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/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 static const struct inode_operations btrfs_dir_inode_operations;
  70 static const struct inode_operations btrfs_symlink_inode_operations;
  71 static const struct inode_operations btrfs_dir_ro_inode_operations;
  72 static const struct inode_operations btrfs_special_inode_operations;
  73 static const struct inode_operations btrfs_file_inode_operations;
  74 static const struct address_space_operations btrfs_aops;
  75 static const struct address_space_operations btrfs_symlink_aops;
  76 static const struct file_operations btrfs_dir_file_operations;
  77 static struct extent_io_ops btrfs_extent_io_ops;
  78
  79 static struct kmem_cache *btrfs_inode_cachep;
  80 static struct kmem_cache *btrfs_delalloc_work_cachep;
  81 struct kmem_cache *btrfs_trans_handle_cachep;
  82 struct kmem_cache *btrfs_transaction_cachep;
  83 struct kmem_cache *btrfs_path_cachep;
  84 struct kmem_cache *btrfs_free_space_cachep;
  85
  86 #define S_SHIFT 12
  87 static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  88         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  89         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  90         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  91         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  92         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  93         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  94         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  95 };
  96
  97 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  98 static int btrfs_truncate(struct inode *inode);
  99 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 100 static noinline int cow_file_range(struct inode *inode,
 101                                    struct page *locked_page,
 102                                    u64 start, u64 end, int *page_started,
 103                                    unsigned long *nr_written, int unlock);
 104 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 105                                            u64 len, u64 orig_start,
 106                                            u64 block_start, u64 block_len,
 107                                            u64 orig_block_len, u64 ram_bytes,
 108                                            int type);
 109
 110 static int btrfs_dirty_inode(struct inode *inode);
 111
 112 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 113 void btrfs_test_inode_set_ops(struct inode *inode)
 114 {
 115         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 116 }
 117 #endif
 118
 119 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 120                                      struct inode *inode,  struct inode *dir,
 121                                      const struct qstr *qstr)
 122 {
 123         int err;
 124
 125         err = btrfs_init_acl(trans, inode, dir);
 126         if (!err)
 127                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 128         return err;
 129 }
 130
 131 /*
 132  * this does all the hard work for inserting an inline extent into
 133  * the btree.  The caller should have done a btrfs_drop_extents so that
 134  * no overlapping inline items exist in the btree
 135  */
 136 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 137                                 struct btrfs_path *path, int extent_inserted,
 138                                 struct btrfs_root *root, struct inode *inode,
 139                                 u64 start, size_t size, size_t compressed_size,
 140                                 int compress_type,
 141                                 struct page **compressed_pages)
 142 {
 143         struct extent_buffer *leaf;
 144         struct page *page = NULL;
 145         char *kaddr;
 146         unsigned long ptr;
 147         struct btrfs_file_extent_item *ei;
 148         int err = 0;
 149         int ret;
 150         size_t cur_size = size;
 151         unsigned long offset;
 152
 153         if (compressed_size && compressed_pages)
 154                 cur_size = compressed_size;
 155
 156         inode_add_bytes(inode, size);
 157
 158         if (!extent_inserted) {
 159                 struct btrfs_key key;
 160                 size_t datasize;
 161
 162                 key.objectid = btrfs_ino(inode);
 163                 key.offset = start;
 164                 key.type = BTRFS_EXTENT_DATA_KEY;
 165
 166                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 167                 path->leave_spinning = 1;
 168                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 169                                               datasize);
 170                 if (ret) {
 171                         err = ret;
 172                         goto fail;
 173                 }
 174         }
 175         leaf = path->nodes[0];
 176         ei = btrfs_item_ptr(leaf, path->slots[0],
 177                             struct btrfs_file_extent_item);
 178         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 179         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 180         btrfs_set_file_extent_encryption(leaf, ei, 0);
 181         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 182         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 183         ptr = btrfs_file_extent_inline_start(ei);
 184
 185         if (compress_type != BTRFS_COMPRESS_NONE) {
 186                 struct page *cpage;
 187                 int i = 0;
 188                 while (compressed_size > 0) {
 189                         cpage = compressed_pages[i];
 190                         cur_size = min_t(unsigned long, compressed_size,
 191                                        PAGE_CACHE_SIZE);
 192
 193                         kaddr = kmap_atomic(cpage);
 194                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 195                         kunmap_atomic(kaddr);
 196
 197                         i++;
 198                         ptr += cur_size;
 199                         compressed_size -= cur_size;
 200                 }
 201                 btrfs_set_file_extent_compression(leaf, ei,
 202                                                   compress_type);
 203         } else {
 204                 page = find_get_page(inode->i_mapping,
 205                                      start >> PAGE_CACHE_SHIFT);
 206                 btrfs_set_file_extent_compression(leaf, ei, 0);
 207                 kaddr = kmap_atomic(page);
 208                 offset = start & (PAGE_CACHE_SIZE - 1);
 209                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 210                 kunmap_atomic(kaddr);
 211                 page_cache_release(page);
 212         }
 213         btrfs_mark_buffer_dirty(leaf);
 214         btrfs_release_path(path);
 215
 216         /*
 217          * we're an inline extent, so nobody can
 218          * extend the file past i_size without locking
 219          * a page we already have locked.
 220          *
 221          * We must do any isize and inode updates
 222          * before we unlock the pages.  Otherwise we
 223          * could end up racing with unlink.
 224          */
 225         BTRFS_I(inode)->disk_i_size = inode->i_size;
 226         ret = btrfs_update_inode(trans, root, inode);
 227
 228         return ret;
 229 fail:
 230         return err;
 231 }
 232
 233
 234 /*
 235  * conditionally insert an inline extent into the file.  This
 236  * does the checks required to make sure the data is small enough
 237  * to fit as an inline extent.
 238  */
 239 static noinline int cow_file_range_inline(struct btrfs_root *root,
 240                                           struct inode *inode, u64 start,
 241                                           u64 end, size_t compressed_size,
 242                                           int compress_type,
 243                                           struct page **compressed_pages)
 244 {
 245         struct btrfs_trans_handle *trans;
 246         u64 isize = i_size_read(inode);
 247         u64 actual_end = min(end + 1, isize);
 248         u64 inline_len = actual_end - start;
 249         u64 aligned_end = ALIGN(end, root->sectorsize);
 250         u64 data_len = inline_len;
 251         int ret;
 252         struct btrfs_path *path;
 253         int extent_inserted = 0;
 254         u32 extent_item_size;
 255
 256         if (compressed_size)
 257                 data_len = compressed_size;
 258
 259         if (start > 0 ||
 260             actual_end > PAGE_CACHE_SIZE ||
 261             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 262             (!compressed_size &&
 263             (actual_end & (root->sectorsize - 1)) == 0) ||
 264             end + 1 < isize ||
 265             data_len > root->fs_info->max_inline) {
 266                 return 1;
 267         }
 268
 269         path = btrfs_alloc_path();
 270         if (!path)
 271                 return -ENOMEM;
 272
 273         trans = btrfs_join_transaction(root);
 274         if (IS_ERR(trans)) {
 275                 btrfs_free_path(path);
 276                 return PTR_ERR(trans);
 277         }
 278         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 279
 280         if (compressed_size && compressed_pages)
 281                 extent_item_size = btrfs_file_extent_calc_inline_size(
 282                    compressed_size);
 283         else
 284                 extent_item_size = btrfs_file_extent_calc_inline_size(
 285                     inline_len);
 286
 287         ret = __btrfs_drop_extents(trans, root, inode, path,
 288                                    start, aligned_end, NULL,
 289                                    1, 1, extent_item_size, &extent_inserted);
 290         if (ret) {
 291                 btrfs_abort_transaction(trans, root, ret);
 292                 goto out;
 293         }
 294
 295         if (isize > actual_end)
 296                 inline_len = min_t(u64, isize, actual_end);
 297         ret = insert_inline_extent(trans, path, extent_inserted,
 298                                    root, inode, start,
 299                                    inline_len, compressed_size,
 300                                    compress_type, compressed_pages);
 301         if (ret && ret != -ENOSPC) {
 302                 btrfs_abort_transaction(trans, root, ret);
 303                 goto out;
 304         } else if (ret == -ENOSPC) {
 305                 ret = 1;
 306                 goto out;
 307         }
 308
 309         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 310         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 311         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 312 out:
 313         /*
 314          * Don't forget to free the reserved space, as for inlined extent
 315          * it won't count as data extent, free them directly here.
 316          * And at reserve time, it's always aligned to page size, so
 317          * just free one page here.
 318          */
 319         btrfs_qgroup_free_data(inode, 0, PAGE_CACHE_SIZE);
 320         btrfs_free_path(path);
 321         btrfs_end_transaction(trans, root);
 322         return ret;
 323 }
 324
 325 struct async_extent {
 326         u64 start;
 327         u64 ram_size;
 328         u64 compressed_size;
 329         struct page **pages;
 330         unsigned long nr_pages;
 331         int compress_type;
 332         struct list_head list;
 333 };
 334
 335 struct async_cow {
 336         struct inode *inode;
 337         struct btrfs_root *root;
 338         struct page *locked_page;
 339         u64 start;
 340         u64 end;
 341         struct list_head extents;
 342         struct btrfs_work work;
 343 };
 344
 345 static noinline int add_async_extent(struct async_cow *cow,
 346                                      u64 start, u64 ram_size,
 347                                      u64 compressed_size,
 348                                      struct page **pages,
 349                                      unsigned long nr_pages,
 350                                      int compress_type)
 351 {
 352         struct async_extent *async_extent;
 353
 354         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 355         BUG_ON(!async_extent); /* -ENOMEM */
 356         async_extent->start = start;
 357         async_extent->ram_size = ram_size;
 358         async_extent->compressed_size = compressed_size;
 359         async_extent->pages = pages;
 360         async_extent->nr_pages = nr_pages;
 361         async_extent->compress_type = compress_type;
 362         list_add_tail(&async_extent->list, &cow->extents);
 363         return 0;
 364 }
 365
 366 static inline int inode_need_compress(struct inode *inode)
 367 {
 368         struct btrfs_root *root = BTRFS_I(inode)->root;
 369
 370         /* force compress */
 371         if (btrfs_test_opt(root, FORCE_COMPRESS))
 372                 return 1;
 373         /* bad compression ratios */
 374         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 375                 return 0;
 376         if (btrfs_test_opt(root, COMPRESS) ||
 377             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 378             BTRFS_I(inode)->force_compress)
 379                 return 1;
 380         return 0;
 381 }
 382
 383 /*
 384  * we create compressed extents in two phases.  The first
 385  * phase compresses a range of pages that have already been
 386  * locked (both pages and state bits are locked).
 387  *
 388  * This is done inside an ordered work queue, and the compression
 389  * is spread across many cpus.  The actual IO submission is step
 390  * two, and the ordered work queue takes care of making sure that
 391  * happens in the same order things were put onto the queue by
 392  * writepages and friends.
 393  *
 394  * If this code finds it can't get good compression, it puts an
 395  * entry onto the work queue to write the uncompressed bytes.  This
 396  * makes sure that both compressed inodes and uncompressed inodes
 397  * are written in the same order that the flusher thread sent them
 398  * down.
 399  */
 400 static noinline void compress_file_range(struct inode *inode,
 401                                         struct page *locked_page,
 402                                         u64 start, u64 end,
 403                                         struct async_cow *async_cow,
 404                                         int *num_added)
 405 {
 406         struct btrfs_root *root = BTRFS_I(inode)->root;
 407         u64 num_bytes;
 408         u64 blocksize = root->sectorsize;
 409         u64 actual_end;
 410         u64 isize = i_size_read(inode);
 411         int ret = 0;
 412         struct page **pages = NULL;
 413         unsigned long nr_pages;
 414         unsigned long nr_pages_ret = 0;
 415         unsigned long total_compressed = 0;
 416         unsigned long total_in = 0;
 417         unsigned long max_compressed = 128 * 1024;
 418         unsigned long max_uncompressed = 128 * 1024;
 419         int i;
 420         int will_compress;
 421         int compress_type = root->fs_info->compress_type;
 422         int redirty = 0;
 423
 424         /* if this is a small write inside eof, kick off a defrag */
 425         if ((end - start + 1) < 16 * 1024 &&
 426             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 427                 btrfs_add_inode_defrag(NULL, inode);
 428
 429         actual_end = min_t(u64, isize, end + 1);
 430 again:
 431         will_compress = 0;
 432         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
 433         nr_pages = min(nr_pages, (128 * 1024UL) / PAGE_CACHE_SIZE);
 434
 435         /*
 436          * we don't want to send crud past the end of i_size through
 437          * compression, that's just a waste of CPU time.  So, if the
 438          * end of the file is before the start of our current
 439          * requested range of bytes, we bail out to the uncompressed
 440          * cleanup code that can deal with all of this.
 441          *
 442          * It isn't really the fastest way to fix things, but this is a
 443          * very uncommon corner.
 444          */
 445         if (actual_end <= start)
 446                 goto cleanup_and_bail_uncompressed;
 447
 448         total_compressed = actual_end - start;
 449
 450         /*
 451          * skip compression for a small file range(<=blocksize) that
 452          * isn't an inline extent, since it dosen't save disk space at all.
 453          */
 454         if (total_compressed <= blocksize &&
 455            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 456                 goto cleanup_and_bail_uncompressed;
 457
 458         /* we want to make sure that amount of ram required to uncompress
 459          * an extent is reasonable, so we limit the total size in ram
 460          * of a compressed extent to 128k.  This is a crucial number
 461          * because it also controls how easily we can spread reads across
 462          * cpus for decompression.
 463          *
 464          * We also want to make sure the amount of IO required to do
 465          * a random read is reasonably small, so we limit the size of
 466          * a compressed extent to 128k.
 467          */
 468         total_compressed = min(total_compressed, max_uncompressed);
 469         num_bytes = ALIGN(end - start + 1, blocksize);
 470         num_bytes = max(blocksize,  num_bytes);
 471         total_in = 0;
 472         ret = 0;
 473
 474         /*
 475          * we do compression for mount -o compress and when the
 476          * inode has not been flagged as nocompress.  This flag can
 477          * change at any time if we discover bad compression ratios.
 478          */
 479         if (inode_need_compress(inode)) {
 480                 WARN_ON(pages);
 481                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 482                 if (!pages) {
 483                         /* just bail out to the uncompressed code */
 484                         goto cont;
 485                 }
 486
 487                 if (BTRFS_I(inode)->force_compress)
 488                         compress_type = BTRFS_I(inode)->force_compress;
 489
 490                 /*
 491                  * we need to call clear_page_dirty_for_io on each
 492                  * page in the range.  Otherwise applications with the file
 493                  * mmap'd can wander in and change the page contents while
 494                  * we are compressing them.
 495                  *
 496                  * If the compression fails for any reason, we set the pages
 497                  * dirty again later on.
 498                  */
 499                 extent_range_clear_dirty_for_io(inode, start, end);
 500                 redirty = 1;
 501                 ret = btrfs_compress_pages(compress_type,
 502                                            inode->i_mapping, start,
 503                                            total_compressed, pages,
 504                                            nr_pages, &nr_pages_ret,
 505                                            &total_in,
 506                                            &total_compressed,
 507                                            max_compressed);
 508
 509                 if (!ret) {
 510                         unsigned long offset = total_compressed &
 511                                 (PAGE_CACHE_SIZE - 1);
 512                         struct page *page = pages[nr_pages_ret - 1];
 513                         char *kaddr;
 514
 515                         /* zero the tail end of the last page, we might be
 516                          * sending it down to disk
 517                          */
 518                         if (offset) {
 519                                 kaddr = kmap_atomic(page);
 520                                 memset(kaddr + offset, 0,
 521                                        PAGE_CACHE_SIZE - offset);
 522                                 kunmap_atomic(kaddr);
 523                         }
 524                         will_compress = 1;
 525                 }
 526         }
 527 cont:
 528         if (start == 0) {
 529                 /* lets try to make an inline extent */
 530                 if (ret || total_in < (actual_end - start)) {
 531                         /* we didn't compress the entire range, try
 532                          * to make an uncompressed inline extent.
 533                          */
 534                         ret = cow_file_range_inline(root, inode, start, end,
 535                                                     0, 0, NULL);
 536                 } else {
 537                         /* try making a compressed inline extent */
 538                         ret = cow_file_range_inline(root, inode, start, end,
 539                                                     total_compressed,
 540                                                     compress_type, pages);
 541                 }
 542                 if (ret <= 0) {
 543                         unsigned long clear_flags = EXTENT_DELALLOC |
 544                                 EXTENT_DEFRAG;
 545                         unsigned long page_error_op;
 546
 547                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 548                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 549
 550                         /*
 551                          * inline extent creation worked or returned error,
 552                          * we don't need to create any more async work items.
 553                          * Unlock and free up our temp pages.
 554                          */
 555                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 556                                                      clear_flags, PAGE_UNLOCK |
 557                                                      PAGE_CLEAR_DIRTY |
 558                                                      PAGE_SET_WRITEBACK |
 559                                                      page_error_op |
 560                                                      PAGE_END_WRITEBACK);
 561                         goto free_pages_out;
 562                 }
 563         }
 564
 565         if (will_compress) {
 566                 /*
 567                  * we aren't doing an inline extent round the compressed size
 568                  * up to a block size boundary so the allocator does sane
 569                  * things
 570                  */
 571                 total_compressed = ALIGN(total_compressed, blocksize);
 572
 573                 /*
 574                  * one last check to make sure the compression is really a
 575                  * win, compare the page count read with the blocks on disk
 576                  */
 577                 total_in = ALIGN(total_in, PAGE_CACHE_SIZE);
 578                 if (total_compressed >= total_in) {
 579                         will_compress = 0;
 580                 } else {
 581                         num_bytes = total_in;
 582                 }
 583         }
 584         if (!will_compress && pages) {
 585                 /*
 586                  * the compression code ran but failed to make things smaller,
 587                  * free any pages it allocated and our page pointer array
 588                  */
 589                 for (i = 0; i < nr_pages_ret; i++) {
 590                         WARN_ON(pages[i]->mapping);
 591                         page_cache_release(pages[i]);
 592                 }
 593                 kfree(pages);
 594                 pages = NULL;
 595                 total_compressed = 0;
 596                 nr_pages_ret = 0;
 597
 598                 /* flag the file so we don't compress in the future */
 599                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 600                     !(BTRFS_I(inode)->force_compress)) {
 601                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 602                 }
 603         }
 604         if (will_compress) {
 605                 *num_added += 1;
 606
 607                 /* the async work queues will take care of doing actual
 608                  * allocation on disk for these compressed pages,
 609                  * and will submit them to the elevator.
 610                  */
 611                 add_async_extent(async_cow, start, num_bytes,
 612                                  total_compressed, pages, nr_pages_ret,
 613                                  compress_type);
 614
 615                 if (start + num_bytes < end) {
 616                         start += num_bytes;
 617                         pages = NULL;
 618                         cond_resched();
 619                         goto again;
 620                 }
 621         } else {
 622 cleanup_and_bail_uncompressed:
 623                 /*
 624                  * No compression, but we still need to write the pages in
 625                  * the file we've been given so far.  redirty the locked
 626                  * page if it corresponds to our extent and set things up
 627                  * for the async work queue to run cow_file_range to do
 628                  * the normal delalloc dance
 629                  */
 630                 if (page_offset(locked_page) >= start &&
 631                     page_offset(locked_page) <= end) {
 632                         __set_page_dirty_nobuffers(locked_page);
 633                         /* unlocked later on in the async handlers */
 634                 }
 635                 if (redirty)
 636                         extent_range_redirty_for_io(inode, start, end);
 637                 add_async_extent(async_cow, start, end - start + 1,
 638                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 639                 *num_added += 1;
 640         }
 641
 642         return;
 643
 644 free_pages_out:
 645         for (i = 0; i < nr_pages_ret; i++) {
 646                 WARN_ON(pages[i]->mapping);
 647                 page_cache_release(pages[i]);
 648         }
 649         kfree(pages);
 650 }
 651
 652 static void free_async_extent_pages(struct async_extent *async_extent)
 653 {
 654         int i;
 655
 656         if (!async_extent->pages)
 657                 return;
 658
 659         for (i = 0; i < async_extent->nr_pages; i++) {
 660                 WARN_ON(async_extent->pages[i]->mapping);
 661                 page_cache_release(async_extent->pages[i]);
 662         }
 663         kfree(async_extent->pages);
 664         async_extent->nr_pages = 0;
 665         async_extent->pages = NULL;
 666 }
 667
 668 /*
 669  * phase two of compressed writeback.  This is the ordered portion
 670  * of the code, which only gets called in the order the work was
 671  * queued.  We walk all the async extents created by compress_file_range
 672  * and send them down to the disk.
 673  */
 674 static noinline void submit_compressed_extents(struct inode *inode,
 675                                               struct async_cow *async_cow)
 676 {
 677         struct async_extent *async_extent;
 678         u64 alloc_hint = 0;
 679         struct btrfs_key ins;
 680         struct extent_map *em;
 681         struct btrfs_root *root = BTRFS_I(inode)->root;
 682         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 683         struct extent_io_tree *io_tree;
 684         int ret = 0;
 685
 686 again:
 687         while (!list_empty(&async_cow->extents)) {
 688                 async_extent = list_entry(async_cow->extents.next,
 689                                           struct async_extent, list);
 690                 list_del(&async_extent->list);
 691
 692                 io_tree = &BTRFS_I(inode)->io_tree;
 693
 694 retry:
 695                 /* did the compression code fall back to uncompressed IO? */
 696                 if (!async_extent->pages) {
 697                         int page_started = 0;
 698                         unsigned long nr_written = 0;
 699
 700                         lock_extent(io_tree, async_extent->start,
 701                                          async_extent->start +
 702                                          async_extent->ram_size - 1);
 703
 704                         /* allocate blocks */
 705                         ret = cow_file_range(inode, async_cow->locked_page,
 706                                              async_extent->start,
 707                                              async_extent->start +
 708                                              async_extent->ram_size - 1,
 709                                              &page_started, &nr_written, 0);
 710
 711                         /* JDM XXX */
 712
 713                         /*
 714                          * if page_started, cow_file_range inserted an
 715                          * inline extent and took care of all the unlocking
 716                          * and IO for us.  Otherwise, we need to submit
 717                          * all those pages down to the drive.
 718                          */
 719                         if (!page_started && !ret)
 720                                 extent_write_locked_range(io_tree,
 721                                                   inode, async_extent->start,
 722                                                   async_extent->start +
 723                                                   async_extent->ram_size - 1,
 724                                                   btrfs_get_extent,
 725                                                   WB_SYNC_ALL);
 726                         else if (ret)
 727                                 unlock_page(async_cow->locked_page);
 728                         kfree(async_extent);
 729                         cond_resched();
 730                         continue;
 731                 }
 732
 733                 lock_extent(io_tree, async_extent->start,
 734                             async_extent->start + async_extent->ram_size - 1);
 735
 736                 ret = btrfs_reserve_extent(root,
 737                                            async_extent->compressed_size,
 738                                            async_extent->compressed_size,
 739                                            0, alloc_hint, &ins, 1, 1);
 740                 if (ret) {
 741                         free_async_extent_pages(async_extent);
 742
 743                         if (ret == -ENOSPC) {
 744                                 unlock_extent(io_tree, async_extent->start,
 745                                               async_extent->start +
 746                                               async_extent->ram_size - 1);
 747
 748                                 /*
 749                                  * we need to redirty the pages if we decide to
 750                                  * fallback to uncompressed IO, otherwise we
 751                                  * will not submit these pages down to lower
 752                                  * layers.
 753                                  */
 754                                 extent_range_redirty_for_io(inode,
 755                                                 async_extent->start,
 756                                                 async_extent->start +
 757                                                 async_extent->ram_size - 1);
 758
 759                                 goto retry;
 760                         }
 761                         goto out_free;
 762                 }
 763                 /*
 764                  * here we're doing allocation and writeback of the
 765                  * compressed pages
 766                  */
 767                 btrfs_drop_extent_cache(inode, async_extent->start,
 768                                         async_extent->start +
 769                                         async_extent->ram_size - 1, 0);
 770
 771                 em = alloc_extent_map();
 772                 if (!em) {
 773                         ret = -ENOMEM;
 774                         goto out_free_reserve;
 775                 }
 776                 em->start = async_extent->start;
 777                 em->len = async_extent->ram_size;
 778                 em->orig_start = em->start;
 779                 em->mod_start = em->start;
 780                 em->mod_len = em->len;
 781
 782                 em->block_start = ins.objectid;
 783                 em->block_len = ins.offset;
 784                 em->orig_block_len = ins.offset;
 785                 em->ram_bytes = async_extent->ram_size;
 786                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 787                 em->compress_type = async_extent->compress_type;
 788                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 789                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 790                 em->generation = -1;
 791
 792                 while (1) {
 793                         write_lock(&em_tree->lock);
 794                         ret = add_extent_mapping(em_tree, em, 1);
 795                         write_unlock(&em_tree->lock);
 796                         if (ret != -EEXIST) {
 797                                 free_extent_map(em);
 798                                 break;
 799                         }
 800                         btrfs_drop_extent_cache(inode, async_extent->start,
 801                                                 async_extent->start +
 802                                                 async_extent->ram_size - 1, 0);
 803                 }
 804
 805                 if (ret)
 806                         goto out_free_reserve;
 807
 808                 ret = btrfs_add_ordered_extent_compress(inode,
 809                                                 async_extent->start,
 810                                                 ins.objectid,
 811                                                 async_extent->ram_size,
 812                                                 ins.offset,
 813                                                 BTRFS_ORDERED_COMPRESSED,
 814                                                 async_extent->compress_type);
 815                 if (ret) {
 816                         btrfs_drop_extent_cache(inode, async_extent->start,
 817                                                 async_extent->start +
 818                                                 async_extent->ram_size - 1, 0);
 819                         goto out_free_reserve;
 820                 }
 821
 822                 /*
 823                  * clear dirty, set writeback and unlock the pages.
 824                  */
 825                 extent_clear_unlock_delalloc(inode, async_extent->start,
 826                                 async_extent->start +
 827                                 async_extent->ram_size - 1,
 828                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 829                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 830                                 PAGE_SET_WRITEBACK);
 831                 ret = btrfs_submit_compressed_write(inode,
 832                                     async_extent->start,
 833                                     async_extent->ram_size,
 834                                     ins.objectid,
 835                                     ins.offset, async_extent->pages,
 836                                     async_extent->nr_pages);
 837                 if (ret) {
 838                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 839                         struct page *p = async_extent->pages[0];
 840                         const u64 start = async_extent->start;
 841                         const u64 end = start + async_extent->ram_size - 1;
 842
 843                         p->mapping = inode->i_mapping;
 844                         tree->ops->writepage_end_io_hook(p, start, end,
 845                                                          NULL, 0);
 846                         p->mapping = NULL;
 847                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 848                                                      PAGE_END_WRITEBACK |
 849                                                      PAGE_SET_ERROR);
 850                         free_async_extent_pages(async_extent);
 851                 }
 852                 alloc_hint = ins.objectid + ins.offset;
 853                 kfree(async_extent);
 854                 cond_resched();
 855         }
 856         return;
 857 out_free_reserve:
 858         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 859 out_free:
 860         extent_clear_unlock_delalloc(inode, async_extent->start,
 861                                      async_extent->start +
 862                                      async_extent->ram_size - 1,
 863                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 864                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 865                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 866                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 867                                      PAGE_SET_ERROR);
 868         free_async_extent_pages(async_extent);
 869         kfree(async_extent);
 870         goto again;
 871 }
 872
 873 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 874                                       u64 num_bytes)
 875 {
 876         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 877         struct extent_map *em;
 878         u64 alloc_hint = 0;
 879
 880         read_lock(&em_tree->lock);
 881         em = search_extent_mapping(em_tree, start, num_bytes);
 882         if (em) {
 883                 /*
 884                  * if block start isn't an actual block number then find the
 885                  * first block in this inode and use that as a hint.  If that
 886                  * block is also bogus then just don't worry about it.
 887                  */
 888                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 889                         free_extent_map(em);
 890                         em = search_extent_mapping(em_tree, 0, 0);
 891                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 892                                 alloc_hint = em->block_start;
 893                         if (em)
 894                                 free_extent_map(em);
 895                 } else {
 896                         alloc_hint = em->block_start;
 897                         free_extent_map(em);
 898                 }
 899         }
 900         read_unlock(&em_tree->lock);
 901
 902         return alloc_hint;
 903 }
 904
 905 /*
 906  * when extent_io.c finds a delayed allocation range in the file,
 907  * the call backs end up in this code.  The basic idea is to
 908  * allocate extents on disk for the range, and create ordered data structs
 909  * in ram to track those extents.
 910  *
 911  * locked_page is the page that writepage had locked already.  We use
 912  * it to make sure we don't do extra locks or unlocks.
 913  *
 914  * *page_started is set to one if we unlock locked_page and do everything
 915  * required to start IO on it.  It may be clean and already done with
 916  * IO when we return.
 917  */
 918 static noinline int cow_file_range(struct inode *inode,
 919                                    struct page *locked_page,
 920                                    u64 start, u64 end, int *page_started,
 921                                    unsigned long *nr_written,
 922                                    int unlock)
 923 {
 924         struct btrfs_root *root = BTRFS_I(inode)->root;
 925         u64 alloc_hint = 0;
 926         u64 num_bytes;
 927         unsigned long ram_size;
 928         u64 disk_num_bytes;
 929         u64 cur_alloc_size;
 930         u64 blocksize = root->sectorsize;
 931         struct btrfs_key ins;
 932         struct extent_map *em;
 933         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 934         int ret = 0;
 935
 936         if (btrfs_is_free_space_inode(inode)) {
 937                 WARN_ON_ONCE(1);
 938                 ret = -EINVAL;
 939                 goto out_unlock;
 940         }
 941
 942         num_bytes = ALIGN(end - start + 1, blocksize);
 943         num_bytes = max(blocksize,  num_bytes);
 944         disk_num_bytes = num_bytes;
 945
 946         /* if this is a small write inside eof, kick off defrag */
 947         if (num_bytes < 64 * 1024 &&
 948             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 949                 btrfs_add_inode_defrag(NULL, inode);
 950
 951         if (start == 0) {
 952                 /* lets try to make an inline extent */
 953                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 954                                             NULL);
 955                 if (ret == 0) {
 956                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 957                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 958                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 959                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 960                                      PAGE_END_WRITEBACK);
 961
 962                         *nr_written = *nr_written +
 963                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
 964                         *page_started = 1;
 965                         goto out;
 966                 } else if (ret < 0) {
 967                         goto out_unlock;
 968                 }
 969         }
 970
 971         BUG_ON(disk_num_bytes >
 972                btrfs_super_total_bytes(root->fs_info->super_copy));
 973
 974         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 975         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 976
 977         while (disk_num_bytes > 0) {
 978                 unsigned long op;
 979
 980                 cur_alloc_size = disk_num_bytes;
 981                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 982                                            root->sectorsize, 0, alloc_hint,
 983                                            &ins, 1, 1);
 984                 if (ret < 0)
 985                         goto out_unlock;
 986
 987                 em = alloc_extent_map();
 988                 if (!em) {
 989                         ret = -ENOMEM;
 990                         goto out_reserve;
 991                 }
 992                 em->start = start;
 993                 em->orig_start = em->start;
 994                 ram_size = ins.offset;
 995                 em->len = ins.offset;
 996                 em->mod_start = em->start;
 997                 em->mod_len = em->len;
 998
 999                 em->block_start = ins.objectid;
1000                 em->block_len = ins.offset;
1001                 em->orig_block_len = ins.offset;
1002                 em->ram_bytes = ram_size;
1003                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1004                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1005                 em->generation = -1;
1006
1007                 while (1) {
1008                         write_lock(&em_tree->lock);
1009                         ret = add_extent_mapping(em_tree, em, 1);
1010                         write_unlock(&em_tree->lock);
1011                         if (ret != -EEXIST) {
1012                                 free_extent_map(em);
1013                                 break;
1014                         }
1015                         btrfs_drop_extent_cache(inode, start,
1016                                                 start + ram_size - 1, 0);
1017                 }
1018                 if (ret)
1019                         goto out_reserve;
1020
1021                 cur_alloc_size = ins.offset;
1022                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1023                                                ram_size, cur_alloc_size, 0);
1024                 if (ret)
1025                         goto out_drop_extent_cache;
1026
1027                 if (root->root_key.objectid ==
1028                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1029                         ret = btrfs_reloc_clone_csums(inode, start,
1030                                                       cur_alloc_size);
1031                         if (ret)
1032                                 goto out_drop_extent_cache;
1033                 }
1034
1035                 if (disk_num_bytes < cur_alloc_size)
1036                         break;
1037
1038                 /* we're not doing compressed IO, don't unlock the first
1039                  * page (which the caller expects to stay locked), don't
1040                  * clear any dirty bits and don't set any writeback bits
1041                  *
1042                  * Do set the Private2 bit so we know this page was properly
1043                  * setup for writepage
1044                  */
1045                 op = unlock ? PAGE_UNLOCK : 0;
1046                 op |= PAGE_SET_PRIVATE2;
1047
1048                 extent_clear_unlock_delalloc(inode, start,
1049                                              start + ram_size - 1, locked_page,
1050                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1051                                              op);
1052                 disk_num_bytes -= cur_alloc_size;
1053                 num_bytes -= cur_alloc_size;
1054                 alloc_hint = ins.objectid + ins.offset;
1055                 start += cur_alloc_size;
1056         }
1057 out:
1058         return ret;
1059
1060 out_drop_extent_cache:
1061         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1062 out_reserve:
1063         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1064 out_unlock:
1065         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1066                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1067                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1068                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1069                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1070         goto out;
1071 }
1072
1073 /*
1074  * work queue call back to started compression on a file and pages
1075  */
1076 static noinline void async_cow_start(struct btrfs_work *work)
1077 {
1078         struct async_cow *async_cow;
1079         int num_added = 0;
1080         async_cow = container_of(work, struct async_cow, work);
1081
1082         compress_file_range(async_cow->inode, async_cow->locked_page,
1083                             async_cow->start, async_cow->end, async_cow,
1084                             &num_added);
1085         if (num_added == 0) {
1086                 btrfs_add_delayed_iput(async_cow->inode);
1087                 async_cow->inode = NULL;
1088         }
1089 }
1090
1091 /*
1092  * work queue call back to submit previously compressed pages
1093  */
1094 static noinline void async_cow_submit(struct btrfs_work *work)
1095 {
1096         struct async_cow *async_cow;
1097         struct btrfs_root *root;
1098         unsigned long nr_pages;
1099
1100         async_cow = container_of(work, struct async_cow, work);
1101
1102         root = async_cow->root;
1103         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
1104                 PAGE_CACHE_SHIFT;
1105
1106         /*
1107          * atomic_sub_return implies a barrier for waitqueue_active
1108          */
1109         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1110             5 * 1024 * 1024 &&
1111             waitqueue_active(&root->fs_info->async_submit_wait))
1112                 wake_up(&root->fs_info->async_submit_wait);
1113
1114         if (async_cow->inode)
1115                 submit_compressed_extents(async_cow->inode, async_cow);
1116 }
1117
1118 static noinline void async_cow_free(struct btrfs_work *work)
1119 {
1120         struct async_cow *async_cow;
1121         async_cow = container_of(work, struct async_cow, work);
1122         if (async_cow->inode)
1123                 btrfs_add_delayed_iput(async_cow->inode);
1124         kfree(async_cow);
1125 }
1126
1127 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1128                                 u64 start, u64 end, int *page_started,
1129                                 unsigned long *nr_written)
1130 {
1131         struct async_cow *async_cow;
1132         struct btrfs_root *root = BTRFS_I(inode)->root;
1133         unsigned long nr_pages;
1134         u64 cur_end;
1135         int limit = 10 * 1024 * 1024;
1136
1137         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1138                          1, 0, NULL, GFP_NOFS);
1139         while (start < end) {
1140                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1141                 BUG_ON(!async_cow); /* -ENOMEM */
1142                 async_cow->inode = igrab(inode);
1143                 async_cow->root = root;
1144                 async_cow->locked_page = locked_page;
1145                 async_cow->start = start;
1146
1147                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1148                     !btrfs_test_opt(root, FORCE_COMPRESS))
1149                         cur_end = end;
1150                 else
1151                         cur_end = min(end, start + 512 * 1024 - 1);
1152
1153                 async_cow->end = cur_end;
1154                 INIT_LIST_HEAD(&async_cow->extents);
1155
1156                 btrfs_init_work(&async_cow->work,
1157                                 btrfs_delalloc_helper,
1158                                 async_cow_start, async_cow_submit,
1159                                 async_cow_free);
1160
1161                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
1162                         PAGE_CACHE_SHIFT;
1163                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1164
1165                 btrfs_queue_work(root->fs_info->delalloc_workers,
1166                                  &async_cow->work);
1167
1168                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1169                         wait_event(root->fs_info->async_submit_wait,
1170                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1171                             limit));
1172                 }
1173
1174                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1175                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1176                         wait_event(root->fs_info->async_submit_wait,
1177                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1178                            0));
1179                 }
1180
1181                 *nr_written += nr_pages;
1182                 start = cur_end + 1;
1183         }
1184         *page_started = 1;
1185         return 0;
1186 }
1187
1188 static noinline int csum_exist_in_range(struct btrfs_root *root,
1189                                         u64 bytenr, u64 num_bytes)
1190 {
1191         int ret;
1192         struct btrfs_ordered_sum *sums;
1193         LIST_HEAD(list);
1194
1195         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1196                                        bytenr + num_bytes - 1, &list, 0);
1197         if (ret == 0 && list_empty(&list))
1198                 return 0;
1199
1200         while (!list_empty(&list)) {
1201                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1202                 list_del(&sums->list);
1203                 kfree(sums);
1204         }
1205         return 1;
1206 }
1207
1208 /*
1209  * when nowcow writeback call back.  This checks for snapshots or COW copies
1210  * of the extents that exist in the file, and COWs the file as required.
1211  *
1212  * If no cow copies or snapshots exist, we write directly to the existing
1213  * blocks on disk
1214  */
1215 static noinline int run_delalloc_nocow(struct inode *inode,
1216                                        struct page *locked_page,
1217                               u64 start, u64 end, int *page_started, int force,
1218                               unsigned long *nr_written)
1219 {
1220         struct btrfs_root *root = BTRFS_I(inode)->root;
1221         struct btrfs_trans_handle *trans;
1222         struct extent_buffer *leaf;
1223         struct btrfs_path *path;
1224         struct btrfs_file_extent_item *fi;
1225         struct btrfs_key found_key;
1226         u64 cow_start;
1227         u64 cur_offset;
1228         u64 extent_end;
1229         u64 extent_offset;
1230         u64 disk_bytenr;
1231         u64 num_bytes;
1232         u64 disk_num_bytes;
1233         u64 ram_bytes;
1234         int extent_type;
1235         int ret, err;
1236         int type;
1237         int nocow;
1238         int check_prev = 1;
1239         bool nolock;
1240         u64 ino = btrfs_ino(inode);
1241
1242         path = btrfs_alloc_path();
1243         if (!path) {
1244                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1245                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1246                                              EXTENT_DO_ACCOUNTING |
1247                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1248                                              PAGE_CLEAR_DIRTY |
1249                                              PAGE_SET_WRITEBACK |
1250                                              PAGE_END_WRITEBACK);
1251                 return -ENOMEM;
1252         }
1253
1254         nolock = btrfs_is_free_space_inode(inode);
1255
1256         if (nolock)
1257                 trans = btrfs_join_transaction_nolock(root);
1258         else
1259                 trans = btrfs_join_transaction(root);
1260
1261         if (IS_ERR(trans)) {
1262                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1263                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1264                                              EXTENT_DO_ACCOUNTING |
1265                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1266                                              PAGE_CLEAR_DIRTY |
1267                                              PAGE_SET_WRITEBACK |
1268                                              PAGE_END_WRITEBACK);
1269                 btrfs_free_path(path);
1270                 return PTR_ERR(trans);
1271         }
1272
1273         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1274
1275         cow_start = (u64)-1;
1276         cur_offset = start;
1277         while (1) {
1278                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1279                                                cur_offset, 0);
1280                 if (ret < 0)
1281                         goto error;
1282                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1283                         leaf = path->nodes[0];
1284                         btrfs_item_key_to_cpu(leaf, &found_key,
1285                                               path->slots[0] - 1);
1286                         if (found_key.objectid == ino &&
1287                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1288                                 path->slots[0]--;
1289                 }
1290                 check_prev = 0;
1291 next_slot:
1292                 leaf = path->nodes[0];
1293                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1294                         ret = btrfs_next_leaf(root, path);
1295                         if (ret < 0)
1296                                 goto error;
1297                         if (ret > 0)
1298                                 break;
1299                         leaf = path->nodes[0];
1300                 }
1301
1302                 nocow = 0;
1303                 disk_bytenr = 0;
1304                 num_bytes = 0;
1305                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1306
1307                 if (found_key.objectid > ino)
1308                         break;
1309                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1310                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1311                         path->slots[0]++;
1312                         goto next_slot;
1313                 }
1314                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1315                     found_key.offset > end)
1316                         break;
1317
1318                 if (found_key.offset > cur_offset) {
1319                         extent_end = found_key.offset;
1320                         extent_type = 0;
1321                         goto out_check;
1322                 }
1323
1324                 fi = btrfs_item_ptr(leaf, path->slots[0],
1325                                     struct btrfs_file_extent_item);
1326                 extent_type = btrfs_file_extent_type(leaf, fi);
1327
1328                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1329                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1330                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1331                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1332                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1333                         extent_end = found_key.offset +
1334                                 btrfs_file_extent_num_bytes(leaf, fi);
1335                         disk_num_bytes =
1336                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1337                         if (extent_end <= start) {
1338                                 path->slots[0]++;
1339                                 goto next_slot;
1340                         }
1341                         if (disk_bytenr == 0)
1342                                 goto out_check;
1343                         if (btrfs_file_extent_compression(leaf, fi) ||
1344                             btrfs_file_extent_encryption(leaf, fi) ||
1345                             btrfs_file_extent_other_encoding(leaf, fi))
1346                                 goto out_check;
1347                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1348                                 goto out_check;
1349                         if (btrfs_extent_readonly(root, disk_bytenr))
1350                                 goto out_check;
1351                         if (btrfs_cross_ref_exist(trans, root, ino,
1352                                                   found_key.offset -
1353                                                   extent_offset, disk_bytenr))
1354                                 goto out_check;
1355                         disk_bytenr += extent_offset;
1356                         disk_bytenr += cur_offset - found_key.offset;
1357                         num_bytes = min(end + 1, extent_end) - cur_offset;
1358                         /*
1359                          * if there are pending snapshots for this root,
1360                          * we fall into common COW way.
1361                          */
1362                         if (!nolock) {
1363                                 err = btrfs_start_write_no_snapshoting(root);
1364                                 if (!err)
1365                                         goto out_check;
1366                         }
1367                         /*
1368                          * force cow if csum exists in the range.
1369                          * this ensure that csum for a given extent are
1370                          * either valid or do not exist.
1371                          */
1372                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1373                                 goto out_check;
1374                         nocow = 1;
1375                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1376                         extent_end = found_key.offset +
1377                                 btrfs_file_extent_inline_len(leaf,
1378                                                      path->slots[0], fi);
1379                         extent_end = ALIGN(extent_end, root->sectorsize);
1380                 } else {
1381                         BUG_ON(1);
1382                 }
1383 out_check:
1384                 if (extent_end <= start) {
1385                         path->slots[0]++;
1386                         if (!nolock && nocow)
1387                                 btrfs_end_write_no_snapshoting(root);
1388                         goto next_slot;
1389                 }
1390                 if (!nocow) {
1391                         if (cow_start == (u64)-1)
1392                                 cow_start = cur_offset;
1393                         cur_offset = extent_end;
1394                         if (cur_offset > end)
1395                                 break;
1396                         path->slots[0]++;
1397                         goto next_slot;
1398                 }
1399
1400                 btrfs_release_path(path);
1401                 if (cow_start != (u64)-1) {
1402                         ret = cow_file_range(inode, locked_page,
1403                                              cow_start, found_key.offset - 1,
1404                                              page_started, nr_written, 1);
1405                         if (ret) {
1406                                 if (!nolock && nocow)
1407                                         btrfs_end_write_no_snapshoting(root);
1408                                 goto error;
1409                         }
1410                         cow_start = (u64)-1;
1411                 }
1412
1413                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1414                         struct extent_map *em;
1415                         struct extent_map_tree *em_tree;
1416                         em_tree = &BTRFS_I(inode)->extent_tree;
1417                         em = alloc_extent_map();
1418                         BUG_ON(!em); /* -ENOMEM */
1419                         em->start = cur_offset;
1420                         em->orig_start = found_key.offset - extent_offset;
1421                         em->len = num_bytes;
1422                         em->block_len = num_bytes;
1423                         em->block_start = disk_bytenr;
1424                         em->orig_block_len = disk_num_bytes;
1425                         em->ram_bytes = ram_bytes;
1426                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1427                         em->mod_start = em->start;
1428                         em->mod_len = em->len;
1429                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1430                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1431                         em->generation = -1;
1432                         while (1) {
1433                                 write_lock(&em_tree->lock);
1434                                 ret = add_extent_mapping(em_tree, em, 1);
1435                                 write_unlock(&em_tree->lock);
1436                                 if (ret != -EEXIST) {
1437                                         free_extent_map(em);
1438                                         break;
1439                                 }
1440                                 btrfs_drop_extent_cache(inode, em->start,
1441                                                 em->start + em->len - 1, 0);
1442                         }
1443                         type = BTRFS_ORDERED_PREALLOC;
1444                 } else {
1445                         type = BTRFS_ORDERED_NOCOW;
1446                 }
1447
1448                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1449                                                num_bytes, num_bytes, type);
1450                 BUG_ON(ret); /* -ENOMEM */
1451
1452                 if (root->root_key.objectid ==
1453                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1454                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1455                                                       num_bytes);
1456                         if (ret) {
1457                                 if (!nolock && nocow)
1458                                         btrfs_end_write_no_snapshoting(root);
1459                                 goto error;
1460                         }
1461                 }
1462
1463                 extent_clear_unlock_delalloc(inode, cur_offset,
1464                                              cur_offset + num_bytes - 1,
1465                                              locked_page, EXTENT_LOCKED |
1466                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1467                                              PAGE_SET_PRIVATE2);
1468                 if (!nolock && nocow)
1469                         btrfs_end_write_no_snapshoting(root);
1470                 cur_offset = extent_end;
1471                 if (cur_offset > end)
1472                         break;
1473         }
1474         btrfs_release_path(path);
1475
1476         if (cur_offset <= end && cow_start == (u64)-1) {
1477                 cow_start = cur_offset;
1478                 cur_offset = end;
1479         }
1480
1481         if (cow_start != (u64)-1) {
1482                 ret = cow_file_range(inode, locked_page, cow_start, end,
1483                                      page_started, nr_written, 1);
1484                 if (ret)
1485                         goto error;
1486         }
1487
1488 error:
1489         err = btrfs_end_transaction(trans, root);
1490         if (!ret)
1491                 ret = err;
1492
1493         if (ret && cur_offset < end)
1494                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1495                                              locked_page, EXTENT_LOCKED |
1496                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1497                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1498                                              PAGE_CLEAR_DIRTY |
1499                                              PAGE_SET_WRITEBACK |
1500                                              PAGE_END_WRITEBACK);
1501         btrfs_free_path(path);
1502         return ret;
1503 }
1504
1505 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1506 {
1507
1508         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1509             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1510                 return 0;
1511
1512         /*
1513          * @defrag_bytes is a hint value, no spinlock held here,
1514          * if is not zero, it means the file is defragging.
1515          * Force cow if given extent needs to be defragged.
1516          */
1517         if (BTRFS_I(inode)->defrag_bytes &&
1518             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1519                            EXTENT_DEFRAG, 0, NULL))
1520                 return 1;
1521
1522         return 0;
1523 }
1524
1525 /*
1526  * extent_io.c call back to do delayed allocation processing
1527  */
1528 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1529                               u64 start, u64 end, int *page_started,
1530                               unsigned long *nr_written)
1531 {
1532         int ret;
1533         int force_cow = need_force_cow(inode, start, end);
1534
1535         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1536                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1537                                          page_started, 1, nr_written);
1538         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1539                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1540                                          page_started, 0, nr_written);
1541         } else if (!inode_need_compress(inode)) {
1542                 ret = cow_file_range(inode, locked_page, start, end,
1543                                       page_started, nr_written, 1);
1544         } else {
1545                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1546                         &BTRFS_I(inode)->runtime_flags);
1547                 ret = cow_file_range_async(inode, locked_page, start, end,
1548                                            page_started, nr_written);
1549         }
1550         return ret;
1551 }
1552
1553 static void btrfs_split_extent_hook(struct inode *inode,
1554                                     struct extent_state *orig, u64 split)
1555 {
1556         u64 size;
1557
1558         /* not delalloc, ignore it */
1559         if (!(orig->state & EXTENT_DELALLOC))
1560                 return;
1561
1562         size = orig->end - orig->start + 1;
1563         if (size > BTRFS_MAX_EXTENT_SIZE) {
1564                 u64 num_extents;
1565                 u64 new_size;
1566
1567                 /*
1568                  * See the explanation in btrfs_merge_extent_hook, the same
1569                  * applies here, just in reverse.
1570                  */
1571                 new_size = orig->end - split + 1;
1572                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1573                                         BTRFS_MAX_EXTENT_SIZE);
1574                 new_size = split - orig->start;
1575                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1576                                         BTRFS_MAX_EXTENT_SIZE);
1577                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1578                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1579                         return;
1580         }
1581
1582         spin_lock(&BTRFS_I(inode)->lock);
1583         BTRFS_I(inode)->outstanding_extents++;
1584         spin_unlock(&BTRFS_I(inode)->lock);
1585 }
1586
1587 /*
1588  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1589  * extents so we can keep track of new extents that are just merged onto old
1590  * extents, such as when we are doing sequential writes, so we can properly
1591  * account for the metadata space we'll need.
1592  */
1593 static void btrfs_merge_extent_hook(struct inode *inode,
1594                                     struct extent_state *new,
1595                                     struct extent_state *other)
1596 {
1597         u64 new_size, old_size;
1598         u64 num_extents;
1599
1600         /* not delalloc, ignore it */
1601         if (!(other->state & EXTENT_DELALLOC))
1602                 return;
1603
1604         if (new->start > other->start)
1605                 new_size = new->end - other->start + 1;
1606         else
1607                 new_size = other->end - new->start + 1;
1608
1609         /* we're not bigger than the max, unreserve the space and go */
1610         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1611                 spin_lock(&BTRFS_I(inode)->lock);
1612                 BTRFS_I(inode)->outstanding_extents--;
1613                 spin_unlock(&BTRFS_I(inode)->lock);
1614                 return;
1615         }
1616
1617         /*
1618          * We have to add up either side to figure out how many extents were
1619          * accounted for before we merged into one big extent.  If the number of
1620          * extents we accounted for is <= the amount we need for the new range
1621          * then we can return, otherwise drop.  Think of it like this
1622          *
1623          * [ 4k][MAX_SIZE]
1624          *
1625          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1626          * need 2 outstanding extents, on one side we have 1 and the other side
1627          * we have 1 so they are == and we can return.  But in this case
1628          *
1629          * [MAX_SIZE+4k][MAX_SIZE+4k]
1630          *
1631          * Each range on their own accounts for 2 extents, but merged together
1632          * they are only 3 extents worth of accounting, so we need to drop in
1633          * this case.
1634          */
1635         old_size = other->end - other->start + 1;
1636         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1637                                 BTRFS_MAX_EXTENT_SIZE);
1638         old_size = new->end - new->start + 1;
1639         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1640                                  BTRFS_MAX_EXTENT_SIZE);
1641
1642         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1643                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1644                 return;
1645
1646         spin_lock(&BTRFS_I(inode)->lock);
1647         BTRFS_I(inode)->outstanding_extents--;
1648         spin_unlock(&BTRFS_I(inode)->lock);
1649 }
1650
1651 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1652                                       struct inode *inode)
1653 {
1654         spin_lock(&root->delalloc_lock);
1655         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1656                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1657                               &root->delalloc_inodes);
1658                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1659                         &BTRFS_I(inode)->runtime_flags);
1660                 root->nr_delalloc_inodes++;
1661                 if (root->nr_delalloc_inodes == 1) {
1662                         spin_lock(&root->fs_info->delalloc_root_lock);
1663                         BUG_ON(!list_empty(&root->delalloc_root));
1664                         list_add_tail(&root->delalloc_root,
1665                                       &root->fs_info->delalloc_roots);
1666                         spin_unlock(&root->fs_info->delalloc_root_lock);
1667                 }
1668         }
1669         spin_unlock(&root->delalloc_lock);
1670 }
1671
1672 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1673                                      struct inode *inode)
1674 {
1675         spin_lock(&root->delalloc_lock);
1676         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1677                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1678                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1679                           &BTRFS_I(inode)->runtime_flags);
1680                 root->nr_delalloc_inodes--;
1681                 if (!root->nr_delalloc_inodes) {
1682                         spin_lock(&root->fs_info->delalloc_root_lock);
1683                         BUG_ON(list_empty(&root->delalloc_root));
1684                         list_del_init(&root->delalloc_root);
1685                         spin_unlock(&root->fs_info->delalloc_root_lock);
1686                 }
1687         }
1688         spin_unlock(&root->delalloc_lock);
1689 }
1690
1691 /*
1692  * extent_io.c set_bit_hook, used to track delayed allocation
1693  * bytes in this file, and to maintain the list of inodes that
1694  * have pending delalloc work to be done.
1695  */
1696 static void btrfs_set_bit_hook(struct inode *inode,
1697                                struct extent_state *state, unsigned *bits)
1698 {
1699
1700         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1701                 WARN_ON(1);
1702         /*
1703          * set_bit and clear bit hooks normally require _irqsave/restore
1704          * but in this case, we are only testing for the DELALLOC
1705          * bit, which is only set or cleared with irqs on
1706          */
1707         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1708                 struct btrfs_root *root = BTRFS_I(inode)->root;
1709                 u64 len = state->end + 1 - state->start;
1710                 bool do_list = !btrfs_is_free_space_inode(inode);
1711
1712                 if (*bits & EXTENT_FIRST_DELALLOC) {
1713                         *bits &= ~EXTENT_FIRST_DELALLOC;
1714                 } else {
1715                         spin_lock(&BTRFS_I(inode)->lock);
1716                         BTRFS_I(inode)->outstanding_extents++;
1717                         spin_unlock(&BTRFS_I(inode)->lock);
1718                 }
1719
1720                 /* For sanity tests */
1721                 if (btrfs_test_is_dummy_root(root))
1722                         return;
1723
1724                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1725                                      root->fs_info->delalloc_batch);
1726                 spin_lock(&BTRFS_I(inode)->lock);
1727                 BTRFS_I(inode)->delalloc_bytes += len;
1728                 if (*bits & EXTENT_DEFRAG)
1729                         BTRFS_I(inode)->defrag_bytes += len;
1730                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1731                                          &BTRFS_I(inode)->runtime_flags))
1732                         btrfs_add_delalloc_inodes(root, inode);
1733                 spin_unlock(&BTRFS_I(inode)->lock);
1734         }
1735 }
1736
1737 /*
1738  * extent_io.c clear_bit_hook, see set_bit_hook for why
1739  */
1740 static void btrfs_clear_bit_hook(struct inode *inode,
1741                                  struct extent_state *state,
1742                                  unsigned *bits)
1743 {
1744         u64 len = state->end + 1 - state->start;
1745         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1746                                     BTRFS_MAX_EXTENT_SIZE);
1747
1748         spin_lock(&BTRFS_I(inode)->lock);
1749         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1750                 BTRFS_I(inode)->defrag_bytes -= len;
1751         spin_unlock(&BTRFS_I(inode)->lock);
1752
1753         /*
1754          * set_bit and clear bit hooks normally require _irqsave/restore
1755          * but in this case, we are only testing for the DELALLOC
1756          * bit, which is only set or cleared with irqs on
1757          */
1758         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1759                 struct btrfs_root *root = BTRFS_I(inode)->root;
1760                 bool do_list = !btrfs_is_free_space_inode(inode);
1761
1762                 if (*bits & EXTENT_FIRST_DELALLOC) {
1763                         *bits &= ~EXTENT_FIRST_DELALLOC;
1764                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1765                         spin_lock(&BTRFS_I(inode)->lock);
1766                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1767                         spin_unlock(&BTRFS_I(inode)->lock);
1768                 }
1769
1770                 /*
1771                  * We don't reserve metadata space for space cache inodes so we
1772                  * don't need to call dellalloc_release_metadata if there is an
1773                  * error.
1774                  */
1775                 if (*bits & EXTENT_DO_ACCOUNTING &&
1776                     root != root->fs_info->tree_root)
1777                         btrfs_delalloc_release_metadata(inode, len);
1778
1779                 /* For sanity tests. */
1780                 if (btrfs_test_is_dummy_root(root))
1781                         return;
1782
1783                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1784                     && do_list && !(state->state & EXTENT_NORESERVE))
1785                         btrfs_free_reserved_data_space_noquota(inode,
1786                                         state->start, len);
1787
1788                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1789                                      root->fs_info->delalloc_batch);
1790                 spin_lock(&BTRFS_I(inode)->lock);
1791                 BTRFS_I(inode)->delalloc_bytes -= len;
1792                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1793                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1794                              &BTRFS_I(inode)->runtime_flags))
1795                         btrfs_del_delalloc_inode(root, inode);
1796                 spin_unlock(&BTRFS_I(inode)->lock);
1797         }
1798 }
1799
1800 /*
1801  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1802  * we don't create bios that span stripes or chunks
1803  */
1804 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1805                          size_t size, struct bio *bio,
1806                          unsigned long bio_flags)
1807 {
1808         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1809         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1810         u64 length = 0;
1811         u64 map_length;
1812         int ret;
1813
1814         if (bio_flags & EXTENT_BIO_COMPRESSED)
1815                 return 0;
1816
1817         length = bio->bi_iter.bi_size;
1818         map_length = length;
1819         ret = btrfs_map_block(root->fs_info, rw, logical,
1820                               &map_length, NULL, 0);
1821         /* Will always return 0 with map_multi == NULL */
1822         BUG_ON(ret < 0);
1823         if (map_length < length + size)
1824                 return 1;
1825         return 0;
1826 }
1827
1828 /*
1829  * in order to insert checksums into the metadata in large chunks,
1830  * we wait until bio submission time.   All the pages in the bio are
1831  * checksummed and sums are attached onto the ordered extent record.
1832  *
1833  * At IO completion time the cums attached on the ordered extent record
1834  * are inserted into the btree
1835  */
1836 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1837                                     struct bio *bio, int mirror_num,
1838                                     unsigned long bio_flags,
1839                                     u64 bio_offset)
1840 {
1841         struct btrfs_root *root = BTRFS_I(inode)->root;
1842         int ret = 0;
1843
1844         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1845         BUG_ON(ret); /* -ENOMEM */
1846         return 0;
1847 }
1848
1849 /*
1850  * in order to insert checksums into the metadata in large chunks,
1851  * we wait until bio submission time.   All the pages in the bio are
1852  * checksummed and sums are attached onto the ordered extent record.
1853  *
1854  * At IO completion time the cums attached on the ordered extent record
1855  * are inserted into the btree
1856  */
1857 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1858                           int mirror_num, unsigned long bio_flags,
1859                           u64 bio_offset)
1860 {
1861         struct btrfs_root *root = BTRFS_I(inode)->root;
1862         int ret;
1863
1864         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1865         if (ret) {
1866                 bio->bi_error = ret;
1867                 bio_endio(bio);
1868         }
1869         return ret;
1870 }
1871
1872 /*
1873  * extent_io.c submission hook. This does the right thing for csum calculation
1874  * on write, or reading the csums from the tree before a read
1875  */
1876 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1877                           int mirror_num, unsigned long bio_flags,
1878                           u64 bio_offset)
1879 {
1880         struct btrfs_root *root = BTRFS_I(inode)->root;
1881         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1882         int ret = 0;
1883         int skip_sum;
1884         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1885
1886         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1887
1888         if (btrfs_is_free_space_inode(inode))
1889                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1890
1891         if (!(rw & REQ_WRITE)) {
1892                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1893                 if (ret)
1894                         goto out;
1895
1896                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1897                         ret = btrfs_submit_compressed_read(inode, bio,
1898                                                            mirror_num,
1899                                                            bio_flags);
1900                         goto out;
1901                 } else if (!skip_sum) {
1902                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1903                         if (ret)
1904                                 goto out;
1905                 }
1906                 goto mapit;
1907         } else if (async && !skip_sum) {
1908                 /* csum items have already been cloned */
1909                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1910                         goto mapit;
1911                 /* we're doing a write, do the async checksumming */
1912                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1913                                    inode, rw, bio, mirror_num,
1914                                    bio_flags, bio_offset,
1915                                    __btrfs_submit_bio_start,
1916                                    __btrfs_submit_bio_done);
1917                 goto out;
1918         } else if (!skip_sum) {
1919                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1920                 if (ret)
1921                         goto out;
1922         }
1923
1924 mapit:
1925         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1926
1927 out:
1928         if (ret < 0) {
1929                 bio->bi_error = ret;
1930                 bio_endio(bio);
1931         }
1932         return ret;
1933 }
1934
1935 /*
1936  * given a list of ordered sums record them in the inode.  This happens
1937  * at IO completion time based on sums calculated at bio submission time.
1938  */
1939 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1940                              struct inode *inode, u64 file_offset,
1941                              struct list_head *list)
1942 {
1943         struct btrfs_ordered_sum *sum;
1944
1945         list_for_each_entry(sum, list, list) {
1946                 trans->adding_csums = 1;
1947                 btrfs_csum_file_blocks(trans,
1948                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1949                 trans->adding_csums = 0;
1950         }
1951         return 0;
1952 }
1953
1954 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1955                               struct extent_state **cached_state)
1956 {
1957         WARN_ON((end & (PAGE_CACHE_SIZE - 1)) == 0);
1958         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1959                                    cached_state, GFP_NOFS);
1960 }
1961
1962 /* see btrfs_writepage_start_hook for details on why this is required */
1963 struct btrfs_writepage_fixup {
1964         struct page *page;
1965         struct btrfs_work work;
1966 };
1967
1968 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1969 {
1970         struct btrfs_writepage_fixup *fixup;
1971         struct btrfs_ordered_extent *ordered;
1972         struct extent_state *cached_state = NULL;
1973         struct page *page;
1974         struct inode *inode;
1975         u64 page_start;
1976         u64 page_end;
1977         int ret;
1978
1979         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1980         page = fixup->page;
1981 again:
1982         lock_page(page);
1983         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1984                 ClearPageChecked(page);
1985                 goto out_page;
1986         }
1987
1988         inode = page->mapping->host;
1989         page_start = page_offset(page);
1990         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1991
1992         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
1993                          &cached_state);
1994
1995         /* already ordered? We're done */
1996         if (PagePrivate2(page))
1997                 goto out;
1998
1999         ordered = btrfs_lookup_ordered_extent(inode, page_start);
2000         if (ordered) {
2001                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2002                                      page_end, &cached_state, GFP_NOFS);
2003                 unlock_page(page);
2004                 btrfs_start_ordered_extent(inode, ordered, 1);
2005                 btrfs_put_ordered_extent(ordered);
2006                 goto again;
2007         }
2008
2009         ret = btrfs_delalloc_reserve_space(inode, page_start,
2010                                            PAGE_CACHE_SIZE);
2011         if (ret) {
2012                 mapping_set_error(page->mapping, ret);
2013                 end_extent_writepage(page, ret, page_start, page_end);
2014                 ClearPageChecked(page);
2015                 goto out;
2016          }
2017
2018         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2019         ClearPageChecked(page);
2020         set_page_dirty(page);
2021 out:
2022         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2023                              &cached_state, GFP_NOFS);
2024 out_page:
2025         unlock_page(page);
2026         page_cache_release(page);
2027         kfree(fixup);
2028 }
2029
2030 /*
2031  * There are a few paths in the higher layers of the kernel that directly
2032  * set the page dirty bit without asking the filesystem if it is a
2033  * good idea.  This causes problems because we want to make sure COW
2034  * properly happens and the data=ordered rules are followed.
2035  *
2036  * In our case any range that doesn't have the ORDERED bit set
2037  * hasn't been properly setup for IO.  We kick off an async process
2038  * to fix it up.  The async helper will wait for ordered extents, set
2039  * the delalloc bit and make it safe to write the page.
2040  */
2041 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2042 {
2043         struct inode *inode = page->mapping->host;
2044         struct btrfs_writepage_fixup *fixup;
2045         struct btrfs_root *root = BTRFS_I(inode)->root;
2046
2047         /* this page is properly in the ordered list */
2048         if (TestClearPagePrivate2(page))
2049                 return 0;
2050
2051         if (PageChecked(page))
2052                 return -EAGAIN;
2053
2054         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2055         if (!fixup)
2056                 return -EAGAIN;
2057
2058         SetPageChecked(page);
2059         page_cache_get(page);
2060         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2061                         btrfs_writepage_fixup_worker, NULL, NULL);
2062         fixup->page = page;
2063         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2064         return -EBUSY;
2065 }
2066
2067 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2068                                        struct inode *inode, u64 file_pos,
2069                                        u64 disk_bytenr, u64 disk_num_bytes,
2070                                        u64 num_bytes, u64 ram_bytes,
2071                                        u8 compression, u8 encryption,
2072                                        u16 other_encoding, int extent_type)
2073 {
2074         struct btrfs_root *root = BTRFS_I(inode)->root;
2075         struct btrfs_file_extent_item *fi;
2076         struct btrfs_path *path;
2077         struct extent_buffer *leaf;
2078         struct btrfs_key ins;
2079         int extent_inserted = 0;
2080         int ret;
2081
2082         path = btrfs_alloc_path();
2083         if (!path)
2084                 return -ENOMEM;
2085
2086         /*
2087          * we may be replacing one extent in the tree with another.
2088          * The new extent is pinned in the extent map, and we don't want
2089          * to drop it from the cache until it is completely in the btree.
2090          *
2091          * So, tell btrfs_drop_extents to leave this extent in the cache.
2092          * the caller is expected to unpin it and allow it to be merged
2093          * with the others.
2094          */
2095         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2096                                    file_pos + num_bytes, NULL, 0,
2097                                    1, sizeof(*fi), &extent_inserted);
2098         if (ret)
2099                 goto out;
2100
2101         if (!extent_inserted) {
2102                 ins.objectid = btrfs_ino(inode);
2103                 ins.offset = file_pos;
2104                 ins.type = BTRFS_EXTENT_DATA_KEY;
2105
2106                 path->leave_spinning = 1;
2107                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2108                                               sizeof(*fi));
2109                 if (ret)
2110                         goto out;
2111         }
2112         leaf = path->nodes[0];
2113         fi = btrfs_item_ptr(leaf, path->slots[0],
2114                             struct btrfs_file_extent_item);
2115         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2116         btrfs_set_file_extent_type(leaf, fi, extent_type);
2117         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2118         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2119         btrfs_set_file_extent_offset(leaf, fi, 0);
2120         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2121         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2122         btrfs_set_file_extent_compression(leaf, fi, compression);
2123         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2124         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2125
2126         btrfs_mark_buffer_dirty(leaf);
2127         btrfs_release_path(path);
2128
2129         inode_add_bytes(inode, num_bytes);
2130
2131         ins.objectid = disk_bytenr;
2132         ins.offset = disk_num_bytes;
2133         ins.type = BTRFS_EXTENT_ITEM_KEY;
2134         ret = btrfs_alloc_reserved_file_extent(trans, root,
2135                                         root->root_key.objectid,
2136                                         btrfs_ino(inode), file_pos,
2137                                         ram_bytes, &ins);
2138         /*
2139          * Release the reserved range from inode dirty range map, as it is
2140          * already moved into delayed_ref_head
2141          */
2142         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2143 out:
2144         btrfs_free_path(path);
2145
2146         return ret;
2147 }
2148
2149 /* snapshot-aware defrag */
2150 struct sa_defrag_extent_backref {
2151         struct rb_node node;
2152         struct old_sa_defrag_extent *old;
2153         u64 root_id;
2154         u64 inum;
2155         u64 file_pos;
2156         u64 extent_offset;
2157         u64 num_bytes;
2158         u64 generation;
2159 };
2160
2161 struct old_sa_defrag_extent {
2162         struct list_head list;
2163         struct new_sa_defrag_extent *new;
2164
2165         u64 extent_offset;
2166         u64 bytenr;
2167         u64 offset;
2168         u64 len;
2169         int count;
2170 };
2171
2172 struct new_sa_defrag_extent {
2173         struct rb_root root;
2174         struct list_head head;
2175         struct btrfs_path *path;
2176         struct inode *inode;
2177         u64 file_pos;
2178         u64 len;
2179         u64 bytenr;
2180         u64 disk_len;
2181         u8 compress_type;
2182 };
2183
2184 static int backref_comp(struct sa_defrag_extent_backref *b1,
2185                         struct sa_defrag_extent_backref *b2)
2186 {
2187         if (b1->root_id < b2->root_id)
2188                 return -1;
2189         else if (b1->root_id > b2->root_id)
2190                 return 1;
2191
2192         if (b1->inum < b2->inum)
2193                 return -1;
2194         else if (b1->inum > b2->inum)
2195                 return 1;
2196
2197         if (b1->file_pos < b2->file_pos)
2198                 return -1;
2199         else if (b1->file_pos > b2->file_pos)
2200                 return 1;
2201
2202         /*
2203          * [------------------------------] ===> (a range of space)
2204          *     |<--->|   |<---->| =============> (fs/file tree A)
2205          * |<---------------------------->| ===> (fs/file tree B)
2206          *
2207          * A range of space can refer to two file extents in one tree while
2208          * refer to only one file extent in another tree.
2209          *
2210          * So we may process a disk offset more than one time(two extents in A)
2211          * and locate at the same extent(one extent in B), then insert two same
2212          * backrefs(both refer to the extent in B).
2213          */
2214         return 0;
2215 }
2216
2217 static void backref_insert(struct rb_root *root,
2218                            struct sa_defrag_extent_backref *backref)
2219 {
2220         struct rb_node **p = &root->rb_node;
2221         struct rb_node *parent = NULL;
2222         struct sa_defrag_extent_backref *entry;
2223         int ret;
2224
2225         while (*p) {
2226                 parent = *p;
2227                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2228
2229                 ret = backref_comp(backref, entry);
2230                 if (ret < 0)
2231                         p = &(*p)->rb_left;
2232                 else
2233                         p = &(*p)->rb_right;
2234         }
2235
2236         rb_link_node(&backref->node, parent, p);
2237         rb_insert_color(&backref->node, root);
2238 }
2239
2240 /*
2241  * Note the backref might has changed, and in this case we just return 0.
2242  */
2243 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2244                                        void *ctx)
2245 {
2246         struct btrfs_file_extent_item *extent;
2247         struct btrfs_fs_info *fs_info;
2248         struct old_sa_defrag_extent *old = ctx;
2249         struct new_sa_defrag_extent *new = old->new;
2250         struct btrfs_path *path = new->path;
2251         struct btrfs_key key;
2252         struct btrfs_root *root;
2253         struct sa_defrag_extent_backref *backref;
2254         struct extent_buffer *leaf;
2255         struct inode *inode = new->inode;
2256         int slot;
2257         int ret;
2258         u64 extent_offset;
2259         u64 num_bytes;
2260
2261         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2262             inum == btrfs_ino(inode))
2263                 return 0;
2264
2265         key.objectid = root_id;
2266         key.type = BTRFS_ROOT_ITEM_KEY;
2267         key.offset = (u64)-1;
2268
2269         fs_info = BTRFS_I(inode)->root->fs_info;
2270         root = btrfs_read_fs_root_no_name(fs_info, &key);
2271         if (IS_ERR(root)) {
2272                 if (PTR_ERR(root) == -ENOENT)
2273                         return 0;
2274                 WARN_ON(1);
2275                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2276                          inum, offset, root_id);
2277                 return PTR_ERR(root);
2278         }
2279
2280         key.objectid = inum;
2281         key.type = BTRFS_EXTENT_DATA_KEY;
2282         if (offset > (u64)-1 << 32)
2283                 key.offset = 0;
2284         else
2285                 key.offset = offset;
2286
2287         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2288         if (WARN_ON(ret < 0))
2289                 return ret;
2290         ret = 0;
2291
2292         while (1) {
2293                 cond_resched();
2294
2295                 leaf = path->nodes[0];
2296                 slot = path->slots[0];
2297
2298                 if (slot >= btrfs_header_nritems(leaf)) {
2299                         ret = btrfs_next_leaf(root, path);
2300                         if (ret < 0) {
2301                                 goto out;
2302                         } else if (ret > 0) {
2303                                 ret = 0;
2304                                 goto out;
2305                         }
2306                         continue;
2307                 }
2308
2309                 path->slots[0]++;
2310
2311                 btrfs_item_key_to_cpu(leaf, &key, slot);
2312
2313                 if (key.objectid > inum)
2314                         goto out;
2315
2316                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2317                         continue;
2318
2319                 extent = btrfs_item_ptr(leaf, slot,
2320                                         struct btrfs_file_extent_item);
2321
2322                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2323                         continue;
2324
2325                 /*
2326                  * 'offset' refers to the exact key.offset,
2327                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2328                  * (key.offset - extent_offset).
2329                  */
2330                 if (key.offset != offset)
2331                         continue;
2332
2333                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2334                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2335
2336                 if (extent_offset >= old->extent_offset + old->offset +
2337                     old->len || extent_offset + num_bytes <=
2338                     old->extent_offset + old->offset)
2339                         continue;
2340                 break;
2341         }
2342
2343         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2344         if (!backref) {
2345                 ret = -ENOENT;
2346                 goto out;
2347         }
2348
2349         backref->root_id = root_id;
2350         backref->inum = inum;
2351         backref->file_pos = offset;
2352         backref->num_bytes = num_bytes;
2353         backref->extent_offset = extent_offset;
2354         backref->generation = btrfs_file_extent_generation(leaf, extent);
2355         backref->old = old;
2356         backref_insert(&new->root, backref);
2357         old->count++;
2358 out:
2359         btrfs_release_path(path);
2360         WARN_ON(ret);
2361         return ret;
2362 }
2363
2364 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2365                                    struct new_sa_defrag_extent *new)
2366 {
2367         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2368         struct old_sa_defrag_extent *old, *tmp;
2369         int ret;
2370
2371         new->path = path;
2372
2373         list_for_each_entry_safe(old, tmp, &new->head, list) {
2374                 ret = iterate_inodes_from_logical(old->bytenr +
2375                                                   old->extent_offset, fs_info,
2376                                                   path, record_one_backref,
2377                                                   old);
2378                 if (ret < 0 && ret != -ENOENT)
2379                         return false;
2380
2381                 /* no backref to be processed for this extent */
2382                 if (!old->count) {
2383                         list_del(&old->list);
2384                         kfree(old);
2385                 }
2386         }
2387
2388         if (list_empty(&new->head))
2389                 return false;
2390
2391         return true;
2392 }
2393
2394 static int relink_is_mergable(struct extent_buffer *leaf,
2395                               struct btrfs_file_extent_item *fi,
2396                               struct new_sa_defrag_extent *new)
2397 {
2398         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2399                 return 0;
2400
2401         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2402                 return 0;
2403
2404         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2405                 return 0;
2406
2407         if (btrfs_file_extent_encryption(leaf, fi) ||
2408             btrfs_file_extent_other_encoding(leaf, fi))
2409                 return 0;
2410
2411         return 1;
2412 }
2413
2414 /*
2415  * Note the backref might has changed, and in this case we just return 0.
2416  */
2417 static noinline int relink_extent_backref(struct btrfs_path *path,
2418                                  struct sa_defrag_extent_backref *prev,
2419                                  struct sa_defrag_extent_backref *backref)
2420 {
2421         struct btrfs_file_extent_item *extent;
2422         struct btrfs_file_extent_item *item;
2423         struct btrfs_ordered_extent *ordered;
2424         struct btrfs_trans_handle *trans;
2425         struct btrfs_fs_info *fs_info;
2426         struct btrfs_root *root;
2427         struct btrfs_key key;
2428         struct extent_buffer *leaf;
2429         struct old_sa_defrag_extent *old = backref->old;
2430         struct new_sa_defrag_extent *new = old->new;
2431         struct inode *src_inode = new->inode;
2432         struct inode *inode;
2433         struct extent_state *cached = NULL;
2434         int ret = 0;
2435         u64 start;
2436         u64 len;
2437         u64 lock_start;
2438         u64 lock_end;
2439         bool merge = false;
2440         int index;
2441
2442         if (prev && prev->root_id == backref->root_id &&
2443             prev->inum == backref->inum &&
2444             prev->file_pos + prev->num_bytes == backref->file_pos)
2445                 merge = true;
2446
2447         /* step 1: get root */
2448         key.objectid = backref->root_id;
2449         key.type = BTRFS_ROOT_ITEM_KEY;
2450         key.offset = (u64)-1;
2451
2452         fs_info = BTRFS_I(src_inode)->root->fs_info;
2453         index = srcu_read_lock(&fs_info->subvol_srcu);
2454
2455         root = btrfs_read_fs_root_no_name(fs_info, &key);
2456         if (IS_ERR(root)) {
2457                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2458                 if (PTR_ERR(root) == -ENOENT)
2459                         return 0;
2460                 return PTR_ERR(root);
2461         }
2462
2463         if (btrfs_root_readonly(root)) {
2464                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2465                 return 0;
2466         }
2467
2468         /* step 2: get inode */
2469         key.objectid = backref->inum;
2470         key.type = BTRFS_INODE_ITEM_KEY;
2471         key.offset = 0;
2472
2473         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2474         if (IS_ERR(inode)) {
2475                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2476                 return 0;
2477         }
2478
2479         srcu_read_unlock(&fs_info->subvol_srcu, index);
2480
2481         /* step 3: relink backref */
2482         lock_start = backref->file_pos;
2483         lock_end = backref->file_pos + backref->num_bytes - 1;
2484         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2485                          &cached);
2486
2487         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2488         if (ordered) {
2489                 btrfs_put_ordered_extent(ordered);
2490                 goto out_unlock;
2491         }
2492
2493         trans = btrfs_join_transaction(root);
2494         if (IS_ERR(trans)) {
2495                 ret = PTR_ERR(trans);
2496                 goto out_unlock;
2497         }
2498
2499         key.objectid = backref->inum;
2500         key.type = BTRFS_EXTENT_DATA_KEY;
2501         key.offset = backref->file_pos;
2502
2503         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2504         if (ret < 0) {
2505                 goto out_free_path;
2506         } else if (ret > 0) {
2507                 ret = 0;
2508                 goto out_free_path;
2509         }
2510
2511         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2512                                 struct btrfs_file_extent_item);
2513
2514         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2515             backref->generation)
2516                 goto out_free_path;
2517
2518         btrfs_release_path(path);
2519
2520         start = backref->file_pos;
2521         if (backref->extent_offset < old->extent_offset + old->offset)
2522                 start += old->extent_offset + old->offset -
2523                          backref->extent_offset;
2524
2525         len = min(backref->extent_offset + backref->num_bytes,
2526                   old->extent_offset + old->offset + old->len);
2527         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2528
2529         ret = btrfs_drop_extents(trans, root, inode, start,
2530                                  start + len, 1);
2531         if (ret)
2532                 goto out_free_path;
2533 again:
2534         key.objectid = btrfs_ino(inode);
2535         key.type = BTRFS_EXTENT_DATA_KEY;
2536         key.offset = start;
2537
2538         path->leave_spinning = 1;
2539         if (merge) {
2540                 struct btrfs_file_extent_item *fi;
2541                 u64 extent_len;
2542                 struct btrfs_key found_key;
2543
2544                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2545                 if (ret < 0)
2546                         goto out_free_path;
2547
2548                 path->slots[0]--;
2549                 leaf = path->nodes[0];
2550                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2551
2552                 fi = btrfs_item_ptr(leaf, path->slots[0],
2553                                     struct btrfs_file_extent_item);
2554                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2555
2556                 if (extent_len + found_key.offset == start &&
2557                     relink_is_mergable(leaf, fi, new)) {
2558                         btrfs_set_file_extent_num_bytes(leaf, fi,
2559                                                         extent_len + len);
2560                         btrfs_mark_buffer_dirty(leaf);
2561                         inode_add_bytes(inode, len);
2562
2563                         ret = 1;
2564                         goto out_free_path;
2565                 } else {
2566                         merge = false;
2567                         btrfs_release_path(path);
2568                         goto again;
2569                 }
2570         }
2571
2572         ret = btrfs_insert_empty_item(trans, root, path, &key,
2573                                         sizeof(*extent));
2574         if (ret) {
2575                 btrfs_abort_transaction(trans, root, ret);
2576                 goto out_free_path;
2577         }
2578
2579         leaf = path->nodes[0];
2580         item = btrfs_item_ptr(leaf, path->slots[0],
2581                                 struct btrfs_file_extent_item);
2582         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2583         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2584         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2585         btrfs_set_file_extent_num_bytes(leaf, item, len);
2586         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2587         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2588         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2589         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2590         btrfs_set_file_extent_encryption(leaf, item, 0);
2591         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2592
2593         btrfs_mark_buffer_dirty(leaf);
2594         inode_add_bytes(inode, len);
2595         btrfs_release_path(path);
2596
2597         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2598                         new->disk_len, 0,
2599                         backref->root_id, backref->inum,
2600                         new->file_pos); /* start - extent_offset */
2601         if (ret) {
2602                 btrfs_abort_transaction(trans, root, ret);
2603                 goto out_free_path;
2604         }
2605
2606         ret = 1;
2607 out_free_path:
2608         btrfs_release_path(path);
2609         path->leave_spinning = 0;
2610         btrfs_end_transaction(trans, root);
2611 out_unlock:
2612         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2613                              &cached, GFP_NOFS);
2614         iput(inode);
2615         return ret;
2616 }
2617
2618 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2619 {
2620         struct old_sa_defrag_extent *old, *tmp;
2621
2622         if (!new)
2623                 return;
2624
2625         list_for_each_entry_safe(old, tmp, &new->head, list) {
2626                 kfree(old);
2627         }
2628         kfree(new);
2629 }
2630
2631 static void relink_file_extents(struct new_sa_defrag_extent *new)
2632 {
2633         struct btrfs_path *path;
2634         struct sa_defrag_extent_backref *backref;
2635         struct sa_defrag_extent_backref *prev = NULL;
2636         struct inode *inode;
2637         struct btrfs_root *root;
2638         struct rb_node *node;
2639         int ret;
2640
2641         inode = new->inode;
2642         root = BTRFS_I(inode)->root;
2643
2644         path = btrfs_alloc_path();
2645         if (!path)
2646                 return;
2647
2648         if (!record_extent_backrefs(path, new)) {
2649                 btrfs_free_path(path);
2650                 goto out;
2651         }
2652         btrfs_release_path(path);
2653
2654         while (1) {
2655                 node = rb_first(&new->root);
2656                 if (!node)
2657                         break;
2658                 rb_erase(node, &new->root);
2659
2660                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2661
2662                 ret = relink_extent_backref(path, prev, backref);
2663                 WARN_ON(ret < 0);
2664
2665                 kfree(prev);
2666
2667                 if (ret == 1)
2668                         prev = backref;
2669                 else
2670                         prev = NULL;
2671                 cond_resched();
2672         }
2673         kfree(prev);
2674
2675         btrfs_free_path(path);
2676 out:
2677         free_sa_defrag_extent(new);
2678
2679         atomic_dec(&root->fs_info->defrag_running);
2680         wake_up(&root->fs_info->transaction_wait);
2681 }
2682
2683 static struct new_sa_defrag_extent *
2684 record_old_file_extents(struct inode *inode,
2685                         struct btrfs_ordered_extent *ordered)
2686 {
2687         struct btrfs_root *root = BTRFS_I(inode)->root;
2688         struct btrfs_path *path;
2689         struct btrfs_key key;
2690         struct old_sa_defrag_extent *old;
2691         struct new_sa_defrag_extent *new;
2692         int ret;
2693
2694         new = kmalloc(sizeof(*new), GFP_NOFS);
2695         if (!new)
2696                 return NULL;
2697
2698         new->inode = inode;
2699         new->file_pos = ordered->file_offset;
2700         new->len = ordered->len;
2701         new->bytenr = ordered->start;
2702         new->disk_len = ordered->disk_len;
2703         new->compress_type = ordered->compress_type;
2704         new->root = RB_ROOT;
2705         INIT_LIST_HEAD(&new->head);
2706
2707         path = btrfs_alloc_path();
2708         if (!path)
2709                 goto out_kfree;
2710
2711         key.objectid = btrfs_ino(inode);
2712         key.type = BTRFS_EXTENT_DATA_KEY;
2713         key.offset = new->file_pos;
2714
2715         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2716         if (ret < 0)
2717                 goto out_free_path;
2718         if (ret > 0 && path->slots[0] > 0)
2719                 path->slots[0]--;
2720
2721         /* find out all the old extents for the file range */
2722         while (1) {
2723                 struct btrfs_file_extent_item *extent;
2724                 struct extent_buffer *l;
2725                 int slot;
2726                 u64 num_bytes;
2727                 u64 offset;
2728                 u64 end;
2729                 u64 disk_bytenr;
2730                 u64 extent_offset;
2731
2732                 l = path->nodes[0];
2733                 slot = path->slots[0];
2734
2735                 if (slot >= btrfs_header_nritems(l)) {
2736                         ret = btrfs_next_leaf(root, path);
2737                         if (ret < 0)
2738                                 goto out_free_path;
2739                         else if (ret > 0)
2740                                 break;
2741                         continue;
2742                 }
2743
2744                 btrfs_item_key_to_cpu(l, &key, slot);
2745
2746                 if (key.objectid != btrfs_ino(inode))
2747                         break;
2748                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2749                         break;
2750                 if (key.offset >= new->file_pos + new->len)
2751                         break;
2752
2753                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2754
2755                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2756                 if (key.offset + num_bytes < new->file_pos)
2757                         goto next;
2758
2759                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2760                 if (!disk_bytenr)
2761                         goto next;
2762
2763                 extent_offset = btrfs_file_extent_offset(l, extent);
2764
2765                 old = kmalloc(sizeof(*old), GFP_NOFS);
2766                 if (!old)
2767                         goto out_free_path;
2768
2769                 offset = max(new->file_pos, key.offset);
2770                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2771
2772                 old->bytenr = disk_bytenr;
2773                 old->extent_offset = extent_offset;
2774                 old->offset = offset - key.offset;
2775                 old->len = end - offset;
2776                 old->new = new;
2777                 old->count = 0;
2778                 list_add_tail(&old->list, &new->head);
2779 next:
2780                 path->slots[0]++;
2781                 cond_resched();
2782         }
2783
2784         btrfs_free_path(path);
2785         atomic_inc(&root->fs_info->defrag_running);
2786
2787         return new;
2788
2789 out_free_path:
2790         btrfs_free_path(path);
2791 out_kfree:
2792         free_sa_defrag_extent(new);
2793         return NULL;
2794 }
2795
2796 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2797                                          u64 start, u64 len)
2798 {
2799         struct btrfs_block_group_cache *cache;
2800
2801         cache = btrfs_lookup_block_group(root->fs_info, start);
2802         ASSERT(cache);
2803
2804         spin_lock(&cache->lock);
2805         cache->delalloc_bytes -= len;
2806         spin_unlock(&cache->lock);
2807
2808         btrfs_put_block_group(cache);
2809 }
2810
2811 /* as ordered data IO finishes, this gets called so we can finish
2812  * an ordered extent if the range of bytes in the file it covers are
2813  * fully written.
2814  */
2815 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2816 {
2817         struct inode *inode = ordered_extent->inode;
2818         struct btrfs_root *root = BTRFS_I(inode)->root;
2819         struct btrfs_trans_handle *trans = NULL;
2820         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2821         struct extent_state *cached_state = NULL;
2822         struct new_sa_defrag_extent *new = NULL;
2823         int compress_type = 0;
2824         int ret = 0;
2825         u64 logical_len = ordered_extent->len;
2826         bool nolock;
2827         bool truncated = false;
2828
2829         nolock = btrfs_is_free_space_inode(inode);
2830
2831         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2832                 ret = -EIO;
2833                 goto out;
2834         }
2835
2836         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2837                                      ordered_extent->file_offset +
2838                                      ordered_extent->len - 1);
2839
2840         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2841                 truncated = true;
2842                 logical_len = ordered_extent->truncated_len;
2843                 /* Truncated the entire extent, don't bother adding */
2844                 if (!logical_len)
2845                         goto out;
2846         }
2847
2848         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2849                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2850
2851                 /*
2852                  * For mwrite(mmap + memset to write) case, we still reserve
2853                  * space for NOCOW range.
2854                  * As NOCOW won't cause a new delayed ref, just free the space
2855                  */
2856                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2857                                        ordered_extent->len);
2858                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2859                 if (nolock)
2860                         trans = btrfs_join_transaction_nolock(root);
2861                 else
2862                         trans = btrfs_join_transaction(root);
2863                 if (IS_ERR(trans)) {
2864                         ret = PTR_ERR(trans);
2865                         trans = NULL;
2866                         goto out;
2867                 }
2868                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2869                 ret = btrfs_update_inode_fallback(trans, root, inode);
2870                 if (ret) /* -ENOMEM or corruption */
2871                         btrfs_abort_transaction(trans, root, ret);
2872                 goto out;
2873         }
2874
2875         lock_extent_bits(io_tree, ordered_extent->file_offset,
2876                          ordered_extent->file_offset + ordered_extent->len - 1,
2877                          &cached_state);
2878
2879         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2880                         ordered_extent->file_offset + ordered_extent->len - 1,
2881                         EXTENT_DEFRAG, 1, cached_state);
2882         if (ret) {
2883                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2884                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2885                         /* the inode is shared */
2886                         new = record_old_file_extents(inode, ordered_extent);
2887
2888                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2889                         ordered_extent->file_offset + ordered_extent->len - 1,
2890                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2891         }
2892
2893         if (nolock)
2894                 trans = btrfs_join_transaction_nolock(root);
2895         else
2896                 trans = btrfs_join_transaction(root);
2897         if (IS_ERR(trans)) {
2898                 ret = PTR_ERR(trans);
2899                 trans = NULL;
2900                 goto out_unlock;
2901         }
2902
2903         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2904
2905         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2906                 compress_type = ordered_extent->compress_type;
2907         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2908                 BUG_ON(compress_type);
2909                 ret = btrfs_mark_extent_written(trans, inode,
2910                                                 ordered_extent->file_offset,
2911                                                 ordered_extent->file_offset +
2912                                                 logical_len);
2913         } else {
2914                 BUG_ON(root == root->fs_info->tree_root);
2915                 ret = insert_reserved_file_extent(trans, inode,
2916                                                 ordered_extent->file_offset,
2917                                                 ordered_extent->start,
2918                                                 ordered_extent->disk_len,
2919                                                 logical_len, logical_len,
2920                                                 compress_type, 0, 0,
2921                                                 BTRFS_FILE_EXTENT_REG);
2922                 if (!ret)
2923                         btrfs_release_delalloc_bytes(root,
2924                                                      ordered_extent->start,
2925                                                      ordered_extent->disk_len);
2926         }
2927         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2928                            ordered_extent->file_offset, ordered_extent->len,
2929                            trans->transid);
2930         if (ret < 0) {
2931                 btrfs_abort_transaction(trans, root, ret);
2932                 goto out_unlock;
2933         }
2934
2935         add_pending_csums(trans, inode, ordered_extent->file_offset,
2936                           &ordered_extent->list);
2937
2938         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2939         ret = btrfs_update_inode_fallback(trans, root, inode);
2940         if (ret) { /* -ENOMEM or corruption */
2941                 btrfs_abort_transaction(trans, root, ret);
2942                 goto out_unlock;
2943         }
2944         ret = 0;
2945 out_unlock:
2946         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2947                              ordered_extent->file_offset +
2948                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2949 out:
2950         if (root != root->fs_info->tree_root)
2951                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2952         if (trans)
2953                 btrfs_end_transaction(trans, root);
2954
2955         if (ret || truncated) {
2956                 u64 start, end;
2957
2958                 if (truncated)
2959                         start = ordered_extent->file_offset + logical_len;
2960                 else
2961                         start = ordered_extent->file_offset;
2962                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2963                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2964
2965                 /* Drop the cache for the part of the extent we didn't write. */
2966                 btrfs_drop_extent_cache(inode, start, end, 0);
2967
2968                 /*
2969                  * If the ordered extent had an IOERR or something else went
2970                  * wrong we need to return the space for this ordered extent
2971                  * back to the allocator.  We only free the extent in the
2972                  * truncated case if we didn't write out the extent at all.
2973                  */
2974                 if ((ret || !logical_len) &&
2975                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2976                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
2977                         btrfs_free_reserved_extent(root, ordered_extent->start,
2978                                                    ordered_extent->disk_len, 1);
2979         }
2980
2981
2982         /*
2983          * This needs to be done to make sure anybody waiting knows we are done
2984          * updating everything for this ordered extent.
2985          */
2986         btrfs_remove_ordered_extent(inode, ordered_extent);
2987
2988         /* for snapshot-aware defrag */
2989         if (new) {
2990                 if (ret) {
2991                         free_sa_defrag_extent(new);
2992                         atomic_dec(&root->fs_info->defrag_running);
2993                 } else {
2994                         relink_file_extents(new);
2995                 }
2996         }
2997
2998         /* once for us */
2999         btrfs_put_ordered_extent(ordered_extent);
3000         /* once for the tree */
3001         btrfs_put_ordered_extent(ordered_extent);
3002
3003         return ret;
3004 }
3005
3006 static void finish_ordered_fn(struct btrfs_work *work)
3007 {
3008         struct btrfs_ordered_extent *ordered_extent;
3009         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3010         btrfs_finish_ordered_io(ordered_extent);
3011 }
3012
3013 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3014                                 struct extent_state *state, int uptodate)
3015 {
3016         struct inode *inode = page->mapping->host;
3017         struct btrfs_root *root = BTRFS_I(inode)->root;
3018         struct btrfs_ordered_extent *ordered_extent = NULL;
3019         struct btrfs_workqueue *wq;
3020         btrfs_work_func_t func;
3021
3022         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3023
3024         ClearPagePrivate2(page);
3025         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3026                                             end - start + 1, uptodate))
3027                 return 0;
3028
3029         if (btrfs_is_free_space_inode(inode)) {
3030                 wq = root->fs_info->endio_freespace_worker;
3031                 func = btrfs_freespace_write_helper;
3032         } else {
3033                 wq = root->fs_info->endio_write_workers;
3034                 func = btrfs_endio_write_helper;
3035         }
3036
3037         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3038                         NULL);
3039         btrfs_queue_work(wq, &ordered_extent->work);
3040
3041         return 0;
3042 }
3043
3044 static int __readpage_endio_check(struct inode *inode,
3045                                   struct btrfs_io_bio *io_bio,
3046                                   int icsum, struct page *page,
3047                                   int pgoff, u64 start, size_t len)
3048 {
3049         char *kaddr;
3050         u32 csum_expected;
3051         u32 csum = ~(u32)0;
3052
3053         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3054
3055         kaddr = kmap_atomic(page);
3056         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3057         btrfs_csum_final(csum, (char *)&csum);
3058         if (csum != csum_expected)
3059                 goto zeroit;
3060
3061         kunmap_atomic(kaddr);
3062         return 0;
3063 zeroit:
3064         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3065                 "csum failed ino %llu off %llu csum %u expected csum %u",
3066                            btrfs_ino(inode), start, csum, csum_expected);
3067         memset(kaddr + pgoff, 1, len);
3068         flush_dcache_page(page);
3069         kunmap_atomic(kaddr);
3070         if (csum_expected == 0)
3071                 return 0;
3072         return -EIO;
3073 }
3074
3075 /*
3076  * when reads are done, we need to check csums to verify the data is correct
3077  * if there's a match, we allow the bio to finish.  If not, the code in
3078  * extent_io.c will try to find good copies for us.
3079  */
3080 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3081                                       u64 phy_offset, struct page *page,
3082                                       u64 start, u64 end, int mirror)
3083 {
3084         size_t offset = start - page_offset(page);
3085         struct inode *inode = page->mapping->host;
3086         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3087         struct btrfs_root *root = BTRFS_I(inode)->root;
3088
3089         if (PageChecked(page)) {
3090                 ClearPageChecked(page);
3091                 return 0;
3092         }
3093
3094         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3095                 return 0;
3096
3097         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3098             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3099                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
3100                                   GFP_NOFS);
3101                 return 0;
3102         }
3103
3104         phy_offset >>= inode->i_sb->s_blocksize_bits;
3105         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3106                                       start, (size_t)(end - start + 1));
3107 }
3108
3109 struct delayed_iput {
3110         struct list_head list;
3111         struct inode *inode;
3112 };
3113
3114 /* JDM: If this is fs-wide, why can't we add a pointer to
3115  * btrfs_inode instead and avoid the allocation? */
3116 void btrfs_add_delayed_iput(struct inode *inode)
3117 {
3118         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3119         struct delayed_iput *delayed;
3120
3121         if (atomic_add_unless(&inode->i_count, -1, 1))
3122                 return;
3123
3124         delayed = kmalloc(sizeof(*delayed), GFP_NOFS | __GFP_NOFAIL);
3125         delayed->inode = inode;
3126
3127         spin_lock(&fs_info->delayed_iput_lock);
3128         list_add_tail(&delayed->list, &fs_info->delayed_iputs);
3129         spin_unlock(&fs_info->delayed_iput_lock);
3130 }
3131
3132 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3133 {
3134         LIST_HEAD(list);
3135         struct btrfs_fs_info *fs_info = root->fs_info;
3136         struct delayed_iput *delayed;
3137         int empty;
3138
3139         spin_lock(&fs_info->delayed_iput_lock);
3140         empty = list_empty(&fs_info->delayed_iputs);
3141         spin_unlock(&fs_info->delayed_iput_lock);
3142         if (empty)
3143                 return;
3144
3145         down_read(&fs_info->delayed_iput_sem);
3146
3147         spin_lock(&fs_info->delayed_iput_lock);
3148         list_splice_init(&fs_info->delayed_iputs, &list);
3149         spin_unlock(&fs_info->delayed_iput_lock);
3150
3151         while (!list_empty(&list)) {
3152                 delayed = list_entry(list.next, struct delayed_iput, list);
3153                 list_del(&delayed->list);
3154                 iput(delayed->inode);
3155                 kfree(delayed);
3156         }
3157
3158         up_read(&root->fs_info->delayed_iput_sem);
3159 }
3160
3161 /*
3162  * This is called in transaction commit time. If there are no orphan
3163  * files in the subvolume, it removes orphan item and frees block_rsv
3164  * structure.
3165  */
3166 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3167                               struct btrfs_root *root)
3168 {
3169         struct btrfs_block_rsv *block_rsv;
3170         int ret;
3171
3172         if (atomic_read(&root->orphan_inodes) ||
3173             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3174                 return;
3175
3176         spin_lock(&root->orphan_lock);
3177         if (atomic_read(&root->orphan_inodes)) {
3178                 spin_unlock(&root->orphan_lock);
3179                 return;
3180         }
3181
3182         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3183                 spin_unlock(&root->orphan_lock);
3184                 return;
3185         }
3186
3187         block_rsv = root->orphan_block_rsv;
3188         root->orphan_block_rsv = NULL;
3189         spin_unlock(&root->orphan_lock);
3190
3191         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3192             btrfs_root_refs(&root->root_item) > 0) {
3193                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3194                                             root->root_key.objectid);
3195                 if (ret)
3196                         btrfs_abort_transaction(trans, root, ret);
3197                 else
3198                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3199                                   &root->state);
3200         }
3201
3202         if (block_rsv) {
3203                 WARN_ON(block_rsv->size > 0);
3204                 btrfs_free_block_rsv(root, block_rsv);
3205         }
3206 }
3207
3208 /*
3209  * This creates an orphan entry for the given inode in case something goes
3210  * wrong in the middle of an unlink/truncate.
3211  *
3212  * NOTE: caller of this function should reserve 5 units of metadata for
3213  *       this function.
3214  */
3215 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3216 {
3217         struct btrfs_root *root = BTRFS_I(inode)->root;
3218         struct btrfs_block_rsv *block_rsv = NULL;
3219         int reserve = 0;
3220         int insert = 0;
3221         int ret;
3222
3223         if (!root->orphan_block_rsv) {
3224                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3225                 if (!block_rsv)
3226                         return -ENOMEM;
3227         }
3228
3229         spin_lock(&root->orphan_lock);
3230         if (!root->orphan_block_rsv) {
3231                 root->orphan_block_rsv = block_rsv;
3232         } else if (block_rsv) {
3233                 btrfs_free_block_rsv(root, block_rsv);
3234                 block_rsv = NULL;
3235         }
3236
3237         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3238                               &BTRFS_I(inode)->runtime_flags)) {
3239 #if 0
3240                 /*
3241                  * For proper ENOSPC handling, we should do orphan
3242                  * cleanup when mounting. But this introduces backward
3243                  * compatibility issue.
3244                  */
3245                 if (!xchg(&root->orphan_item_inserted, 1))
3246                         insert = 2;
3247                 else
3248                         insert = 1;
3249 #endif
3250                 insert = 1;
3251                 atomic_inc(&root->orphan_inodes);
3252         }
3253
3254         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3255                               &BTRFS_I(inode)->runtime_flags))
3256                 reserve = 1;
3257         spin_unlock(&root->orphan_lock);
3258
3259         /* grab metadata reservation from transaction handle */
3260         if (reserve) {
3261                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3262                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
3263         }
3264
3265         /* insert an orphan item to track this unlinked/truncated file */
3266         if (insert >= 1) {
3267                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3268                 if (ret) {
3269                         atomic_dec(&root->orphan_inodes);
3270                         if (reserve) {
3271                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3272                                           &BTRFS_I(inode)->runtime_flags);
3273                                 btrfs_orphan_release_metadata(inode);
3274                         }
3275                         if (ret != -EEXIST) {
3276                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3277                                           &BTRFS_I(inode)->runtime_flags);
3278                                 btrfs_abort_transaction(trans, root, ret);
3279                                 return ret;
3280                         }
3281                 }
3282                 ret = 0;
3283         }
3284
3285         /* insert an orphan item to track subvolume contains orphan files */
3286         if (insert >= 2) {
3287                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3288                                                root->root_key.objectid);
3289                 if (ret && ret != -EEXIST) {
3290                         btrfs_abort_transaction(trans, root, ret);
3291                         return ret;
3292                 }
3293         }
3294         return 0;
3295 }
3296
3297 /*
3298  * We have done the truncate/delete so we can go ahead and remove the orphan
3299  * item for this particular inode.
3300  */
3301 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3302                             struct inode *inode)
3303 {
3304         struct btrfs_root *root = BTRFS_I(inode)->root;
3305         int delete_item = 0;
3306         int release_rsv = 0;
3307         int ret = 0;
3308
3309         spin_lock(&root->orphan_lock);
3310         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3311                                &BTRFS_I(inode)->runtime_flags))
3312                 delete_item = 1;
3313
3314         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3315                                &BTRFS_I(inode)->runtime_flags))
3316                 release_rsv = 1;
3317         spin_unlock(&root->orphan_lock);
3318
3319         if (delete_item) {
3320                 atomic_dec(&root->orphan_inodes);
3321                 if (trans)
3322                         ret = btrfs_del_orphan_item(trans, root,
3323                                                     btrfs_ino(inode));
3324         }
3325
3326         if (release_rsv)
3327                 btrfs_orphan_release_metadata(inode);
3328
3329         return ret;
3330 }
3331
3332 /*
3333  * this cleans up any orphans that may be left on the list from the last use
3334  * of this root.
3335  */
3336 int btrfs_orphan_cleanup(struct btrfs_root *root)
3337 {
3338         struct btrfs_path *path;
3339         struct extent_buffer *leaf;
3340         struct btrfs_key key, found_key;
3341         struct btrfs_trans_handle *trans;
3342         struct inode *inode;
3343         u64 last_objectid = 0;
3344         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3345
3346         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3347                 return 0;
3348
3349         path = btrfs_alloc_path();
3350         if (!path) {
3351                 ret = -ENOMEM;
3352                 goto out;
3353         }
3354         path->reada = -1;
3355
3356         key.objectid = BTRFS_ORPHAN_OBJECTID;
3357         key.type = BTRFS_ORPHAN_ITEM_KEY;
3358         key.offset = (u64)-1;
3359
3360         while (1) {
3361                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3362                 if (ret < 0)
3363                         goto out;
3364
3365                 /*
3366                  * if ret == 0 means we found what we were searching for, which
3367                  * is weird, but possible, so only screw with path if we didn't
3368                  * find the key and see if we have stuff that matches
3369                  */
3370                 if (ret > 0) {
3371                         ret = 0;
3372                         if (path->slots[0] == 0)
3373                                 break;
3374                         path->slots[0]--;
3375                 }
3376
3377                 /* pull out the item */
3378                 leaf = path->nodes[0];
3379                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3380
3381                 /* make sure the item matches what we want */
3382                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3383                         break;
3384                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3385                         break;
3386
3387                 /* release the path since we're done with it */
3388                 btrfs_release_path(path);
3389
3390                 /*
3391                  * this is where we are basically btrfs_lookup, without the
3392                  * crossing root thing.  we store the inode number in the
3393                  * offset of the orphan item.
3394                  */
3395
3396                 if (found_key.offset == last_objectid) {
3397                         btrfs_err(root->fs_info,
3398                                 "Error removing orphan entry, stopping orphan cleanup");
3399                         ret = -EINVAL;
3400                         goto out;
3401                 }
3402
3403                 last_objectid = found_key.offset;
3404
3405                 found_key.objectid = found_key.offset;
3406                 found_key.type = BTRFS_INODE_ITEM_KEY;
3407                 found_key.offset = 0;
3408                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3409                 ret = PTR_ERR_OR_ZERO(inode);
3410                 if (ret && ret != -ESTALE)
3411                         goto out;
3412
3413                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3414                         struct btrfs_root *dead_root;
3415                         struct btrfs_fs_info *fs_info = root->fs_info;
3416                         int is_dead_root = 0;
3417
3418                         /*
3419                          * this is an orphan in the tree root. Currently these
3420                          * could come from 2 sources:
3421                          *  a) a snapshot deletion in progress
3422                          *  b) a free space cache inode
3423                          * We need to distinguish those two, as the snapshot
3424                          * orphan must not get deleted.
3425                          * find_dead_roots already ran before us, so if this
3426                          * is a snapshot deletion, we should find the root
3427                          * in the dead_roots list
3428                          */
3429                         spin_lock(&fs_info->trans_lock);
3430                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3431                                             root_list) {
3432                                 if (dead_root->root_key.objectid ==
3433                                     found_key.objectid) {
3434                                         is_dead_root = 1;
3435                                         break;
3436                                 }
3437                         }
3438                         spin_unlock(&fs_info->trans_lock);
3439                         if (is_dead_root) {
3440                                 /* prevent this orphan from being found again */
3441                                 key.offset = found_key.objectid - 1;
3442                                 continue;
3443                         }
3444                 }
3445                 /*
3446                  * Inode is already gone but the orphan item is still there,
3447                  * kill the orphan item.
3448                  */
3449                 if (ret == -ESTALE) {
3450                         trans = btrfs_start_transaction(root, 1);
3451                         if (IS_ERR(trans)) {
3452                                 ret = PTR_ERR(trans);
3453                                 goto out;
3454                         }
3455                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3456                                 found_key.objectid);
3457                         ret = btrfs_del_orphan_item(trans, root,
3458                                                     found_key.objectid);
3459                         btrfs_end_transaction(trans, root);
3460                         if (ret)
3461                                 goto out;
3462                         continue;
3463                 }
3464
3465                 /*
3466                  * add this inode to the orphan list so btrfs_orphan_del does
3467                  * the proper thing when we hit it
3468                  */
3469                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3470                         &BTRFS_I(inode)->runtime_flags);
3471                 atomic_inc(&root->orphan_inodes);
3472
3473                 /* if we have links, this was a truncate, lets do that */
3474                 if (inode->i_nlink) {
3475                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3476                                 iput(inode);
3477                                 continue;
3478                         }
3479                         nr_truncate++;
3480
3481                         /* 1 for the orphan item deletion. */
3482                         trans = btrfs_start_transaction(root, 1);
3483                         if (IS_ERR(trans)) {
3484                                 iput(inode);
3485                                 ret = PTR_ERR(trans);
3486                                 goto out;
3487                         }
3488                         ret = btrfs_orphan_add(trans, inode);
3489                         btrfs_end_transaction(trans, root);
3490                         if (ret) {
3491                                 iput(inode);
3492                                 goto out;
3493                         }
3494
3495                         ret = btrfs_truncate(inode);
3496                         if (ret)
3497                                 btrfs_orphan_del(NULL, inode);
3498                 } else {
3499                         nr_unlink++;
3500                 }
3501
3502                 /* this will do delete_inode and everything for us */
3503                 iput(inode);
3504                 if (ret)
3505                         goto out;
3506         }
3507         /* release the path since we're done with it */
3508         btrfs_release_path(path);
3509
3510         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3511
3512         if (root->orphan_block_rsv)
3513                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3514                                         (u64)-1);
3515
3516         if (root->orphan_block_rsv ||
3517             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3518                 trans = btrfs_join_transaction(root);
3519                 if (!IS_ERR(trans))
3520                         btrfs_end_transaction(trans, root);
3521         }
3522
3523         if (nr_unlink)
3524                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3525         if (nr_truncate)
3526                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3527
3528 out:
3529         if (ret)
3530                 btrfs_err(root->fs_info,
3531                         "could not do orphan cleanup %d", ret);
3532         btrfs_free_path(path);
3533         return ret;
3534 }
3535
3536 /*
3537  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3538  * don't find any xattrs, we know there can't be any acls.
3539  *
3540  * slot is the slot the inode is in, objectid is the objectid of the inode
3541  */
3542 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3543                                           int slot, u64 objectid,
3544                                           int *first_xattr_slot)
3545 {
3546         u32 nritems = btrfs_header_nritems(leaf);
3547         struct btrfs_key found_key;
3548         static u64 xattr_access = 0;
3549         static u64 xattr_default = 0;
3550         int scanned = 0;
3551
3552         if (!xattr_access) {
3553                 xattr_access = btrfs_name_hash(POSIX_ACL_XATTR_ACCESS,
3554                                         strlen(POSIX_ACL_XATTR_ACCESS));
3555                 xattr_default = btrfs_name_hash(POSIX_ACL_XATTR_DEFAULT,
3556                                         strlen(POSIX_ACL_XATTR_DEFAULT));
3557         }
3558
3559         slot++;
3560         *first_xattr_slot = -1;
3561         while (slot < nritems) {
3562                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3563
3564                 /* we found a different objectid, there must not be acls */
3565                 if (found_key.objectid != objectid)
3566                         return 0;
3567
3568                 /* we found an xattr, assume we've got an acl */
3569                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3570                         if (*first_xattr_slot == -1)
3571                                 *first_xattr_slot = slot;
3572                         if (found_key.offset == xattr_access ||
3573                             found_key.offset == xattr_default)
3574                                 return 1;
3575                 }
3576
3577                 /*
3578                  * we found a key greater than an xattr key, there can't
3579                  * be any acls later on
3580                  */
3581                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3582                         return 0;
3583
3584                 slot++;
3585                 scanned++;
3586
3587                 /*
3588                  * it goes inode, inode backrefs, xattrs, extents,
3589                  * so if there are a ton of hard links to an inode there can
3590                  * be a lot of backrefs.  Don't waste time searching too hard,
3591                  * this is just an optimization
3592                  */
3593                 if (scanned >= 8)
3594                         break;
3595         }
3596         /* we hit the end of the leaf before we found an xattr or
3597          * something larger than an xattr.  We have to assume the inode
3598          * has acls
3599          */
3600         if (*first_xattr_slot == -1)
3601                 *first_xattr_slot = slot;
3602         return 1;
3603 }
3604
3605 /*
3606  * read an inode from the btree into the in-memory inode
3607  */
3608 static void btrfs_read_locked_inode(struct inode *inode)
3609 {
3610         struct btrfs_path *path;
3611         struct extent_buffer *leaf;
3612         struct btrfs_inode_item *inode_item;
3613         struct btrfs_root *root = BTRFS_I(inode)->root;
3614         struct btrfs_key location;
3615         unsigned long ptr;
3616         int maybe_acls;
3617         u32 rdev;
3618         int ret;
3619         bool filled = false;
3620         int first_xattr_slot;
3621
3622         ret = btrfs_fill_inode(inode, &rdev);
3623         if (!ret)
3624                 filled = true;
3625
3626         path = btrfs_alloc_path();
3627         if (!path)
3628                 goto make_bad;
3629
3630         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3631
3632         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3633         if (ret)
3634                 goto make_bad;
3635
3636         leaf = path->nodes[0];
3637
3638         if (filled)
3639                 goto cache_index;
3640
3641         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3642                                     struct btrfs_inode_item);
3643         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3644         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3645         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3646         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3647         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3648
3649         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3650         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3651
3652         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3653         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3654
3655         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3656         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3657
3658         BTRFS_I(inode)->i_otime.tv_sec =
3659                 btrfs_timespec_sec(leaf, &inode_item->otime);
3660         BTRFS_I(inode)->i_otime.tv_nsec =
3661                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3662
3663         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3664         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3665         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3666
3667         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3668         inode->i_generation = BTRFS_I(inode)->generation;
3669         inode->i_rdev = 0;
3670         rdev = btrfs_inode_rdev(leaf, inode_item);
3671
3672         BTRFS_I(inode)->index_cnt = (u64)-1;
3673         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3674
3675 cache_index:
3676         /*
3677          * If we were modified in the current generation and evicted from memory
3678          * and then re-read we need to do a full sync since we don't have any
3679          * idea about which extents were modified before we were evicted from
3680          * cache.
3681          *
3682          * This is required for both inode re-read from disk and delayed inode
3683          * in delayed_nodes_tree.
3684          */
3685         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3686                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3687                         &BTRFS_I(inode)->runtime_flags);
3688
3689         /*
3690          * We don't persist the id of the transaction where an unlink operation
3691          * against the inode was last made. So here we assume the inode might
3692          * have been evicted, and therefore the exact value of last_unlink_trans
3693          * lost, and set it to last_trans to avoid metadata inconsistencies
3694          * between the inode and its parent if the inode is fsync'ed and the log
3695          * replayed. For example, in the scenario:
3696          *
3697          * touch mydir/foo
3698          * ln mydir/foo mydir/bar
3699          * sync
3700          * unlink mydir/bar
3701          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3702          * xfs_io -c fsync mydir/foo
3703          * <power failure>
3704          * mount fs, triggers fsync log replay
3705          *
3706          * We must make sure that when we fsync our inode foo we also log its
3707          * parent inode, otherwise after log replay the parent still has the
3708          * dentry with the "bar" name but our inode foo has a link count of 1
3709          * and doesn't have an inode ref with the name "bar" anymore.
3710          *
3711          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3712          * but it guarantees correctness at the expense of ocassional full
3713          * transaction commits on fsync if our inode is a directory, or if our
3714          * inode is not a directory, logging its parent unnecessarily.
3715          */
3716         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3717
3718         path->slots[0]++;
3719         if (inode->i_nlink != 1 ||
3720             path->slots[0] >= btrfs_header_nritems(leaf))
3721                 goto cache_acl;
3722
3723         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3724         if (location.objectid != btrfs_ino(inode))
3725                 goto cache_acl;
3726
3727         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3728         if (location.type == BTRFS_INODE_REF_KEY) {
3729                 struct btrfs_inode_ref *ref;
3730
3731                 ref = (struct btrfs_inode_ref *)ptr;
3732                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3733         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3734                 struct btrfs_inode_extref *extref;
3735
3736                 extref = (struct btrfs_inode_extref *)ptr;
3737                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3738                                                                      extref);
3739         }
3740 cache_acl:
3741         /*
3742          * try to precache a NULL acl entry for files that don't have
3743          * any xattrs or acls
3744          */
3745         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3746                                            btrfs_ino(inode), &first_xattr_slot);
3747         if (first_xattr_slot != -1) {
3748                 path->slots[0] = first_xattr_slot;
3749                 ret = btrfs_load_inode_props(inode, path);
3750                 if (ret)
3751                         btrfs_err(root->fs_info,
3752                                   "error loading props for ino %llu (root %llu): %d",
3753                                   btrfs_ino(inode),
3754                                   root->root_key.objectid, ret);
3755         }
3756         btrfs_free_path(path);
3757
3758         if (!maybe_acls)
3759                 cache_no_acl(inode);
3760
3761         switch (inode->i_mode & S_IFMT) {
3762         case S_IFREG:
3763                 inode->i_mapping->a_ops = &btrfs_aops;
3764                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3765                 inode->i_fop = &btrfs_file_operations;
3766                 inode->i_op = &btrfs_file_inode_operations;
3767                 break;
3768         case S_IFDIR:
3769                 inode->i_fop = &btrfs_dir_file_operations;
3770                 if (root == root->fs_info->tree_root)
3771                         inode->i_op = &btrfs_dir_ro_inode_operations;
3772                 else
3773                         inode->i_op = &btrfs_dir_inode_operations;
3774                 break;
3775         case S_IFLNK:
3776                 inode->i_op = &btrfs_symlink_inode_operations;
3777                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3778                 break;
3779         default:
3780                 inode->i_op = &btrfs_special_inode_operations;
3781                 init_special_inode(inode, inode->i_mode, rdev);
3782                 break;
3783         }
3784
3785         btrfs_update_iflags(inode);
3786         return;
3787
3788 make_bad:
3789         btrfs_free_path(path);
3790         make_bad_inode(inode);
3791 }
3792
3793 /*
3794  * given a leaf and an inode, copy the inode fields into the leaf
3795  */
3796 static void fill_inode_item(struct btrfs_trans_handle *trans,
3797                             struct extent_buffer *leaf,
3798                             struct btrfs_inode_item *item,
3799                             struct inode *inode)
3800 {
3801         struct btrfs_map_token token;
3802
3803         btrfs_init_map_token(&token);
3804
3805         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3806         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3807         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3808                                    &token);
3809         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3810         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3811
3812         btrfs_set_token_timespec_sec(leaf, &item->atime,
3813                                      inode->i_atime.tv_sec, &token);
3814         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3815                                       inode->i_atime.tv_nsec, &token);
3816
3817         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3818                                      inode->i_mtime.tv_sec, &token);
3819         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3820                                       inode->i_mtime.tv_nsec, &token);
3821
3822         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3823                                      inode->i_ctime.tv_sec, &token);
3824         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3825                                       inode->i_ctime.tv_nsec, &token);
3826
3827         btrfs_set_token_timespec_sec(leaf, &item->otime,
3828                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3829         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3830                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3831
3832         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3833                                      &token);
3834         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3835                                          &token);
3836         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3837         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3838         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3839         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3840         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3841 }
3842
3843 /*
3844  * copy everything in the in-memory inode into the btree.
3845  */
3846 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3847                                 struct btrfs_root *root, struct inode *inode)
3848 {
3849         struct btrfs_inode_item *inode_item;
3850         struct btrfs_path *path;
3851         struct extent_buffer *leaf;
3852         int ret;
3853
3854         path = btrfs_alloc_path();
3855         if (!path)
3856                 return -ENOMEM;
3857
3858         path->leave_spinning = 1;
3859         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3860                                  1);
3861         if (ret) {
3862                 if (ret > 0)
3863                         ret = -ENOENT;
3864                 goto failed;
3865         }
3866
3867         leaf = path->nodes[0];
3868         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3869                                     struct btrfs_inode_item);
3870
3871         fill_inode_item(trans, leaf, inode_item, inode);
3872         btrfs_mark_buffer_dirty(leaf);
3873         btrfs_set_inode_last_trans(trans, inode);
3874         ret = 0;
3875 failed:
3876         btrfs_free_path(path);
3877         return ret;
3878 }
3879
3880 /*
3881  * copy everything in the in-memory inode into the btree.
3882  */
3883 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3884                                 struct btrfs_root *root, struct inode *inode)
3885 {
3886         int ret;
3887
3888         /*
3889          * If the inode is a free space inode, we can deadlock during commit
3890          * if we put it into the delayed code.
3891          *
3892          * The data relocation inode should also be directly updated
3893          * without delay
3894          */
3895         if (!btrfs_is_free_space_inode(inode)
3896             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3897             && !root->fs_info->log_root_recovering) {
3898                 btrfs_update_root_times(trans, root);
3899
3900                 ret = btrfs_delayed_update_inode(trans, root, inode);
3901                 if (!ret)
3902                         btrfs_set_inode_last_trans(trans, inode);
3903                 return ret;
3904         }
3905
3906         return btrfs_update_inode_item(trans, root, inode);
3907 }
3908
3909 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3910                                          struct btrfs_root *root,
3911                                          struct inode *inode)
3912 {
3913         int ret;
3914
3915         ret = btrfs_update_inode(trans, root, inode);
3916         if (ret == -ENOSPC)
3917                 return btrfs_update_inode_item(trans, root, inode);
3918         return ret;
3919 }
3920
3921 /*
3922  * unlink helper that gets used here in inode.c and in the tree logging
3923  * recovery code.  It remove a link in a directory with a given name, and
3924  * also drops the back refs in the inode to the directory
3925  */
3926 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3927                                 struct btrfs_root *root,
3928                                 struct inode *dir, struct inode *inode,
3929                                 const char *name, int name_len)
3930 {
3931         struct btrfs_path *path;
3932         int ret = 0;
3933         struct extent_buffer *leaf;
3934         struct btrfs_dir_item *di;
3935         struct btrfs_key key;
3936         u64 index;
3937         u64 ino = btrfs_ino(inode);
3938         u64 dir_ino = btrfs_ino(dir);
3939
3940         path = btrfs_alloc_path();
3941         if (!path) {
3942                 ret = -ENOMEM;
3943                 goto out;
3944         }
3945
3946         path->leave_spinning = 1;
3947         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3948                                     name, name_len, -1);
3949         if (IS_ERR(di)) {
3950                 ret = PTR_ERR(di);
3951                 goto err;
3952         }
3953         if (!di) {
3954                 ret = -ENOENT;
3955                 goto err;
3956         }
3957         leaf = path->nodes[0];
3958         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3959         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3960         if (ret)
3961                 goto err;
3962         btrfs_release_path(path);
3963
3964         /*
3965          * If we don't have dir index, we have to get it by looking up
3966          * the inode ref, since we get the inode ref, remove it directly,
3967          * it is unnecessary to do delayed deletion.
3968          *
3969          * But if we have dir index, needn't search inode ref to get it.
3970          * Since the inode ref is close to the inode item, it is better
3971          * that we delay to delete it, and just do this deletion when
3972          * we update the inode item.
3973          */
3974         if (BTRFS_I(inode)->dir_index) {
3975                 ret = btrfs_delayed_delete_inode_ref(inode);
3976                 if (!ret) {
3977                         index = BTRFS_I(inode)->dir_index;
3978                         goto skip_backref;
3979                 }
3980         }
3981
3982         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3983                                   dir_ino, &index);
3984         if (ret) {
3985                 btrfs_info(root->fs_info,
3986                         "failed to delete reference to %.*s, inode %llu parent %llu",
3987                         name_len, name, ino, dir_ino);
3988                 btrfs_abort_transaction(trans, root, ret);
3989                 goto err;
3990         }
3991 skip_backref:
3992         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3993         if (ret) {
3994                 btrfs_abort_transaction(trans, root, ret);
3995                 goto err;
3996         }
3997
3998         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
3999                                          inode, dir_ino);
4000         if (ret != 0 && ret != -ENOENT) {
4001                 btrfs_abort_transaction(trans, root, ret);
4002                 goto err;
4003         }
4004
4005         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4006                                            dir, index);
4007         if (ret == -ENOENT)
4008                 ret = 0;
4009         else if (ret)
4010                 btrfs_abort_transaction(trans, root, ret);
4011 err:
4012         btrfs_free_path(path);
4013         if (ret)
4014                 goto out;
4015
4016         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4017         inode_inc_iversion(inode);
4018         inode_inc_iversion(dir);
4019         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4020         ret = btrfs_update_inode(trans, root, dir);
4021 out:
4022         return ret;
4023 }
4024
4025 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4026                        struct btrfs_root *root,
4027                        struct inode *dir, struct inode *inode,
4028                        const char *name, int name_len)
4029 {
4030         int ret;
4031         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4032         if (!ret) {
4033                 drop_nlink(inode);
4034                 ret = btrfs_update_inode(trans, root, inode);
4035         }
4036         return ret;
4037 }
4038
4039 /*
4040  * helper to start transaction for unlink and rmdir.
4041  *
4042  * unlink and rmdir are special in btrfs, they do not always free space, so
4043  * if we cannot make our reservations the normal way try and see if there is
4044  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4045  * allow the unlink to occur.
4046  */
4047 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4048 {
4049         struct btrfs_trans_handle *trans;
4050         struct btrfs_root *root = BTRFS_I(dir)->root;
4051         int ret;
4052
4053         /*
4054          * 1 for the possible orphan item
4055          * 1 for the dir item
4056          * 1 for the dir index
4057          * 1 for the inode ref
4058          * 1 for the inode
4059          */
4060         trans = btrfs_start_transaction(root, 5);
4061         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
4062                 return trans;
4063
4064         if (PTR_ERR(trans) == -ENOSPC) {
4065                 u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4066
4067                 trans = btrfs_start_transaction(root, 0);
4068                 if (IS_ERR(trans))
4069                         return trans;
4070                 ret = btrfs_cond_migrate_bytes(root->fs_info,
4071                                                &root->fs_info->trans_block_rsv,
4072                                                num_bytes, 5);
4073                 if (ret) {
4074                         btrfs_end_transaction(trans, root);
4075                         return ERR_PTR(ret);
4076                 }
4077                 trans->block_rsv = &root->fs_info->trans_block_rsv;
4078                 trans->bytes_reserved = num_bytes;
4079         }
4080         return trans;
4081 }
4082
4083 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4084 {
4085         struct btrfs_root *root = BTRFS_I(dir)->root;
4086         struct btrfs_trans_handle *trans;
4087         struct inode *inode = d_inode(dentry);
4088         int ret;
4089
4090         trans = __unlink_start_trans(dir);
4091         if (IS_ERR(trans))
4092                 return PTR_ERR(trans);
4093
4094         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4095
4096         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4097                                  dentry->d_name.name, dentry->d_name.len);
4098         if (ret)
4099                 goto out;
4100
4101         if (inode->i_nlink == 0) {
4102                 ret = btrfs_orphan_add(trans, inode);
4103                 if (ret)
4104                         goto out;
4105         }
4106
4107 out:
4108         btrfs_end_transaction(trans, root);
4109         btrfs_btree_balance_dirty(root);
4110         return ret;
4111 }
4112
4113 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4114                         struct btrfs_root *root,
4115                         struct inode *dir, u64 objectid,
4116                         const char *name, int name_len)
4117 {
4118         struct btrfs_path *path;
4119         struct extent_buffer *leaf;
4120         struct btrfs_dir_item *di;
4121         struct btrfs_key key;
4122         u64 index;
4123         int ret;
4124         u64 dir_ino = btrfs_ino(dir);
4125
4126         path = btrfs_alloc_path();
4127         if (!path)
4128                 return -ENOMEM;
4129
4130         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4131                                    name, name_len, -1);
4132         if (IS_ERR_OR_NULL(di)) {
4133                 if (!di)
4134                         ret = -ENOENT;
4135                 else
4136                         ret = PTR_ERR(di);
4137                 goto out;
4138         }
4139
4140         leaf = path->nodes[0];
4141         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4142         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4143         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4144         if (ret) {
4145                 btrfs_abort_transaction(trans, root, ret);
4146                 goto out;
4147         }
4148         btrfs_release_path(path);
4149
4150         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4151                                  objectid, root->root_key.objectid,
4152                                  dir_ino, &index, name, name_len);
4153         if (ret < 0) {
4154                 if (ret != -ENOENT) {
4155                         btrfs_abort_transaction(trans, root, ret);
4156                         goto out;
4157                 }
4158                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4159                                                  name, name_len);
4160                 if (IS_ERR_OR_NULL(di)) {
4161                         if (!di)
4162                                 ret = -ENOENT;
4163                         else
4164                                 ret = PTR_ERR(di);
4165                         btrfs_abort_transaction(trans, root, ret);
4166                         goto out;
4167                 }
4168
4169                 leaf = path->nodes[0];
4170                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4171                 btrfs_release_path(path);
4172                 index = key.offset;
4173         }
4174         btrfs_release_path(path);
4175
4176         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4177         if (ret) {
4178                 btrfs_abort_transaction(trans, root, ret);
4179                 goto out;
4180         }
4181
4182         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4183         inode_inc_iversion(dir);
4184         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4185         ret = btrfs_update_inode_fallback(trans, root, dir);
4186         if (ret)
4187                 btrfs_abort_transaction(trans, root, ret);
4188 out:
4189         btrfs_free_path(path);
4190         return ret;
4191 }
4192
4193 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4194 {
4195         struct inode *inode = d_inode(dentry);
4196         int err = 0;
4197         struct btrfs_root *root = BTRFS_I(dir)->root;
4198         struct btrfs_trans_handle *trans;
4199
4200         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4201                 return -ENOTEMPTY;
4202         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4203                 return -EPERM;
4204
4205         trans = __unlink_start_trans(dir);
4206         if (IS_ERR(trans))
4207                 return PTR_ERR(trans);
4208
4209         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4210                 err = btrfs_unlink_subvol(trans, root, dir,
4211                                           BTRFS_I(inode)->location.objectid,
4212                                           dentry->d_name.name,
4213                                           dentry->d_name.len);
4214                 goto out;
4215         }
4216
4217         err = btrfs_orphan_add(trans, inode);
4218         if (err)
4219                 goto out;
4220
4221         /* now the directory is empty */
4222         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4223                                  dentry->d_name.name, dentry->d_name.len);
4224         if (!err)
4225                 btrfs_i_size_write(inode, 0);
4226 out:
4227         btrfs_end_transaction(trans, root);
4228         btrfs_btree_balance_dirty(root);
4229
4230         return err;
4231 }
4232
4233 static int truncate_space_check(struct btrfs_trans_handle *trans,
4234                                 struct btrfs_root *root,
4235                                 u64 bytes_deleted)
4236 {
4237         int ret;
4238
4239         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4240         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4241                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4242         if (!ret)
4243                 trans->bytes_reserved += bytes_deleted;
4244         return ret;
4245
4246 }
4247
4248 static int truncate_inline_extent(struct inode *inode,
4249                                   struct btrfs_path *path,
4250                                   struct btrfs_key *found_key,
4251                                   const u64 item_end,
4252                                   const u64 new_size)
4253 {
4254         struct extent_buffer *leaf = path->nodes[0];
4255         int slot = path->slots[0];
4256         struct btrfs_file_extent_item *fi;
4257         u32 size = (u32)(new_size - found_key->offset);
4258         struct btrfs_root *root = BTRFS_I(inode)->root;
4259
4260         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4261
4262         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4263                 loff_t offset = new_size;
4264                 loff_t page_end = ALIGN(offset, PAGE_CACHE_SIZE);
4265
4266                 /*
4267                  * Zero out the remaining of the last page of our inline extent,
4268                  * instead of directly truncating our inline extent here - that
4269                  * would be much more complex (decompressing all the data, then
4270                  * compressing the truncated data, which might be bigger than
4271                  * the size of the inline extent, resize the extent, etc).
4272                  * We release the path because to get the page we might need to
4273                  * read the extent item from disk (data not in the page cache).
4274                  */
4275                 btrfs_release_path(path);
4276                 return btrfs_truncate_page(inode, offset, page_end - offset, 0);
4277         }
4278
4279         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4280         size = btrfs_file_extent_calc_inline_size(size);
4281         btrfs_truncate_item(root, path, size, 1);
4282
4283         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4284                 inode_sub_bytes(inode, item_end + 1 - new_size);
4285
4286         return 0;
4287 }
4288
4289 /*
4290  * this can truncate away extent items, csum items and directory items.
4291  * It starts at a high offset and removes keys until it can't find
4292  * any higher than new_size
4293  *
4294  * csum items that cross the new i_size are truncated to the new size
4295  * as well.
4296  *
4297  * min_type is the minimum key type to truncate down to.  If set to 0, this
4298  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4299  */
4300 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4301                                struct btrfs_root *root,
4302                                struct inode *inode,
4303                                u64 new_size, u32 min_type)
4304 {
4305         struct btrfs_path *path;
4306         struct extent_buffer *leaf;
4307         struct btrfs_file_extent_item *fi;
4308         struct btrfs_key key;
4309         struct btrfs_key found_key;
4310         u64 extent_start = 0;
4311         u64 extent_num_bytes = 0;
4312         u64 extent_offset = 0;
4313         u64 item_end = 0;
4314         u64 last_size = new_size;
4315         u32 found_type = (u8)-1;
4316         int found_extent;
4317         int del_item;
4318         int pending_del_nr = 0;
4319         int pending_del_slot = 0;
4320         int extent_type = -1;
4321         int ret;
4322         int err = 0;
4323         u64 ino = btrfs_ino(inode);
4324         u64 bytes_deleted = 0;
4325         bool be_nice = 0;
4326         bool should_throttle = 0;
4327         bool should_end = 0;
4328
4329         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4330
4331         /*
4332          * for non-free space inodes and ref cows, we want to back off from
4333          * time to time
4334          */
4335         if (!btrfs_is_free_space_inode(inode) &&
4336             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4337                 be_nice = 1;
4338
4339         path = btrfs_alloc_path();
4340         if (!path)
4341                 return -ENOMEM;
4342         path->reada = -1;
4343
4344         /*
4345          * We want to drop from the next block forward in case this new size is
4346          * not block aligned since we will be keeping the last block of the
4347          * extent just the way it is.
4348          */
4349         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4350             root == root->fs_info->tree_root)
4351                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4352                                         root->sectorsize), (u64)-1, 0);
4353
4354         /*
4355          * This function is also used to drop the items in the log tree before
4356          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4357          * it is used to drop the loged items. So we shouldn't kill the delayed
4358          * items.
4359          */
4360         if (min_type == 0 && root == BTRFS_I(inode)->root)
4361                 btrfs_kill_delayed_inode_items(inode);
4362
4363         key.objectid = ino;
4364         key.offset = (u64)-1;
4365         key.type = (u8)-1;
4366
4367 search_again:
4368         /*
4369          * with a 16K leaf size and 128MB extents, you can actually queue
4370          * up a huge file in a single leaf.  Most of the time that
4371          * bytes_deleted is > 0, it will be huge by the time we get here
4372          */
4373         if (be_nice && bytes_deleted > 32 * 1024 * 1024) {
4374                 if (btrfs_should_end_transaction(trans, root)) {
4375                         err = -EAGAIN;
4376                         goto error;
4377                 }
4378         }
4379
4380
4381         path->leave_spinning = 1;
4382         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4383         if (ret < 0) {
4384                 err = ret;
4385                 goto out;
4386         }
4387
4388         if (ret > 0) {
4389                 /* there are no items in the tree for us to truncate, we're
4390                  * done
4391                  */
4392                 if (path->slots[0] == 0)
4393                         goto out;
4394                 path->slots[0]--;
4395         }
4396
4397         while (1) {
4398                 fi = NULL;
4399                 leaf = path->nodes[0];
4400                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4401                 found_type = found_key.type;
4402
4403                 if (found_key.objectid != ino)
4404                         break;
4405
4406                 if (found_type < min_type)
4407                         break;
4408
4409                 item_end = found_key.offset;
4410                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4411                         fi = btrfs_item_ptr(leaf, path->slots[0],
4412                                             struct btrfs_file_extent_item);
4413                         extent_type = btrfs_file_extent_type(leaf, fi);
4414                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4415                                 item_end +=
4416                                     btrfs_file_extent_num_bytes(leaf, fi);
4417                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4418                                 item_end += btrfs_file_extent_inline_len(leaf,
4419                                                          path->slots[0], fi);
4420                         }
4421                         item_end--;
4422                 }
4423                 if (found_type > min_type) {
4424                         del_item = 1;
4425                 } else {
4426                         if (item_end < new_size)
4427                                 break;
4428                         if (found_key.offset >= new_size)
4429                                 del_item = 1;
4430                         else
4431                                 del_item = 0;
4432                 }
4433                 found_extent = 0;
4434                 /* FIXME, shrink the extent if the ref count is only 1 */
4435                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4436                         goto delete;
4437
4438                 if (del_item)
4439                         last_size = found_key.offset;
4440                 else
4441                         last_size = new_size;
4442
4443                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4444                         u64 num_dec;
4445                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4446                         if (!del_item) {
4447                                 u64 orig_num_bytes =
4448                                         btrfs_file_extent_num_bytes(leaf, fi);
4449                                 extent_num_bytes = ALIGN(new_size -
4450                                                 found_key.offset,
4451                                                 root->sectorsize);
4452                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4453                                                          extent_num_bytes);
4454                                 num_dec = (orig_num_bytes -
4455                                            extent_num_bytes);
4456                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4457                                              &root->state) &&
4458                                     extent_start != 0)
4459                                         inode_sub_bytes(inode, num_dec);
4460                                 btrfs_mark_buffer_dirty(leaf);
4461                         } else {
4462                                 extent_num_bytes =
4463                                         btrfs_file_extent_disk_num_bytes(leaf,
4464                                                                          fi);
4465                                 extent_offset = found_key.offset -
4466                                         btrfs_file_extent_offset(leaf, fi);
4467
4468                                 /* FIXME blocksize != 4096 */
4469                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4470                                 if (extent_start != 0) {
4471                                         found_extent = 1;
4472                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4473                                                      &root->state))
4474                                                 inode_sub_bytes(inode, num_dec);
4475                                 }
4476                         }
4477                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4478                         /*
4479                          * we can't truncate inline items that have had
4480                          * special encodings
4481                          */
4482                         if (!del_item &&
4483                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4484                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4485
4486                                 /*
4487                                  * Need to release path in order to truncate a
4488                                  * compressed extent. So delete any accumulated
4489                                  * extent items so far.
4490                                  */
4491                                 if (btrfs_file_extent_compression(leaf, fi) !=
4492                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4493                                         err = btrfs_del_items(trans, root, path,
4494                                                               pending_del_slot,
4495                                                               pending_del_nr);
4496                                         if (err) {
4497                                                 btrfs_abort_transaction(trans,
4498                                                                         root,
4499                                                                         err);
4500                                                 goto error;
4501                                         }
4502                                         pending_del_nr = 0;
4503                                 }
4504
4505                                 err = truncate_inline_extent(inode, path,
4506                                                              &found_key,
4507                                                              item_end,
4508                                                              new_size);
4509                                 if (err) {
4510                                         btrfs_abort_transaction(trans,
4511                                                                 root, err);
4512                                         goto error;
4513                                 }
4514                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4515                                             &root->state)) {
4516                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4517                         }
4518                 }
4519 delete:
4520                 if (del_item) {
4521                         if (!pending_del_nr) {
4522                                 /* no pending yet, add ourselves */
4523                                 pending_del_slot = path->slots[0];
4524                                 pending_del_nr = 1;
4525                         } else if (pending_del_nr &&
4526                                    path->slots[0] + 1 == pending_del_slot) {
4527                                 /* hop on the pending chunk */
4528                                 pending_del_nr++;
4529                                 pending_del_slot = path->slots[0];
4530                         } else {
4531                                 BUG();
4532                         }
4533                 } else {
4534                         break;
4535                 }
4536                 should_throttle = 0;
4537
4538                 if (found_extent &&
4539                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4540                      root == root->fs_info->tree_root)) {
4541                         btrfs_set_path_blocking(path);
4542                         bytes_deleted += extent_num_bytes;
4543                         ret = btrfs_free_extent(trans, root, extent_start,
4544                                                 extent_num_bytes, 0,
4545                                                 btrfs_header_owner(leaf),
4546                                                 ino, extent_offset);
4547                         BUG_ON(ret);
4548                         if (btrfs_should_throttle_delayed_refs(trans, root))
4549                                 btrfs_async_run_delayed_refs(root,
4550                                         trans->delayed_ref_updates * 2, 0);
4551                         if (be_nice) {
4552                                 if (truncate_space_check(trans, root,
4553                                                          extent_num_bytes)) {
4554                                         should_end = 1;
4555                                 }
4556                                 if (btrfs_should_throttle_delayed_refs(trans,
4557                                                                        root)) {
4558                                         should_throttle = 1;
4559                                 }
4560                         }
4561                 }
4562
4563                 if (found_type == BTRFS_INODE_ITEM_KEY)
4564                         break;
4565
4566                 if (path->slots[0] == 0 ||
4567                     path->slots[0] != pending_del_slot ||
4568                     should_throttle || should_end) {
4569                         if (pending_del_nr) {
4570                                 ret = btrfs_del_items(trans, root, path,
4571                                                 pending_del_slot,
4572                                                 pending_del_nr);
4573                                 if (ret) {
4574                                         btrfs_abort_transaction(trans,
4575                                                                 root, ret);
4576                                         goto error;
4577                                 }
4578                                 pending_del_nr = 0;
4579                         }
4580                         btrfs_release_path(path);
4581                         if (should_throttle) {
4582                                 unsigned long updates = trans->delayed_ref_updates;
4583                                 if (updates) {
4584                                         trans->delayed_ref_updates = 0;
4585                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4586                                         if (ret && !err)
4587                                                 err = ret;
4588                                 }
4589                         }
4590                         /*
4591                          * if we failed to refill our space rsv, bail out
4592                          * and let the transaction restart
4593                          */
4594                         if (should_end) {
4595                                 err = -EAGAIN;
4596                                 goto error;
4597                         }
4598                         goto search_again;
4599                 } else {
4600                         path->slots[0]--;
4601                 }
4602         }
4603 out:
4604         if (pending_del_nr) {
4605                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4606                                       pending_del_nr);
4607                 if (ret)
4608                         btrfs_abort_transaction(trans, root, ret);
4609         }
4610 error:
4611         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4612                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4613
4614         btrfs_free_path(path);
4615
4616         if (be_nice && bytes_deleted > 32 * 1024 * 1024) {
4617                 unsigned long updates = trans->delayed_ref_updates;
4618                 if (updates) {
4619                         trans->delayed_ref_updates = 0;
4620                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4621                         if (ret && !err)
4622                                 err = ret;
4623                 }
4624         }
4625         return err;
4626 }
4627
4628 /*
4629  * btrfs_truncate_page - read, zero a chunk and write a page
4630  * @inode - inode that we're zeroing
4631  * @from - the offset to start zeroing
4632  * @len - the length to zero, 0 to zero the entire range respective to the
4633  *      offset
4634  * @front - zero up to the offset instead of from the offset on
4635  *
4636  * This will find the page for the "from" offset and cow the page and zero the
4637  * part we want to zero.  This is used with truncate and hole punching.
4638  */
4639 int btrfs_truncate_page(struct inode *inode, loff_t from, loff_t len,
4640                         int front)
4641 {
4642         struct address_space *mapping = inode->i_mapping;
4643         struct btrfs_root *root = BTRFS_I(inode)->root;
4644         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4645         struct btrfs_ordered_extent *ordered;
4646         struct extent_state *cached_state = NULL;
4647         char *kaddr;
4648         u32 blocksize = root->sectorsize;
4649         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4650         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4651         struct page *page;
4652         gfp_t mask = btrfs_alloc_write_mask(mapping);
4653         int ret = 0;
4654         u64 page_start;
4655         u64 page_end;
4656
4657         if ((offset & (blocksize - 1)) == 0 &&
4658             (!len || ((len & (blocksize - 1)) == 0)))
4659                 goto out;
4660         ret = btrfs_delalloc_reserve_space(inode,
4661                         round_down(from, PAGE_CACHE_SIZE), PAGE_CACHE_SIZE);
4662         if (ret)
4663                 goto out;
4664
4665 again:
4666         page = find_or_create_page(mapping, index, mask);
4667         if (!page) {
4668                 btrfs_delalloc_release_space(inode,
4669                                 round_down(from, PAGE_CACHE_SIZE),
4670                                 PAGE_CACHE_SIZE);
4671                 ret = -ENOMEM;
4672                 goto out;
4673         }
4674
4675         page_start = page_offset(page);
4676         page_end = page_start + PAGE_CACHE_SIZE - 1;
4677
4678         if (!PageUptodate(page)) {
4679                 ret = btrfs_readpage(NULL, page);
4680                 lock_page(page);
4681                 if (page->mapping != mapping) {
4682                         unlock_page(page);
4683                         page_cache_release(page);
4684                         goto again;
4685                 }
4686                 if (!PageUptodate(page)) {
4687                         ret = -EIO;
4688                         goto out_unlock;
4689                 }
4690         }
4691         wait_on_page_writeback(page);
4692
4693         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
4694         set_page_extent_mapped(page);
4695
4696         ordered = btrfs_lookup_ordered_extent(inode, page_start);
4697         if (ordered) {
4698                 unlock_extent_cached(io_tree, page_start, page_end,
4699                                      &cached_state, GFP_NOFS);
4700                 unlock_page(page);
4701                 page_cache_release(page);
4702                 btrfs_start_ordered_extent(inode, ordered, 1);
4703                 btrfs_put_ordered_extent(ordered);
4704                 goto again;
4705         }
4706
4707         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
4708                           EXTENT_DIRTY | EXTENT_DELALLOC |
4709                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4710                           0, 0, &cached_state, GFP_NOFS);
4711
4712         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
4713                                         &cached_state);
4714         if (ret) {
4715                 unlock_extent_cached(io_tree, page_start, page_end,
4716                                      &cached_state, GFP_NOFS);
4717                 goto out_unlock;
4718         }
4719
4720         if (offset != PAGE_CACHE_SIZE) {
4721                 if (!len)
4722                         len = PAGE_CACHE_SIZE - offset;
4723                 kaddr = kmap(page);
4724                 if (front)
4725                         memset(kaddr, 0, offset);
4726                 else
4727                         memset(kaddr + offset, 0, len);
4728                 flush_dcache_page(page);
4729                 kunmap(page);
4730         }
4731         ClearPageChecked(page);
4732         set_page_dirty(page);
4733         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
4734                              GFP_NOFS);
4735
4736 out_unlock:
4737         if (ret)
4738                 btrfs_delalloc_release_space(inode, page_start,
4739                                              PAGE_CACHE_SIZE);
4740         unlock_page(page);
4741         page_cache_release(page);
4742 out:
4743         return ret;
4744 }
4745
4746 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4747                              u64 offset, u64 len)
4748 {
4749         struct btrfs_trans_handle *trans;
4750         int ret;
4751
4752         /*
4753          * Still need to make sure the inode looks like it's been updated so
4754          * that any holes get logged if we fsync.
4755          */
4756         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4757                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4758                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4759                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4760                 return 0;
4761         }
4762
4763         /*
4764          * 1 - for the one we're dropping
4765          * 1 - for the one we're adding
4766          * 1 - for updating the inode.
4767          */
4768         trans = btrfs_start_transaction(root, 3);
4769         if (IS_ERR(trans))
4770                 return PTR_ERR(trans);
4771
4772         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4773         if (ret) {
4774                 btrfs_abort_transaction(trans, root, ret);
4775                 btrfs_end_transaction(trans, root);
4776                 return ret;
4777         }
4778
4779         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4780                                        0, 0, len, 0, len, 0, 0, 0);
4781         if (ret)
4782                 btrfs_abort_transaction(trans, root, ret);
4783         else
4784                 btrfs_update_inode(trans, root, inode);
4785         btrfs_end_transaction(trans, root);
4786         return ret;
4787 }
4788
4789 /*
4790  * This function puts in dummy file extents for the area we're creating a hole
4791  * for.  So if we are truncating this file to a larger size we need to insert
4792  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4793  * the range between oldsize and size
4794  */
4795 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4796 {
4797         struct btrfs_root *root = BTRFS_I(inode)->root;
4798         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4799         struct extent_map *em = NULL;
4800         struct extent_state *cached_state = NULL;
4801         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4802         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4803         u64 block_end = ALIGN(size, root->sectorsize);
4804         u64 last_byte;
4805         u64 cur_offset;
4806         u64 hole_size;
4807         int err = 0;
4808
4809         /*
4810          * If our size started in the middle of a page we need to zero out the
4811          * rest of the page before we expand the i_size, otherwise we could
4812          * expose stale data.
4813          */
4814         err = btrfs_truncate_page(inode, oldsize, 0, 0);
4815         if (err)
4816                 return err;
4817
4818         if (size <= hole_start)
4819                 return 0;
4820
4821         while (1) {
4822                 struct btrfs_ordered_extent *ordered;
4823
4824                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4825                                  &cached_state);
4826                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4827                                                      block_end - hole_start);
4828                 if (!ordered)
4829                         break;
4830                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4831                                      &cached_state, GFP_NOFS);
4832                 btrfs_start_ordered_extent(inode, ordered, 1);
4833                 btrfs_put_ordered_extent(ordered);
4834         }
4835
4836         cur_offset = hole_start;
4837         while (1) {
4838                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4839                                 block_end - cur_offset, 0);
4840                 if (IS_ERR(em)) {
4841                         err = PTR_ERR(em);
4842                         em = NULL;
4843                         break;
4844                 }
4845                 last_byte = min(extent_map_end(em), block_end);
4846                 last_byte = ALIGN(last_byte , root->sectorsize);
4847                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4848                         struct extent_map *hole_em;
4849                         hole_size = last_byte - cur_offset;
4850
4851                         err = maybe_insert_hole(root, inode, cur_offset,
4852                                                 hole_size);
4853                         if (err)
4854                                 break;
4855                         btrfs_drop_extent_cache(inode, cur_offset,
4856                                                 cur_offset + hole_size - 1, 0);
4857                         hole_em = alloc_extent_map();
4858                         if (!hole_em) {
4859                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4860                                         &BTRFS_I(inode)->runtime_flags);
4861                                 goto next;
4862                         }
4863                         hole_em->start = cur_offset;
4864                         hole_em->len = hole_size;
4865                         hole_em->orig_start = cur_offset;
4866
4867                         hole_em->block_start = EXTENT_MAP_HOLE;
4868                         hole_em->block_len = 0;
4869                         hole_em->orig_block_len = 0;
4870                         hole_em->ram_bytes = hole_size;
4871                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4872                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4873                         hole_em->generation = root->fs_info->generation;
4874
4875                         while (1) {
4876                                 write_lock(&em_tree->lock);
4877                                 err = add_extent_mapping(em_tree, hole_em, 1);
4878                                 write_unlock(&em_tree->lock);
4879                                 if (err != -EEXIST)
4880                                         break;
4881                                 btrfs_drop_extent_cache(inode, cur_offset,
4882                                                         cur_offset +
4883                                                         hole_size - 1, 0);
4884                         }
4885                         free_extent_map(hole_em);
4886                 }
4887 next:
4888                 free_extent_map(em);
4889                 em = NULL;
4890                 cur_offset = last_byte;
4891                 if (cur_offset >= block_end)
4892                         break;
4893         }
4894         free_extent_map(em);
4895         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4896                              GFP_NOFS);
4897         return err;
4898 }
4899
4900 static int wait_snapshoting_atomic_t(atomic_t *a)
4901 {
4902         schedule();
4903         return 0;
4904 }
4905
4906 static void wait_for_snapshot_creation(struct btrfs_root *root)
4907 {
4908         while (true) {
4909                 int ret;
4910
4911                 ret = btrfs_start_write_no_snapshoting(root);
4912                 if (ret)
4913                         break;
4914                 wait_on_atomic_t(&root->will_be_snapshoted,
4915                                  wait_snapshoting_atomic_t,
4916                                  TASK_UNINTERRUPTIBLE);
4917         }
4918 }
4919
4920 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4921 {
4922         struct btrfs_root *root = BTRFS_I(inode)->root;
4923         struct btrfs_trans_handle *trans;
4924         loff_t oldsize = i_size_read(inode);
4925         loff_t newsize = attr->ia_size;
4926         int mask = attr->ia_valid;
4927         int ret;
4928
4929         /*
4930          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4931          * special case where we need to update the times despite not having
4932          * these flags set.  For all other operations the VFS set these flags
4933          * explicitly if it wants a timestamp update.
4934          */
4935         if (newsize != oldsize) {
4936                 inode_inc_iversion(inode);
4937                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4938                         inode->i_ctime = inode->i_mtime =
4939                                 current_fs_time(inode->i_sb);
4940         }
4941
4942         if (newsize > oldsize) {
4943                 truncate_pagecache(inode, newsize);
4944                 /*
4945                  * Don't do an expanding truncate while snapshoting is ongoing.
4946                  * This is to ensure the snapshot captures a fully consistent
4947                  * state of this file - if the snapshot captures this expanding
4948                  * truncation, it must capture all writes that happened before
4949                  * this truncation.
4950                  */
4951                 wait_for_snapshot_creation(root);
4952                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4953                 if (ret) {
4954                         btrfs_end_write_no_snapshoting(root);
4955                         return ret;
4956                 }
4957
4958                 trans = btrfs_start_transaction(root, 1);
4959                 if (IS_ERR(trans)) {
4960                         btrfs_end_write_no_snapshoting(root);
4961                         return PTR_ERR(trans);
4962                 }
4963
4964                 i_size_write(inode, newsize);
4965                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4966                 ret = btrfs_update_inode(trans, root, inode);
4967                 btrfs_end_write_no_snapshoting(root);
4968                 btrfs_end_transaction(trans, root);
4969         } else {
4970
4971                 /*
4972                  * We're truncating a file that used to have good data down to
4973                  * zero. Make sure it gets into the ordered flush list so that
4974                  * any new writes get down to disk quickly.
4975                  */
4976                 if (newsize == 0)
4977                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4978                                 &BTRFS_I(inode)->runtime_flags);
4979
4980                 /*
4981                  * 1 for the orphan item we're going to add
4982                  * 1 for the orphan item deletion.
4983                  */
4984                 trans = btrfs_start_transaction(root, 2);
4985                 if (IS_ERR(trans))
4986                         return PTR_ERR(trans);
4987
4988                 /*
4989                  * We need to do this in case we fail at _any_ point during the
4990                  * actual truncate.  Once we do the truncate_setsize we could
4991                  * invalidate pages which forces any outstanding ordered io to
4992                  * be instantly completed which will give us extents that need
4993                  * to be truncated.  If we fail to get an orphan inode down we
4994                  * could have left over extents that were never meant to live,
4995                  * so we need to garuntee from this point on that everything
4996                  * will be consistent.
4997                  */
4998                 ret = btrfs_orphan_add(trans, inode);
4999                 btrfs_end_transaction(trans, root);
5000                 if (ret)
5001                         return ret;
5002
5003                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5004                 truncate_setsize(inode, newsize);
5005
5006                 /* Disable nonlocked read DIO to avoid the end less truncate */
5007                 btrfs_inode_block_unlocked_dio(inode);
5008                 inode_dio_wait(inode);
5009                 btrfs_inode_resume_unlocked_dio(inode);
5010
5011                 ret = btrfs_truncate(inode);
5012                 if (ret && inode->i_nlink) {
5013                         int err;
5014
5015                         /*
5016                          * failed to truncate, disk_i_size is only adjusted down
5017                          * as we remove extents, so it should represent the true
5018                          * size of the inode, so reset the in memory size and
5019                          * delete our orphan entry.
5020                          */
5021                         trans = btrfs_join_transaction(root);
5022                         if (IS_ERR(trans)) {
5023                                 btrfs_orphan_del(NULL, inode);
5024                                 return ret;
5025                         }
5026                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5027                         err = btrfs_orphan_del(trans, inode);
5028                         if (err)
5029                                 btrfs_abort_transaction(trans, root, err);
5030                         btrfs_end_transaction(trans, root);
5031                 }
5032         }
5033
5034         return ret;
5035 }
5036
5037 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5038 {
5039         struct inode *inode = d_inode(dentry);
5040         struct btrfs_root *root = BTRFS_I(inode)->root;
5041         int err;
5042
5043         if (btrfs_root_readonly(root))
5044                 return -EROFS;
5045
5046         err = inode_change_ok(inode, attr);
5047         if (err)
5048                 return err;
5049
5050         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5051                 err = btrfs_setsize(inode, attr);
5052                 if (err)
5053                         return err;
5054         }
5055
5056         if (attr->ia_valid) {
5057                 setattr_copy(inode, attr);
5058                 inode_inc_iversion(inode);
5059                 err = btrfs_dirty_inode(inode);
5060
5061                 if (!err && attr->ia_valid & ATTR_MODE)
5062                         err = posix_acl_chmod(inode, inode->i_mode);
5063         }
5064
5065         return err;
5066 }
5067
5068 /*
5069  * While truncating the inode pages during eviction, we get the VFS calling
5070  * btrfs_invalidatepage() against each page of the inode. This is slow because
5071  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5072  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5073  * extent_state structures over and over, wasting lots of time.
5074  *
5075  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5076  * those expensive operations on a per page basis and do only the ordered io
5077  * finishing, while we release here the extent_map and extent_state structures,
5078  * without the excessive merging and splitting.
5079  */
5080 static void evict_inode_truncate_pages(struct inode *inode)
5081 {
5082         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5083         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5084         struct rb_node *node;
5085
5086         ASSERT(inode->i_state & I_FREEING);
5087         truncate_inode_pages_final(&inode->i_data);
5088
5089         write_lock(&map_tree->lock);
5090         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5091                 struct extent_map *em;
5092
5093                 node = rb_first(&map_tree->map);
5094                 em = rb_entry(node, struct extent_map, rb_node);
5095                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5096                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5097                 remove_extent_mapping(map_tree, em);
5098                 free_extent_map(em);
5099                 if (need_resched()) {
5100                         write_unlock(&map_tree->lock);
5101                         cond_resched();
5102                         write_lock(&map_tree->lock);
5103                 }
5104         }
5105         write_unlock(&map_tree->lock);
5106
5107         /*
5108          * Keep looping until we have no more ranges in the io tree.
5109          * We can have ongoing bios started by readpages (called from readahead)
5110          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5111          * still in progress (unlocked the pages in the bio but did not yet
5112          * unlocked the ranges in the io tree). Therefore this means some
5113          * ranges can still be locked and eviction started because before
5114          * submitting those bios, which are executed by a separate task (work
5115          * queue kthread), inode references (inode->i_count) were not taken
5116          * (which would be dropped in the end io callback of each bio).
5117          * Therefore here we effectively end up waiting for those bios and
5118          * anyone else holding locked ranges without having bumped the inode's
5119          * reference count - if we don't do it, when they access the inode's
5120          * io_tree to unlock a range it may be too late, leading to an
5121          * use-after-free issue.
5122          */
5123         spin_lock(&io_tree->lock);
5124         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5125                 struct extent_state *state;
5126                 struct extent_state *cached_state = NULL;
5127                 u64 start;
5128                 u64 end;
5129
5130                 node = rb_first(&io_tree->state);
5131                 state = rb_entry(node, struct extent_state, rb_node);
5132                 start = state->start;
5133                 end = state->end;
5134                 spin_unlock(&io_tree->lock);
5135
5136                 lock_extent_bits(io_tree, start, end, &cached_state);
5137
5138                 /*
5139                  * If still has DELALLOC flag, the extent didn't reach disk,
5140                  * and its reserved space won't be freed by delayed_ref.
5141                  * So we need to free its reserved space here.
5142                  * (Refer to comment in btrfs_invalidatepage, case 2)
5143                  *
5144                  * Note, end is the bytenr of last byte, so we need + 1 here.
5145                  */
5146                 if (state->state & EXTENT_DELALLOC)
5147                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5148
5149                 clear_extent_bit(io_tree, start, end,
5150                                  EXTENT_LOCKED | EXTENT_DIRTY |
5151                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5152                                  EXTENT_DEFRAG, 1, 1,
5153                                  &cached_state, GFP_NOFS);
5154
5155                 cond_resched();
5156                 spin_lock(&io_tree->lock);
5157         }
5158         spin_unlock(&io_tree->lock);
5159 }
5160
5161 void btrfs_evict_inode(struct inode *inode)
5162 {
5163         struct btrfs_trans_handle *trans;
5164         struct btrfs_root *root = BTRFS_I(inode)->root;
5165         struct btrfs_block_rsv *rsv, *global_rsv;
5166         int steal_from_global = 0;
5167         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5168         int ret;
5169
5170         trace_btrfs_inode_evict(inode);
5171
5172         evict_inode_truncate_pages(inode);
5173
5174         if (inode->i_nlink &&
5175             ((btrfs_root_refs(&root->root_item) != 0 &&
5176               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5177              btrfs_is_free_space_inode(inode)))
5178                 goto no_delete;
5179
5180         if (is_bad_inode(inode)) {
5181                 btrfs_orphan_del(NULL, inode);
5182                 goto no_delete;
5183         }
5184         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5185         if (!special_file(inode->i_mode))
5186                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5187
5188         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5189
5190         if (root->fs_info->log_root_recovering) {
5191                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5192                                  &BTRFS_I(inode)->runtime_flags));
5193                 goto no_delete;
5194         }
5195
5196         if (inode->i_nlink > 0) {
5197                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5198                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5199                 goto no_delete;
5200         }
5201
5202         ret = btrfs_commit_inode_delayed_inode(inode);
5203         if (ret) {
5204                 btrfs_orphan_del(NULL, inode);
5205                 goto no_delete;
5206         }
5207
5208         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5209         if (!rsv) {
5210                 btrfs_orphan_del(NULL, inode);
5211                 goto no_delete;
5212         }
5213         rsv->size = min_size;
5214         rsv->failfast = 1;
5215         global_rsv = &root->fs_info->global_block_rsv;
5216
5217         btrfs_i_size_write(inode, 0);
5218
5219         /*
5220          * This is a bit simpler than btrfs_truncate since we've already
5221          * reserved our space for our orphan item in the unlink, so we just
5222          * need to reserve some slack space in case we add bytes and update
5223          * inode item when doing the truncate.
5224          */
5225         while (1) {
5226                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5227                                              BTRFS_RESERVE_FLUSH_LIMIT);
5228
5229                 /*
5230                  * Try and steal from the global reserve since we will
5231                  * likely not use this space anyway, we want to try as
5232                  * hard as possible to get this to work.
5233                  */
5234                 if (ret)
5235                         steal_from_global++;
5236                 else
5237                         steal_from_global = 0;
5238                 ret = 0;
5239
5240                 /*
5241                  * steal_from_global == 0: we reserved stuff, hooray!
5242                  * steal_from_global == 1: we didn't reserve stuff, boo!
5243                  * steal_from_global == 2: we've committed, still not a lot of
5244                  * room but maybe we'll have room in the global reserve this
5245                  * time.
5246                  * steal_from_global == 3: abandon all hope!
5247                  */
5248                 if (steal_from_global > 2) {
5249                         btrfs_warn(root->fs_info,
5250                                 "Could not get space for a delete, will truncate on mount %d",
5251                                 ret);
5252                         btrfs_orphan_del(NULL, inode);
5253                         btrfs_free_block_rsv(root, rsv);
5254                         goto no_delete;
5255                 }
5256
5257                 trans = btrfs_join_transaction(root);
5258                 if (IS_ERR(trans)) {
5259                         btrfs_orphan_del(NULL, inode);
5260                         btrfs_free_block_rsv(root, rsv);
5261                         goto no_delete;
5262                 }
5263
5264                 /*
5265                  * We can't just steal from the global reserve, we need tomake
5266                  * sure there is room to do it, if not we need to commit and try
5267                  * again.
5268                  */
5269                 if (steal_from_global) {
5270                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5271                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5272                                                               min_size);
5273                         else
5274                                 ret = -ENOSPC;
5275                 }
5276
5277                 /*
5278                  * Couldn't steal from the global reserve, we have too much
5279                  * pending stuff built up, commit the transaction and try it
5280                  * again.
5281                  */
5282                 if (ret) {
5283                         ret = btrfs_commit_transaction(trans, root);
5284                         if (ret) {
5285                                 btrfs_orphan_del(NULL, inode);
5286                                 btrfs_free_block_rsv(root, rsv);
5287                                 goto no_delete;
5288                         }
5289                         continue;
5290                 } else {
5291                         steal_from_global = 0;
5292                 }
5293
5294                 trans->block_rsv = rsv;
5295
5296                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5297                 if (ret != -ENOSPC && ret != -EAGAIN)
5298                         break;
5299
5300                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5301                 btrfs_end_transaction(trans, root);
5302                 trans = NULL;
5303                 btrfs_btree_balance_dirty(root);
5304         }
5305
5306         btrfs_free_block_rsv(root, rsv);
5307
5308         /*
5309          * Errors here aren't a big deal, it just means we leave orphan items
5310          * in the tree.  They will be cleaned up on the next mount.
5311          */
5312         if (ret == 0) {
5313                 trans->block_rsv = root->orphan_block_rsv;
5314                 btrfs_orphan_del(trans, inode);
5315         } else {
5316                 btrfs_orphan_del(NULL, inode);
5317         }
5318
5319         trans->block_rsv = &root->fs_info->trans_block_rsv;
5320         if (!(root == root->fs_info->tree_root ||
5321               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5322                 btrfs_return_ino(root, btrfs_ino(inode));
5323
5324         btrfs_end_transaction(trans, root);
5325         btrfs_btree_balance_dirty(root);
5326 no_delete:
5327         btrfs_remove_delayed_node(inode);
5328         clear_inode(inode);
5329         return;
5330 }
5331
5332 /*
5333  * this returns the key found in the dir entry in the location pointer.
5334  * If no dir entries were found, location->objectid is 0.
5335  */
5336 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5337                                struct btrfs_key *location)
5338 {
5339         const char *name = dentry->d_name.name;
5340         int namelen = dentry->d_name.len;
5341         struct btrfs_dir_item *di;
5342         struct btrfs_path *path;
5343         struct btrfs_root *root = BTRFS_I(dir)->root;
5344         int ret = 0;
5345
5346         path = btrfs_alloc_path();
5347         if (!path)
5348                 return -ENOMEM;
5349
5350         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5351                                     namelen, 0);
5352         if (IS_ERR(di))
5353                 ret = PTR_ERR(di);
5354
5355         if (IS_ERR_OR_NULL(di))
5356                 goto out_err;
5357
5358         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5359 out:
5360         btrfs_free_path(path);
5361         return ret;
5362 out_err:
5363         location->objectid = 0;
5364         goto out;
5365 }
5366
5367 /*
5368  * when we hit a tree root in a directory, the btrfs part of the inode
5369  * needs to be changed to reflect the root directory of the tree root.  This
5370  * is kind of like crossing a mount point.
5371  */
5372 static int fixup_tree_root_location(struct btrfs_root *root,
5373                                     struct inode *dir,
5374                                     struct dentry *dentry,
5375                                     struct btrfs_key *location,
5376                                     struct btrfs_root **sub_root)
5377 {
5378         struct btrfs_path *path;
5379         struct btrfs_root *new_root;
5380         struct btrfs_root_ref *ref;
5381         struct extent_buffer *leaf;
5382         struct btrfs_key key;
5383         int ret;
5384         int err = 0;
5385
5386         path = btrfs_alloc_path();
5387         if (!path) {
5388                 err = -ENOMEM;
5389                 goto out;
5390         }
5391
5392         err = -ENOENT;
5393         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5394         key.type = BTRFS_ROOT_REF_KEY;
5395         key.offset = location->objectid;
5396
5397         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5398                                 0, 0);
5399         if (ret) {
5400                 if (ret < 0)
5401                         err = ret;
5402                 goto out;
5403         }
5404
5405         leaf = path->nodes[0];
5406         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5407         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5408             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5409                 goto out;
5410
5411         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5412                                    (unsigned long)(ref + 1),
5413                                    dentry->d_name.len);
5414         if (ret)
5415                 goto out;
5416
5417         btrfs_release_path(path);
5418
5419         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5420         if (IS_ERR(new_root)) {
5421                 err = PTR_ERR(new_root);
5422                 goto out;
5423         }
5424
5425         *sub_root = new_root;
5426         location->objectid = btrfs_root_dirid(&new_root->root_item);
5427         location->type = BTRFS_INODE_ITEM_KEY;
5428         location->offset = 0;
5429         err = 0;
5430 out:
5431         btrfs_free_path(path);
5432         return err;
5433 }
5434
5435 static void inode_tree_add(struct inode *inode)
5436 {
5437         struct btrfs_root *root = BTRFS_I(inode)->root;
5438         struct btrfs_inode *entry;
5439         struct rb_node **p;
5440         struct rb_node *parent;
5441         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5442         u64 ino = btrfs_ino(inode);
5443
5444         if (inode_unhashed(inode))
5445                 return;
5446         parent = NULL;
5447         spin_lock(&root->inode_lock);
5448         p = &root->inode_tree.rb_node;
5449         while (*p) {
5450                 parent = *p;
5451                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5452
5453                 if (ino < btrfs_ino(&entry->vfs_inode))
5454                         p = &parent->rb_left;
5455                 else if (ino > btrfs_ino(&entry->vfs_inode))
5456                         p = &parent->rb_right;
5457                 else {
5458                         WARN_ON(!(entry->vfs_inode.i_state &
5459                                   (I_WILL_FREE | I_FREEING)));
5460                         rb_replace_node(parent, new, &root->inode_tree);
5461                         RB_CLEAR_NODE(parent);
5462                         spin_unlock(&root->inode_lock);
5463                         return;
5464                 }
5465         }
5466         rb_link_node(new, parent, p);
5467         rb_insert_color(new, &root->inode_tree);
5468         spin_unlock(&root->inode_lock);
5469 }
5470
5471 static void inode_tree_del(struct inode *inode)
5472 {
5473         struct btrfs_root *root = BTRFS_I(inode)->root;
5474         int empty = 0;
5475
5476         spin_lock(&root->inode_lock);
5477         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5478                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5479                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5480                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5481         }
5482         spin_unlock(&root->inode_lock);
5483
5484         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5485                 synchronize_srcu(&root->fs_info->subvol_srcu);
5486                 spin_lock(&root->inode_lock);
5487                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5488                 spin_unlock(&root->inode_lock);
5489                 if (empty)
5490                         btrfs_add_dead_root(root);
5491         }
5492 }
5493
5494 void btrfs_invalidate_inodes(struct btrfs_root *root)
5495 {
5496         struct rb_node *node;
5497         struct rb_node *prev;
5498         struct btrfs_inode *entry;
5499         struct inode *inode;
5500         u64 objectid = 0;
5501
5502         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5503                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5504
5505         spin_lock(&root->inode_lock);
5506 again:
5507         node = root->inode_tree.rb_node;
5508         prev = NULL;
5509         while (node) {
5510                 prev = node;
5511                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5512
5513                 if (objectid < btrfs_ino(&entry->vfs_inode))
5514                         node = node->rb_left;
5515                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5516                         node = node->rb_right;
5517                 else
5518                         break;
5519         }
5520         if (!node) {
5521                 while (prev) {
5522                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5523                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5524                                 node = prev;
5525                                 break;
5526                         }
5527                         prev = rb_next(prev);
5528                 }
5529         }
5530         while (node) {
5531                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5532                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5533                 inode = igrab(&entry->vfs_inode);
5534                 if (inode) {
5535                         spin_unlock(&root->inode_lock);
5536                         if (atomic_read(&inode->i_count) > 1)
5537                                 d_prune_aliases(inode);
5538                         /*
5539                          * btrfs_drop_inode will have it removed from
5540                          * the inode cache when its usage count
5541                          * hits zero.
5542                          */
5543                         iput(inode);
5544                         cond_resched();
5545                         spin_lock(&root->inode_lock);
5546                         goto again;
5547                 }
5548
5549                 if (cond_resched_lock(&root->inode_lock))
5550                         goto again;
5551
5552                 node = rb_next(node);
5553         }
5554         spin_unlock(&root->inode_lock);
5555 }
5556
5557 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5558 {
5559         struct btrfs_iget_args *args = p;
5560         inode->i_ino = args->location->objectid;
5561         memcpy(&BTRFS_I(inode)->location, args->location,
5562                sizeof(*args->location));
5563         BTRFS_I(inode)->root = args->root;
5564         return 0;
5565 }
5566
5567 static int btrfs_find_actor(struct inode *inode, void *opaque)
5568 {
5569         struct btrfs_iget_args *args = opaque;
5570         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5571                 args->root == BTRFS_I(inode)->root;
5572 }
5573
5574 static struct inode *btrfs_iget_locked(struct super_block *s,
5575                                        struct btrfs_key *location,
5576                                        struct btrfs_root *root)
5577 {
5578         struct inode *inode;
5579         struct btrfs_iget_args args;
5580         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5581
5582         args.location = location;
5583         args.root = root;
5584
5585         inode = iget5_locked(s, hashval, btrfs_find_actor,
5586                              btrfs_init_locked_inode,
5587                              (void *)&args);
5588         return inode;
5589 }
5590
5591 /* Get an inode object given its location and corresponding root.
5592  * Returns in *is_new if the inode was read from disk
5593  */
5594 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5595                          struct btrfs_root *root, int *new)
5596 {
5597         struct inode *inode;
5598
5599         inode = btrfs_iget_locked(s, location, root);
5600         if (!inode)
5601                 return ERR_PTR(-ENOMEM);
5602
5603         if (inode->i_state & I_NEW) {
5604                 btrfs_read_locked_inode(inode);
5605                 if (!is_bad_inode(inode)) {
5606                         inode_tree_add(inode);
5607                         unlock_new_inode(inode);
5608                         if (new)
5609                                 *new = 1;
5610                 } else {
5611                         unlock_new_inode(inode);
5612                         iput(inode);
5613                         inode = ERR_PTR(-ESTALE);
5614                 }
5615         }
5616
5617         return inode;
5618 }
5619
5620 static struct inode *new_simple_dir(struct super_block *s,
5621                                     struct btrfs_key *key,
5622                                     struct btrfs_root *root)
5623 {
5624         struct inode *inode = new_inode(s);
5625
5626         if (!inode)
5627                 return ERR_PTR(-ENOMEM);
5628
5629         BTRFS_I(inode)->root = root;
5630         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5631         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5632
5633         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5634         inode->i_op = &btrfs_dir_ro_inode_operations;
5635         inode->i_fop = &simple_dir_operations;
5636         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5637         inode->i_mtime = CURRENT_TIME;
5638         inode->i_atime = inode->i_mtime;
5639         inode->i_ctime = inode->i_mtime;
5640         BTRFS_I(inode)->i_otime = inode->i_mtime;
5641
5642         return inode;
5643 }
5644
5645 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5646 {
5647         struct inode *inode;
5648         struct btrfs_root *root = BTRFS_I(dir)->root;
5649         struct btrfs_root *sub_root = root;
5650         struct btrfs_key location;
5651         int index;
5652         int ret = 0;
5653
5654         if (dentry->d_name.len > BTRFS_NAME_LEN)
5655                 return ERR_PTR(-ENAMETOOLONG);
5656
5657         ret = btrfs_inode_by_name(dir, dentry, &location);
5658         if (ret < 0)
5659                 return ERR_PTR(ret);
5660
5661         if (location.objectid == 0)
5662                 return ERR_PTR(-ENOENT);
5663
5664         if (location.type == BTRFS_INODE_ITEM_KEY) {
5665                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5666                 return inode;
5667         }
5668
5669         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5670
5671         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5672         ret = fixup_tree_root_location(root, dir, dentry,
5673                                        &location, &sub_root);
5674         if (ret < 0) {
5675                 if (ret != -ENOENT)
5676                         inode = ERR_PTR(ret);
5677                 else
5678                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5679         } else {
5680                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5681         }
5682         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5683
5684         if (!IS_ERR(inode) && root != sub_root) {
5685                 down_read(&root->fs_info->cleanup_work_sem);
5686                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5687                         ret = btrfs_orphan_cleanup(sub_root);
5688                 up_read(&root->fs_info->cleanup_work_sem);
5689                 if (ret) {
5690                         iput(inode);
5691                         inode = ERR_PTR(ret);
5692                 }
5693         }
5694
5695         return inode;
5696 }
5697
5698 static int btrfs_dentry_delete(const struct dentry *dentry)
5699 {
5700         struct btrfs_root *root;
5701         struct inode *inode = d_inode(dentry);
5702
5703         if (!inode && !IS_ROOT(dentry))
5704                 inode = d_inode(dentry->d_parent);
5705
5706         if (inode) {
5707                 root = BTRFS_I(inode)->root;
5708                 if (btrfs_root_refs(&root->root_item) == 0)
5709                         return 1;
5710
5711                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5712                         return 1;
5713         }
5714         return 0;
5715 }
5716
5717 static void btrfs_dentry_release(struct dentry *dentry)
5718 {
5719         kfree(dentry->d_fsdata);
5720 }
5721
5722 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5723                                    unsigned int flags)
5724 {
5725         struct inode *inode;
5726
5727         inode = btrfs_lookup_dentry(dir, dentry);
5728         if (IS_ERR(inode)) {
5729                 if (PTR_ERR(inode) == -ENOENT)
5730                         inode = NULL;
5731                 else
5732                         return ERR_CAST(inode);
5733         }
5734
5735         return d_splice_alias(inode, dentry);
5736 }
5737
5738 unsigned char btrfs_filetype_table[] = {
5739         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5740 };
5741
5742 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5743 {
5744         struct inode *inode = file_inode(file);
5745         struct btrfs_root *root = BTRFS_I(inode)->root;
5746         struct btrfs_item *item;
5747         struct btrfs_dir_item *di;
5748         struct btrfs_key key;
5749         struct btrfs_key found_key;
5750         struct btrfs_path *path;
5751         struct list_head ins_list;
5752         struct list_head del_list;
5753         int ret;
5754         struct extent_buffer *leaf;
5755         int slot;
5756         unsigned char d_type;
5757         int over = 0;
5758         u32 di_cur;
5759         u32 di_total;
5760         u32 di_len;
5761         int key_type = BTRFS_DIR_INDEX_KEY;
5762         char tmp_name[32];
5763         char *name_ptr;
5764         int name_len;
5765         int is_curr = 0;        /* ctx->pos points to the current index? */
5766
5767         /* FIXME, use a real flag for deciding about the key type */
5768         if (root->fs_info->tree_root == root)
5769                 key_type = BTRFS_DIR_ITEM_KEY;
5770
5771         if (!dir_emit_dots(file, ctx))
5772                 return 0;
5773
5774         path = btrfs_alloc_path();
5775         if (!path)
5776                 return -ENOMEM;
5777
5778         path->reada = 1;
5779
5780         if (key_type == BTRFS_DIR_INDEX_KEY) {
5781                 INIT_LIST_HEAD(&ins_list);
5782                 INIT_LIST_HEAD(&del_list);
5783                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5784         }
5785
5786         key.type = key_type;
5787         key.offset = ctx->pos;
5788         key.objectid = btrfs_ino(inode);
5789
5790         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5791         if (ret < 0)
5792                 goto err;
5793
5794         while (1) {
5795                 leaf = path->nodes[0];
5796                 slot = path->slots[0];
5797                 if (slot >= btrfs_header_nritems(leaf)) {
5798                         ret = btrfs_next_leaf(root, path);
5799                         if (ret < 0)
5800                                 goto err;
5801                         else if (ret > 0)
5802                                 break;
5803                         continue;
5804                 }
5805
5806                 item = btrfs_item_nr(slot);
5807                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5808
5809                 if (found_key.objectid != key.objectid)
5810                         break;
5811                 if (found_key.type != key_type)
5812                         break;
5813                 if (found_key.offset < ctx->pos)
5814                         goto next;
5815                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5816                     btrfs_should_delete_dir_index(&del_list,
5817                                                   found_key.offset))
5818                         goto next;
5819
5820                 ctx->pos = found_key.offset;
5821                 is_curr = 1;
5822
5823                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5824                 di_cur = 0;
5825                 di_total = btrfs_item_size(leaf, item);
5826
5827                 while (di_cur < di_total) {
5828                         struct btrfs_key location;
5829
5830                         if (verify_dir_item(root, leaf, di))
5831                                 break;
5832
5833                         name_len = btrfs_dir_name_len(leaf, di);
5834                         if (name_len <= sizeof(tmp_name)) {
5835                                 name_ptr = tmp_name;
5836                         } else {
5837                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5838                                 if (!name_ptr) {
5839                                         ret = -ENOMEM;
5840                                         goto err;
5841                                 }
5842                         }
5843                         read_extent_buffer(leaf, name_ptr,
5844                                            (unsigned long)(di + 1), name_len);
5845
5846                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5847                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5848
5849
5850                         /* is this a reference to our own snapshot? If so
5851                          * skip it.
5852                          *
5853                          * In contrast to old kernels, we insert the snapshot's
5854                          * dir item and dir index after it has been created, so
5855                          * we won't find a reference to our own snapshot. We
5856                          * still keep the following code for backward
5857                          * compatibility.
5858                          */
5859                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5860                             location.objectid == root->root_key.objectid) {
5861                                 over = 0;
5862                                 goto skip;
5863                         }
5864                         over = !dir_emit(ctx, name_ptr, name_len,
5865                                        location.objectid, d_type);
5866
5867 skip:
5868                         if (name_ptr != tmp_name)
5869                                 kfree(name_ptr);
5870
5871                         if (over)
5872                                 goto nopos;
5873                         di_len = btrfs_dir_name_len(leaf, di) +
5874                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5875                         di_cur += di_len;
5876                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5877                 }
5878 next:
5879                 path->slots[0]++;
5880         }
5881
5882         if (key_type == BTRFS_DIR_INDEX_KEY) {
5883                 if (is_curr)
5884                         ctx->pos++;
5885                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5886                 if (ret)
5887                         goto nopos;
5888         }
5889
5890         /* Reached end of directory/root. Bump pos past the last item. */
5891         ctx->pos++;
5892
5893         /*
5894          * Stop new entries from being returned after we return the last
5895          * entry.
5896          *
5897          * New directory entries are assigned a strictly increasing
5898          * offset.  This means that new entries created during readdir
5899          * are *guaranteed* to be seen in the future by that readdir.
5900          * This has broken buggy programs which operate on names as
5901          * they're returned by readdir.  Until we re-use freed offsets
5902          * we have this hack to stop new entries from being returned
5903          * under the assumption that they'll never reach this huge
5904          * offset.
5905          *
5906          * This is being careful not to overflow 32bit loff_t unless the
5907          * last entry requires it because doing so has broken 32bit apps
5908          * in the past.
5909          */
5910         if (key_type == BTRFS_DIR_INDEX_KEY) {
5911                 if (ctx->pos >= INT_MAX)
5912                         ctx->pos = LLONG_MAX;
5913                 else
5914                         ctx->pos = INT_MAX;
5915         }
5916 nopos:
5917         ret = 0;
5918 err:
5919         if (key_type == BTRFS_DIR_INDEX_KEY)
5920                 btrfs_put_delayed_items(&ins_list, &del_list);
5921         btrfs_free_path(path);
5922         return ret;
5923 }
5924
5925 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5926 {
5927         struct btrfs_root *root = BTRFS_I(inode)->root;
5928         struct btrfs_trans_handle *trans;
5929         int ret = 0;
5930         bool nolock = false;
5931
5932         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5933                 return 0;
5934
5935         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5936                 nolock = true;
5937
5938         if (wbc->sync_mode == WB_SYNC_ALL) {
5939                 if (nolock)
5940                         trans = btrfs_join_transaction_nolock(root);
5941                 else
5942                         trans = btrfs_join_transaction(root);
5943                 if (IS_ERR(trans))
5944                         return PTR_ERR(trans);
5945                 ret = btrfs_commit_transaction(trans, root);
5946         }
5947         return ret;
5948 }
5949
5950 /*
5951  * This is somewhat expensive, updating the tree every time the
5952  * inode changes.  But, it is most likely to find the inode in cache.
5953  * FIXME, needs more benchmarking...there are no reasons other than performance
5954  * to keep or drop this code.
5955  */
5956 static int btrfs_dirty_inode(struct inode *inode)
5957 {
5958         struct btrfs_root *root = BTRFS_I(inode)->root;
5959         struct btrfs_trans_handle *trans;
5960         int ret;
5961
5962         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5963                 return 0;
5964
5965         trans = btrfs_join_transaction(root);
5966         if (IS_ERR(trans))
5967                 return PTR_ERR(trans);
5968
5969         ret = btrfs_update_inode(trans, root, inode);
5970         if (ret && ret == -ENOSPC) {
5971                 /* whoops, lets try again with the full transaction */
5972                 btrfs_end_transaction(trans, root);
5973                 trans = btrfs_start_transaction(root, 1);
5974                 if (IS_ERR(trans))
5975                         return PTR_ERR(trans);
5976
5977                 ret = btrfs_update_inode(trans, root, inode);
5978         }
5979         btrfs_end_transaction(trans, root);
5980         if (BTRFS_I(inode)->delayed_node)
5981                 btrfs_balance_delayed_items(root);
5982
5983         return ret;
5984 }
5985
5986 /*
5987  * This is a copy of file_update_time.  We need this so we can return error on
5988  * ENOSPC for updating the inode in the case of file write and mmap writes.
5989  */
5990 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5991                              int flags)
5992 {
5993         struct btrfs_root *root = BTRFS_I(inode)->root;
5994
5995         if (btrfs_root_readonly(root))
5996                 return -EROFS;
5997
5998         if (flags & S_VERSION)
5999                 inode_inc_iversion(inode);
6000         if (flags & S_CTIME)
6001                 inode->i_ctime = *now;
6002         if (flags & S_MTIME)
6003                 inode->i_mtime = *now;
6004         if (flags & S_ATIME)
6005                 inode->i_atime = *now;
6006         return btrfs_dirty_inode(inode);
6007 }
6008
6009 /*
6010  * find the highest existing sequence number in a directory
6011  * and then set the in-memory index_cnt variable to reflect
6012  * free sequence numbers
6013  */
6014 static int btrfs_set_inode_index_count(struct inode *inode)
6015 {
6016         struct btrfs_root *root = BTRFS_I(inode)->root;
6017         struct btrfs_key key, found_key;
6018         struct btrfs_path *path;
6019         struct extent_buffer *leaf;
6020         int ret;
6021
6022         key.objectid = btrfs_ino(inode);
6023         key.type = BTRFS_DIR_INDEX_KEY;
6024         key.offset = (u64)-1;
6025
6026         path = btrfs_alloc_path();
6027         if (!path)
6028                 return -ENOMEM;
6029
6030         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6031         if (ret < 0)
6032                 goto out;
6033         /* FIXME: we should be able to handle this */
6034         if (ret == 0)
6035                 goto out;
6036         ret = 0;
6037
6038         /*
6039          * MAGIC NUMBER EXPLANATION:
6040          * since we search a directory based on f_pos we have to start at 2
6041          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6042          * else has to start at 2
6043          */
6044         if (path->slots[0] == 0) {
6045                 BTRFS_I(inode)->index_cnt = 2;
6046                 goto out;
6047         }
6048
6049         path->slots[0]--;
6050
6051         leaf = path->nodes[0];
6052         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6053
6054         if (found_key.objectid != btrfs_ino(inode) ||
6055             found_key.type != BTRFS_DIR_INDEX_KEY) {
6056                 BTRFS_I(inode)->index_cnt = 2;
6057                 goto out;
6058         }
6059
6060         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6061 out:
6062         btrfs_free_path(path);
6063         return ret;
6064 }
6065
6066 /*
6067  * helper to find a free sequence number in a given directory.  This current
6068  * code is very simple, later versions will do smarter things in the btree
6069  */
6070 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6071 {
6072         int ret = 0;
6073
6074         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6075                 ret = btrfs_inode_delayed_dir_index_count(dir);
6076                 if (ret) {
6077                         ret = btrfs_set_inode_index_count(dir);
6078                         if (ret)
6079                                 return ret;
6080                 }
6081         }
6082
6083         *index = BTRFS_I(dir)->index_cnt;
6084         BTRFS_I(dir)->index_cnt++;
6085
6086         return ret;
6087 }
6088
6089 static int btrfs_insert_inode_locked(struct inode *inode)
6090 {
6091         struct btrfs_iget_args args;
6092         args.location = &BTRFS_I(inode)->location;
6093         args.root = BTRFS_I(inode)->root;
6094
6095         return insert_inode_locked4(inode,
6096                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6097                    btrfs_find_actor, &args);
6098 }
6099
6100 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6101                                      struct btrfs_root *root,
6102                                      struct inode *dir,
6103                                      const char *name, int name_len,
6104                                      u64 ref_objectid, u64 objectid,
6105                                      umode_t mode, u64 *index)
6106 {
6107         struct inode *inode;
6108         struct btrfs_inode_item *inode_item;
6109         struct btrfs_key *location;
6110         struct btrfs_path *path;
6111         struct btrfs_inode_ref *ref;
6112         struct btrfs_key key[2];
6113         u32 sizes[2];
6114         int nitems = name ? 2 : 1;
6115         unsigned long ptr;
6116         int ret;
6117
6118         path = btrfs_alloc_path();
6119         if (!path)
6120                 return ERR_PTR(-ENOMEM);
6121
6122         inode = new_inode(root->fs_info->sb);
6123         if (!inode) {
6124                 btrfs_free_path(path);
6125                 return ERR_PTR(-ENOMEM);
6126         }
6127
6128         /*
6129          * O_TMPFILE, set link count to 0, so that after this point,
6130          * we fill in an inode item with the correct link count.
6131          */
6132         if (!name)
6133                 set_nlink(inode, 0);
6134
6135         /*
6136          * we have to initialize this early, so we can reclaim the inode
6137          * number if we fail afterwards in this function.
6138          */
6139         inode->i_ino = objectid;
6140
6141         if (dir && name) {
6142                 trace_btrfs_inode_request(dir);
6143
6144                 ret = btrfs_set_inode_index(dir, index);
6145                 if (ret) {
6146                         btrfs_free_path(path);
6147                         iput(inode);
6148                         return ERR_PTR(ret);
6149                 }
6150         } else if (dir) {
6151                 *index = 0;
6152         }
6153         /*
6154          * index_cnt is ignored for everything but a dir,
6155          * btrfs_get_inode_index_count has an explanation for the magic
6156          * number
6157          */
6158         BTRFS_I(inode)->index_cnt = 2;
6159         BTRFS_I(inode)->dir_index = *index;
6160         BTRFS_I(inode)->root = root;
6161         BTRFS_I(inode)->generation = trans->transid;
6162         inode->i_generation = BTRFS_I(inode)->generation;
6163
6164         /*
6165          * We could have gotten an inode number from somebody who was fsynced
6166          * and then removed in this same transaction, so let's just set full
6167          * sync since it will be a full sync anyway and this will blow away the
6168          * old info in the log.
6169          */
6170         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6171
6172         key[0].objectid = objectid;
6173         key[0].type = BTRFS_INODE_ITEM_KEY;
6174         key[0].offset = 0;
6175
6176         sizes[0] = sizeof(struct btrfs_inode_item);
6177
6178         if (name) {
6179                 /*
6180                  * Start new inodes with an inode_ref. This is slightly more
6181                  * efficient for small numbers of hard links since they will
6182                  * be packed into one item. Extended refs will kick in if we
6183                  * add more hard links than can fit in the ref item.
6184                  */
6185                 key[1].objectid = objectid;
6186                 key[1].type = BTRFS_INODE_REF_KEY;
6187                 key[1].offset = ref_objectid;
6188
6189                 sizes[1] = name_len + sizeof(*ref);
6190         }
6191
6192         location = &BTRFS_I(inode)->location;
6193         location->objectid = objectid;
6194         location->offset = 0;
6195         location->type = BTRFS_INODE_ITEM_KEY;
6196
6197         ret = btrfs_insert_inode_locked(inode);
6198         if (ret < 0)
6199                 goto fail;
6200
6201         path->leave_spinning = 1;
6202         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6203         if (ret != 0)
6204                 goto fail_unlock;
6205
6206         inode_init_owner(inode, dir, mode);
6207         inode_set_bytes(inode, 0);
6208
6209         inode->i_mtime = CURRENT_TIME;
6210         inode->i_atime = inode->i_mtime;
6211         inode->i_ctime = inode->i_mtime;
6212         BTRFS_I(inode)->i_otime = inode->i_mtime;
6213
6214         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6215                                   struct btrfs_inode_item);
6216         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6217                              sizeof(*inode_item));
6218         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6219
6220         if (name) {
6221                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6222                                      struct btrfs_inode_ref);
6223                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6224                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6225                 ptr = (unsigned long)(ref + 1);
6226                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6227         }
6228
6229         btrfs_mark_buffer_dirty(path->nodes[0]);
6230         btrfs_free_path(path);
6231
6232         btrfs_inherit_iflags(inode, dir);
6233
6234         if (S_ISREG(mode)) {
6235                 if (btrfs_test_opt(root, NODATASUM))
6236                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6237                 if (btrfs_test_opt(root, NODATACOW))
6238                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6239                                 BTRFS_INODE_NODATASUM;
6240         }
6241
6242         inode_tree_add(inode);
6243
6244         trace_btrfs_inode_new(inode);
6245         btrfs_set_inode_last_trans(trans, inode);
6246
6247         btrfs_update_root_times(trans, root);
6248
6249         ret = btrfs_inode_inherit_props(trans, inode, dir);
6250         if (ret)
6251                 btrfs_err(root->fs_info,
6252                           "error inheriting props for ino %llu (root %llu): %d",
6253                           btrfs_ino(inode), root->root_key.objectid, ret);
6254
6255         return inode;
6256
6257 fail_unlock:
6258         unlock_new_inode(inode);
6259 fail:
6260         if (dir && name)
6261                 BTRFS_I(dir)->index_cnt--;
6262         btrfs_free_path(path);
6263         iput(inode);
6264         return ERR_PTR(ret);
6265 }
6266
6267 static inline u8 btrfs_inode_type(struct inode *inode)
6268 {
6269         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6270 }
6271
6272 /*
6273  * utility function to add 'inode' into 'parent_inode' with
6274  * a give name and a given sequence number.
6275  * if 'add_backref' is true, also insert a backref from the
6276  * inode to the parent directory.
6277  */
6278 int btrfs_add_link(struct btrfs_trans_handle *trans,
6279                    struct inode *parent_inode, struct inode *inode,
6280                    const char *name, int name_len, int add_backref, u64 index)
6281 {
6282         int ret = 0;
6283         struct btrfs_key key;
6284         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6285         u64 ino = btrfs_ino(inode);
6286         u64 parent_ino = btrfs_ino(parent_inode);
6287
6288         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6289                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6290         } else {
6291                 key.objectid = ino;
6292                 key.type = BTRFS_INODE_ITEM_KEY;
6293                 key.offset = 0;
6294         }
6295
6296         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6297                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6298                                          key.objectid, root->root_key.objectid,
6299                                          parent_ino, index, name, name_len);
6300         } else if (add_backref) {
6301                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6302                                              parent_ino, index);
6303         }
6304
6305         /* Nothing to clean up yet */
6306         if (ret)
6307                 return ret;
6308
6309         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6310                                     parent_inode, &key,
6311                                     btrfs_inode_type(inode), index);
6312         if (ret == -EEXIST || ret == -EOVERFLOW)
6313                 goto fail_dir_item;
6314         else if (ret) {
6315                 btrfs_abort_transaction(trans, root, ret);
6316                 return ret;
6317         }
6318
6319         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6320                            name_len * 2);
6321         inode_inc_iversion(parent_inode);
6322         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
6323         ret = btrfs_update_inode(trans, root, parent_inode);
6324         if (ret)
6325                 btrfs_abort_transaction(trans, root, ret);
6326         return ret;
6327
6328 fail_dir_item:
6329         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6330                 u64 local_index;
6331                 int err;
6332                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6333                                  key.objectid, root->root_key.objectid,
6334                                  parent_ino, &local_index, name, name_len);
6335
6336         } else if (add_backref) {
6337                 u64 local_index;
6338                 int err;
6339
6340                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6341                                           ino, parent_ino, &local_index);
6342         }
6343         return ret;
6344 }
6345
6346 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6347                             struct inode *dir, struct dentry *dentry,
6348                             struct inode *inode, int backref, u64 index)
6349 {
6350         int err = btrfs_add_link(trans, dir, inode,
6351                                  dentry->d_name.name, dentry->d_name.len,
6352                                  backref, index);
6353         if (err > 0)
6354                 err = -EEXIST;
6355         return err;
6356 }
6357
6358 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6359                         umode_t mode, dev_t rdev)
6360 {
6361         struct btrfs_trans_handle *trans;
6362         struct btrfs_root *root = BTRFS_I(dir)->root;
6363         struct inode *inode = NULL;
6364         int err;
6365         int drop_inode = 0;
6366         u64 objectid;
6367         u64 index = 0;
6368
6369         /*
6370          * 2 for inode item and ref
6371          * 2 for dir items
6372          * 1 for xattr if selinux is on
6373          */
6374         trans = btrfs_start_transaction(root, 5);
6375         if (IS_ERR(trans))
6376                 return PTR_ERR(trans);
6377
6378         err = btrfs_find_free_ino(root, &objectid);
6379         if (err)
6380                 goto out_unlock;
6381
6382         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6383                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6384                                 mode, &index);
6385         if (IS_ERR(inode)) {
6386                 err = PTR_ERR(inode);
6387                 goto out_unlock;
6388         }
6389
6390         /*
6391         * If the active LSM wants to access the inode during
6392         * d_instantiate it needs these. Smack checks to see
6393         * if the filesystem supports xattrs by looking at the
6394         * ops vector.
6395         */
6396         inode->i_op = &btrfs_special_inode_operations;
6397         init_special_inode(inode, inode->i_mode, rdev);
6398
6399         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6400         if (err)
6401                 goto out_unlock_inode;
6402
6403         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6404         if (err) {
6405                 goto out_unlock_inode;
6406         } else {
6407                 btrfs_update_inode(trans, root, inode);
6408                 unlock_new_inode(inode);
6409                 d_instantiate(dentry, inode);
6410         }
6411
6412 out_unlock:
6413         btrfs_end_transaction(trans, root);
6414         btrfs_balance_delayed_items(root);
6415         btrfs_btree_balance_dirty(root);
6416         if (drop_inode) {
6417                 inode_dec_link_count(inode);
6418                 iput(inode);
6419         }
6420         return err;
6421
6422 out_unlock_inode:
6423         drop_inode = 1;
6424         unlock_new_inode(inode);
6425         goto out_unlock;
6426
6427 }
6428
6429 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6430                         umode_t mode, bool excl)
6431 {
6432         struct btrfs_trans_handle *trans;
6433         struct btrfs_root *root = BTRFS_I(dir)->root;
6434         struct inode *inode = NULL;
6435         int drop_inode_on_err = 0;
6436         int err;
6437         u64 objectid;
6438         u64 index = 0;
6439
6440         /*
6441          * 2 for inode item and ref
6442          * 2 for dir items
6443          * 1 for xattr if selinux is on
6444          */
6445         trans = btrfs_start_transaction(root, 5);
6446         if (IS_ERR(trans))
6447                 return PTR_ERR(trans);
6448
6449         err = btrfs_find_free_ino(root, &objectid);
6450         if (err)
6451                 goto out_unlock;
6452
6453         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6454                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6455                                 mode, &index);
6456         if (IS_ERR(inode)) {
6457                 err = PTR_ERR(inode);
6458                 goto out_unlock;
6459         }
6460         drop_inode_on_err = 1;
6461         /*
6462         * If the active LSM wants to access the inode during
6463         * d_instantiate it needs these. Smack checks to see
6464         * if the filesystem supports xattrs by looking at the
6465         * ops vector.
6466         */
6467         inode->i_fop = &btrfs_file_operations;
6468         inode->i_op = &btrfs_file_inode_operations;
6469         inode->i_mapping->a_ops = &btrfs_aops;
6470
6471         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6472         if (err)
6473                 goto out_unlock_inode;
6474
6475         err = btrfs_update_inode(trans, root, inode);
6476         if (err)
6477                 goto out_unlock_inode;
6478
6479         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6480         if (err)
6481                 goto out_unlock_inode;
6482
6483         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6484         unlock_new_inode(inode);
6485         d_instantiate(dentry, inode);
6486
6487 out_unlock:
6488         btrfs_end_transaction(trans, root);
6489         if (err && drop_inode_on_err) {
6490                 inode_dec_link_count(inode);
6491                 iput(inode);
6492         }
6493         btrfs_balance_delayed_items(root);
6494         btrfs_btree_balance_dirty(root);
6495         return err;
6496
6497 out_unlock_inode:
6498         unlock_new_inode(inode);
6499         goto out_unlock;
6500
6501 }
6502
6503 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6504                       struct dentry *dentry)
6505 {
6506         struct btrfs_trans_handle *trans;
6507         struct btrfs_root *root = BTRFS_I(dir)->root;
6508         struct inode *inode = d_inode(old_dentry);
6509         u64 index;
6510         int err;
6511         int drop_inode = 0;
6512
6513         /* do not allow sys_link's with other subvols of the same device */
6514         if (root->objectid != BTRFS_I(inode)->root->objectid)
6515                 return -EXDEV;
6516
6517         if (inode->i_nlink >= BTRFS_LINK_MAX)
6518                 return -EMLINK;
6519
6520         err = btrfs_set_inode_index(dir, &index);
6521         if (err)
6522                 goto fail;
6523
6524         /*
6525          * 2 items for inode and inode ref
6526          * 2 items for dir items
6527          * 1 item for parent inode
6528          */
6529         trans = btrfs_start_transaction(root, 5);
6530         if (IS_ERR(trans)) {
6531                 err = PTR_ERR(trans);
6532                 goto fail;
6533         }
6534
6535         /* There are several dir indexes for this inode, clear the cache. */
6536         BTRFS_I(inode)->dir_index = 0ULL;
6537         inc_nlink(inode);
6538         inode_inc_iversion(inode);
6539         inode->i_ctime = CURRENT_TIME;
6540         ihold(inode);
6541         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6542
6543         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6544
6545         if (err) {
6546                 drop_inode = 1;
6547         } else {
6548                 struct dentry *parent = dentry->d_parent;
6549                 err = btrfs_update_inode(trans, root, inode);
6550                 if (err)
6551                         goto fail;
6552                 if (inode->i_nlink == 1) {
6553                         /*
6554                          * If new hard link count is 1, it's a file created
6555                          * with open(2) O_TMPFILE flag.
6556                          */
6557                         err = btrfs_orphan_del(trans, inode);
6558                         if (err)
6559                                 goto fail;
6560                 }
6561                 d_instantiate(dentry, inode);
6562                 btrfs_log_new_name(trans, inode, NULL, parent);
6563         }
6564
6565         btrfs_end_transaction(trans, root);
6566         btrfs_balance_delayed_items(root);
6567 fail:
6568         if (drop_inode) {
6569                 inode_dec_link_count(inode);
6570                 iput(inode);
6571         }
6572         btrfs_btree_balance_dirty(root);
6573         return err;
6574 }
6575
6576 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6577 {
6578         struct inode *inode = NULL;
6579         struct btrfs_trans_handle *trans;
6580         struct btrfs_root *root = BTRFS_I(dir)->root;
6581         int err = 0;
6582         int drop_on_err = 0;
6583         u64 objectid = 0;
6584         u64 index = 0;
6585
6586         /*
6587          * 2 items for inode and ref
6588          * 2 items for dir items
6589          * 1 for xattr if selinux is on
6590          */
6591         trans = btrfs_start_transaction(root, 5);
6592         if (IS_ERR(trans))
6593                 return PTR_ERR(trans);
6594
6595         err = btrfs_find_free_ino(root, &objectid);
6596         if (err)
6597                 goto out_fail;
6598
6599         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6600                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6601                                 S_IFDIR | mode, &index);
6602         if (IS_ERR(inode)) {
6603                 err = PTR_ERR(inode);
6604                 goto out_fail;
6605         }
6606
6607         drop_on_err = 1;
6608         /* these must be set before we unlock the inode */
6609         inode->i_op = &btrfs_dir_inode_operations;
6610         inode->i_fop = &btrfs_dir_file_operations;
6611
6612         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6613         if (err)
6614                 goto out_fail_inode;
6615
6616         btrfs_i_size_write(inode, 0);
6617         err = btrfs_update_inode(trans, root, inode);
6618         if (err)
6619                 goto out_fail_inode;
6620
6621         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6622                              dentry->d_name.len, 0, index);
6623         if (err)
6624                 goto out_fail_inode;
6625
6626         d_instantiate(dentry, inode);
6627         /*
6628          * mkdir is special.  We're unlocking after we call d_instantiate
6629          * to avoid a race with nfsd calling d_instantiate.
6630          */
6631         unlock_new_inode(inode);
6632         drop_on_err = 0;
6633
6634 out_fail:
6635         btrfs_end_transaction(trans, root);
6636         if (drop_on_err) {
6637                 inode_dec_link_count(inode);
6638                 iput(inode);
6639         }
6640         btrfs_balance_delayed_items(root);
6641         btrfs_btree_balance_dirty(root);
6642         return err;
6643
6644 out_fail_inode:
6645         unlock_new_inode(inode);
6646         goto out_fail;
6647 }
6648
6649 /* Find next extent map of a given extent map, caller needs to ensure locks */
6650 static struct extent_map *next_extent_map(struct extent_map *em)
6651 {
6652         struct rb_node *next;
6653
6654         next = rb_next(&em->rb_node);
6655         if (!next)
6656                 return NULL;
6657         return container_of(next, struct extent_map, rb_node);
6658 }
6659
6660 static struct extent_map *prev_extent_map(struct extent_map *em)
6661 {
6662         struct rb_node *prev;
6663
6664         prev = rb_prev(&em->rb_node);
6665         if (!prev)
6666                 return NULL;
6667         return container_of(prev, struct extent_map, rb_node);
6668 }
6669
6670 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6671  * the existing extent is the nearest extent to map_start,
6672  * and an extent that you want to insert, deal with overlap and insert
6673  * the best fitted new extent into the tree.
6674  */
6675 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6676                                 struct extent_map *existing,
6677                                 struct extent_map *em,
6678                                 u64 map_start)
6679 {
6680         struct extent_map *prev;
6681         struct extent_map *next;
6682         u64 start;
6683         u64 end;
6684         u64 start_diff;
6685
6686         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6687
6688         if (existing->start > map_start) {
6689                 next = existing;
6690                 prev = prev_extent_map(next);
6691         } else {
6692                 prev = existing;
6693                 next = next_extent_map(prev);
6694         }
6695
6696         start = prev ? extent_map_end(prev) : em->start;
6697         start = max_t(u64, start, em->start);
6698         end = next ? next->start : extent_map_end(em);
6699         end = min_t(u64, end, extent_map_end(em));
6700         start_diff = start - em->start;
6701         em->start = start;
6702         em->len = end - start;
6703         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6704             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6705                 em->block_start += start_diff;
6706                 em->block_len -= start_diff;
6707         }
6708         return add_extent_mapping(em_tree, em, 0);
6709 }
6710
6711 static noinline int uncompress_inline(struct btrfs_path *path,
6712                                       struct inode *inode, struct page *page,
6713                                       size_t pg_offset, u64 extent_offset,
6714                                       struct btrfs_file_extent_item *item)
6715 {
6716         int ret;
6717         struct extent_buffer *leaf = path->nodes[0];
6718         char *tmp;
6719         size_t max_size;
6720         unsigned long inline_size;
6721         unsigned long ptr;
6722         int compress_type;
6723
6724         WARN_ON(pg_offset != 0);
6725         compress_type = btrfs_file_extent_compression(leaf, item);
6726         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6727         inline_size = btrfs_file_extent_inline_item_len(leaf,
6728                                         btrfs_item_nr(path->slots[0]));
6729         tmp = kmalloc(inline_size, GFP_NOFS);
6730         if (!tmp)
6731                 return -ENOMEM;
6732         ptr = btrfs_file_extent_inline_start(item);
6733
6734         read_extent_buffer(leaf, tmp, ptr, inline_size);
6735
6736         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
6737         ret = btrfs_decompress(compress_type, tmp, page,
6738                                extent_offset, inline_size, max_size);
6739         kfree(tmp);
6740         return ret;
6741 }
6742
6743 /*
6744  * a bit scary, this does extent mapping from logical file offset to the disk.
6745  * the ugly parts come from merging extents from the disk with the in-ram
6746  * representation.  This gets more complex because of the data=ordered code,
6747  * where the in-ram extents might be locked pending data=ordered completion.
6748  *
6749  * This also copies inline extents directly into the page.
6750  */
6751
6752 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6753                                     size_t pg_offset, u64 start, u64 len,
6754                                     int create)
6755 {
6756         int ret;
6757         int err = 0;
6758         u64 extent_start = 0;
6759         u64 extent_end = 0;
6760         u64 objectid = btrfs_ino(inode);
6761         u32 found_type;
6762         struct btrfs_path *path = NULL;
6763         struct btrfs_root *root = BTRFS_I(inode)->root;
6764         struct btrfs_file_extent_item *item;
6765         struct extent_buffer *leaf;
6766         struct btrfs_key found_key;
6767         struct extent_map *em = NULL;
6768         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6769         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6770         struct btrfs_trans_handle *trans = NULL;
6771         const bool new_inline = !page || create;
6772
6773 again:
6774         read_lock(&em_tree->lock);
6775         em = lookup_extent_mapping(em_tree, start, len);
6776         if (em)
6777                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6778         read_unlock(&em_tree->lock);
6779
6780         if (em) {
6781                 if (em->start > start || em->start + em->len <= start)
6782                         free_extent_map(em);
6783                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6784                         free_extent_map(em);
6785                 else
6786                         goto out;
6787         }
6788         em = alloc_extent_map();
6789         if (!em) {
6790                 err = -ENOMEM;
6791                 goto out;
6792         }
6793         em->bdev = root->fs_info->fs_devices->latest_bdev;
6794         em->start = EXTENT_MAP_HOLE;
6795         em->orig_start = EXTENT_MAP_HOLE;
6796         em->len = (u64)-1;
6797         em->block_len = (u64)-1;
6798
6799         if (!path) {
6800                 path = btrfs_alloc_path();
6801                 if (!path) {
6802                         err = -ENOMEM;
6803                         goto out;
6804                 }
6805                 /*
6806                  * Chances are we'll be called again, so go ahead and do
6807                  * readahead
6808                  */
6809                 path->reada = 1;
6810         }
6811
6812         ret = btrfs_lookup_file_extent(trans, root, path,
6813                                        objectid, start, trans != NULL);
6814         if (ret < 0) {
6815                 err = ret;
6816                 goto out;
6817         }
6818
6819         if (ret != 0) {
6820                 if (path->slots[0] == 0)
6821                         goto not_found;
6822                 path->slots[0]--;
6823         }
6824
6825         leaf = path->nodes[0];
6826         item = btrfs_item_ptr(leaf, path->slots[0],
6827                               struct btrfs_file_extent_item);
6828         /* are we inside the extent that was found? */
6829         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6830         found_type = found_key.type;
6831         if (found_key.objectid != objectid ||
6832             found_type != BTRFS_EXTENT_DATA_KEY) {
6833                 /*
6834                  * If we backup past the first extent we want to move forward
6835                  * and see if there is an extent in front of us, otherwise we'll
6836                  * say there is a hole for our whole search range which can
6837                  * cause problems.
6838                  */
6839                 extent_end = start;
6840                 goto next;
6841         }
6842
6843         found_type = btrfs_file_extent_type(leaf, item);
6844         extent_start = found_key.offset;
6845         if (found_type == BTRFS_FILE_EXTENT_REG ||
6846             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6847                 extent_end = extent_start +
6848                        btrfs_file_extent_num_bytes(leaf, item);
6849         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6850                 size_t size;
6851                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6852                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6853         }
6854 next:
6855         if (start >= extent_end) {
6856                 path->slots[0]++;
6857                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6858                         ret = btrfs_next_leaf(root, path);
6859                         if (ret < 0) {
6860                                 err = ret;
6861                                 goto out;
6862                         }
6863                         if (ret > 0)
6864                                 goto not_found;
6865                         leaf = path->nodes[0];
6866                 }
6867                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6868                 if (found_key.objectid != objectid ||
6869                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6870                         goto not_found;
6871                 if (start + len <= found_key.offset)
6872                         goto not_found;
6873                 if (start > found_key.offset)
6874                         goto next;
6875                 em->start = start;
6876                 em->orig_start = start;
6877                 em->len = found_key.offset - start;
6878                 goto not_found_em;
6879         }
6880
6881         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6882
6883         if (found_type == BTRFS_FILE_EXTENT_REG ||
6884             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6885                 goto insert;
6886         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6887                 unsigned long ptr;
6888                 char *map;
6889                 size_t size;
6890                 size_t extent_offset;
6891                 size_t copy_size;
6892
6893                 if (new_inline)
6894                         goto out;
6895
6896                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6897                 extent_offset = page_offset(page) + pg_offset - extent_start;
6898                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6899                                 size - extent_offset);
6900                 em->start = extent_start + extent_offset;
6901                 em->len = ALIGN(copy_size, root->sectorsize);
6902                 em->orig_block_len = em->len;
6903                 em->orig_start = em->start;
6904                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6905                 if (create == 0 && !PageUptodate(page)) {
6906                         if (btrfs_file_extent_compression(leaf, item) !=
6907                             BTRFS_COMPRESS_NONE) {
6908                                 ret = uncompress_inline(path, inode, page,
6909                                                         pg_offset,
6910                                                         extent_offset, item);
6911                                 if (ret) {
6912                                         err = ret;
6913                                         goto out;
6914                                 }
6915                         } else {
6916                                 map = kmap(page);
6917                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6918                                                    copy_size);
6919                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6920                                         memset(map + pg_offset + copy_size, 0,
6921                                                PAGE_CACHE_SIZE - pg_offset -
6922                                                copy_size);
6923                                 }
6924                                 kunmap(page);
6925                         }
6926                         flush_dcache_page(page);
6927                 } else if (create && PageUptodate(page)) {
6928                         BUG();
6929                         if (!trans) {
6930                                 kunmap(page);
6931                                 free_extent_map(em);
6932                                 em = NULL;
6933
6934                                 btrfs_release_path(path);
6935                                 trans = btrfs_join_transaction(root);
6936
6937                                 if (IS_ERR(trans))
6938                                         return ERR_CAST(trans);
6939                                 goto again;
6940                         }
6941                         map = kmap(page);
6942                         write_extent_buffer(leaf, map + pg_offset, ptr,
6943                                             copy_size);
6944                         kunmap(page);
6945                         btrfs_mark_buffer_dirty(leaf);
6946                 }
6947                 set_extent_uptodate(io_tree, em->start,
6948                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6949                 goto insert;
6950         }
6951 not_found:
6952         em->start = start;
6953         em->orig_start = start;
6954         em->len = len;
6955 not_found_em:
6956         em->block_start = EXTENT_MAP_HOLE;
6957         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6958 insert:
6959         btrfs_release_path(path);
6960         if (em->start > start || extent_map_end(em) <= start) {
6961                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6962                         em->start, em->len, start, len);
6963                 err = -EIO;
6964                 goto out;
6965         }
6966
6967         err = 0;
6968         write_lock(&em_tree->lock);
6969         ret = add_extent_mapping(em_tree, em, 0);
6970         /* it is possible that someone inserted the extent into the tree
6971          * while we had the lock dropped.  It is also possible that
6972          * an overlapping map exists in the tree
6973          */
6974         if (ret == -EEXIST) {
6975                 struct extent_map *existing;
6976
6977                 ret = 0;
6978
6979                 existing = search_extent_mapping(em_tree, start, len);
6980                 /*
6981                  * existing will always be non-NULL, since there must be
6982                  * extent causing the -EEXIST.
6983                  */
6984                 if (start >= extent_map_end(existing) ||
6985                     start <= existing->start) {
6986                         /*
6987                          * The existing extent map is the one nearest to
6988                          * the [start, start + len) range which overlaps
6989                          */
6990                         err = merge_extent_mapping(em_tree, existing,
6991                                                    em, start);
6992                         free_extent_map(existing);
6993                         if (err) {
6994                                 free_extent_map(em);
6995                                 em = NULL;
6996                         }
6997                 } else {
6998                         free_extent_map(em);
6999                         em = existing;
7000                         err = 0;
7001                 }
7002         }
7003         write_unlock(&em_tree->lock);
7004 out:
7005
7006         trace_btrfs_get_extent(root, em);
7007
7008         btrfs_free_path(path);
7009         if (trans) {
7010                 ret = btrfs_end_transaction(trans, root);
7011                 if (!err)
7012                         err = ret;
7013         }
7014         if (err) {
7015                 free_extent_map(em);
7016                 return ERR_PTR(err);
7017         }
7018         BUG_ON(!em); /* Error is always set */
7019         return em;
7020 }
7021
7022 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7023                                            size_t pg_offset, u64 start, u64 len,
7024                                            int create)
7025 {
7026         struct extent_map *em;
7027         struct extent_map *hole_em = NULL;
7028         u64 range_start = start;
7029         u64 end;
7030         u64 found;
7031         u64 found_end;
7032         int err = 0;
7033
7034         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7035         if (IS_ERR(em))
7036                 return em;
7037         if (em) {
7038                 /*
7039                  * if our em maps to
7040                  * -  a hole or
7041                  * -  a pre-alloc extent,
7042                  * there might actually be delalloc bytes behind it.
7043                  */
7044                 if (em->block_start != EXTENT_MAP_HOLE &&
7045                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7046                         return em;
7047                 else
7048                         hole_em = em;
7049         }
7050
7051         /* check to see if we've wrapped (len == -1 or similar) */
7052         end = start + len;
7053         if (end < start)
7054                 end = (u64)-1;
7055         else
7056                 end -= 1;
7057
7058         em = NULL;
7059
7060         /* ok, we didn't find anything, lets look for delalloc */
7061         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7062                                  end, len, EXTENT_DELALLOC, 1);
7063         found_end = range_start + found;
7064         if (found_end < range_start)
7065                 found_end = (u64)-1;
7066
7067         /*
7068          * we didn't find anything useful, return
7069          * the original results from get_extent()
7070          */
7071         if (range_start > end || found_end <= start) {
7072                 em = hole_em;
7073                 hole_em = NULL;
7074                 goto out;
7075         }
7076
7077         /* adjust the range_start to make sure it doesn't
7078          * go backwards from the start they passed in
7079          */
7080         range_start = max(start, range_start);
7081         found = found_end - range_start;
7082
7083         if (found > 0) {
7084                 u64 hole_start = start;
7085                 u64 hole_len = len;
7086
7087                 em = alloc_extent_map();
7088                 if (!em) {
7089                         err = -ENOMEM;
7090                         goto out;
7091                 }
7092                 /*
7093                  * when btrfs_get_extent can't find anything it
7094                  * returns one huge hole
7095                  *
7096                  * make sure what it found really fits our range, and
7097                  * adjust to make sure it is based on the start from
7098                  * the caller
7099                  */
7100                 if (hole_em) {
7101                         u64 calc_end = extent_map_end(hole_em);
7102
7103                         if (calc_end <= start || (hole_em->start > end)) {
7104                                 free_extent_map(hole_em);
7105                                 hole_em = NULL;
7106                         } else {
7107                                 hole_start = max(hole_em->start, start);
7108                                 hole_len = calc_end - hole_start;
7109                         }
7110                 }
7111                 em->bdev = NULL;
7112                 if (hole_em && range_start > hole_start) {
7113                         /* our hole starts before our delalloc, so we
7114                          * have to return just the parts of the hole
7115                          * that go until  the delalloc starts
7116                          */
7117                         em->len = min(hole_len,
7118                                       range_start - hole_start);
7119                         em->start = hole_start;
7120                         em->orig_start = hole_start;
7121                         /*
7122                          * don't adjust block start at all,
7123                          * it is fixed at EXTENT_MAP_HOLE
7124                          */
7125                         em->block_start = hole_em->block_start;
7126                         em->block_len = hole_len;
7127                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7128                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7129                 } else {
7130                         em->start = range_start;
7131                         em->len = found;
7132                         em->orig_start = range_start;
7133                         em->block_start = EXTENT_MAP_DELALLOC;
7134                         em->block_len = found;
7135                 }
7136         } else if (hole_em) {
7137                 return hole_em;
7138         }
7139 out:
7140
7141         free_extent_map(hole_em);
7142         if (err) {
7143                 free_extent_map(em);
7144                 return ERR_PTR(err);
7145         }
7146         return em;
7147 }
7148
7149 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7150                                                   u64 start, u64 len)
7151 {
7152         struct btrfs_root *root = BTRFS_I(inode)->root;
7153         struct extent_map *em;
7154         struct btrfs_key ins;
7155         u64 alloc_hint;
7156         int ret;
7157
7158         alloc_hint = get_extent_allocation_hint(inode, start, len);
7159         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7160                                    alloc_hint, &ins, 1, 1);
7161         if (ret)
7162                 return ERR_PTR(ret);
7163
7164         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
7165                               ins.offset, ins.offset, ins.offset, 0);
7166         if (IS_ERR(em)) {
7167                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7168                 return em;
7169         }
7170
7171         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
7172                                            ins.offset, ins.offset, 0);
7173         if (ret) {
7174                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7175                 free_extent_map(em);
7176                 return ERR_PTR(ret);
7177         }
7178
7179         return em;
7180 }
7181
7182 /*
7183  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7184  * block must be cow'd
7185  */
7186 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7187                               u64 *orig_start, u64 *orig_block_len,
7188                               u64 *ram_bytes)
7189 {
7190         struct btrfs_trans_handle *trans;
7191         struct btrfs_path *path;
7192         int ret;
7193         struct extent_buffer *leaf;
7194         struct btrfs_root *root = BTRFS_I(inode)->root;
7195         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7196         struct btrfs_file_extent_item *fi;
7197         struct btrfs_key key;
7198         u64 disk_bytenr;
7199         u64 backref_offset;
7200         u64 extent_end;
7201         u64 num_bytes;
7202         int slot;
7203         int found_type;
7204         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7205
7206         path = btrfs_alloc_path();
7207         if (!path)
7208                 return -ENOMEM;
7209
7210         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7211                                        offset, 0);
7212         if (ret < 0)
7213                 goto out;
7214
7215         slot = path->slots[0];
7216         if (ret == 1) {
7217                 if (slot == 0) {
7218                         /* can't find the item, must cow */
7219                         ret = 0;
7220                         goto out;
7221                 }
7222                 slot--;
7223         }
7224         ret = 0;
7225         leaf = path->nodes[0];
7226         btrfs_item_key_to_cpu(leaf, &key, slot);
7227         if (key.objectid != btrfs_ino(inode) ||
7228             key.type != BTRFS_EXTENT_DATA_KEY) {
7229                 /* not our file or wrong item type, must cow */
7230                 goto out;
7231         }
7232
7233         if (key.offset > offset) {
7234                 /* Wrong offset, must cow */
7235                 goto out;
7236         }
7237
7238         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7239         found_type = btrfs_file_extent_type(leaf, fi);
7240         if (found_type != BTRFS_FILE_EXTENT_REG &&
7241             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7242                 /* not a regular extent, must cow */
7243                 goto out;
7244         }
7245
7246         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7247                 goto out;
7248
7249         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7250         if (extent_end <= offset)
7251                 goto out;
7252
7253         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7254         if (disk_bytenr == 0)
7255                 goto out;
7256
7257         if (btrfs_file_extent_compression(leaf, fi) ||
7258             btrfs_file_extent_encryption(leaf, fi) ||
7259             btrfs_file_extent_other_encoding(leaf, fi))
7260                 goto out;
7261
7262         backref_offset = btrfs_file_extent_offset(leaf, fi);
7263
7264         if (orig_start) {
7265                 *orig_start = key.offset - backref_offset;
7266                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7267                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7268         }
7269
7270         if (btrfs_extent_readonly(root, disk_bytenr))
7271                 goto out;
7272
7273         num_bytes = min(offset + *len, extent_end) - offset;
7274         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7275                 u64 range_end;
7276
7277                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7278                 ret = test_range_bit(io_tree, offset, range_end,
7279                                      EXTENT_DELALLOC, 0, NULL);
7280                 if (ret) {
7281                         ret = -EAGAIN;
7282                         goto out;
7283                 }
7284         }
7285
7286         btrfs_release_path(path);
7287
7288         /*
7289          * look for other files referencing this extent, if we
7290          * find any we must cow
7291          */
7292         trans = btrfs_join_transaction(root);
7293         if (IS_ERR(trans)) {
7294                 ret = 0;
7295                 goto out;
7296         }
7297
7298         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7299                                     key.offset - backref_offset, disk_bytenr);
7300         btrfs_end_transaction(trans, root);
7301         if (ret) {
7302                 ret = 0;
7303                 goto out;
7304         }
7305
7306         /*
7307          * adjust disk_bytenr and num_bytes to cover just the bytes
7308          * in this extent we are about to write.  If there
7309          * are any csums in that range we have to cow in order
7310          * to keep the csums correct
7311          */
7312         disk_bytenr += backref_offset;
7313         disk_bytenr += offset - key.offset;
7314         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7315                                 goto out;
7316         /*
7317          * all of the above have passed, it is safe to overwrite this extent
7318          * without cow
7319          */
7320         *len = num_bytes;
7321         ret = 1;
7322 out:
7323         btrfs_free_path(path);
7324         return ret;
7325 }
7326
7327 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7328 {
7329         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7330         int found = false;
7331         void **pagep = NULL;
7332         struct page *page = NULL;
7333         int start_idx;
7334         int end_idx;
7335
7336         start_idx = start >> PAGE_CACHE_SHIFT;
7337
7338         /*
7339          * end is the last byte in the last page.  end == start is legal
7340          */
7341         end_idx = end >> PAGE_CACHE_SHIFT;
7342
7343         rcu_read_lock();
7344
7345         /* Most of the code in this while loop is lifted from
7346          * find_get_page.  It's been modified to begin searching from a
7347          * page and return just the first page found in that range.  If the
7348          * found idx is less than or equal to the end idx then we know that
7349          * a page exists.  If no pages are found or if those pages are
7350          * outside of the range then we're fine (yay!) */
7351         while (page == NULL &&
7352                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7353                 page = radix_tree_deref_slot(pagep);
7354                 if (unlikely(!page))
7355                         break;
7356
7357                 if (radix_tree_exception(page)) {
7358                         if (radix_tree_deref_retry(page)) {
7359                                 page = NULL;
7360                                 continue;
7361                         }
7362                         /*
7363                          * Otherwise, shmem/tmpfs must be storing a swap entry
7364                          * here as an exceptional entry: so return it without
7365                          * attempting to raise page count.
7366                          */
7367                         page = NULL;
7368                         break; /* TODO: Is this relevant for this use case? */
7369                 }
7370
7371                 if (!page_cache_get_speculative(page)) {
7372                         page = NULL;
7373                         continue;
7374                 }
7375
7376                 /*
7377                  * Has the page moved?
7378                  * This is part of the lockless pagecache protocol. See
7379                  * include/linux/pagemap.h for details.
7380                  */
7381                 if (unlikely(page != *pagep)) {
7382                         page_cache_release(page);
7383                         page = NULL;
7384                 }
7385         }
7386
7387         if (page) {
7388                 if (page->index <= end_idx)
7389                         found = true;
7390                 page_cache_release(page);
7391         }
7392
7393         rcu_read_unlock();
7394         return found;
7395 }
7396
7397 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7398                               struct extent_state **cached_state, int writing)
7399 {
7400         struct btrfs_ordered_extent *ordered;
7401         int ret = 0;
7402
7403         while (1) {
7404                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7405                                  cached_state);
7406                 /*
7407                  * We're concerned with the entire range that we're going to be
7408                  * doing DIO to, so we need to make sure theres no ordered
7409                  * extents in this range.
7410                  */
7411                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7412                                                      lockend - lockstart + 1);
7413
7414                 /*
7415                  * We need to make sure there are no buffered pages in this
7416                  * range either, we could have raced between the invalidate in
7417                  * generic_file_direct_write and locking the extent.  The
7418                  * invalidate needs to happen so that reads after a write do not
7419                  * get stale data.
7420                  */
7421                 if (!ordered &&
7422                     (!writing ||
7423                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7424                         break;
7425
7426                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7427                                      cached_state, GFP_NOFS);
7428
7429                 if (ordered) {
7430                         btrfs_start_ordered_extent(inode, ordered, 1);
7431                         btrfs_put_ordered_extent(ordered);
7432                 } else {
7433                         /* Screw you mmap */
7434                         ret = btrfs_fdatawrite_range(inode, lockstart, lockend);
7435                         if (ret)
7436                                 break;
7437                         ret = filemap_fdatawait_range(inode->i_mapping,
7438                                                       lockstart,
7439                                                       lockend);
7440                         if (ret)
7441                                 break;
7442
7443                         /*
7444                          * If we found a page that couldn't be invalidated just
7445                          * fall back to buffered.
7446                          */
7447                         ret = invalidate_inode_pages2_range(inode->i_mapping,
7448                                         lockstart >> PAGE_CACHE_SHIFT,
7449                                         lockend >> PAGE_CACHE_SHIFT);
7450                         if (ret)
7451                                 break;
7452                 }
7453
7454                 cond_resched();
7455         }
7456
7457         return ret;
7458 }
7459
7460 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7461                                            u64 len, u64 orig_start,
7462                                            u64 block_start, u64 block_len,
7463                                            u64 orig_block_len, u64 ram_bytes,
7464                                            int type)
7465 {
7466         struct extent_map_tree *em_tree;
7467         struct extent_map *em;
7468         struct btrfs_root *root = BTRFS_I(inode)->root;
7469         int ret;
7470
7471         em_tree = &BTRFS_I(inode)->extent_tree;
7472         em = alloc_extent_map();
7473         if (!em)
7474                 return ERR_PTR(-ENOMEM);
7475
7476         em->start = start;
7477         em->orig_start = orig_start;
7478         em->mod_start = start;
7479         em->mod_len = len;
7480         em->len = len;
7481         em->block_len = block_len;
7482         em->block_start = block_start;
7483         em->bdev = root->fs_info->fs_devices->latest_bdev;
7484         em->orig_block_len = orig_block_len;
7485         em->ram_bytes = ram_bytes;
7486         em->generation = -1;
7487         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7488         if (type == BTRFS_ORDERED_PREALLOC)
7489                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7490
7491         do {
7492                 btrfs_drop_extent_cache(inode, em->start,
7493                                 em->start + em->len - 1, 0);
7494                 write_lock(&em_tree->lock);
7495                 ret = add_extent_mapping(em_tree, em, 1);
7496                 write_unlock(&em_tree->lock);
7497         } while (ret == -EEXIST);
7498
7499         if (ret) {
7500                 free_extent_map(em);
7501                 return ERR_PTR(ret);
7502         }
7503
7504         return em;
7505 }
7506
7507 struct btrfs_dio_data {
7508         u64 outstanding_extents;
7509         u64 reserve;
7510 };
7511
7512 static void adjust_dio_outstanding_extents(struct inode *inode,
7513                                            struct btrfs_dio_data *dio_data,
7514                                            const u64 len)
7515 {
7516         unsigned num_extents;
7517
7518         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7519                                            BTRFS_MAX_EXTENT_SIZE);
7520         /*
7521          * If we have an outstanding_extents count still set then we're
7522          * within our reservation, otherwise we need to adjust our inode
7523          * counter appropriately.
7524          */
7525         if (dio_data->outstanding_extents) {
7526                 dio_data->outstanding_extents -= num_extents;
7527         } else {
7528                 spin_lock(&BTRFS_I(inode)->lock);
7529                 BTRFS_I(inode)->outstanding_extents += num_extents;
7530                 spin_unlock(&BTRFS_I(inode)->lock);
7531         }
7532 }
7533
7534 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7535                                    struct buffer_head *bh_result, int create)
7536 {
7537         struct extent_map *em;
7538         struct btrfs_root *root = BTRFS_I(inode)->root;
7539         struct extent_state *cached_state = NULL;
7540         struct btrfs_dio_data *dio_data = NULL;
7541         u64 start = iblock << inode->i_blkbits;
7542         u64 lockstart, lockend;
7543         u64 len = bh_result->b_size;
7544         int unlock_bits = EXTENT_LOCKED;
7545         int ret = 0;
7546
7547         if (create)
7548                 unlock_bits |= EXTENT_DIRTY;
7549         else
7550                 len = min_t(u64, len, root->sectorsize);
7551
7552         lockstart = start;
7553         lockend = start + len - 1;
7554
7555         if (current->journal_info) {
7556                 /*
7557                  * Need to pull our outstanding extents and set journal_info to NULL so
7558                  * that anything that needs to check if there's a transction doesn't get
7559                  * confused.
7560                  */
7561                 dio_data = current->journal_info;
7562                 current->journal_info = NULL;
7563         }
7564
7565         /*
7566          * If this errors out it's because we couldn't invalidate pagecache for
7567          * this range and we need to fallback to buffered.
7568          */
7569         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7570                                create)) {
7571                 ret = -ENOTBLK;
7572                 goto err;
7573         }
7574
7575         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7576         if (IS_ERR(em)) {
7577                 ret = PTR_ERR(em);
7578                 goto unlock_err;
7579         }
7580
7581         /*
7582          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7583          * io.  INLINE is special, and we could probably kludge it in here, but
7584          * it's still buffered so for safety lets just fall back to the generic
7585          * buffered path.
7586          *
7587          * For COMPRESSED we _have_ to read the entire extent in so we can
7588          * decompress it, so there will be buffering required no matter what we
7589          * do, so go ahead and fallback to buffered.
7590          *
7591          * We return -ENOTBLK because thats what makes DIO go ahead and go back
7592          * to buffered IO.  Don't blame me, this is the price we pay for using
7593          * the generic code.
7594          */
7595         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7596             em->block_start == EXTENT_MAP_INLINE) {
7597                 free_extent_map(em);
7598                 ret = -ENOTBLK;
7599                 goto unlock_err;
7600         }
7601
7602         /* Just a good old fashioned hole, return */
7603         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7604                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7605                 free_extent_map(em);
7606                 goto unlock_err;
7607         }
7608
7609         /*
7610          * We don't allocate a new extent in the following cases
7611          *
7612          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7613          * existing extent.
7614          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7615          * just use the extent.
7616          *
7617          */
7618         if (!create) {
7619                 len = min(len, em->len - (start - em->start));
7620                 lockstart = start + len;
7621                 goto unlock;
7622         }
7623
7624         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7625             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7626              em->block_start != EXTENT_MAP_HOLE)) {
7627                 int type;
7628                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7629
7630                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7631                         type = BTRFS_ORDERED_PREALLOC;
7632                 else
7633                         type = BTRFS_ORDERED_NOCOW;
7634                 len = min(len, em->len - (start - em->start));
7635                 block_start = em->block_start + (start - em->start);
7636
7637                 if (can_nocow_extent(inode, start, &len, &orig_start,
7638                                      &orig_block_len, &ram_bytes) == 1) {
7639                         if (type == BTRFS_ORDERED_PREALLOC) {
7640                                 free_extent_map(em);
7641                                 em = create_pinned_em(inode, start, len,
7642                                                        orig_start,
7643                                                        block_start, len,
7644                                                        orig_block_len,
7645                                                        ram_bytes, type);
7646                                 if (IS_ERR(em)) {
7647                                         ret = PTR_ERR(em);
7648                                         goto unlock_err;
7649                                 }
7650                         }
7651
7652                         ret = btrfs_add_ordered_extent_dio(inode, start,
7653                                            block_start, len, len, type);
7654                         if (ret) {
7655                                 free_extent_map(em);
7656                                 goto unlock_err;
7657                         }
7658                         goto unlock;
7659                 }
7660         }
7661
7662         /*
7663          * this will cow the extent, reset the len in case we changed
7664          * it above
7665          */
7666         len = bh_result->b_size;
7667         free_extent_map(em);
7668         em = btrfs_new_extent_direct(inode, start, len);
7669         if (IS_ERR(em)) {
7670                 ret = PTR_ERR(em);
7671                 goto unlock_err;
7672         }
7673         len = min(len, em->len - (start - em->start));
7674 unlock:
7675         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7676                 inode->i_blkbits;
7677         bh_result->b_size = len;
7678         bh_result->b_bdev = em->bdev;
7679         set_buffer_mapped(bh_result);
7680         if (create) {
7681                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7682                         set_buffer_new(bh_result);
7683
7684                 /*
7685                  * Need to update the i_size under the extent lock so buffered
7686                  * readers will get the updated i_size when we unlock.
7687                  */
7688                 if (start + len > i_size_read(inode))
7689                         i_size_write(inode, start + len);
7690
7691                 adjust_dio_outstanding_extents(inode, dio_data, len);
7692                 btrfs_free_reserved_data_space(inode, start, len);
7693                 WARN_ON(dio_data->reserve < len);
7694                 dio_data->reserve -= len;
7695                 current->journal_info = dio_data;
7696         }
7697
7698         /*
7699          * In the case of write we need to clear and unlock the entire range,
7700          * in the case of read we need to unlock only the end area that we
7701          * aren't using if there is any left over space.
7702          */
7703         if (lockstart < lockend) {
7704                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7705                                  lockend, unlock_bits, 1, 0,
7706                                  &cached_state, GFP_NOFS);
7707         } else {
7708                 free_extent_state(cached_state);
7709         }
7710
7711         free_extent_map(em);
7712
7713         return 0;
7714
7715 unlock_err:
7716         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7717                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7718 err:
7719         if (dio_data)
7720                 current->journal_info = dio_data;
7721         /*
7722          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7723          * write less data then expected, so that we don't underflow our inode's
7724          * outstanding extents counter.
7725          */
7726         if (create && dio_data)
7727                 adjust_dio_outstanding_extents(inode, dio_data, len);
7728
7729         return ret;
7730 }
7731
7732 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7733                                         int rw, int mirror_num)
7734 {
7735         struct btrfs_root *root = BTRFS_I(inode)->root;
7736         int ret;
7737
7738         BUG_ON(rw & REQ_WRITE);
7739
7740         bio_get(bio);
7741
7742         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7743                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7744         if (ret)
7745                 goto err;
7746
7747         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7748 err:
7749         bio_put(bio);
7750         return ret;
7751 }
7752
7753 static int btrfs_check_dio_repairable(struct inode *inode,
7754                                       struct bio *failed_bio,
7755                                       struct io_failure_record *failrec,
7756                                       int failed_mirror)
7757 {
7758         int num_copies;
7759
7760         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7761                                       failrec->logical, failrec->len);
7762         if (num_copies == 1) {
7763                 /*
7764                  * we only have a single copy of the data, so don't bother with
7765                  * all the retry and error correction code that follows. no
7766                  * matter what the error is, it is very likely to persist.
7767                  */
7768                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7769                          num_copies, failrec->this_mirror, failed_mirror);
7770                 return 0;
7771         }
7772
7773         failrec->failed_mirror = failed_mirror;
7774         failrec->this_mirror++;
7775         if (failrec->this_mirror == failed_mirror)
7776                 failrec->this_mirror++;
7777
7778         if (failrec->this_mirror > num_copies) {
7779                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7780                          num_copies, failrec->this_mirror, failed_mirror);
7781                 return 0;
7782         }
7783
7784         return 1;
7785 }
7786
7787 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7788                           struct page *page, u64 start, u64 end,
7789                           int failed_mirror, bio_end_io_t *repair_endio,
7790                           void *repair_arg)
7791 {
7792         struct io_failure_record *failrec;
7793         struct bio *bio;
7794         int isector;
7795         int read_mode;
7796         int ret;
7797
7798         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7799
7800         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7801         if (ret)
7802                 return ret;
7803
7804         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7805                                          failed_mirror);
7806         if (!ret) {
7807                 free_io_failure(inode, failrec);
7808                 return -EIO;
7809         }
7810
7811         if (failed_bio->bi_vcnt > 1)
7812                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7813         else
7814                 read_mode = READ_SYNC;
7815
7816         isector = start - btrfs_io_bio(failed_bio)->logical;
7817         isector >>= inode->i_sb->s_blocksize_bits;
7818         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7819                                       0, isector, repair_endio, repair_arg);
7820         if (!bio) {
7821                 free_io_failure(inode, failrec);
7822                 return -EIO;
7823         }
7824
7825         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7826                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7827                     read_mode, failrec->this_mirror, failrec->in_validation);
7828
7829         ret = submit_dio_repair_bio(inode, bio, read_mode,
7830                                     failrec->this_mirror);
7831         if (ret) {
7832                 free_io_failure(inode, failrec);
7833                 bio_put(bio);
7834         }
7835
7836         return ret;
7837 }
7838
7839 struct btrfs_retry_complete {
7840         struct completion done;
7841         struct inode *inode;
7842         u64 start;
7843         int uptodate;
7844 };
7845
7846 static void btrfs_retry_endio_nocsum(struct bio *bio)
7847 {
7848         struct btrfs_retry_complete *done = bio->bi_private;
7849         struct bio_vec *bvec;
7850         int i;
7851
7852         if (bio->bi_error)
7853                 goto end;
7854
7855         done->uptodate = 1;
7856         bio_for_each_segment_all(bvec, bio, i)
7857                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7858 end:
7859         complete(&done->done);
7860         bio_put(bio);
7861 }
7862
7863 static int __btrfs_correct_data_nocsum(struct inode *inode,
7864                                        struct btrfs_io_bio *io_bio)
7865 {
7866         struct bio_vec *bvec;
7867         struct btrfs_retry_complete done;
7868         u64 start;
7869         int i;
7870         int ret;
7871
7872         start = io_bio->logical;
7873         done.inode = inode;
7874
7875         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7876 try_again:
7877                 done.uptodate = 0;
7878                 done.start = start;
7879                 init_completion(&done.done);
7880
7881                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7882                                      start + bvec->bv_len - 1,
7883                                      io_bio->mirror_num,
7884                                      btrfs_retry_endio_nocsum, &done);
7885                 if (ret)
7886                         return ret;
7887
7888                 wait_for_completion(&done.done);
7889
7890                 if (!done.uptodate) {
7891                         /* We might have another mirror, so try again */
7892                         goto try_again;
7893                 }
7894
7895                 start += bvec->bv_len;
7896         }
7897
7898         return 0;
7899 }
7900
7901 static void btrfs_retry_endio(struct bio *bio)
7902 {
7903         struct btrfs_retry_complete *done = bio->bi_private;
7904         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7905         struct bio_vec *bvec;
7906         int uptodate;
7907         int ret;
7908         int i;
7909
7910         if (bio->bi_error)
7911                 goto end;
7912
7913         uptodate = 1;
7914         bio_for_each_segment_all(bvec, bio, i) {
7915                 ret = __readpage_endio_check(done->inode, io_bio, i,
7916                                              bvec->bv_page, 0,
7917                                              done->start, bvec->bv_len);
7918                 if (!ret)
7919                         clean_io_failure(done->inode, done->start,
7920                                          bvec->bv_page, 0);
7921                 else
7922                         uptodate = 0;
7923         }
7924
7925         done->uptodate = uptodate;
7926 end:
7927         complete(&done->done);
7928         bio_put(bio);
7929 }
7930
7931 static int __btrfs_subio_endio_read(struct inode *inode,
7932                                     struct btrfs_io_bio *io_bio, int err)
7933 {
7934         struct bio_vec *bvec;
7935         struct btrfs_retry_complete done;
7936         u64 start;
7937         u64 offset = 0;
7938         int i;
7939         int ret;
7940
7941         err = 0;
7942         start = io_bio->logical;
7943         done.inode = inode;
7944
7945         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7946                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7947                                              0, start, bvec->bv_len);
7948                 if (likely(!ret))
7949                         goto next;
7950 try_again:
7951                 done.uptodate = 0;
7952                 done.start = start;
7953                 init_completion(&done.done);
7954
7955                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7956                                      start + bvec->bv_len - 1,
7957                                      io_bio->mirror_num,
7958                                      btrfs_retry_endio, &done);
7959                 if (ret) {
7960                         err = ret;
7961                         goto next;
7962                 }
7963
7964                 wait_for_completion(&done.done);
7965
7966                 if (!done.uptodate) {
7967                         /* We might have another mirror, so try again */
7968                         goto try_again;
7969                 }
7970 next:
7971                 offset += bvec->bv_len;
7972                 start += bvec->bv_len;
7973         }
7974
7975         return err;
7976 }
7977
7978 static int btrfs_subio_endio_read(struct inode *inode,
7979                                   struct btrfs_io_bio *io_bio, int err)
7980 {
7981         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7982
7983         if (skip_csum) {
7984                 if (unlikely(err))
7985                         return __btrfs_correct_data_nocsum(inode, io_bio);
7986                 else
7987                         return 0;
7988         } else {
7989                 return __btrfs_subio_endio_read(inode, io_bio, err);
7990         }
7991 }
7992
7993 static void btrfs_endio_direct_read(struct bio *bio)
7994 {
7995         struct btrfs_dio_private *dip = bio->bi_private;
7996         struct inode *inode = dip->inode;
7997         struct bio *dio_bio;
7998         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7999         int err = bio->bi_error;
8000
8001         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8002                 err = btrfs_subio_endio_read(inode, io_bio, err);
8003
8004         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8005                       dip->logical_offset + dip->bytes - 1);
8006         dio_bio = dip->dio_bio;
8007
8008         kfree(dip);
8009
8010         dio_end_io(dio_bio, bio->bi_error);
8011
8012         if (io_bio->end_io)
8013                 io_bio->end_io(io_bio, err);
8014         bio_put(bio);
8015 }
8016
8017 static void btrfs_endio_direct_write(struct bio *bio)
8018 {
8019         struct btrfs_dio_private *dip = bio->bi_private;
8020         struct inode *inode = dip->inode;
8021         struct btrfs_root *root = BTRFS_I(inode)->root;
8022         struct btrfs_ordered_extent *ordered = NULL;
8023         u64 ordered_offset = dip->logical_offset;
8024         u64 ordered_bytes = dip->bytes;
8025         struct bio *dio_bio;
8026         int ret;
8027
8028 again:
8029         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8030                                                    &ordered_offset,
8031                                                    ordered_bytes,
8032                                                    !bio->bi_error);
8033         if (!ret)
8034                 goto out_test;
8035
8036         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8037                         finish_ordered_fn, NULL, NULL);
8038         btrfs_queue_work(root->fs_info->endio_write_workers,
8039                          &ordered->work);
8040 out_test:
8041         /*
8042          * our bio might span multiple ordered extents.  If we haven't
8043          * completed the accounting for the whole dio, go back and try again
8044          */
8045         if (ordered_offset < dip->logical_offset + dip->bytes) {
8046                 ordered_bytes = dip->logical_offset + dip->bytes -
8047                         ordered_offset;
8048                 ordered = NULL;
8049                 goto again;
8050         }
8051         dio_bio = dip->dio_bio;
8052
8053         kfree(dip);
8054
8055         dio_end_io(dio_bio, bio->bi_error);
8056         bio_put(bio);
8057 }
8058
8059 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8060                                     struct bio *bio, int mirror_num,
8061                                     unsigned long bio_flags, u64 offset)
8062 {
8063         int ret;
8064         struct btrfs_root *root = BTRFS_I(inode)->root;
8065         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8066         BUG_ON(ret); /* -ENOMEM */
8067         return 0;
8068 }
8069
8070 static void btrfs_end_dio_bio(struct bio *bio)
8071 {
8072         struct btrfs_dio_private *dip = bio->bi_private;
8073         int err = bio->bi_error;
8074
8075         if (err)
8076                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8077                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8078                            btrfs_ino(dip->inode), bio->bi_rw,
8079                            (unsigned long long)bio->bi_iter.bi_sector,
8080                            bio->bi_iter.bi_size, err);
8081
8082         if (dip->subio_endio)
8083                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8084
8085         if (err) {
8086                 dip->errors = 1;
8087
8088                 /*
8089                  * before atomic variable goto zero, we must make sure
8090                  * dip->errors is perceived to be set.
8091                  */
8092                 smp_mb__before_atomic();
8093         }
8094
8095         /* if there are more bios still pending for this dio, just exit */
8096         if (!atomic_dec_and_test(&dip->pending_bios))
8097                 goto out;
8098
8099         if (dip->errors) {
8100                 bio_io_error(dip->orig_bio);
8101         } else {
8102                 dip->dio_bio->bi_error = 0;
8103                 bio_endio(dip->orig_bio);
8104         }
8105 out:
8106         bio_put(bio);
8107 }
8108
8109 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8110                                        u64 first_sector, gfp_t gfp_flags)
8111 {
8112         struct bio *bio;
8113         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8114         if (bio)
8115                 bio_associate_current(bio);
8116         return bio;
8117 }
8118
8119 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8120                                                  struct inode *inode,
8121                                                  struct btrfs_dio_private *dip,
8122                                                  struct bio *bio,
8123                                                  u64 file_offset)
8124 {
8125         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8126         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8127         int ret;
8128
8129         /*
8130          * We load all the csum data we need when we submit
8131          * the first bio to reduce the csum tree search and
8132          * contention.
8133          */
8134         if (dip->logical_offset == file_offset) {
8135                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8136                                                 file_offset);
8137                 if (ret)
8138                         return ret;
8139         }
8140
8141         if (bio == dip->orig_bio)
8142                 return 0;
8143
8144         file_offset -= dip->logical_offset;
8145         file_offset >>= inode->i_sb->s_blocksize_bits;
8146         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8147
8148         return 0;
8149 }
8150
8151 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8152                                          int rw, u64 file_offset, int skip_sum,
8153                                          int async_submit)
8154 {
8155         struct btrfs_dio_private *dip = bio->bi_private;
8156         int write = rw & REQ_WRITE;
8157         struct btrfs_root *root = BTRFS_I(inode)->root;
8158         int ret;
8159
8160         if (async_submit)
8161                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8162
8163         bio_get(bio);
8164
8165         if (!write) {
8166                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8167                                 BTRFS_WQ_ENDIO_DATA);
8168                 if (ret)
8169                         goto err;
8170         }
8171
8172         if (skip_sum)
8173                 goto map;
8174
8175         if (write && async_submit) {
8176                 ret = btrfs_wq_submit_bio(root->fs_info,
8177                                    inode, rw, bio, 0, 0,
8178                                    file_offset,
8179                                    __btrfs_submit_bio_start_direct_io,
8180                                    __btrfs_submit_bio_done);
8181                 goto err;
8182         } else if (write) {
8183                 /*
8184                  * If we aren't doing async submit, calculate the csum of the
8185                  * bio now.
8186                  */
8187                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8188                 if (ret)
8189                         goto err;
8190         } else {
8191                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8192                                                      file_offset);
8193                 if (ret)
8194                         goto err;
8195         }
8196 map:
8197         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8198 err:
8199         bio_put(bio);
8200         return ret;
8201 }
8202
8203 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8204                                     int skip_sum)
8205 {
8206         struct inode *inode = dip->inode;
8207         struct btrfs_root *root = BTRFS_I(inode)->root;
8208         struct bio *bio;
8209         struct bio *orig_bio = dip->orig_bio;
8210         struct bio_vec *bvec = orig_bio->bi_io_vec;
8211         u64 start_sector = orig_bio->bi_iter.bi_sector;
8212         u64 file_offset = dip->logical_offset;
8213         u64 submit_len = 0;
8214         u64 map_length;
8215         int nr_pages = 0;
8216         int ret;
8217         int async_submit = 0;
8218
8219         map_length = orig_bio->bi_iter.bi_size;
8220         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8221                               &map_length, NULL, 0);
8222         if (ret)
8223                 return -EIO;
8224
8225         if (map_length >= orig_bio->bi_iter.bi_size) {
8226                 bio = orig_bio;
8227                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8228                 goto submit;
8229         }
8230
8231         /* async crcs make it difficult to collect full stripe writes. */
8232         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8233                 async_submit = 0;
8234         else
8235                 async_submit = 1;
8236
8237         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8238         if (!bio)
8239                 return -ENOMEM;
8240
8241         bio->bi_private = dip;
8242         bio->bi_end_io = btrfs_end_dio_bio;
8243         btrfs_io_bio(bio)->logical = file_offset;
8244         atomic_inc(&dip->pending_bios);
8245
8246         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8247                 if (map_length < submit_len + bvec->bv_len ||
8248                     bio_add_page(bio, bvec->bv_page, bvec->bv_len,
8249                                  bvec->bv_offset) < bvec->bv_len) {
8250                         /*
8251                          * inc the count before we submit the bio so
8252                          * we know the end IO handler won't happen before
8253                          * we inc the count. Otherwise, the dip might get freed
8254                          * before we're done setting it up
8255                          */
8256                         atomic_inc(&dip->pending_bios);
8257                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8258                                                      file_offset, skip_sum,
8259                                                      async_submit);
8260                         if (ret) {
8261                                 bio_put(bio);
8262                                 atomic_dec(&dip->pending_bios);
8263                                 goto out_err;
8264                         }
8265
8266                         start_sector += submit_len >> 9;
8267                         file_offset += submit_len;
8268
8269                         submit_len = 0;
8270                         nr_pages = 0;
8271
8272                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8273                                                   start_sector, GFP_NOFS);
8274                         if (!bio)
8275                                 goto out_err;
8276                         bio->bi_private = dip;
8277                         bio->bi_end_io = btrfs_end_dio_bio;
8278                         btrfs_io_bio(bio)->logical = file_offset;
8279
8280                         map_length = orig_bio->bi_iter.bi_size;
8281                         ret = btrfs_map_block(root->fs_info, rw,
8282                                               start_sector << 9,
8283                                               &map_length, NULL, 0);
8284                         if (ret) {
8285                                 bio_put(bio);
8286                                 goto out_err;
8287                         }
8288                 } else {
8289                         submit_len += bvec->bv_len;
8290                         nr_pages++;
8291                         bvec++;
8292                 }
8293         }
8294
8295 submit:
8296         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8297                                      async_submit);
8298         if (!ret)
8299                 return 0;
8300
8301         bio_put(bio);
8302 out_err:
8303         dip->errors = 1;
8304         /*
8305          * before atomic variable goto zero, we must
8306          * make sure dip->errors is perceived to be set.
8307          */
8308         smp_mb__before_atomic();
8309         if (atomic_dec_and_test(&dip->pending_bios))
8310                 bio_io_error(dip->orig_bio);
8311
8312         /* bio_end_io() will handle error, so we needn't return it */
8313         return 0;
8314 }
8315
8316 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8317                                 struct inode *inode, loff_t file_offset)
8318 {
8319         struct btrfs_dio_private *dip = NULL;
8320         struct bio *io_bio = NULL;
8321         struct btrfs_io_bio *btrfs_bio;
8322         int skip_sum;
8323         int write = rw & REQ_WRITE;
8324         int ret = 0;
8325
8326         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8327
8328         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8329         if (!io_bio) {
8330                 ret = -ENOMEM;
8331                 goto free_ordered;
8332         }
8333
8334         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8335         if (!dip) {
8336                 ret = -ENOMEM;
8337                 goto free_ordered;
8338         }
8339
8340         dip->private = dio_bio->bi_private;
8341         dip->inode = inode;
8342         dip->logical_offset = file_offset;
8343         dip->bytes = dio_bio->bi_iter.bi_size;
8344         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8345         io_bio->bi_private = dip;
8346         dip->orig_bio = io_bio;
8347         dip->dio_bio = dio_bio;
8348         atomic_set(&dip->pending_bios, 0);
8349         btrfs_bio = btrfs_io_bio(io_bio);
8350         btrfs_bio->logical = file_offset;
8351
8352         if (write) {
8353                 io_bio->bi_end_io = btrfs_endio_direct_write;
8354         } else {
8355                 io_bio->bi_end_io = btrfs_endio_direct_read;
8356                 dip->subio_endio = btrfs_subio_endio_read;
8357         }
8358
8359         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8360         if (!ret)
8361                 return;
8362
8363         if (btrfs_bio->end_io)
8364                 btrfs_bio->end_io(btrfs_bio, ret);
8365
8366 free_ordered:
8367         /*
8368          * If we arrived here it means either we failed to submit the dip
8369          * or we either failed to clone the dio_bio or failed to allocate the
8370          * dip. If we cloned the dio_bio and allocated the dip, we can just
8371          * call bio_endio against our io_bio so that we get proper resource
8372          * cleanup if we fail to submit the dip, otherwise, we must do the
8373          * same as btrfs_endio_direct_[write|read] because we can't call these
8374          * callbacks - they require an allocated dip and a clone of dio_bio.
8375          */
8376         if (io_bio && dip) {
8377                 io_bio->bi_error = -EIO;
8378                 bio_endio(io_bio);
8379                 /*
8380                  * The end io callbacks free our dip, do the final put on io_bio
8381                  * and all the cleanup and final put for dio_bio (through
8382                  * dio_end_io()).
8383                  */
8384                 dip = NULL;
8385                 io_bio = NULL;
8386         } else {
8387                 if (write) {
8388                         struct btrfs_ordered_extent *ordered;
8389
8390                         ordered = btrfs_lookup_ordered_extent(inode,
8391                                                               file_offset);
8392                         set_bit(BTRFS_ORDERED_IOERR, &ordered->flags);
8393                         /*
8394                          * Decrements our ref on the ordered extent and removes
8395                          * the ordered extent from the inode's ordered tree,
8396                          * doing all the proper resource cleanup such as for the
8397                          * reserved space and waking up any waiters for this
8398                          * ordered extent (through btrfs_remove_ordered_extent).
8399                          */
8400                         btrfs_finish_ordered_io(ordered);
8401                 } else {
8402                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8403                               file_offset + dio_bio->bi_iter.bi_size - 1);
8404                 }
8405                 dio_bio->bi_error = -EIO;
8406                 /*
8407                  * Releases and cleans up our dio_bio, no need to bio_put()
8408                  * nor bio_endio()/bio_io_error() against dio_bio.
8409                  */
8410                 dio_end_io(dio_bio, ret);
8411         }
8412         if (io_bio)
8413                 bio_put(io_bio);
8414         kfree(dip);
8415 }
8416
8417 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8418                         const struct iov_iter *iter, loff_t offset)
8419 {
8420         int seg;
8421         int i;
8422         unsigned blocksize_mask = root->sectorsize - 1;
8423         ssize_t retval = -EINVAL;
8424
8425         if (offset & blocksize_mask)
8426                 goto out;
8427
8428         if (iov_iter_alignment(iter) & blocksize_mask)
8429                 goto out;
8430
8431         /* If this is a write we don't need to check anymore */
8432         if (iov_iter_rw(iter) == WRITE)
8433                 return 0;
8434         /*
8435          * Check to make sure we don't have duplicate iov_base's in this
8436          * iovec, if so return EINVAL, otherwise we'll get csum errors
8437          * when reading back.
8438          */
8439         for (seg = 0; seg < iter->nr_segs; seg++) {
8440                 for (i = seg + 1; i < iter->nr_segs; i++) {
8441                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8442                                 goto out;
8443                 }
8444         }
8445         retval = 0;
8446 out:
8447         return retval;
8448 }
8449
8450 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
8451                                loff_t offset)
8452 {
8453         struct file *file = iocb->ki_filp;
8454         struct inode *inode = file->f_mapping->host;
8455         struct btrfs_root *root = BTRFS_I(inode)->root;
8456         struct btrfs_dio_data dio_data = { 0 };
8457         size_t count = 0;
8458         int flags = 0;
8459         bool wakeup = true;
8460         bool relock = false;
8461         ssize_t ret;
8462
8463         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8464                 return 0;
8465
8466         inode_dio_begin(inode);
8467         smp_mb__after_atomic();
8468
8469         /*
8470          * The generic stuff only does filemap_write_and_wait_range, which
8471          * isn't enough if we've written compressed pages to this area, so
8472          * we need to flush the dirty pages again to make absolutely sure
8473          * that any outstanding dirty pages are on disk.
8474          */
8475         count = iov_iter_count(iter);
8476         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8477                      &BTRFS_I(inode)->runtime_flags))
8478                 filemap_fdatawrite_range(inode->i_mapping, offset,
8479                                          offset + count - 1);
8480
8481         if (iov_iter_rw(iter) == WRITE) {
8482                 /*
8483                  * If the write DIO is beyond the EOF, we need update
8484                  * the isize, but it is protected by i_mutex. So we can
8485                  * not unlock the i_mutex at this case.
8486                  */
8487                 if (offset + count <= inode->i_size) {
8488                         mutex_unlock(&inode->i_mutex);
8489                         relock = true;
8490                 }
8491                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8492                 if (ret)
8493                         goto out;
8494                 dio_data.outstanding_extents = div64_u64(count +
8495                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8496                                                 BTRFS_MAX_EXTENT_SIZE);
8497
8498                 /*
8499                  * We need to know how many extents we reserved so that we can
8500                  * do the accounting properly if we go over the number we
8501                  * originally calculated.  Abuse current->journal_info for this.
8502                  */
8503                 dio_data.reserve = round_up(count, root->sectorsize);
8504                 current->journal_info = &dio_data;
8505         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8506                                      &BTRFS_I(inode)->runtime_flags)) {
8507                 inode_dio_end(inode);
8508                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8509                 wakeup = false;
8510         }
8511
8512         ret = __blockdev_direct_IO(iocb, inode,
8513                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8514                                    iter, offset, btrfs_get_blocks_direct, NULL,
8515                                    btrfs_submit_direct, flags);
8516         if (iov_iter_rw(iter) == WRITE) {
8517                 current->journal_info = NULL;
8518                 if (ret < 0 && ret != -EIOCBQUEUED) {
8519                         if (dio_data.reserve)
8520                                 btrfs_delalloc_release_space(inode, offset,
8521                                                              dio_data.reserve);
8522                 } else if (ret >= 0 && (size_t)ret < count)
8523                         btrfs_delalloc_release_space(inode, offset,
8524                                                      count - (size_t)ret);
8525         }
8526 out:
8527         if (wakeup)
8528                 inode_dio_end(inode);
8529         if (relock)
8530                 mutex_lock(&inode->i_mutex);
8531
8532         return ret;
8533 }
8534
8535 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8536
8537 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8538                 __u64 start, __u64 len)
8539 {
8540         int     ret;
8541
8542         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8543         if (ret)
8544                 return ret;
8545
8546         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8547 }
8548
8549 int btrfs_readpage(struct file *file, struct page *page)
8550 {
8551         struct extent_io_tree *tree;
8552         tree = &BTRFS_I(page->mapping->host)->io_tree;
8553         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8554 }
8555
8556 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8557 {
8558         struct extent_io_tree *tree;
8559
8560
8561         if (current->flags & PF_MEMALLOC) {
8562                 redirty_page_for_writepage(wbc, page);
8563                 unlock_page(page);
8564                 return 0;
8565         }
8566         tree = &BTRFS_I(page->mapping->host)->io_tree;
8567         return extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8568 }
8569
8570 static int btrfs_writepages(struct address_space *mapping,
8571                             struct writeback_control *wbc)
8572 {
8573         struct extent_io_tree *tree;
8574
8575         tree = &BTRFS_I(mapping->host)->io_tree;
8576         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8577 }
8578
8579 static int
8580 btrfs_readpages(struct file *file, struct address_space *mapping,
8581                 struct list_head *pages, unsigned nr_pages)
8582 {
8583         struct extent_io_tree *tree;
8584         tree = &BTRFS_I(mapping->host)->io_tree;
8585         return extent_readpages(tree, mapping, pages, nr_pages,
8586                                 btrfs_get_extent);
8587 }
8588 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8589 {
8590         struct extent_io_tree *tree;
8591         struct extent_map_tree *map;
8592         int ret;
8593
8594         tree = &BTRFS_I(page->mapping->host)->io_tree;
8595         map = &BTRFS_I(page->mapping->host)->extent_tree;
8596         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8597         if (ret == 1) {
8598                 ClearPagePrivate(page);
8599                 set_page_private(page, 0);
8600                 page_cache_release(page);
8601         }
8602         return ret;
8603 }
8604
8605 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8606 {
8607         if (PageWriteback(page) || PageDirty(page))
8608                 return 0;
8609         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8610 }
8611
8612 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8613                                  unsigned int length)
8614 {
8615         struct inode *inode = page->mapping->host;
8616         struct extent_io_tree *tree;
8617         struct btrfs_ordered_extent *ordered;
8618         struct extent_state *cached_state = NULL;
8619         u64 page_start = page_offset(page);
8620         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
8621         int inode_evicting = inode->i_state & I_FREEING;
8622
8623         /*
8624          * we have the page locked, so new writeback can't start,
8625          * and the dirty bit won't be cleared while we are here.
8626          *
8627          * Wait for IO on this page so that we can safely clear
8628          * the PagePrivate2 bit and do ordered accounting
8629          */
8630         wait_on_page_writeback(page);
8631
8632         tree = &BTRFS_I(inode)->io_tree;
8633         if (offset) {
8634                 btrfs_releasepage(page, GFP_NOFS);
8635                 return;
8636         }
8637
8638         if (!inode_evicting)
8639                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8640         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8641         if (ordered) {
8642                 /*
8643                  * IO on this page will never be started, so we need
8644                  * to account for any ordered extents now
8645                  */
8646                 if (!inode_evicting)
8647                         clear_extent_bit(tree, page_start, page_end,
8648                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8649                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8650                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8651                                          GFP_NOFS);
8652                 /*
8653                  * whoever cleared the private bit is responsible
8654                  * for the finish_ordered_io
8655                  */
8656                 if (TestClearPagePrivate2(page)) {
8657                         struct btrfs_ordered_inode_tree *tree;
8658                         u64 new_len;
8659
8660                         tree = &BTRFS_I(inode)->ordered_tree;
8661
8662                         spin_lock_irq(&tree->lock);
8663                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8664                         new_len = page_start - ordered->file_offset;
8665                         if (new_len < ordered->truncated_len)
8666                                 ordered->truncated_len = new_len;
8667                         spin_unlock_irq(&tree->lock);
8668
8669                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8670                                                            page_start,
8671                                                            PAGE_CACHE_SIZE, 1))
8672                                 btrfs_finish_ordered_io(ordered);
8673                 }
8674                 btrfs_put_ordered_extent(ordered);
8675                 if (!inode_evicting) {
8676                         cached_state = NULL;
8677                         lock_extent_bits(tree, page_start, page_end,
8678                                          &cached_state);
8679                 }
8680         }
8681
8682         /*
8683          * Qgroup reserved space handler
8684          * Page here will be either
8685          * 1) Already written to disk
8686          *    In this case, its reserved space is released from data rsv map
8687          *    and will be freed by delayed_ref handler finally.
8688          *    So even we call qgroup_free_data(), it won't decrease reserved
8689          *    space.
8690          * 2) Not written to disk
8691          *    This means the reserved space should be freed here.
8692          */
8693         btrfs_qgroup_free_data(inode, page_start, PAGE_CACHE_SIZE);
8694         if (!inode_evicting) {
8695                 clear_extent_bit(tree, page_start, page_end,
8696                                  EXTENT_LOCKED | EXTENT_DIRTY |
8697                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8698                                  EXTENT_DEFRAG, 1, 1,
8699                                  &cached_state, GFP_NOFS);
8700
8701                 __btrfs_releasepage(page, GFP_NOFS);
8702         }
8703
8704         ClearPageChecked(page);
8705         if (PagePrivate(page)) {
8706                 ClearPagePrivate(page);
8707                 set_page_private(page, 0);
8708                 page_cache_release(page);
8709         }
8710 }
8711
8712 /*
8713  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8714  * called from a page fault handler when a page is first dirtied. Hence we must
8715  * be careful to check for EOF conditions here. We set the page up correctly
8716  * for a written page which means we get ENOSPC checking when writing into
8717  * holes and correct delalloc and unwritten extent mapping on filesystems that
8718  * support these features.
8719  *
8720  * We are not allowed to take the i_mutex here so we have to play games to
8721  * protect against truncate races as the page could now be beyond EOF.  Because
8722  * vmtruncate() writes the inode size before removing pages, once we have the
8723  * page lock we can determine safely if the page is beyond EOF. If it is not
8724  * beyond EOF, then the page is guaranteed safe against truncation until we
8725  * unlock the page.
8726  */
8727 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8728 {
8729         struct page *page = vmf->page;
8730         struct inode *inode = file_inode(vma->vm_file);
8731         struct btrfs_root *root = BTRFS_I(inode)->root;
8732         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8733         struct btrfs_ordered_extent *ordered;
8734         struct extent_state *cached_state = NULL;
8735         char *kaddr;
8736         unsigned long zero_start;
8737         loff_t size;
8738         int ret;
8739         int reserved = 0;
8740         u64 page_start;
8741         u64 page_end;
8742
8743         sb_start_pagefault(inode->i_sb);
8744         page_start = page_offset(page);
8745         page_end = page_start + PAGE_CACHE_SIZE - 1;
8746
8747         ret = btrfs_delalloc_reserve_space(inode, page_start,
8748                                            PAGE_CACHE_SIZE);
8749         if (!ret) {
8750                 ret = file_update_time(vma->vm_file);
8751                 reserved = 1;
8752         }
8753         if (ret) {
8754                 if (ret == -ENOMEM)
8755                         ret = VM_FAULT_OOM;
8756                 else /* -ENOSPC, -EIO, etc */
8757                         ret = VM_FAULT_SIGBUS;
8758                 if (reserved)
8759                         goto out;
8760                 goto out_noreserve;
8761         }
8762
8763         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8764 again:
8765         lock_page(page);
8766         size = i_size_read(inode);
8767
8768         if ((page->mapping != inode->i_mapping) ||
8769             (page_start >= size)) {
8770                 /* page got truncated out from underneath us */
8771                 goto out_unlock;
8772         }
8773         wait_on_page_writeback(page);
8774
8775         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8776         set_page_extent_mapped(page);
8777
8778         /*
8779          * we can't set the delalloc bits if there are pending ordered
8780          * extents.  Drop our locks and wait for them to finish
8781          */
8782         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8783         if (ordered) {
8784                 unlock_extent_cached(io_tree, page_start, page_end,
8785                                      &cached_state, GFP_NOFS);
8786                 unlock_page(page);
8787                 btrfs_start_ordered_extent(inode, ordered, 1);
8788                 btrfs_put_ordered_extent(ordered);
8789                 goto again;
8790         }
8791
8792         /*
8793          * XXX - page_mkwrite gets called every time the page is dirtied, even
8794          * if it was already dirty, so for space accounting reasons we need to
8795          * clear any delalloc bits for the range we are fixing to save.  There
8796          * is probably a better way to do this, but for now keep consistent with
8797          * prepare_pages in the normal write path.
8798          */
8799         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
8800                           EXTENT_DIRTY | EXTENT_DELALLOC |
8801                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8802                           0, 0, &cached_state, GFP_NOFS);
8803
8804         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
8805                                         &cached_state);
8806         if (ret) {
8807                 unlock_extent_cached(io_tree, page_start, page_end,
8808                                      &cached_state, GFP_NOFS);
8809                 ret = VM_FAULT_SIGBUS;
8810                 goto out_unlock;
8811         }
8812         ret = 0;
8813
8814         /* page is wholly or partially inside EOF */
8815         if (page_start + PAGE_CACHE_SIZE > size)
8816                 zero_start = size & ~PAGE_CACHE_MASK;
8817         else
8818                 zero_start = PAGE_CACHE_SIZE;
8819
8820         if (zero_start != PAGE_CACHE_SIZE) {
8821                 kaddr = kmap(page);
8822                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
8823                 flush_dcache_page(page);
8824                 kunmap(page);
8825         }
8826         ClearPageChecked(page);
8827         set_page_dirty(page);
8828         SetPageUptodate(page);
8829
8830         BTRFS_I(inode)->last_trans = root->fs_info->generation;
8831         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8832         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8833
8834         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
8835
8836 out_unlock:
8837         if (!ret) {
8838                 sb_end_pagefault(inode->i_sb);
8839                 return VM_FAULT_LOCKED;
8840         }
8841         unlock_page(page);
8842 out:
8843         btrfs_delalloc_release_space(inode, page_start, PAGE_CACHE_SIZE);
8844 out_noreserve:
8845         sb_end_pagefault(inode->i_sb);
8846         return ret;
8847 }
8848
8849 static int btrfs_truncate(struct inode *inode)
8850 {
8851         struct btrfs_root *root = BTRFS_I(inode)->root;
8852         struct btrfs_block_rsv *rsv;
8853         int ret = 0;
8854         int err = 0;
8855         struct btrfs_trans_handle *trans;
8856         u64 mask = root->sectorsize - 1;
8857         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
8858
8859         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8860                                        (u64)-1);
8861         if (ret)
8862                 return ret;
8863
8864         /*
8865          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
8866          * 3 things going on here
8867          *
8868          * 1) We need to reserve space for our orphan item and the space to
8869          * delete our orphan item.  Lord knows we don't want to have a dangling
8870          * orphan item because we didn't reserve space to remove it.
8871          *
8872          * 2) We need to reserve space to update our inode.
8873          *
8874          * 3) We need to have something to cache all the space that is going to
8875          * be free'd up by the truncate operation, but also have some slack
8876          * space reserved in case it uses space during the truncate (thank you
8877          * very much snapshotting).
8878          *
8879          * And we need these to all be seperate.  The fact is we can use alot of
8880          * space doing the truncate, and we have no earthly idea how much space
8881          * we will use, so we need the truncate reservation to be seperate so it
8882          * doesn't end up using space reserved for updating the inode or
8883          * removing the orphan item.  We also need to be able to stop the
8884          * transaction and start a new one, which means we need to be able to
8885          * update the inode several times, and we have no idea of knowing how
8886          * many times that will be, so we can't just reserve 1 item for the
8887          * entirety of the opration, so that has to be done seperately as well.
8888          * Then there is the orphan item, which does indeed need to be held on
8889          * to for the whole operation, and we need nobody to touch this reserved
8890          * space except the orphan code.
8891          *
8892          * So that leaves us with
8893          *
8894          * 1) root->orphan_block_rsv - for the orphan deletion.
8895          * 2) rsv - for the truncate reservation, which we will steal from the
8896          * transaction reservation.
8897          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
8898          * updating the inode.
8899          */
8900         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
8901         if (!rsv)
8902                 return -ENOMEM;
8903         rsv->size = min_size;
8904         rsv->failfast = 1;
8905
8906         /*
8907          * 1 for the truncate slack space
8908          * 1 for updating the inode.
8909          */
8910         trans = btrfs_start_transaction(root, 2);
8911         if (IS_ERR(trans)) {
8912                 err = PTR_ERR(trans);
8913                 goto out;
8914         }
8915
8916         /* Migrate the slack space for the truncate to our reserve */
8917         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
8918                                       min_size);
8919         BUG_ON(ret);
8920
8921         /*
8922          * So if we truncate and then write and fsync we normally would just
8923          * write the extents that changed, which is a problem if we need to
8924          * first truncate that entire inode.  So set this flag so we write out
8925          * all of the extents in the inode to the sync log so we're completely
8926          * safe.
8927          */
8928         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8929         trans->block_rsv = rsv;
8930
8931         while (1) {
8932                 ret = btrfs_truncate_inode_items(trans, root, inode,
8933                                                  inode->i_size,
8934                                                  BTRFS_EXTENT_DATA_KEY);
8935                 if (ret != -ENOSPC && ret != -EAGAIN) {
8936                         err = ret;
8937                         break;
8938                 }
8939
8940                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8941                 ret = btrfs_update_inode(trans, root, inode);
8942                 if (ret) {
8943                         err = ret;
8944                         break;
8945                 }
8946
8947                 btrfs_end_transaction(trans, root);
8948                 btrfs_btree_balance_dirty(root);
8949
8950                 trans = btrfs_start_transaction(root, 2);
8951                 if (IS_ERR(trans)) {
8952                         ret = err = PTR_ERR(trans);
8953                         trans = NULL;
8954                         break;
8955                 }
8956
8957                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
8958                                               rsv, min_size);
8959                 BUG_ON(ret);    /* shouldn't happen */
8960                 trans->block_rsv = rsv;
8961         }
8962
8963         if (ret == 0 && inode->i_nlink > 0) {
8964                 trans->block_rsv = root->orphan_block_rsv;
8965                 ret = btrfs_orphan_del(trans, inode);
8966                 if (ret)
8967                         err = ret;
8968         }
8969
8970         if (trans) {
8971                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8972                 ret = btrfs_update_inode(trans, root, inode);
8973                 if (ret && !err)
8974                         err = ret;
8975
8976                 ret = btrfs_end_transaction(trans, root);
8977                 btrfs_btree_balance_dirty(root);
8978         }
8979
8980 out:
8981         btrfs_free_block_rsv(root, rsv);
8982
8983         if (ret && !err)
8984                 err = ret;
8985
8986         return err;
8987 }
8988
8989 /*
8990  * create a new subvolume directory/inode (helper for the ioctl).
8991  */
8992 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
8993                              struct btrfs_root *new_root,
8994                              struct btrfs_root *parent_root,
8995                              u64 new_dirid)
8996 {
8997         struct inode *inode;
8998         int err;
8999         u64 index = 0;
9000
9001         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9002                                 new_dirid, new_dirid,
9003                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9004                                 &index);
9005         if (IS_ERR(inode))
9006                 return PTR_ERR(inode);
9007         inode->i_op = &btrfs_dir_inode_operations;
9008         inode->i_fop = &btrfs_dir_file_operations;
9009
9010         set_nlink(inode, 1);
9011         btrfs_i_size_write(inode, 0);
9012         unlock_new_inode(inode);
9013
9014         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9015         if (err)
9016                 btrfs_err(new_root->fs_info,
9017                           "error inheriting subvolume %llu properties: %d",
9018                           new_root->root_key.objectid, err);
9019
9020         err = btrfs_update_inode(trans, new_root, inode);
9021
9022         iput(inode);
9023         return err;
9024 }
9025
9026 struct inode *btrfs_alloc_inode(struct super_block *sb)
9027 {
9028         struct btrfs_inode *ei;
9029         struct inode *inode;
9030
9031         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9032         if (!ei)
9033                 return NULL;
9034
9035         ei->root = NULL;
9036         ei->generation = 0;
9037         ei->last_trans = 0;
9038         ei->last_sub_trans = 0;
9039         ei->logged_trans = 0;
9040         ei->delalloc_bytes = 0;
9041         ei->defrag_bytes = 0;
9042         ei->disk_i_size = 0;
9043         ei->flags = 0;
9044         ei->csum_bytes = 0;
9045         ei->index_cnt = (u64)-1;
9046         ei->dir_index = 0;
9047         ei->last_unlink_trans = 0;
9048         ei->last_log_commit = 0;
9049
9050         spin_lock_init(&ei->lock);
9051         ei->outstanding_extents = 0;
9052         ei->reserved_extents = 0;
9053
9054         ei->runtime_flags = 0;
9055         ei->force_compress = BTRFS_COMPRESS_NONE;
9056
9057         ei->delayed_node = NULL;
9058
9059         ei->i_otime.tv_sec = 0;
9060         ei->i_otime.tv_nsec = 0;
9061
9062         inode = &ei->vfs_inode;
9063         extent_map_tree_init(&ei->extent_tree);
9064         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9065         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9066         ei->io_tree.track_uptodate = 1;
9067         ei->io_failure_tree.track_uptodate = 1;
9068         atomic_set(&ei->sync_writers, 0);
9069         mutex_init(&ei->log_mutex);
9070         mutex_init(&ei->delalloc_mutex);
9071         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9072         INIT_LIST_HEAD(&ei->delalloc_inodes);
9073         RB_CLEAR_NODE(&ei->rb_node);
9074
9075         return inode;
9076 }
9077
9078 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9079 void btrfs_test_destroy_inode(struct inode *inode)
9080 {
9081         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9082         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9083 }
9084 #endif
9085
9086 static void btrfs_i_callback(struct rcu_head *head)
9087 {
9088         struct inode *inode = container_of(head, struct inode, i_rcu);
9089         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9090 }
9091
9092 void btrfs_destroy_inode(struct inode *inode)
9093 {
9094         struct btrfs_ordered_extent *ordered;
9095         struct btrfs_root *root = BTRFS_I(inode)->root;
9096
9097         WARN_ON(!hlist_empty(&inode->i_dentry));
9098         WARN_ON(inode->i_data.nrpages);
9099         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9100         WARN_ON(BTRFS_I(inode)->reserved_extents);
9101         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9102         WARN_ON(BTRFS_I(inode)->csum_bytes);
9103         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9104
9105         /*
9106          * This can happen where we create an inode, but somebody else also
9107          * created the same inode and we need to destroy the one we already
9108          * created.
9109          */
9110         if (!root)
9111                 goto free;
9112
9113         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9114                      &BTRFS_I(inode)->runtime_flags)) {
9115                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9116                         btrfs_ino(inode));
9117                 atomic_dec(&root->orphan_inodes);
9118         }
9119
9120         while (1) {
9121                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9122                 if (!ordered)
9123                         break;
9124                 else {
9125                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9126                                 ordered->file_offset, ordered->len);
9127                         btrfs_remove_ordered_extent(inode, ordered);
9128                         btrfs_put_ordered_extent(ordered);
9129                         btrfs_put_ordered_extent(ordered);
9130                 }
9131         }
9132         btrfs_qgroup_check_reserved_leak(inode);
9133         inode_tree_del(inode);
9134         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9135 free:
9136         call_rcu(&inode->i_rcu, btrfs_i_callback);
9137 }
9138
9139 int btrfs_drop_inode(struct inode *inode)
9140 {
9141         struct btrfs_root *root = BTRFS_I(inode)->root;
9142
9143         if (root == NULL)
9144                 return 1;
9145
9146         /* the snap/subvol tree is on deleting */
9147         if (btrfs_root_refs(&root->root_item) == 0)
9148                 return 1;
9149         else
9150                 return generic_drop_inode(inode);
9151 }
9152
9153 static void init_once(void *foo)
9154 {
9155         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9156
9157         inode_init_once(&ei->vfs_inode);
9158 }
9159
9160 void btrfs_destroy_cachep(void)
9161 {
9162         /*
9163          * Make sure all delayed rcu free inodes are flushed before we
9164          * destroy cache.
9165          */
9166         rcu_barrier();
9167         if (btrfs_inode_cachep)
9168                 kmem_cache_destroy(btrfs_inode_cachep);
9169         if (btrfs_trans_handle_cachep)
9170                 kmem_cache_destroy(btrfs_trans_handle_cachep);
9171         if (btrfs_transaction_cachep)
9172                 kmem_cache_destroy(btrfs_transaction_cachep);
9173         if (btrfs_path_cachep)
9174                 kmem_cache_destroy(btrfs_path_cachep);
9175         if (btrfs_free_space_cachep)
9176                 kmem_cache_destroy(btrfs_free_space_cachep);
9177         if (btrfs_delalloc_work_cachep)
9178                 kmem_cache_destroy(btrfs_delalloc_work_cachep);
9179 }
9180
9181 int btrfs_init_cachep(void)
9182 {
9183         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9184                         sizeof(struct btrfs_inode), 0,
9185                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, init_once);
9186         if (!btrfs_inode_cachep)
9187                 goto fail;
9188
9189         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9190                         sizeof(struct btrfs_trans_handle), 0,
9191                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9192         if (!btrfs_trans_handle_cachep)
9193                 goto fail;
9194
9195         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9196                         sizeof(struct btrfs_transaction), 0,
9197                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9198         if (!btrfs_transaction_cachep)
9199                 goto fail;
9200
9201         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9202                         sizeof(struct btrfs_path), 0,
9203                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9204         if (!btrfs_path_cachep)
9205                 goto fail;
9206
9207         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9208                         sizeof(struct btrfs_free_space), 0,
9209                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9210         if (!btrfs_free_space_cachep)
9211                 goto fail;
9212
9213         btrfs_delalloc_work_cachep = kmem_cache_create("btrfs_delalloc_work",
9214                         sizeof(struct btrfs_delalloc_work), 0,
9215                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
9216                         NULL);
9217         if (!btrfs_delalloc_work_cachep)
9218                 goto fail;
9219
9220         return 0;
9221 fail:
9222         btrfs_destroy_cachep();
9223         return -ENOMEM;
9224 }
9225
9226 static int btrfs_getattr(struct vfsmount *mnt,
9227                          struct dentry *dentry, struct kstat *stat)
9228 {
9229         u64 delalloc_bytes;
9230         struct inode *inode = d_inode(dentry);
9231         u32 blocksize = inode->i_sb->s_blocksize;
9232
9233         generic_fillattr(inode, stat);
9234         stat->dev = BTRFS_I(inode)->root->anon_dev;
9235         stat->blksize = PAGE_CACHE_SIZE;
9236
9237         spin_lock(&BTRFS_I(inode)->lock);
9238         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9239         spin_unlock(&BTRFS_I(inode)->lock);
9240         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9241                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9242         return 0;
9243 }
9244
9245 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9246                            struct inode *new_dir, struct dentry *new_dentry)
9247 {
9248         struct btrfs_trans_handle *trans;
9249         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9250         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9251         struct inode *new_inode = d_inode(new_dentry);
9252         struct inode *old_inode = d_inode(old_dentry);
9253         struct timespec ctime = CURRENT_TIME;
9254         u64 index = 0;
9255         u64 root_objectid;
9256         int ret;
9257         u64 old_ino = btrfs_ino(old_inode);
9258
9259         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9260                 return -EPERM;
9261
9262         /* we only allow rename subvolume link between subvolumes */
9263         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9264                 return -EXDEV;
9265
9266         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9267             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9268                 return -ENOTEMPTY;
9269
9270         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9271             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9272                 return -ENOTEMPTY;
9273
9274
9275         /* check for collisions, even if the  name isn't there */
9276         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9277                              new_dentry->d_name.name,
9278                              new_dentry->d_name.len);
9279
9280         if (ret) {
9281                 if (ret == -EEXIST) {
9282                         /* we shouldn't get
9283                          * eexist without a new_inode */
9284                         if (WARN_ON(!new_inode)) {
9285                                 return ret;
9286                         }
9287                 } else {
9288                         /* maybe -EOVERFLOW */
9289                         return ret;
9290                 }
9291         }
9292         ret = 0;
9293
9294         /*
9295          * we're using rename to replace one file with another.  Start IO on it
9296          * now so  we don't add too much work to the end of the transaction
9297          */
9298         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9299                 filemap_flush(old_inode->i_mapping);
9300
9301         /* close the racy window with snapshot create/destroy ioctl */
9302         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9303                 down_read(&root->fs_info->subvol_sem);
9304         /*
9305          * We want to reserve the absolute worst case amount of items.  So if
9306          * both inodes are subvols and we need to unlink them then that would
9307          * require 4 item modifications, but if they are both normal inodes it
9308          * would require 5 item modifications, so we'll assume their normal
9309          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9310          * should cover the worst case number of items we'll modify.
9311          */
9312         trans = btrfs_start_transaction(root, 11);
9313         if (IS_ERR(trans)) {
9314                 ret = PTR_ERR(trans);
9315                 goto out_notrans;
9316         }
9317
9318         if (dest != root)
9319                 btrfs_record_root_in_trans(trans, dest);
9320
9321         ret = btrfs_set_inode_index(new_dir, &index);
9322         if (ret)
9323                 goto out_fail;
9324
9325         BTRFS_I(old_inode)->dir_index = 0ULL;
9326         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9327                 /* force full log commit if subvolume involved. */
9328                 btrfs_set_log_full_commit(root->fs_info, trans);
9329         } else {
9330                 ret = btrfs_insert_inode_ref(trans, dest,
9331                                              new_dentry->d_name.name,
9332                                              new_dentry->d_name.len,
9333                                              old_ino,
9334                                              btrfs_ino(new_dir), index);
9335                 if (ret)
9336                         goto out_fail;
9337                 /*
9338                  * this is an ugly little race, but the rename is required
9339                  * to make sure that if we crash, the inode is either at the
9340                  * old name or the new one.  pinning the log transaction lets
9341                  * us make sure we don't allow a log commit to come in after
9342                  * we unlink the name but before we add the new name back in.
9343                  */
9344                 btrfs_pin_log_trans(root);
9345         }
9346
9347         inode_inc_iversion(old_dir);
9348         inode_inc_iversion(new_dir);
9349         inode_inc_iversion(old_inode);
9350         old_dir->i_ctime = old_dir->i_mtime = ctime;
9351         new_dir->i_ctime = new_dir->i_mtime = ctime;
9352         old_inode->i_ctime = ctime;
9353
9354         if (old_dentry->d_parent != new_dentry->d_parent)
9355                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9356
9357         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9358                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9359                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9360                                         old_dentry->d_name.name,
9361                                         old_dentry->d_name.len);
9362         } else {
9363                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9364                                         d_inode(old_dentry),
9365                                         old_dentry->d_name.name,
9366                                         old_dentry->d_name.len);
9367                 if (!ret)
9368                         ret = btrfs_update_inode(trans, root, old_inode);
9369         }
9370         if (ret) {
9371                 btrfs_abort_transaction(trans, root, ret);
9372                 goto out_fail;
9373         }
9374
9375         if (new_inode) {
9376                 inode_inc_iversion(new_inode);
9377                 new_inode->i_ctime = CURRENT_TIME;
9378                 if (unlikely(btrfs_ino(new_inode) ==
9379                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9380                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9381                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9382                                                 root_objectid,
9383                                                 new_dentry->d_name.name,
9384                                                 new_dentry->d_name.len);
9385                         BUG_ON(new_inode->i_nlink == 0);
9386                 } else {
9387                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9388                                                  d_inode(new_dentry),
9389                                                  new_dentry->d_name.name,
9390                                                  new_dentry->d_name.len);
9391                 }
9392                 if (!ret && new_inode->i_nlink == 0)
9393                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9394                 if (ret) {
9395                         btrfs_abort_transaction(trans, root, ret);
9396                         goto out_fail;
9397                 }
9398         }
9399
9400         ret = btrfs_add_link(trans, new_dir, old_inode,
9401                              new_dentry->d_name.name,
9402                              new_dentry->d_name.len, 0, index);
9403         if (ret) {
9404                 btrfs_abort_transaction(trans, root, ret);
9405                 goto out_fail;
9406         }
9407
9408         if (old_inode->i_nlink == 1)
9409                 BTRFS_I(old_inode)->dir_index = index;
9410
9411         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
9412                 struct dentry *parent = new_dentry->d_parent;
9413                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9414                 btrfs_end_log_trans(root);
9415         }
9416 out_fail:
9417         btrfs_end_transaction(trans, root);
9418 out_notrans:
9419         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9420                 up_read(&root->fs_info->subvol_sem);
9421
9422         return ret;
9423 }
9424
9425 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9426                          struct inode *new_dir, struct dentry *new_dentry,
9427                          unsigned int flags)
9428 {
9429         if (flags & ~RENAME_NOREPLACE)
9430                 return -EINVAL;
9431
9432         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry);
9433 }
9434
9435 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9436 {
9437         struct btrfs_delalloc_work *delalloc_work;
9438         struct inode *inode;
9439
9440         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9441                                      work);
9442         inode = delalloc_work->inode;
9443         if (delalloc_work->wait) {
9444                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
9445         } else {
9446                 filemap_flush(inode->i_mapping);
9447                 if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9448                              &BTRFS_I(inode)->runtime_flags))
9449                         filemap_flush(inode->i_mapping);
9450         }
9451
9452         if (delalloc_work->delay_iput)
9453                 btrfs_add_delayed_iput(inode);
9454         else
9455                 iput(inode);
9456         complete(&delalloc_work->completion);
9457 }
9458
9459 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9460                                                     int wait, int delay_iput)
9461 {
9462         struct btrfs_delalloc_work *work;
9463
9464         work = kmem_cache_zalloc(btrfs_delalloc_work_cachep, GFP_NOFS);
9465         if (!work)
9466                 return NULL;
9467
9468         init_completion(&work->completion);
9469         INIT_LIST_HEAD(&work->list);
9470         work->inode = inode;
9471         work->wait = wait;
9472         work->delay_iput = delay_iput;
9473         WARN_ON_ONCE(!inode);
9474         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9475                         btrfs_run_delalloc_work, NULL, NULL);
9476
9477         return work;
9478 }
9479
9480 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9481 {
9482         wait_for_completion(&work->completion);
9483         kmem_cache_free(btrfs_delalloc_work_cachep, work);
9484 }
9485
9486 /*
9487  * some fairly slow code that needs optimization. This walks the list
9488  * of all the inodes with pending delalloc and forces them to disk.
9489  */
9490 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9491                                    int nr)
9492 {
9493         struct btrfs_inode *binode;
9494         struct inode *inode;
9495         struct btrfs_delalloc_work *work, *next;
9496         struct list_head works;
9497         struct list_head splice;
9498         int ret = 0;
9499
9500         INIT_LIST_HEAD(&works);
9501         INIT_LIST_HEAD(&splice);
9502
9503         mutex_lock(&root->delalloc_mutex);
9504         spin_lock(&root->delalloc_lock);
9505         list_splice_init(&root->delalloc_inodes, &splice);
9506         while (!list_empty(&splice)) {
9507                 binode = list_entry(splice.next, struct btrfs_inode,
9508                                     delalloc_inodes);
9509
9510                 list_move_tail(&binode->delalloc_inodes,
9511                                &root->delalloc_inodes);
9512                 inode = igrab(&binode->vfs_inode);
9513                 if (!inode) {
9514                         cond_resched_lock(&root->delalloc_lock);
9515                         continue;
9516                 }
9517                 spin_unlock(&root->delalloc_lock);
9518
9519                 work = btrfs_alloc_delalloc_work(inode, 0, delay_iput);
9520                 if (!work) {
9521                         if (delay_iput)
9522                                 btrfs_add_delayed_iput(inode);
9523                         else
9524                                 iput(inode);
9525                         ret = -ENOMEM;
9526                         goto out;
9527                 }
9528                 list_add_tail(&work->list, &works);
9529                 btrfs_queue_work(root->fs_info->flush_workers,
9530                                  &work->work);
9531                 ret++;
9532                 if (nr != -1 && ret >= nr)
9533                         goto out;
9534                 cond_resched();
9535                 spin_lock(&root->delalloc_lock);
9536         }
9537         spin_unlock(&root->delalloc_lock);
9538
9539 out:
9540         list_for_each_entry_safe(work, next, &works, list) {
9541                 list_del_init(&work->list);
9542                 btrfs_wait_and_free_delalloc_work(work);
9543         }
9544
9545         if (!list_empty_careful(&splice)) {
9546                 spin_lock(&root->delalloc_lock);
9547                 list_splice_tail(&splice, &root->delalloc_inodes);
9548                 spin_unlock(&root->delalloc_lock);
9549         }
9550         mutex_unlock(&root->delalloc_mutex);
9551         return ret;
9552 }
9553
9554 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
9555 {
9556         int ret;
9557
9558         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
9559                 return -EROFS;
9560
9561         ret = __start_delalloc_inodes(root, delay_iput, -1);
9562         if (ret > 0)
9563                 ret = 0;
9564         /*
9565          * the filemap_flush will queue IO into the worker threads, but
9566          * we have to make sure the IO is actually started and that
9567          * ordered extents get created before we return
9568          */
9569         atomic_inc(&root->fs_info->async_submit_draining);
9570         while (atomic_read(&root->fs_info->nr_async_submits) ||
9571               atomic_read(&root->fs_info->async_delalloc_pages)) {
9572                 wait_event(root->fs_info->async_submit_wait,
9573                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
9574                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
9575         }
9576         atomic_dec(&root->fs_info->async_submit_draining);
9577         return ret;
9578 }
9579
9580 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
9581                                int nr)
9582 {
9583         struct btrfs_root *root;
9584         struct list_head splice;
9585         int ret;
9586
9587         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9588                 return -EROFS;
9589
9590         INIT_LIST_HEAD(&splice);
9591
9592         mutex_lock(&fs_info->delalloc_root_mutex);
9593         spin_lock(&fs_info->delalloc_root_lock);
9594         list_splice_init(&fs_info->delalloc_roots, &splice);
9595         while (!list_empty(&splice) && nr) {
9596                 root = list_first_entry(&splice, struct btrfs_root,
9597                                         delalloc_root);
9598                 root = btrfs_grab_fs_root(root);
9599                 BUG_ON(!root);
9600                 list_move_tail(&root->delalloc_root,
9601                                &fs_info->delalloc_roots);
9602                 spin_unlock(&fs_info->delalloc_root_lock);
9603
9604                 ret = __start_delalloc_inodes(root, delay_iput, nr);
9605                 btrfs_put_fs_root(root);
9606                 if (ret < 0)
9607                         goto out;
9608
9609                 if (nr != -1) {
9610                         nr -= ret;
9611                         WARN_ON(nr < 0);
9612                 }
9613                 spin_lock(&fs_info->delalloc_root_lock);
9614         }
9615         spin_unlock(&fs_info->delalloc_root_lock);
9616
9617         ret = 0;
9618         atomic_inc(&fs_info->async_submit_draining);
9619         while (atomic_read(&fs_info->nr_async_submits) ||
9620               atomic_read(&fs_info->async_delalloc_pages)) {
9621                 wait_event(fs_info->async_submit_wait,
9622                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
9623                     atomic_read(&fs_info->async_delalloc_pages) == 0));
9624         }
9625         atomic_dec(&fs_info->async_submit_draining);
9626 out:
9627         if (!list_empty_careful(&splice)) {
9628                 spin_lock(&fs_info->delalloc_root_lock);
9629                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9630                 spin_unlock(&fs_info->delalloc_root_lock);
9631         }
9632         mutex_unlock(&fs_info->delalloc_root_mutex);
9633         return ret;
9634 }
9635
9636 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9637                          const char *symname)
9638 {
9639         struct btrfs_trans_handle *trans;
9640         struct btrfs_root *root = BTRFS_I(dir)->root;
9641         struct btrfs_path *path;
9642         struct btrfs_key key;
9643         struct inode *inode = NULL;
9644         int err;
9645         int drop_inode = 0;
9646         u64 objectid;
9647         u64 index = 0;
9648         int name_len;
9649         int datasize;
9650         unsigned long ptr;
9651         struct btrfs_file_extent_item *ei;
9652         struct extent_buffer *leaf;
9653
9654         name_len = strlen(symname);
9655         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
9656                 return -ENAMETOOLONG;
9657
9658         /*
9659          * 2 items for inode item and ref
9660          * 2 items for dir items
9661          * 1 item for xattr if selinux is on
9662          */
9663         trans = btrfs_start_transaction(root, 5);
9664         if (IS_ERR(trans))
9665                 return PTR_ERR(trans);
9666
9667         err = btrfs_find_free_ino(root, &objectid);
9668         if (err)
9669                 goto out_unlock;
9670
9671         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9672                                 dentry->d_name.len, btrfs_ino(dir), objectid,
9673                                 S_IFLNK|S_IRWXUGO, &index);
9674         if (IS_ERR(inode)) {
9675                 err = PTR_ERR(inode);
9676                 goto out_unlock;
9677         }
9678
9679         /*
9680         * If the active LSM wants to access the inode during
9681         * d_instantiate it needs these. Smack checks to see
9682         * if the filesystem supports xattrs by looking at the
9683         * ops vector.
9684         */
9685         inode->i_fop = &btrfs_file_operations;
9686         inode->i_op = &btrfs_file_inode_operations;
9687         inode->i_mapping->a_ops = &btrfs_aops;
9688         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9689
9690         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9691         if (err)
9692                 goto out_unlock_inode;
9693
9694         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
9695         if (err)
9696                 goto out_unlock_inode;
9697
9698         path = btrfs_alloc_path();
9699         if (!path) {
9700                 err = -ENOMEM;
9701                 goto out_unlock_inode;
9702         }
9703         key.objectid = btrfs_ino(inode);
9704         key.offset = 0;
9705         key.type = BTRFS_EXTENT_DATA_KEY;
9706         datasize = btrfs_file_extent_calc_inline_size(name_len);
9707         err = btrfs_insert_empty_item(trans, root, path, &key,
9708                                       datasize);
9709         if (err) {
9710                 btrfs_free_path(path);
9711                 goto out_unlock_inode;
9712         }
9713         leaf = path->nodes[0];
9714         ei = btrfs_item_ptr(leaf, path->slots[0],
9715                             struct btrfs_file_extent_item);
9716         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9717         btrfs_set_file_extent_type(leaf, ei,
9718                                    BTRFS_FILE_EXTENT_INLINE);
9719         btrfs_set_file_extent_encryption(leaf, ei, 0);
9720         btrfs_set_file_extent_compression(leaf, ei, 0);
9721         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9722         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9723
9724         ptr = btrfs_file_extent_inline_start(ei);
9725         write_extent_buffer(leaf, symname, ptr, name_len);
9726         btrfs_mark_buffer_dirty(leaf);
9727         btrfs_free_path(path);
9728
9729         inode->i_op = &btrfs_symlink_inode_operations;
9730         inode->i_mapping->a_ops = &btrfs_symlink_aops;
9731         inode_set_bytes(inode, name_len);
9732         btrfs_i_size_write(inode, name_len);
9733         err = btrfs_update_inode(trans, root, inode);
9734         if (err) {
9735                 drop_inode = 1;
9736                 goto out_unlock_inode;
9737         }
9738
9739         unlock_new_inode(inode);
9740         d_instantiate(dentry, inode);
9741
9742 out_unlock:
9743         btrfs_end_transaction(trans, root);
9744         if (drop_inode) {
9745                 inode_dec_link_count(inode);
9746                 iput(inode);
9747         }
9748         btrfs_btree_balance_dirty(root);
9749         return err;
9750
9751 out_unlock_inode:
9752         drop_inode = 1;
9753         unlock_new_inode(inode);
9754         goto out_unlock;
9755 }
9756
9757 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9758                                        u64 start, u64 num_bytes, u64 min_size,
9759                                        loff_t actual_len, u64 *alloc_hint,
9760                                        struct btrfs_trans_handle *trans)
9761 {
9762         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9763         struct extent_map *em;
9764         struct btrfs_root *root = BTRFS_I(inode)->root;
9765         struct btrfs_key ins;
9766         u64 cur_offset = start;
9767         u64 i_size;
9768         u64 cur_bytes;
9769         u64 last_alloc = (u64)-1;
9770         int ret = 0;
9771         bool own_trans = true;
9772
9773         if (trans)
9774                 own_trans = false;
9775         while (num_bytes > 0) {
9776                 if (own_trans) {
9777                         trans = btrfs_start_transaction(root, 3);
9778                         if (IS_ERR(trans)) {
9779                                 ret = PTR_ERR(trans);
9780                                 break;
9781                         }
9782                 }
9783
9784                 cur_bytes = min(num_bytes, 256ULL * 1024 * 1024);
9785                 cur_bytes = max(cur_bytes, min_size);
9786                 /*
9787                  * If we are severely fragmented we could end up with really
9788                  * small allocations, so if the allocator is returning small
9789                  * chunks lets make its job easier by only searching for those
9790                  * sized chunks.
9791                  */
9792                 cur_bytes = min(cur_bytes, last_alloc);
9793                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
9794                                            *alloc_hint, &ins, 1, 0);
9795                 if (ret) {
9796                         if (own_trans)
9797                                 btrfs_end_transaction(trans, root);
9798                         break;
9799                 }
9800
9801                 last_alloc = ins.offset;
9802                 ret = insert_reserved_file_extent(trans, inode,
9803                                                   cur_offset, ins.objectid,
9804                                                   ins.offset, ins.offset,
9805                                                   ins.offset, 0, 0, 0,
9806                                                   BTRFS_FILE_EXTENT_PREALLOC);
9807                 if (ret) {
9808                         btrfs_free_reserved_extent(root, ins.objectid,
9809                                                    ins.offset, 0);
9810                         btrfs_abort_transaction(trans, root, ret);
9811                         if (own_trans)
9812                                 btrfs_end_transaction(trans, root);
9813                         break;
9814                 }
9815
9816                 btrfs_drop_extent_cache(inode, cur_offset,
9817                                         cur_offset + ins.offset -1, 0);
9818
9819                 em = alloc_extent_map();
9820                 if (!em) {
9821                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9822                                 &BTRFS_I(inode)->runtime_flags);
9823                         goto next;
9824                 }
9825
9826                 em->start = cur_offset;
9827                 em->orig_start = cur_offset;
9828                 em->len = ins.offset;
9829                 em->block_start = ins.objectid;
9830                 em->block_len = ins.offset;
9831                 em->orig_block_len = ins.offset;
9832                 em->ram_bytes = ins.offset;
9833                 em->bdev = root->fs_info->fs_devices->latest_bdev;
9834                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9835                 em->generation = trans->transid;
9836
9837                 while (1) {
9838                         write_lock(&em_tree->lock);
9839                         ret = add_extent_mapping(em_tree, em, 1);
9840                         write_unlock(&em_tree->lock);
9841                         if (ret != -EEXIST)
9842                                 break;
9843                         btrfs_drop_extent_cache(inode, cur_offset,
9844                                                 cur_offset + ins.offset - 1,
9845                                                 0);
9846                 }
9847                 free_extent_map(em);
9848 next:
9849                 num_bytes -= ins.offset;
9850                 cur_offset += ins.offset;
9851                 *alloc_hint = ins.objectid + ins.offset;
9852
9853                 inode_inc_iversion(inode);
9854                 inode->i_ctime = CURRENT_TIME;
9855                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9856                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9857                     (actual_len > inode->i_size) &&
9858                     (cur_offset > inode->i_size)) {
9859                         if (cur_offset > actual_len)
9860                                 i_size = actual_len;
9861                         else
9862                                 i_size = cur_offset;
9863                         i_size_write(inode, i_size);
9864                         btrfs_ordered_update_i_size(inode, i_size, NULL);
9865                 }
9866
9867                 ret = btrfs_update_inode(trans, root, inode);
9868
9869                 if (ret) {
9870                         btrfs_abort_transaction(trans, root, ret);
9871                         if (own_trans)
9872                                 btrfs_end_transaction(trans, root);
9873                         break;
9874                 }
9875
9876                 if (own_trans)
9877                         btrfs_end_transaction(trans, root);
9878         }
9879         return ret;
9880 }
9881
9882 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9883                               u64 start, u64 num_bytes, u64 min_size,
9884                               loff_t actual_len, u64 *alloc_hint)
9885 {
9886         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9887                                            min_size, actual_len, alloc_hint,
9888                                            NULL);
9889 }
9890
9891 int btrfs_prealloc_file_range_trans(struct inode *inode,
9892                                     struct btrfs_trans_handle *trans, int mode,
9893                                     u64 start, u64 num_bytes, u64 min_size,
9894                                     loff_t actual_len, u64 *alloc_hint)
9895 {
9896         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9897                                            min_size, actual_len, alloc_hint, trans);
9898 }
9899
9900 static int btrfs_set_page_dirty(struct page *page)
9901 {
9902         return __set_page_dirty_nobuffers(page);
9903 }
9904
9905 static int btrfs_permission(struct inode *inode, int mask)
9906 {
9907         struct btrfs_root *root = BTRFS_I(inode)->root;
9908         umode_t mode = inode->i_mode;
9909
9910         if (mask & MAY_WRITE &&
9911             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
9912                 if (btrfs_root_readonly(root))
9913                         return -EROFS;
9914                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
9915                         return -EACCES;
9916         }
9917         return generic_permission(inode, mask);
9918 }
9919
9920 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
9921 {
9922         struct btrfs_trans_handle *trans;
9923         struct btrfs_root *root = BTRFS_I(dir)->root;
9924         struct inode *inode = NULL;
9925         u64 objectid;
9926         u64 index;
9927         int ret = 0;
9928
9929         /*
9930          * 5 units required for adding orphan entry
9931          */
9932         trans = btrfs_start_transaction(root, 5);
9933         if (IS_ERR(trans))
9934                 return PTR_ERR(trans);
9935
9936         ret = btrfs_find_free_ino(root, &objectid);
9937         if (ret)
9938                 goto out;
9939
9940         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
9941                                 btrfs_ino(dir), objectid, mode, &index);
9942         if (IS_ERR(inode)) {
9943                 ret = PTR_ERR(inode);
9944                 inode = NULL;
9945                 goto out;
9946         }
9947
9948         inode->i_fop = &btrfs_file_operations;
9949         inode->i_op = &btrfs_file_inode_operations;
9950
9951         inode->i_mapping->a_ops = &btrfs_aops;
9952         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9953
9954         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
9955         if (ret)
9956                 goto out_inode;
9957
9958         ret = btrfs_update_inode(trans, root, inode);
9959         if (ret)
9960                 goto out_inode;
9961         ret = btrfs_orphan_add(trans, inode);
9962         if (ret)
9963                 goto out_inode;
9964
9965         /*
9966          * We set number of links to 0 in btrfs_new_inode(), and here we set
9967          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
9968          * through:
9969          *
9970          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
9971          */
9972         set_nlink(inode, 1);
9973         unlock_new_inode(inode);
9974         d_tmpfile(dentry, inode);
9975         mark_inode_dirty(inode);
9976
9977 out:
9978         btrfs_end_transaction(trans, root);
9979         if (ret)
9980                 iput(inode);
9981         btrfs_balance_delayed_items(root);
9982         btrfs_btree_balance_dirty(root);
9983         return ret;
9984
9985 out_inode:
9986         unlock_new_inode(inode);
9987         goto out;
9988
9989 }
9990
9991 /* Inspired by filemap_check_errors() */
9992 int btrfs_inode_check_errors(struct inode *inode)
9993 {
9994         int ret = 0;
9995
9996         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
9997             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
9998                 ret = -ENOSPC;
9999         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10000             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10001                 ret = -EIO;
10002
10003         return ret;
10004 }
10005
10006 static const struct inode_operations btrfs_dir_inode_operations = {
10007         .getattr        = btrfs_getattr,
10008         .lookup         = btrfs_lookup,
10009         .create         = btrfs_create,
10010         .unlink         = btrfs_unlink,
10011         .link           = btrfs_link,
10012         .mkdir          = btrfs_mkdir,
10013         .rmdir          = btrfs_rmdir,
10014         .rename2        = btrfs_rename2,
10015         .symlink        = btrfs_symlink,
10016         .setattr        = btrfs_setattr,
10017         .mknod          = btrfs_mknod,
10018         .setxattr       = btrfs_setxattr,
10019         .getxattr       = btrfs_getxattr,
10020         .listxattr      = btrfs_listxattr,
10021         .removexattr    = btrfs_removexattr,
10022         .permission     = btrfs_permission,
10023         .get_acl        = btrfs_get_acl,
10024         .set_acl        = btrfs_set_acl,
10025         .update_time    = btrfs_update_time,
10026         .tmpfile        = btrfs_tmpfile,
10027 };
10028 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10029         .lookup         = btrfs_lookup,
10030         .permission     = btrfs_permission,
10031         .get_acl        = btrfs_get_acl,
10032         .set_acl        = btrfs_set_acl,
10033         .update_time    = btrfs_update_time,
10034 };
10035
10036 static const struct file_operations btrfs_dir_file_operations = {
10037         .llseek         = generic_file_llseek,
10038         .read           = generic_read_dir,
10039         .iterate        = btrfs_real_readdir,
10040         .unlocked_ioctl = btrfs_ioctl,
10041 #ifdef CONFIG_COMPAT
10042         .compat_ioctl   = btrfs_ioctl,
10043 #endif
10044         .release        = btrfs_release_file,
10045         .fsync          = btrfs_sync_file,
10046 };
10047
10048 static struct extent_io_ops btrfs_extent_io_ops = {
10049         .fill_delalloc = run_delalloc_range,
10050         .submit_bio_hook = btrfs_submit_bio_hook,
10051         .merge_bio_hook = btrfs_merge_bio_hook,
10052         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10053         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10054         .writepage_start_hook = btrfs_writepage_start_hook,
10055         .set_bit_hook = btrfs_set_bit_hook,
10056         .clear_bit_hook = btrfs_clear_bit_hook,
10057         .merge_extent_hook = btrfs_merge_extent_hook,
10058         .split_extent_hook = btrfs_split_extent_hook,
10059 };
10060
10061 /*
10062  * btrfs doesn't support the bmap operation because swapfiles
10063  * use bmap to make a mapping of extents in the file.  They assume
10064  * these extents won't change over the life of the file and they
10065  * use the bmap result to do IO directly to the drive.
10066  *
10067  * the btrfs bmap call would return logical addresses that aren't
10068  * suitable for IO and they also will change frequently as COW
10069  * operations happen.  So, swapfile + btrfs == corruption.
10070  *
10071  * For now we're avoiding this by dropping bmap.
10072  */
10073 static const struct address_space_operations btrfs_aops = {
10074         .readpage       = btrfs_readpage,
10075         .writepage      = btrfs_writepage,
10076         .writepages     = btrfs_writepages,
10077         .readpages      = btrfs_readpages,
10078         .direct_IO      = btrfs_direct_IO,
10079         .invalidatepage = btrfs_invalidatepage,
10080         .releasepage    = btrfs_releasepage,
10081         .set_page_dirty = btrfs_set_page_dirty,
10082         .error_remove_page = generic_error_remove_page,
10083 };
10084
10085 static const struct address_space_operations btrfs_symlink_aops = {
10086         .readpage       = btrfs_readpage,
10087         .writepage      = btrfs_writepage,
10088         .invalidatepage = btrfs_invalidatepage,
10089         .releasepage    = btrfs_releasepage,
10090 };
10091
10092 static const struct inode_operations btrfs_file_inode_operations = {
10093         .getattr        = btrfs_getattr,
10094         .setattr        = btrfs_setattr,
10095         .setxattr       = btrfs_setxattr,
10096         .getxattr       = btrfs_getxattr,
10097         .listxattr      = btrfs_listxattr,
10098         .removexattr    = btrfs_removexattr,
10099         .permission     = btrfs_permission,
10100         .fiemap         = btrfs_fiemap,
10101         .get_acl        = btrfs_get_acl,
10102         .set_acl        = btrfs_set_acl,
10103         .update_time    = btrfs_update_time,
10104 };
10105 static const struct inode_operations btrfs_special_inode_operations = {
10106         .getattr        = btrfs_getattr,
10107         .setattr        = btrfs_setattr,
10108         .permission     = btrfs_permission,
10109         .setxattr       = btrfs_setxattr,
10110         .getxattr       = btrfs_getxattr,
10111         .listxattr      = btrfs_listxattr,
10112         .removexattr    = btrfs_removexattr,
10113         .get_acl        = btrfs_get_acl,
10114         .set_acl        = btrfs_set_acl,
10115         .update_time    = btrfs_update_time,
10116 };
10117 static const struct inode_operations btrfs_symlink_inode_operations = {
10118         .readlink       = generic_readlink,
10119         .follow_link    = page_follow_link_light,
10120         .put_link       = page_put_link,
10121         .getattr        = btrfs_getattr,
10122         .setattr        = btrfs_setattr,
10123         .permission     = btrfs_permission,
10124         .setxattr       = btrfs_setxattr,
10125         .getxattr       = btrfs_getxattr,
10126         .listxattr      = btrfs_listxattr,
10127         .removexattr    = btrfs_removexattr,
10128         .update_time    = btrfs_update_time,
10129 };
10130
10131 const struct dentry_operations btrfs_dentry_operations = {
10132         .d_delete       = btrfs_dentry_delete,
10133         .d_release      = btrfs_dentry_release,
10134 };