lock(&fs_info->delalloc_root_lock);
10159         }
10160         spin_unlock(&fs_info->delalloc_root_lock);
10161
10162         ret = 0;
10163 out:
10164         if (!list_empty_careful(&splice)) {
10165                 spin_lock(&fs_info->delalloc_root_lock);
10166                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10167                 spin_unlock(&fs_info->delalloc_root_lock);
10168         }
10169         mutex_unlock(&fs_info->delalloc_root_mutex);
10170         return ret;
10171 }
10172
10173 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10174                          const char *symname)
10175 {
10176         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10177         struct btrfs_trans_handle *trans;
10178         struct btrfs_root *root = BTRFS_I(dir)->root;
10179         struct btrfs_path *path;
10180         struct btrfs_key key;
10181         struct inode *inode = NULL;
10182         int err;
10183         int drop_inode = 0;
10184         u64 objectid;
10185         u64 index = 0;
10186         int name_len;
10187         int datasize;
10188         unsigned long ptr;
10189         struct btrfs_file_extent_item *ei;
10190         struct extent_buffer *leaf;
10191
10192         name_len = strlen(symname);
10193         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10194                 return -ENAMETOOLONG;
10195
10196         /*
10197          * 2 items for inode item and ref
10198          * 2 items for dir items
10199          * 1 item for updating parent inode item
10200          * 1 item for the inline extent item
10201          * 1 item for xattr if selinux is on
10202          */
10203         trans = btrfs_start_transaction(root, 7);
10204         if (IS_ERR(trans))
10205                 return PTR_ERR(trans);
10206
10207         err = btrfs_find_free_ino(root, &objectid);
10208         if (err)
10209                 goto out_unlock;
10210
10211         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10212                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10213                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10214         if (IS_ERR(inode)) {
10215                 err = PTR_ERR(inode);
10216                 goto out_unlock;
10217         }
10218
10219         /*
10220         * If the active LSM wants to access the inode during
10221         * d_instantiate it needs these. Smack checks to see
10222         * if the filesystem supports xattrs by looking at the
10223         * ops vector.
10224         */
10225         inode->i_fop = &btrfs_file_operations;
10226         inode->i_op = &btrfs_file_inode_operations;
10227         inode->i_mapping->a_ops = &btrfs_aops;
10228         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10229
10230         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10231         if (err)
10232                 goto out_unlock_inode;
10233
10234         path = btrfs_alloc_path();
10235         if (!path) {
10236                 err = -ENOMEM;
10237                 goto out_unlock_inode;
10238         }
10239         key.objectid = btrfs_ino(BTRFS_I(inode));
10240         key.offset = 0;
10241         key.type = BTRFS_EXTENT_DATA_KEY;
10242         datasize = btrfs_file_extent_calc_inline_size(name_len);
10243         err = btrfs_insert_empty_item(trans, root, path, &key,
10244                                       datasize);
10245         if (err) {
10246                 btrfs_free_path(path);
10247                 goto out_unlock_inode;
10248         }
10249         leaf = path->nodes[0];
10250         ei = btrfs_item_ptr(leaf, path->slots[0],
10251                             struct btrfs_file_extent_item);
10252         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10253         btrfs_set_file_extent_type(leaf, ei,
10254                                    BTRFS_FILE_EXTENT_INLINE);
10255         btrfs_set_file_extent_encryption(leaf, ei, 0);
10256         btrfs_set_file_extent_compression(leaf, ei, 0);
10257         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10258         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10259
10260         ptr = btrfs_file_extent_inline_start(ei);
10261         write_extent_buffer(leaf, symname, ptr, name_len);
10262         btrfs_mark_buffer_dirty(leaf);
10263         btrfs_free_path(path);
10264
10265         inode->i_op = &btrfs_symlink_inode_operations;
10266         inode_nohighmem(inode);
10267         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10268         inode_set_bytes(inode, name_len);
10269         btrfs_i_size_write(BTRFS_I(inode), name_len);
10270         err = btrfs_update_inode(trans, root, inode);
10271         /*
10272          * Last step, add directory indexes for our symlink inode. This is the
10273          * last step to avoid extra cleanup of these indexes if an error happens
10274          * elsewhere above.
10275          */
10276         if (!err)
10277                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10278                                 BTRFS_I(inode), 0, index);
10279         if (err) {
10280                 drop_inode = 1;
10281                 goto out_unlock_inode;
10282         }
10283
10284         unlock_new_inode(inode);
10285         d_instantiate(dentry, inode);
10286
10287 out_unlock:
10288         btrfs_end_transaction(trans);
10289         if (drop_inode) {
10290                 inode_dec_link_count(inode);
10291                 iput(inode);
10292         }
10293         btrfs_btree_balance_dirty(fs_info);
10294         return err;
10295
10296 out_unlock_inode:
10297         drop_inode = 1;
10298         unlock_new_inode(inode);
10299         goto out_unlock;
10300 }
10301
10302 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10303                                        u64 start, u64 num_bytes, u64 min_size,
10304                                        loff_t actual_len, u64 *alloc_hint,
10305                                        struct btrfs_trans_handle *trans)
10306 {
10307         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10308         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10309         struct extent_map *em;
10310         struct btrfs_root *root = BTRFS_I(inode)->root;
10311         struct btrfs_key ins;
10312         u64 cur_offset = start;
10313         u64 i_size;
10314         u64 cur_bytes;
10315         u64 last_alloc = (u64)-1;
10316         int ret = 0;
10317         bool own_trans = true;
10318         u64 end = start + num_bytes - 1;
10319
10320         if (trans)
10321                 own_trans = false;
10322         while (num_bytes > 0) {
10323                 if (own_trans) {
10324                         trans = btrfs_start_transaction(root, 3);
10325                         if (IS_ERR(trans)) {
10326                                 ret = PTR_ERR(trans);
10327                                 break;
10328                         }
10329                 }
10330
10331                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10332                 cur_bytes = max(cur_bytes, min_size);
10333                 /*
10334                  * If we are severely fragmented we could end up with really
10335                  * small allocations, so if the allocator is returning small
10336                  * chunks lets make its job easier by only searching for those
10337                  * sized chunks.
10338                  */
10339                 cur_bytes = min(cur_bytes, last_alloc);
10340                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10341                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10342                 if (ret) {
10343                         if (own_trans)
10344                                 btrfs_end_transaction(trans);
10345                         break;
10346                 }
10347                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10348
10349                 last_alloc = ins.offset;
10350                 ret = insert_reserved_file_extent(trans, inode,
10351                                                   cur_offset, ins.objectid,
10352                                                   ins.offset, ins.offset,
10353                                                   ins.offset, 0, 0, 0,
10354                                                   BTRFS_FILE_EXTENT_PREALLOC);
10355                 if (ret) {
10356                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10357                                                    ins.offset, 0);
10358                         btrfs_abort_transaction(trans, ret);
10359                         if (own_trans)
10360                                 btrfs_end_transaction(trans);
10361                         break;
10362                 }
10363
10364                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10365                                         cur_offset + ins.offset -1, 0);
10366
10367                 em = alloc_extent_map();
10368                 if (!em) {
10369                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10370                                 &BTRFS_I(inode)->runtime_flags);
10371                         goto next;
10372                 }
10373
10374                 em->start = cur_offset;
10375                 em->orig_start = cur_offset;
10376                 em->len = ins.offset;
10377                 em->block_start = ins.objectid;
10378                 em->block_len = ins.offset;
10379                 em->orig_block_len = ins.offset;
10380                 em->ram_bytes = ins.offset;
10381                 em->bdev = fs_info->fs_devices->latest_bdev;
10382                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10383                 em->generation = trans->transid;
10384
10385                 while (1) {
10386                         write_lock(&em_tree->lock);
10387                         ret = add_extent_mapping(em_tree, em, 1);
10388                         write_unlock(&em_tree->lock);
10389                         if (ret != -EEXIST)
10390                                 break;
10391                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10392                                                 cur_offset + ins.offset - 1,
10393                                                 0);
10394                 }
10395                 free_extent_map(em);
10396 next:
10397                 num_bytes -= ins.offset;
10398                 cur_offset += ins.offset;
10399                 *alloc_hint = ins.objectid + ins.offset;
10400
10401                 inode_inc_iversion(inode);
10402                 inode->i_ctime = current_time(inode);
10403                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10404                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10405                     (actual_len > inode->i_size) &&
10406                     (cur_offset > inode->i_size)) {
10407                         if (cur_offset > actual_len)
10408                                 i_size = actual_len;
10409                         else
10410                                 i_size = cur_offset;
10411                         i_size_write(inode, i_size);
10412                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10413                 }
10414
10415                 ret = btrfs_update_inode(trans, root, inode);
10416
10417                 if (ret) {
10418                         btrfs_abort_transaction(trans, ret);
10419                         if (own_trans)
10420                                 btrfs_end_transaction(trans);
10421                         break;
10422                 }
10423
10424                 if (own_trans)
10425                         btrfs_end_transaction(trans);
10426         }
10427         if (cur_offset < end)
10428                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10429                         end - cur_offset + 1);
10430         return ret;
10431 }
10432
10433 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10434                               u64 start, u64 num_bytes, u64 min_size,
10435                               loff_t actual_len, u64 *alloc_hint)
10436 {
10437         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10438                                            min_size, actual_len, alloc_hint,
10439                                            NULL);
10440 }
10441
10442 int btrfs_prealloc_file_range_trans(struct inode *inode,
10443                                     struct btrfs_trans_handle *trans, int mode,
10444                                     u64 start, u64 num_bytes, u64 min_size,
10445                                     loff_t actual_len, u64 *alloc_hint)
10446 {
10447         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10448                                            min_size, actual_len, alloc_hint, trans);
10449 }
10450
10451 static int btrfs_set_page_dirty(struct page *page)
10452 {
10453         return __set_page_dirty_nobuffers(page);
10454 }
10455
10456 static int btrfs_permission(struct inode *inode, int mask)
10457 {
10458         struct btrfs_root *root = BTRFS_I(inode)->root;
10459         umode_t mode = inode->i_mode;
10460
10461         if (mask & MAY_WRITE &&
10462             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10463                 if (btrfs_root_readonly(root))
10464                         return -EROFS;
10465                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10466                         return -EACCES;
10467         }
10468         return generic_permission(inode, mask);
10469 }
10470
10471 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10472 {
10473         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10474         struct btrfs_trans_handle *trans;
10475         struct btrfs_root *root = BTRFS_I(dir)->root;
10476         struct inode *inode = NULL;
10477         u64 objectid;
10478         u64 index;
10479         int ret = 0;
10480
10481         /*
10482          * 5 units required for adding orphan entry
10483          */
10484         trans = btrfs_start_transaction(root, 5);
10485         if (IS_ERR(trans))
10486                 return PTR_ERR(trans);
10487
10488         ret = btrfs_find_free_ino(root, &objectid);
10489         if (ret)
10490                 goto out;
10491
10492         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10493                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10494         if (IS_ERR(inode)) {
10495                 ret = PTR_ERR(inode);
10496                 inode = NULL;
10497                 goto out;
10498         }
10499
10500         inode->i_fop = &btrfs_file_operations;
10501         inode->i_op = &btrfs_file_inode_operations;
10502
10503         inode->i_mapping->a_ops = &btrfs_aops;
10504         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10505
10506         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10507         if (ret)
10508                 goto out_inode;
10509
10510         ret = btrfs_update_inode(trans, root, inode);
10511         if (ret)
10512                 goto out_inode;
10513         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10514         if (ret)
10515                 goto out_inode;
10516
10517         /*
10518          * We set number of links to 0 in btrfs_new_inode(), and here we set
10519          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10520          * through:
10521          *
10522          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10523          */
10524         set_nlink(inode, 1);
10525         unlock_new_inode(inode);
10526         d_tmpfile(dentry, inode);
10527         mark_inode_dirty(inode);
10528
10529 out:
10530         btrfs_end_transaction(trans);
10531         if (ret)
10532                 iput(inode);
10533         btrfs_btree_balance_dirty(fs_info);
10534         return ret;
10535
10536 out_inode:
10537         unlock_new_inode(inode);
10538         goto out;
10539
10540 }
10541
10542 __attribute__((const))
10543 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10544 {
10545         return -EAGAIN;
10546 }
10547
10548 static struct btrfs_fs_info *iotree_fs_info(void *private_data)
10549 {
10550         struct inode *inode = private_data;
10551         return btrfs_sb(inode->i_sb);
10552 }
10553
10554 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10555                                         u64 start, u64 end)
10556 {
10557         struct inode *inode = private_data;
10558         u64 isize;
10559
10560         isize = i_size_read(inode);
10561         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10562                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10563                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10564                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10565         }
10566 }
10567
10568 void btrfs_set_range_writeback(void *private_data, u64 start, u64 end)
10569 {
10570         struct inode *inode = private_data;
10571         unsigned long index = start >> PAGE_SHIFT;
10572         unsigned long end_index = end >> PAGE_SHIFT;
10573         struct page *page;
10574
10575         while (index <= end_index) {
10576                 page = find_get_page(inode->i_mapping, index);
10577                 ASSERT(page); /* Pages should be in the extent_io_tree */
10578                 set_page_writeback(page);
10579                 put_page(page);
10580                 index++;
10581         }
10582 }
10583
10584 static const struct inode_operations btrfs_dir_inode_operations = {
10585         .getattr        = btrfs_getattr,
10586         .lookup         = btrfs_lookup,
10587         .create         = btrfs_create,
10588         .unlink         = btrfs_unlink,
10589         .link           = btrfs_link,
10590         .mkdir          = btrfs_mkdir,
10591         .rmdir          = btrfs_rmdir,
10592         .rename         = btrfs_rename2,
10593         .symlink        = btrfs_symlink,
10594         .setattr        = btrfs_setattr,
10595         .mknod          = btrfs_mknod,
10596         .listxattr      = btrfs_listxattr,
10597         .permission     = btrfs_permission,
10598         .get_acl        = btrfs_get_acl,
10599         .set_acl        = btrfs_set_acl,
10600         .update_time    = btrfs_update_time,
10601         .tmpfile        = btrfs_tmpfile,
10602 };
10603 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10604         .lookup         = btrfs_lookup,
10605         .permission     = btrfs_permission,
10606         .update_time    = btrfs_update_time,
10607 };
10608
10609 static const struct file_operations btrfs_dir_file_operations = {
10610         .llseek         = generic_file_llseek,
10611         .read           = generic_read_dir,
10612         .iterate_shared = btrfs_real_readdir,
10613         .open           = btrfs_opendir,
10614         .unlocked_ioctl = btrfs_ioctl,
10615 #ifdef CONFIG_COMPAT
10616         .compat_ioctl   = btrfs_compat_ioctl,
10617 #endif
10618         .release        = btrfs_release_file,
10619         .fsync          = btrfs_sync_file,
10620 };
10621
10622 static const struct extent_io_ops btrfs_extent_io_ops = {
10623         /* mandatory callbacks */
10624         .submit_bio_hook = btrfs_submit_bio_hook,
10625         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10626         .merge_bio_hook = btrfs_merge_bio_hook,
10627         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10628         .tree_fs_info = iotree_fs_info,
10629         .set_range_writeback = btrfs_set_range_writeback,
10630
10631         /* optional callbacks */
10632         .fill_delalloc = run_delalloc_range,
10633         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10634         .writepage_start_hook = btrfs_writepage_start_hook,
10635         .set_bit_hook = btrfs_set_bit_hook,
10636         .clear_bit_hook = btrfs_clear_bit_hook,
10637         .merge_extent_hook = btrfs_merge_extent_hook,
10638         .split_extent_hook = btrfs_split_extent_hook,
10639         .check_extent_io_range = btrfs_check_extent_io_range,
10640 };
10641
10642 /*
10643  * btrfs doesn't support the bmap operation because swapfiles
10644  * use bmap to make a mapping of extents in the file.  They assume
10645  * these extents won't change over the life of the file and they
10646  * use the bmap result to do IO directly to the drive.
10647  *
10648  * the btrfs bmap call would return logical addresses that aren't
10649  * suitable for IO and they also will change frequently as COW
10650  * operations happen.  So, swapfile + btrfs == corruption.
10651  *
10652  * For now we're avoiding this by dropping bmap.
10653  */
10654 static const struct address_space_operations btrfs_aops = {
10655         .readpage       = btrfs_readpage,
10656         .writepage      = btrfs_writepage,
10657         .writepages     = btrfs_writepages,
10658         .readpages      = btrfs_readpages,
10659         .direct_IO      = btrfs_direct_IO,
10660         .invalidatepage = btrfs_invalidatepage,
10661         .releasepage    = btrfs_releasepage,
10662         .set_page_dirty = btrfs_set_page_dirty,
10663         .error_remove_page = generic_error_remove_page,
10664 };
10665
10666 static const struct address_space_operations btrfs_symlink_aops = {
10667         .readpage       = btrfs_readpage,
10668         .writepage      = btrfs_writepage,
10669         .invalidatepage = btrfs_invalidatepage,
10670         .releasepage    = btrfs_releasepage,
10671 };
10672
10673 static const struct inode_operations btrfs_file_inode_operations = {
10674         .getattr        = btrfs_getattr,
10675         .setattr        = btrfs_setattr,
10676         .listxattr      = btrfs_listxattr,
10677         .permission     = btrfs_permission,
10678         .fiemap         = btrfs_fiemap,
10679         .get_acl        = btrfs_get_acl,
10680         .set_acl        = btrfs_set_acl,
10681         .update_time    = btrfs_update_time,
10682 };
10683 static const struct inode_operations btrfs_special_inode_operations = {
10684         .getattr        = btrfs_getattr,
10685         .setattr        = btrfs_setattr,
10686         .permission     = btrfs_permission,
10687         .listxattr      = btrfs_listxattr,
10688         .get_acl        = btrfs_get_acl,
10689         .set_acl        = btrfs_set_acl,
10690         .update_time    = btrfs_update_time,
10691 };
10692 static const struct inode_operations btrfs_symlink_inode_operations = {
10693         .get_link       = page_get_link,
10694         .getattr        = btrfs_getattr,
10695         .setattr        = btrfs_setattr,
10696         .permission     = btrfs_permission,
10697         .listxattr      = btrfs_listxattr,
10698         .update_time    = btrfs_update_time,
10699 };
10700
10701 const struct dentry_operations btrfs_dentry_operations = {
10702         .d_delete       = btrfs_dentry_delete,
10703         .d_release      = btrfs_dentry_release,
10704 };