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, 0,
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                          0, &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                          0, &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(XATTR_NAME_POSIX_ACL_ACCESS,
3554                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3555                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3556                                         strlen(XATTR_NAME_POSIX_ACL_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_nohighmem(inode);
3778                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3779                 break;
3780         default:
3781                 inode->i_op = &btrfs_special_inode_operations;
3782                 init_special_inode(inode, inode->i_mode, rdev);
3783                 break;
3784         }
3785
3786         btrfs_update_iflags(inode);
3787         return;
3788
3789 make_bad:
3790         btrfs_free_path(path);
3791         make_bad_inode(inode);
3792 }
3793
3794 /*
3795  * given a leaf and an inode, copy the inode fields into the leaf
3796  */
3797 static void fill_inode_item(struct btrfs_trans_handle *trans,
3798                             struct extent_buffer *leaf,
3799                             struct btrfs_inode_item *item,
3800                             struct inode *inode)
3801 {
3802         struct btrfs_map_token token;
3803
3804         btrfs_init_map_token(&token);
3805
3806         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3807         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3808         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3809                                    &token);
3810         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3811         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3812
3813         btrfs_set_token_timespec_sec(leaf, &item->atime,
3814                                      inode->i_atime.tv_sec, &token);
3815         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3816                                       inode->i_atime.tv_nsec, &token);
3817
3818         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3819                                      inode->i_mtime.tv_sec, &token);
3820         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3821                                       inode->i_mtime.tv_nsec, &token);
3822
3823         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3824                                      inode->i_ctime.tv_sec, &token);
3825         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3826                                       inode->i_ctime.tv_nsec, &token);
3827
3828         btrfs_set_token_timespec_sec(leaf, &item->otime,
3829                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3830         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3831                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3832
3833         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3834                                      &token);
3835         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3836                                          &token);
3837         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3838         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3839         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3840         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3841         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3842 }
3843
3844 /*
3845  * copy everything in the in-memory inode into the btree.
3846  */
3847 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3848                                 struct btrfs_root *root, struct inode *inode)
3849 {
3850         struct btrfs_inode_item *inode_item;
3851         struct btrfs_path *path;
3852         struct extent_buffer *leaf;
3853         int ret;
3854
3855         path = btrfs_alloc_path();
3856         if (!path)
3857                 return -ENOMEM;
3858
3859         path->leave_spinning = 1;
3860         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3861                                  1);
3862         if (ret) {
3863                 if (ret > 0)
3864                         ret = -ENOENT;
3865                 goto failed;
3866         }
3867
3868         leaf = path->nodes[0];
3869         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3870                                     struct btrfs_inode_item);
3871
3872         fill_inode_item(trans, leaf, inode_item, inode);
3873         btrfs_mark_buffer_dirty(leaf);
3874         btrfs_set_inode_last_trans(trans, inode);
3875         ret = 0;
3876 failed:
3877         btrfs_free_path(path);
3878         return ret;
3879 }
3880
3881 /*
3882  * copy everything in the in-memory inode into the btree.
3883  */
3884 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3885                                 struct btrfs_root *root, struct inode *inode)
3886 {
3887         int ret;
3888
3889         /*
3890          * If the inode is a free space inode, we can deadlock during commit
3891          * if we put it into the delayed code.
3892          *
3893          * The data relocation inode should also be directly updated
3894          * without delay
3895          */
3896         if (!btrfs_is_free_space_inode(inode)
3897             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3898             && !root->fs_info->log_root_recovering) {
3899                 btrfs_update_root_times(trans, root);
3900
3901                 ret = btrfs_delayed_update_inode(trans, root, inode);
3902                 if (!ret)
3903                         btrfs_set_inode_last_trans(trans, inode);
3904                 return ret;
3905         }
3906
3907         return btrfs_update_inode_item(trans, root, inode);
3908 }
3909
3910 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3911                                          struct btrfs_root *root,
3912                                          struct inode *inode)
3913 {
3914         int ret;
3915
3916         ret = btrfs_update_inode(trans, root, inode);
3917         if (ret == -ENOSPC)
3918                 return btrfs_update_inode_item(trans, root, inode);
3919         return ret;
3920 }
3921
3922 /*
3923  * unlink helper that gets used here in inode.c and in the tree logging
3924  * recovery code.  It remove a link in a directory with a given name, and
3925  * also drops the back refs in the inode to the directory
3926  */
3927 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3928                                 struct btrfs_root *root,
3929                                 struct inode *dir, struct inode *inode,
3930                                 const char *name, int name_len)
3931 {
3932         struct btrfs_path *path;
3933         int ret = 0;
3934         struct extent_buffer *leaf;
3935         struct btrfs_dir_item *di;
3936         struct btrfs_key key;
3937         u64 index;
3938         u64 ino = btrfs_ino(inode);
3939         u64 dir_ino = btrfs_ino(dir);
3940
3941         path = btrfs_alloc_path();
3942         if (!path) {
3943                 ret = -ENOMEM;
3944                 goto out;
3945         }
3946
3947         path->leave_spinning = 1;
3948         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3949                                     name, name_len, -1);
3950         if (IS_ERR(di)) {
3951                 ret = PTR_ERR(di);
3952                 goto err;
3953         }
3954         if (!di) {
3955                 ret = -ENOENT;
3956                 goto err;
3957         }
3958         leaf = path->nodes[0];
3959         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3960         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3961         if (ret)
3962                 goto err;
3963         btrfs_release_path(path);
3964
3965         /*
3966          * If we don't have dir index, we have to get it by looking up
3967          * the inode ref, since we get the inode ref, remove it directly,
3968          * it is unnecessary to do delayed deletion.
3969          *
3970          * But if we have dir index, needn't search inode ref to get it.
3971          * Since the inode ref is close to the inode item, it is better
3972          * that we delay to delete it, and just do this deletion when
3973          * we update the inode item.
3974          */
3975         if (BTRFS_I(inode)->dir_index) {
3976                 ret = btrfs_delayed_delete_inode_ref(inode);
3977                 if (!ret) {
3978                         index = BTRFS_I(inode)->dir_index;
3979                         goto skip_backref;
3980                 }
3981         }
3982
3983         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3984                                   dir_ino, &index);
3985         if (ret) {
3986                 btrfs_info(root->fs_info,
3987                         "failed to delete reference to %.*s, inode %llu parent %llu",
3988                         name_len, name, ino, dir_ino);
3989                 btrfs_abort_transaction(trans, root, ret);
3990                 goto err;
3991         }
3992 skip_backref:
3993         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3994         if (ret) {
3995                 btrfs_abort_transaction(trans, root, ret);
3996                 goto err;
3997         }
3998
3999         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4000                                          inode, dir_ino);
4001         if (ret != 0 && ret != -ENOENT) {
4002                 btrfs_abort_transaction(trans, root, ret);
4003                 goto err;
4004         }
4005
4006         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4007                                            dir, index);
4008         if (ret == -ENOENT)
4009                 ret = 0;
4010         else if (ret)
4011                 btrfs_abort_transaction(trans, root, ret);
4012 err:
4013         btrfs_free_path(path);
4014         if (ret)
4015                 goto out;
4016
4017         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4018         inode_inc_iversion(inode);
4019         inode_inc_iversion(dir);
4020         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4021         ret = btrfs_update_inode(trans, root, dir);
4022 out:
4023         return ret;
4024 }
4025
4026 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4027                        struct btrfs_root *root,
4028                        struct inode *dir, struct inode *inode,
4029                        const char *name, int name_len)
4030 {
4031         int ret;
4032         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4033         if (!ret) {
4034                 drop_nlink(inode);
4035                 ret = btrfs_update_inode(trans, root, inode);
4036         }
4037         return ret;
4038 }
4039
4040 /*
4041  * helper to start transaction for unlink and rmdir.
4042  *
4043  * unlink and rmdir are special in btrfs, they do not always free space, so
4044  * if we cannot make our reservations the normal way try and see if there is
4045  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4046  * allow the unlink to occur.
4047  */
4048 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4049 {
4050         struct btrfs_root *root = BTRFS_I(dir)->root;
4051
4052         /*
4053          * 1 for the possible orphan item
4054          * 1 for the dir item
4055          * 1 for the dir index
4056          * 1 for the inode ref
4057          * 1 for the inode
4058          */
4059         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4060 }
4061
4062 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4063 {
4064         struct btrfs_root *root = BTRFS_I(dir)->root;
4065         struct btrfs_trans_handle *trans;
4066         struct inode *inode = d_inode(dentry);
4067         int ret;
4068
4069         trans = __unlink_start_trans(dir);
4070         if (IS_ERR(trans))
4071                 return PTR_ERR(trans);
4072
4073         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4074
4075         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4076                                  dentry->d_name.name, dentry->d_name.len);
4077         if (ret)
4078                 goto out;
4079
4080         if (inode->i_nlink == 0) {
4081                 ret = btrfs_orphan_add(trans, inode);
4082                 if (ret)
4083                         goto out;
4084         }
4085
4086 out:
4087         btrfs_end_transaction(trans, root);
4088         btrfs_btree_balance_dirty(root);
4089         return ret;
4090 }
4091
4092 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4093                         struct btrfs_root *root,
4094                         struct inode *dir, u64 objectid,
4095                         const char *name, int name_len)
4096 {
4097         struct btrfs_path *path;
4098         struct extent_buffer *leaf;
4099         struct btrfs_dir_item *di;
4100         struct btrfs_key key;
4101         u64 index;
4102         int ret;
4103         u64 dir_ino = btrfs_ino(dir);
4104
4105         path = btrfs_alloc_path();
4106         if (!path)
4107                 return -ENOMEM;
4108
4109         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4110                                    name, name_len, -1);
4111         if (IS_ERR_OR_NULL(di)) {
4112                 if (!di)
4113                         ret = -ENOENT;
4114                 else
4115                         ret = PTR_ERR(di);
4116                 goto out;
4117         }
4118
4119         leaf = path->nodes[0];
4120         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4121         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4122         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4123         if (ret) {
4124                 btrfs_abort_transaction(trans, root, ret);
4125                 goto out;
4126         }
4127         btrfs_release_path(path);
4128
4129         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4130                                  objectid, root->root_key.objectid,
4131                                  dir_ino, &index, name, name_len);
4132         if (ret < 0) {
4133                 if (ret != -ENOENT) {
4134                         btrfs_abort_transaction(trans, root, ret);
4135                         goto out;
4136                 }
4137                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4138                                                  name, name_len);
4139                 if (IS_ERR_OR_NULL(di)) {
4140                         if (!di)
4141                                 ret = -ENOENT;
4142                         else
4143                                 ret = PTR_ERR(di);
4144                         btrfs_abort_transaction(trans, root, ret);
4145                         goto out;
4146                 }
4147
4148                 leaf = path->nodes[0];
4149                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4150                 btrfs_release_path(path);
4151                 index = key.offset;
4152         }
4153         btrfs_release_path(path);
4154
4155         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4156         if (ret) {
4157                 btrfs_abort_transaction(trans, root, ret);
4158                 goto out;
4159         }
4160
4161         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4162         inode_inc_iversion(dir);
4163         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4164         ret = btrfs_update_inode_fallback(trans, root, dir);
4165         if (ret)
4166                 btrfs_abort_transaction(trans, root, ret);
4167 out:
4168         btrfs_free_path(path);
4169         return ret;
4170 }
4171
4172 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4173 {
4174         struct inode *inode = d_inode(dentry);
4175         int err = 0;
4176         struct btrfs_root *root = BTRFS_I(dir)->root;
4177         struct btrfs_trans_handle *trans;
4178
4179         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4180                 return -ENOTEMPTY;
4181         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4182                 return -EPERM;
4183
4184         trans = __unlink_start_trans(dir);
4185         if (IS_ERR(trans))
4186                 return PTR_ERR(trans);
4187
4188         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4189                 err = btrfs_unlink_subvol(trans, root, dir,
4190                                           BTRFS_I(inode)->location.objectid,
4191                                           dentry->d_name.name,
4192                                           dentry->d_name.len);
4193                 goto out;
4194         }
4195
4196         err = btrfs_orphan_add(trans, inode);
4197         if (err)
4198                 goto out;
4199
4200         /* now the directory is empty */
4201         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4202                                  dentry->d_name.name, dentry->d_name.len);
4203         if (!err)
4204                 btrfs_i_size_write(inode, 0);
4205 out:
4206         btrfs_end_transaction(trans, root);
4207         btrfs_btree_balance_dirty(root);
4208
4209         return err;
4210 }
4211
4212 static int truncate_space_check(struct btrfs_trans_handle *trans,
4213                                 struct btrfs_root *root,
4214                                 u64 bytes_deleted)
4215 {
4216         int ret;
4217
4218         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4219         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4220                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4221         if (!ret)
4222                 trans->bytes_reserved += bytes_deleted;
4223         return ret;
4224
4225 }
4226
4227 static int truncate_inline_extent(struct inode *inode,
4228                                   struct btrfs_path *path,
4229                                   struct btrfs_key *found_key,
4230                                   const u64 item_end,
4231                                   const u64 new_size)
4232 {
4233         struct extent_buffer *leaf = path->nodes[0];
4234         int slot = path->slots[0];
4235         struct btrfs_file_extent_item *fi;
4236         u32 size = (u32)(new_size - found_key->offset);
4237         struct btrfs_root *root = BTRFS_I(inode)->root;
4238
4239         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4240
4241         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4242                 loff_t offset = new_size;
4243                 loff_t page_end = ALIGN(offset, PAGE_CACHE_SIZE);
4244
4245                 /*
4246                  * Zero out the remaining of the last page of our inline extent,
4247                  * instead of directly truncating our inline extent here - that
4248                  * would be much more complex (decompressing all the data, then
4249                  * compressing the truncated data, which might be bigger than
4250                  * the size of the inline extent, resize the extent, etc).
4251                  * We release the path because to get the page we might need to
4252                  * read the extent item from disk (data not in the page cache).
4253                  */
4254                 btrfs_release_path(path);
4255                 return btrfs_truncate_page(inode, offset, page_end - offset, 0);
4256         }
4257
4258         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4259         size = btrfs_file_extent_calc_inline_size(size);
4260         btrfs_truncate_item(root, path, size, 1);
4261
4262         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4263                 inode_sub_bytes(inode, item_end + 1 - new_size);
4264
4265         return 0;
4266 }
4267
4268 /*
4269  * this can truncate away extent items, csum items and directory items.
4270  * It starts at a high offset and removes keys until it can't find
4271  * any higher than new_size
4272  *
4273  * csum items that cross the new i_size are truncated to the new size
4274  * as well.
4275  *
4276  * min_type is the minimum key type to truncate down to.  If set to 0, this
4277  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4278  */
4279 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4280                                struct btrfs_root *root,
4281                                struct inode *inode,
4282                                u64 new_size, u32 min_type)
4283 {
4284         struct btrfs_path *path;
4285         struct extent_buffer *leaf;
4286         struct btrfs_file_extent_item *fi;
4287         struct btrfs_key key;
4288         struct btrfs_key found_key;
4289         u64 extent_start = 0;
4290         u64 extent_num_bytes = 0;
4291         u64 extent_offset = 0;
4292         u64 item_end = 0;
4293         u64 last_size = new_size;
4294         u32 found_type = (u8)-1;
4295         int found_extent;
4296         int del_item;
4297         int pending_del_nr = 0;
4298         int pending_del_slot = 0;
4299         int extent_type = -1;
4300         int ret;
4301         int err = 0;
4302         u64 ino = btrfs_ino(inode);
4303         u64 bytes_deleted = 0;
4304         bool be_nice = 0;
4305         bool should_throttle = 0;
4306         bool should_end = 0;
4307
4308         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4309
4310         /*
4311          * for non-free space inodes and ref cows, we want to back off from
4312          * time to time
4313          */
4314         if (!btrfs_is_free_space_inode(inode) &&
4315             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4316                 be_nice = 1;
4317
4318         path = btrfs_alloc_path();
4319         if (!path)
4320                 return -ENOMEM;
4321         path->reada = -1;
4322
4323         /*
4324          * We want to drop from the next block forward in case this new size is
4325          * not block aligned since we will be keeping the last block of the
4326          * extent just the way it is.
4327          */
4328         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4329             root == root->fs_info->tree_root)
4330                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4331                                         root->sectorsize), (u64)-1, 0);
4332
4333         /*
4334          * This function is also used to drop the items in the log tree before
4335          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4336          * it is used to drop the loged items. So we shouldn't kill the delayed
4337          * items.
4338          */
4339         if (min_type == 0 && root == BTRFS_I(inode)->root)
4340                 btrfs_kill_delayed_inode_items(inode);
4341
4342         key.objectid = ino;
4343         key.offset = (u64)-1;
4344         key.type = (u8)-1;
4345
4346 search_again:
4347         /*
4348          * with a 16K leaf size and 128MB extents, you can actually queue
4349          * up a huge file in a single leaf.  Most of the time that
4350          * bytes_deleted is > 0, it will be huge by the time we get here
4351          */
4352         if (be_nice && bytes_deleted > 32 * 1024 * 1024) {
4353                 if (btrfs_should_end_transaction(trans, root)) {
4354                         err = -EAGAIN;
4355                         goto error;
4356                 }
4357         }
4358
4359
4360         path->leave_spinning = 1;
4361         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4362         if (ret < 0) {
4363                 err = ret;
4364                 goto out;
4365         }
4366
4367         if (ret > 0) {
4368                 /* there are no items in the tree for us to truncate, we're
4369                  * done
4370                  */
4371                 if (path->slots[0] == 0)
4372                         goto out;
4373                 path->slots[0]--;
4374         }
4375
4376         while (1) {
4377                 fi = NULL;
4378                 leaf = path->nodes[0];
4379                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4380                 found_type = found_key.type;
4381
4382                 if (found_key.objectid != ino)
4383                         break;
4384
4385                 if (found_type < min_type)
4386                         break;
4387
4388                 item_end = found_key.offset;
4389                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4390                         fi = btrfs_item_ptr(leaf, path->slots[0],
4391                                             struct btrfs_file_extent_item);
4392                         extent_type = btrfs_file_extent_type(leaf, fi);
4393                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4394                                 item_end +=
4395                                     btrfs_file_extent_num_bytes(leaf, fi);
4396                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4397                                 item_end += btrfs_file_extent_inline_len(leaf,
4398                                                          path->slots[0], fi);
4399                         }
4400                         item_end--;
4401                 }
4402                 if (found_type > min_type) {
4403                         del_item = 1;
4404                 } else {
4405                         if (item_end < new_size)
4406                                 break;
4407                         if (found_key.offset >= new_size)
4408                                 del_item = 1;
4409                         else
4410                                 del_item = 0;
4411                 }
4412                 found_extent = 0;
4413                 /* FIXME, shrink the extent if the ref count is only 1 */
4414                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4415                         goto delete;
4416
4417                 if (del_item)
4418                         last_size = found_key.offset;
4419                 else
4420                         last_size = new_size;
4421
4422                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4423                         u64 num_dec;
4424                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4425                         if (!del_item) {
4426                                 u64 orig_num_bytes =
4427                                         btrfs_file_extent_num_bytes(leaf, fi);
4428                                 extent_num_bytes = ALIGN(new_size -
4429                                                 found_key.offset,
4430                                                 root->sectorsize);
4431                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4432                                                          extent_num_bytes);
4433                                 num_dec = (orig_num_bytes -
4434                                            extent_num_bytes);
4435                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4436                                              &root->state) &&
4437                                     extent_start != 0)
4438                                         inode_sub_bytes(inode, num_dec);
4439                                 btrfs_mark_buffer_dirty(leaf);
4440                         } else {
4441                                 extent_num_bytes =
4442                                         btrfs_file_extent_disk_num_bytes(leaf,
4443                                                                          fi);
4444                                 extent_offset = found_key.offset -
4445                                         btrfs_file_extent_offset(leaf, fi);
4446
4447                                 /* FIXME blocksize != 4096 */
4448                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4449                                 if (extent_start != 0) {
4450                                         found_extent = 1;
4451                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4452                                                      &root->state))
4453                                                 inode_sub_bytes(inode, num_dec);
4454                                 }
4455                         }
4456                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4457                         /*
4458                          * we can't truncate inline items that have had
4459                          * special encodings
4460                          */
4461                         if (!del_item &&
4462                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4463                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4464
4465                                 /*
4466                                  * Need to release path in order to truncate a
4467                                  * compressed extent. So delete any accumulated
4468                                  * extent items so far.
4469                                  */
4470                                 if (btrfs_file_extent_compression(leaf, fi) !=
4471                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4472                                         err = btrfs_del_items(trans, root, path,
4473                                                               pending_del_slot,
4474                                                               pending_del_nr);
4475                                         if (err) {
4476                                                 btrfs_abort_transaction(trans,
4477                                                                         root,
4478                                                                         err);
4479                                                 goto error;
4480                                         }
4481                                         pending_del_nr = 0;
4482                                 }
4483
4484                                 err = truncate_inline_extent(inode, path,
4485                                                              &found_key,
4486                                                              item_end,
4487                                                              new_size);
4488                                 if (err) {
4489                                         btrfs_abort_transaction(trans,
4490                                                                 root, err);
4491                                         goto error;
4492                                 }
4493                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4494                                             &root->state)) {
4495                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4496                         }
4497                 }
4498 delete:
4499                 if (del_item) {
4500                         if (!pending_del_nr) {
4501                                 /* no pending yet, add ourselves */
4502                                 pending_del_slot = path->slots[0];
4503                                 pending_del_nr = 1;
4504                         } else if (pending_del_nr &&
4505                                    path->slots[0] + 1 == pending_del_slot) {
4506                                 /* hop on the pending chunk */
4507                                 pending_del_nr++;
4508                                 pending_del_slot = path->slots[0];
4509                         } else {
4510                                 BUG();
4511                         }
4512                 } else {
4513                         break;
4514                 }
4515                 should_throttle = 0;
4516
4517                 if (found_extent &&
4518                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4519                      root == root->fs_info->tree_root)) {
4520                         btrfs_set_path_blocking(path);
4521                         bytes_deleted += extent_num_bytes;
4522                         ret = btrfs_free_extent(trans, root, extent_start,
4523                                                 extent_num_bytes, 0,
4524                                                 btrfs_header_owner(leaf),
4525                                                 ino, extent_offset);
4526                         BUG_ON(ret);
4527                         if (btrfs_should_throttle_delayed_refs(trans, root))
4528                                 btrfs_async_run_delayed_refs(root,
4529                                         trans->delayed_ref_updates * 2, 0);
4530                         if (be_nice) {
4531                                 if (truncate_space_check(trans, root,
4532                                                          extent_num_bytes)) {
4533                                         should_end = 1;
4534                                 }
4535                                 if (btrfs_should_throttle_delayed_refs(trans,
4536                                                                        root)) {
4537                                         should_throttle = 1;
4538                                 }
4539                         }
4540                 }
4541
4542                 if (found_type == BTRFS_INODE_ITEM_KEY)
4543                         break;
4544
4545                 if (path->slots[0] == 0 ||
4546                     path->slots[0] != pending_del_slot ||
4547                     should_throttle || should_end) {
4548                         if (pending_del_nr) {
4549                                 ret = btrfs_del_items(trans, root, path,
4550                                                 pending_del_slot,
4551                                                 pending_del_nr);
4552                                 if (ret) {
4553                                         btrfs_abort_transaction(trans,
4554                                                                 root, ret);
4555                                         goto error;
4556                                 }
4557                                 pending_del_nr = 0;
4558                         }
4559                         btrfs_release_path(path);
4560                         if (should_throttle) {
4561                                 unsigned long updates = trans->delayed_ref_updates;
4562                                 if (updates) {
4563                                         trans->delayed_ref_updates = 0;
4564                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4565                                         if (ret && !err)
4566                                                 err = ret;
4567                                 }
4568                         }
4569                         /*
4570                          * if we failed to refill our space rsv, bail out
4571                          * and let the transaction restart
4572                          */
4573                         if (should_end) {
4574                                 err = -EAGAIN;
4575                                 goto error;
4576                         }
4577                         goto search_again;
4578                 } else {
4579                         path->slots[0]--;
4580                 }
4581         }
4582 out:
4583         if (pending_del_nr) {
4584                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4585                                       pending_del_nr);
4586                 if (ret)
4587                         btrfs_abort_transaction(trans, root, ret);
4588         }
4589 error:
4590         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4591                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4592
4593         btrfs_free_path(path);
4594
4595         if (be_nice && bytes_deleted > 32 * 1024 * 1024) {
4596                 unsigned long updates = trans->delayed_ref_updates;
4597                 if (updates) {
4598                         trans->delayed_ref_updates = 0;
4599                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4600                         if (ret && !err)
4601                                 err = ret;
4602                 }
4603         }
4604         return err;
4605 }
4606
4607 /*
4608  * btrfs_truncate_page - read, zero a chunk and write a page
4609  * @inode - inode that we're zeroing
4610  * @from - the offset to start zeroing
4611  * @len - the length to zero, 0 to zero the entire range respective to the
4612  *      offset
4613  * @front - zero up to the offset instead of from the offset on
4614  *
4615  * This will find the page for the "from" offset and cow the page and zero the
4616  * part we want to zero.  This is used with truncate and hole punching.
4617  */
4618 int btrfs_truncate_page(struct inode *inode, loff_t from, loff_t len,
4619                         int front)
4620 {
4621         struct address_space *mapping = inode->i_mapping;
4622         struct btrfs_root *root = BTRFS_I(inode)->root;
4623         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4624         struct btrfs_ordered_extent *ordered;
4625         struct extent_state *cached_state = NULL;
4626         char *kaddr;
4627         u32 blocksize = root->sectorsize;
4628         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4629         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4630         struct page *page;
4631         gfp_t mask = btrfs_alloc_write_mask(mapping);
4632         int ret = 0;
4633         u64 page_start;
4634         u64 page_end;
4635
4636         if ((offset & (blocksize - 1)) == 0 &&
4637             (!len || ((len & (blocksize - 1)) == 0)))
4638                 goto out;
4639         ret = btrfs_delalloc_reserve_space(inode,
4640                         round_down(from, PAGE_CACHE_SIZE), PAGE_CACHE_SIZE);
4641         if (ret)
4642                 goto out;
4643
4644 again:
4645         page = find_or_create_page(mapping, index, mask);
4646         if (!page) {
4647                 btrfs_delalloc_release_space(inode,
4648                                 round_down(from, PAGE_CACHE_SIZE),
4649                                 PAGE_CACHE_SIZE);
4650                 ret = -ENOMEM;
4651                 goto out;
4652         }
4653
4654         page_start = page_offset(page);
4655         page_end = page_start + PAGE_CACHE_SIZE - 1;
4656
4657         if (!PageUptodate(page)) {
4658                 ret = btrfs_readpage(NULL, page);
4659                 lock_page(page);
4660                 if (page->mapping != mapping) {
4661                         unlock_page(page);
4662                         page_cache_release(page);
4663                         goto again;
4664                 }
4665                 if (!PageUptodate(page)) {
4666                         ret = -EIO;
4667                         goto out_unlock;
4668                 }
4669         }
4670         wait_on_page_writeback(page);
4671
4672         lock_extent_bits(io_tree, page_start, page_end, 0, &cached_state);
4673         set_page_extent_mapped(page);
4674
4675         ordered = btrfs_lookup_ordered_extent(inode, page_start);
4676         if (ordered) {
4677                 unlock_extent_cached(io_tree, page_start, page_end,
4678                                      &cached_state, GFP_NOFS);
4679                 unlock_page(page);
4680                 page_cache_release(page);
4681                 btrfs_start_ordered_extent(inode, ordered, 1);
4682                 btrfs_put_ordered_extent(ordered);
4683                 goto again;
4684         }
4685
4686         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
4687                           EXTENT_DIRTY | EXTENT_DELALLOC |
4688                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4689                           0, 0, &cached_state, GFP_NOFS);
4690
4691         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
4692                                         &cached_state);
4693         if (ret) {
4694                 unlock_extent_cached(io_tree, page_start, page_end,
4695                                      &cached_state, GFP_NOFS);
4696                 goto out_unlock;
4697         }
4698
4699         if (offset != PAGE_CACHE_SIZE) {
4700                 if (!len)
4701                         len = PAGE_CACHE_SIZE - offset;
4702                 kaddr = kmap(page);
4703                 if (front)
4704                         memset(kaddr, 0, offset);
4705                 else
4706                         memset(kaddr + offset, 0, len);
4707                 flush_dcache_page(page);
4708                 kunmap(page);
4709         }
4710         ClearPageChecked(page);
4711         set_page_dirty(page);
4712         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
4713                              GFP_NOFS);
4714
4715 out_unlock:
4716         if (ret)
4717                 btrfs_delalloc_release_space(inode, page_start,
4718                                              PAGE_CACHE_SIZE);
4719         unlock_page(page);
4720         page_cache_release(page);
4721 out:
4722         return ret;
4723 }
4724
4725 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4726                              u64 offset, u64 len)
4727 {
4728         struct btrfs_trans_handle *trans;
4729         int ret;
4730
4731         /*
4732          * Still need to make sure the inode looks like it's been updated so
4733          * that any holes get logged if we fsync.
4734          */
4735         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4736                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4737                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4738                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4739                 return 0;
4740         }
4741
4742         /*
4743          * 1 - for the one we're dropping
4744          * 1 - for the one we're adding
4745          * 1 - for updating the inode.
4746          */
4747         trans = btrfs_start_transaction(root, 3);
4748         if (IS_ERR(trans))
4749                 return PTR_ERR(trans);
4750
4751         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4752         if (ret) {
4753                 btrfs_abort_transaction(trans, root, ret);
4754                 btrfs_end_transaction(trans, root);
4755                 return ret;
4756         }
4757
4758         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4759                                        0, 0, len, 0, len, 0, 0, 0);
4760         if (ret)
4761                 btrfs_abort_transaction(trans, root, ret);
4762         else
4763                 btrfs_update_inode(trans, root, inode);
4764         btrfs_end_transaction(trans, root);
4765         return ret;
4766 }
4767
4768 /*
4769  * This function puts in dummy file extents for the area we're creating a hole
4770  * for.  So if we are truncating this file to a larger size we need to insert
4771  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4772  * the range between oldsize and size
4773  */
4774 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4775 {
4776         struct btrfs_root *root = BTRFS_I(inode)->root;
4777         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4778         struct extent_map *em = NULL;
4779         struct extent_state *cached_state = NULL;
4780         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4781         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4782         u64 block_end = ALIGN(size, root->sectorsize);
4783         u64 last_byte;
4784         u64 cur_offset;
4785         u64 hole_size;
4786         int err = 0;
4787
4788         /*
4789          * If our size started in the middle of a page we need to zero out the
4790          * rest of the page before we expand the i_size, otherwise we could
4791          * expose stale data.
4792          */
4793         err = btrfs_truncate_page(inode, oldsize, 0, 0);
4794         if (err)
4795                 return err;
4796
4797         if (size <= hole_start)
4798                 return 0;
4799
4800         while (1) {
4801                 struct btrfs_ordered_extent *ordered;
4802
4803                 lock_extent_bits(io_tree, hole_start, block_end - 1, 0,
4804                                  &cached_state);
4805                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4806                                                      block_end - hole_start);
4807                 if (!ordered)
4808                         break;
4809                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4810                                      &cached_state, GFP_NOFS);
4811                 btrfs_start_ordered_extent(inode, ordered, 1);
4812                 btrfs_put_ordered_extent(ordered);
4813         }
4814
4815         cur_offset = hole_start;
4816         while (1) {
4817                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4818                                 block_end - cur_offset, 0);
4819                 if (IS_ERR(em)) {
4820                         err = PTR_ERR(em);
4821                         em = NULL;
4822                         break;
4823                 }
4824                 last_byte = min(extent_map_end(em), block_end);
4825                 last_byte = ALIGN(last_byte , root->sectorsize);
4826                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4827                         struct extent_map *hole_em;
4828                         hole_size = last_byte - cur_offset;
4829
4830                         err = maybe_insert_hole(root, inode, cur_offset,
4831                                                 hole_size);
4832                         if (err)
4833                                 break;
4834                         btrfs_drop_extent_cache(inode, cur_offset,
4835                                                 cur_offset + hole_size - 1, 0);
4836                         hole_em = alloc_extent_map();
4837                         if (!hole_em) {
4838                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4839                                         &BTRFS_I(inode)->runtime_flags);
4840                                 goto next;
4841                         }
4842                         hole_em->start = cur_offset;
4843                         hole_em->len = hole_size;
4844                         hole_em->orig_start = cur_offset;
4845
4846                         hole_em->block_start = EXTENT_MAP_HOLE;
4847                         hole_em->block_len = 0;
4848                         hole_em->orig_block_len = 0;
4849                         hole_em->ram_bytes = hole_size;
4850                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4851                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4852                         hole_em->generation = root->fs_info->generation;
4853
4854                         while (1) {
4855                                 write_lock(&em_tree->lock);
4856                                 err = add_extent_mapping(em_tree, hole_em, 1);
4857                                 write_unlock(&em_tree->lock);
4858                                 if (err != -EEXIST)
4859                                         break;
4860                                 btrfs_drop_extent_cache(inode, cur_offset,
4861                                                         cur_offset +
4862                                                         hole_size - 1, 0);
4863                         }
4864                         free_extent_map(hole_em);
4865                 }
4866 next:
4867                 free_extent_map(em);
4868                 em = NULL;
4869                 cur_offset = last_byte;
4870                 if (cur_offset >= block_end)
4871                         break;
4872         }
4873         free_extent_map(em);
4874         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4875                              GFP_NOFS);
4876         return err;
4877 }
4878
4879 static int wait_snapshoting_atomic_t(atomic_t *a)
4880 {
4881         schedule();
4882         return 0;
4883 }
4884
4885 static void wait_for_snapshot_creation(struct btrfs_root *root)
4886 {
4887         while (true) {
4888                 int ret;
4889
4890                 ret = btrfs_start_write_no_snapshoting(root);
4891                 if (ret)
4892                         break;
4893                 wait_on_atomic_t(&root->will_be_snapshoted,
4894                                  wait_snapshoting_atomic_t,
4895                                  TASK_UNINTERRUPTIBLE);
4896         }
4897 }
4898
4899 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4900 {
4901         struct btrfs_root *root = BTRFS_I(inode)->root;
4902         struct btrfs_trans_handle *trans;
4903         loff_t oldsize = i_size_read(inode);
4904         loff_t newsize = attr->ia_size;
4905         int mask = attr->ia_valid;
4906         int ret;
4907
4908         /*
4909          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4910          * special case where we need to update the times despite not having
4911          * these flags set.  For all other operations the VFS set these flags
4912          * explicitly if it wants a timestamp update.
4913          */
4914         if (newsize != oldsize) {
4915                 inode_inc_iversion(inode);
4916                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4917                         inode->i_ctime = inode->i_mtime =
4918                                 current_fs_time(inode->i_sb);
4919         }
4920
4921         if (newsize > oldsize) {
4922                 truncate_pagecache(inode, newsize);
4923                 /*
4924                  * Don't do an expanding truncate while snapshoting is ongoing.
4925                  * This is to ensure the snapshot captures a fully consistent
4926                  * state of this file - if the snapshot captures this expanding
4927                  * truncation, it must capture all writes that happened before
4928                  * this truncation.
4929                  */
4930                 wait_for_snapshot_creation(root);
4931                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4932                 if (ret) {
4933                         btrfs_end_write_no_snapshoting(root);
4934                         return ret;
4935                 }
4936
4937                 trans = btrfs_start_transaction(root, 1);
4938                 if (IS_ERR(trans)) {
4939                         btrfs_end_write_no_snapshoting(root);
4940                         return PTR_ERR(trans);
4941                 }
4942
4943                 i_size_write(inode, newsize);
4944                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4945                 ret = btrfs_update_inode(trans, root, inode);
4946                 btrfs_end_write_no_snapshoting(root);
4947                 btrfs_end_transaction(trans, root);
4948         } else {
4949
4950                 /*
4951                  * We're truncating a file that used to have good data down to
4952                  * zero. Make sure it gets into the ordered flush list so that
4953                  * any new writes get down to disk quickly.
4954                  */
4955                 if (newsize == 0)
4956                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4957                                 &BTRFS_I(inode)->runtime_flags);
4958
4959                 /*
4960                  * 1 for the orphan item we're going to add
4961                  * 1 for the orphan item deletion.
4962                  */
4963                 trans = btrfs_start_transaction(root, 2);
4964                 if (IS_ERR(trans))
4965                         return PTR_ERR(trans);
4966
4967                 /*
4968                  * We need to do this in case we fail at _any_ point during the
4969                  * actual truncate.  Once we do the truncate_setsize we could
4970                  * invalidate pages which forces any outstanding ordered io to
4971                  * be instantly completed which will give us extents that need
4972                  * to be truncated.  If we fail to get an orphan inode down we
4973                  * could have left over extents that were never meant to live,
4974                  * so we need to garuntee from this point on that everything
4975                  * will be consistent.
4976                  */
4977                 ret = btrfs_orphan_add(trans, inode);
4978                 btrfs_end_transaction(trans, root);
4979                 if (ret)
4980                         return ret;
4981
4982                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4983                 truncate_setsize(inode, newsize);
4984
4985                 /* Disable nonlocked read DIO to avoid the end less truncate */
4986                 btrfs_inode_block_unlocked_dio(inode);
4987                 inode_dio_wait(inode);
4988                 btrfs_inode_resume_unlocked_dio(inode);
4989
4990                 ret = btrfs_truncate(inode);
4991                 if (ret && inode->i_nlink) {
4992                         int err;
4993
4994                         /*
4995                          * failed to truncate, disk_i_size is only adjusted down
4996                          * as we remove extents, so it should represent the true
4997                          * size of the inode, so reset the in memory size and
4998                          * delete our orphan entry.
4999                          */
5000                         trans = btrfs_join_transaction(root);
5001                         if (IS_ERR(trans)) {
5002                                 btrfs_orphan_del(NULL, inode);
5003                                 return ret;
5004                         }
5005                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5006                         err = btrfs_orphan_del(trans, inode);
5007                         if (err)
5008                                 btrfs_abort_transaction(trans, root, err);
5009                         btrfs_end_transaction(trans, root);
5010                 }
5011         }
5012
5013         return ret;
5014 }
5015
5016 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5017 {
5018         struct inode *inode = d_inode(dentry);
5019         struct btrfs_root *root = BTRFS_I(inode)->root;
5020         int err;
5021
5022         if (btrfs_root_readonly(root))
5023                 return -EROFS;
5024
5025         err = inode_change_ok(inode, attr);
5026         if (err)
5027                 return err;
5028
5029         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5030                 err = btrfs_setsize(inode, attr);
5031                 if (err)
5032                         return err;
5033         }
5034
5035         if (attr->ia_valid) {
5036                 setattr_copy(inode, attr);
5037                 inode_inc_iversion(inode);
5038                 err = btrfs_dirty_inode(inode);
5039
5040                 if (!err && attr->ia_valid & ATTR_MODE)
5041                         err = posix_acl_chmod(inode, inode->i_mode);
5042         }
5043
5044         return err;
5045 }
5046
5047 /*
5048  * While truncating the inode pages during eviction, we get the VFS calling
5049  * btrfs_invalidatepage() against each page of the inode. This is slow because
5050  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5051  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5052  * extent_state structures over and over, wasting lots of time.
5053  *
5054  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5055  * those expensive operations on a per page basis and do only the ordered io
5056  * finishing, while we release here the extent_map and extent_state structures,
5057  * without the excessive merging and splitting.
5058  */
5059 static void evict_inode_truncate_pages(struct inode *inode)
5060 {
5061         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5062         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5063         struct rb_node *node;
5064
5065         ASSERT(inode->i_state & I_FREEING);
5066         truncate_inode_pages_final(&inode->i_data);
5067
5068         write_lock(&map_tree->lock);
5069         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5070                 struct extent_map *em;
5071
5072                 node = rb_first(&map_tree->map);
5073                 em = rb_entry(node, struct extent_map, rb_node);
5074                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5075                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5076                 remove_extent_mapping(map_tree, em);
5077                 free_extent_map(em);
5078                 if (need_resched()) {
5079                         write_unlock(&map_tree->lock);
5080                         cond_resched();
5081                         write_lock(&map_tree->lock);
5082                 }
5083         }
5084         write_unlock(&map_tree->lock);
5085
5086         /*
5087          * Keep looping until we have no more ranges in the io tree.
5088          * We can have ongoing bios started by readpages (called from readahead)
5089          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5090          * still in progress (unlocked the pages in the bio but did not yet
5091          * unlocked the ranges in the io tree). Therefore this means some
5092          * ranges can still be locked and eviction started because before
5093          * submitting those bios, which are executed by a separate task (work
5094          * queue kthread), inode references (inode->i_count) were not taken
5095          * (which would be dropped in the end io callback of each bio).
5096          * Therefore here we effectively end up waiting for those bios and
5097          * anyone else holding locked ranges without having bumped the inode's
5098          * reference count - if we don't do it, when they access the inode's
5099          * io_tree to unlock a range it may be too late, leading to an
5100          * use-after-free issue.
5101          */
5102         spin_lock(&io_tree->lock);
5103         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5104                 struct extent_state *state;
5105                 struct extent_state *cached_state = NULL;
5106                 u64 start;
5107                 u64 end;
5108
5109                 node = rb_first(&io_tree->state);
5110                 state = rb_entry(node, struct extent_state, rb_node);
5111                 start = state->start;
5112                 end = state->end;
5113                 spin_unlock(&io_tree->lock);
5114
5115                 lock_extent_bits(io_tree, start, end, 0, &cached_state);
5116
5117                 /*
5118                  * If still has DELALLOC flag, the extent didn't reach disk,
5119                  * and its reserved space won't be freed by delayed_ref.
5120                  * So we need to free its reserved space here.
5121                  * (Refer to comment in btrfs_invalidatepage, case 2)
5122                  *
5123                  * Note, end is the bytenr of last byte, so we need + 1 here.
5124                  */
5125                 if (state->state & EXTENT_DELALLOC)
5126                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5127
5128                 clear_extent_bit(io_tree, start, end,
5129                                  EXTENT_LOCKED | EXTENT_DIRTY |
5130                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5131                                  EXTENT_DEFRAG, 1, 1,
5132                                  &cached_state, GFP_NOFS);
5133
5134                 cond_resched();
5135                 spin_lock(&io_tree->lock);
5136         }
5137         spin_unlock(&io_tree->lock);
5138 }
5139
5140 void btrfs_evict_inode(struct inode *inode)
5141 {
5142         struct btrfs_trans_handle *trans;
5143         struct btrfs_root *root = BTRFS_I(inode)->root;
5144         struct btrfs_block_rsv *rsv, *global_rsv;
5145         int steal_from_global = 0;
5146         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5147         int ret;
5148
5149         trace_btrfs_inode_evict(inode);
5150
5151         evict_inode_truncate_pages(inode);
5152
5153         if (inode->i_nlink &&
5154             ((btrfs_root_refs(&root->root_item) != 0 &&
5155               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5156              btrfs_is_free_space_inode(inode)))
5157                 goto no_delete;
5158
5159         if (is_bad_inode(inode)) {
5160                 btrfs_orphan_del(NULL, inode);
5161                 goto no_delete;
5162         }
5163         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5164         if (!special_file(inode->i_mode))
5165                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5166
5167         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5168
5169         if (root->fs_info->log_root_recovering) {
5170                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5171                                  &BTRFS_I(inode)->runtime_flags));
5172                 goto no_delete;
5173         }
5174
5175         if (inode->i_nlink > 0) {
5176                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5177                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5178                 goto no_delete;
5179         }
5180
5181         ret = btrfs_commit_inode_delayed_inode(inode);
5182         if (ret) {
5183                 btrfs_orphan_del(NULL, inode);
5184                 goto no_delete;
5185         }
5186
5187         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5188         if (!rsv) {
5189                 btrfs_orphan_del(NULL, inode);
5190                 goto no_delete;
5191         }
5192         rsv->size = min_size;
5193         rsv->failfast = 1;
5194         global_rsv = &root->fs_info->global_block_rsv;
5195
5196         btrfs_i_size_write(inode, 0);
5197
5198         /*
5199          * This is a bit simpler than btrfs_truncate since we've already
5200          * reserved our space for our orphan item in the unlink, so we just
5201          * need to reserve some slack space in case we add bytes and update
5202          * inode item when doing the truncate.
5203          */
5204         while (1) {
5205                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5206                                              BTRFS_RESERVE_FLUSH_LIMIT);
5207
5208                 /*
5209                  * Try and steal from the global reserve since we will
5210                  * likely not use this space anyway, we want to try as
5211                  * hard as possible to get this to work.
5212                  */
5213                 if (ret)
5214                         steal_from_global++;
5215                 else
5216                         steal_from_global = 0;
5217                 ret = 0;
5218
5219                 /*
5220                  * steal_from_global == 0: we reserved stuff, hooray!
5221                  * steal_from_global == 1: we didn't reserve stuff, boo!
5222                  * steal_from_global == 2: we've committed, still not a lot of
5223                  * room but maybe we'll have room in the global reserve this
5224                  * time.
5225                  * steal_from_global == 3: abandon all hope!
5226                  */
5227                 if (steal_from_global > 2) {
5228                         btrfs_warn(root->fs_info,
5229                                 "Could not get space for a delete, will truncate on mount %d",
5230                                 ret);
5231                         btrfs_orphan_del(NULL, inode);
5232                         btrfs_free_block_rsv(root, rsv);
5233                         goto no_delete;
5234                 }
5235
5236                 trans = btrfs_join_transaction(root);
5237                 if (IS_ERR(trans)) {
5238                         btrfs_orphan_del(NULL, inode);
5239                         btrfs_free_block_rsv(root, rsv);
5240                         goto no_delete;
5241                 }
5242
5243                 /*
5244                  * We can't just steal from the global reserve, we need tomake
5245                  * sure there is room to do it, if not we need to commit and try
5246                  * again.
5247                  */
5248                 if (steal_from_global) {
5249                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5250                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5251                                                               min_size);
5252                         else
5253                                 ret = -ENOSPC;
5254                 }
5255
5256                 /*
5257                  * Couldn't steal from the global reserve, we have too much
5258                  * pending stuff built up, commit the transaction and try it
5259                  * again.
5260                  */
5261                 if (ret) {
5262                         ret = btrfs_commit_transaction(trans, root);
5263                         if (ret) {
5264                                 btrfs_orphan_del(NULL, inode);
5265                                 btrfs_free_block_rsv(root, rsv);
5266                                 goto no_delete;
5267                         }
5268                         continue;
5269                 } else {
5270                         steal_from_global = 0;
5271                 }
5272
5273                 trans->block_rsv = rsv;
5274
5275                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5276                 if (ret != -ENOSPC && ret != -EAGAIN)
5277                         break;
5278
5279                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5280                 btrfs_end_transaction(trans, root);
5281                 trans = NULL;
5282                 btrfs_btree_balance_dirty(root);
5283         }
5284
5285         btrfs_free_block_rsv(root, rsv);
5286
5287         /*
5288          * Errors here aren't a big deal, it just means we leave orphan items
5289          * in the tree.  They will be cleaned up on the next mount.
5290          */
5291         if (ret == 0) {
5292                 trans->block_rsv = root->orphan_block_rsv;
5293                 btrfs_orphan_del(trans, inode);
5294         } else {
5295                 btrfs_orphan_del(NULL, inode);
5296         }
5297
5298         trans->block_rsv = &root->fs_info->trans_block_rsv;
5299         if (!(root == root->fs_info->tree_root ||
5300               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5301                 btrfs_return_ino(root, btrfs_ino(inode));
5302
5303         btrfs_end_transaction(trans, root);
5304         btrfs_btree_balance_dirty(root);
5305 no_delete:
5306         btrfs_remove_delayed_node(inode);
5307         clear_inode(inode);
5308         return;
5309 }
5310
5311 /*
5312  * this returns the key found in the dir entry in the location pointer.
5313  * If no dir entries were found, location->objectid is 0.
5314  */
5315 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5316                                struct btrfs_key *location)
5317 {
5318         const char *name = dentry->d_name.name;
5319         int namelen = dentry->d_name.len;
5320         struct btrfs_dir_item *di;
5321         struct btrfs_path *path;
5322         struct btrfs_root *root = BTRFS_I(dir)->root;
5323         int ret = 0;
5324
5325         path = btrfs_alloc_path();
5326         if (!path)
5327                 return -ENOMEM;
5328
5329         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5330                                     namelen, 0);
5331         if (IS_ERR(di))
5332                 ret = PTR_ERR(di);
5333
5334         if (IS_ERR_OR_NULL(di))
5335                 goto out_err;
5336
5337         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5338 out:
5339         btrfs_free_path(path);
5340         return ret;
5341 out_err:
5342         location->objectid = 0;
5343         goto out;
5344 }
5345
5346 /*
5347  * when we hit a tree root in a directory, the btrfs part of the inode
5348  * needs to be changed to reflect the root directory of the tree root.  This
5349  * is kind of like crossing a mount point.
5350  */
5351 static int fixup_tree_root_location(struct btrfs_root *root,
5352                                     struct inode *dir,
5353                                     struct dentry *dentry,
5354                                     struct btrfs_key *location,
5355                                     struct btrfs_root **sub_root)
5356 {
5357         struct btrfs_path *path;
5358         struct btrfs_root *new_root;
5359         struct btrfs_root_ref *ref;
5360         struct extent_buffer *leaf;
5361         struct btrfs_key key;
5362         int ret;
5363         int err = 0;
5364
5365         path = btrfs_alloc_path();
5366         if (!path) {
5367                 err = -ENOMEM;
5368                 goto out;
5369         }
5370
5371         err = -ENOENT;
5372         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5373         key.type = BTRFS_ROOT_REF_KEY;
5374         key.offset = location->objectid;
5375
5376         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5377                                 0, 0);
5378         if (ret) {
5379                 if (ret < 0)
5380                         err = ret;
5381                 goto out;
5382         }
5383
5384         leaf = path->nodes[0];
5385         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5386         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5387             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5388                 goto out;
5389
5390         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5391                                    (unsigned long)(ref + 1),
5392                                    dentry->d_name.len);
5393         if (ret)
5394                 goto out;
5395
5396         btrfs_release_path(path);
5397
5398         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5399         if (IS_ERR(new_root)) {
5400                 err = PTR_ERR(new_root);
5401                 goto out;
5402         }
5403
5404         *sub_root = new_root;
5405         location->objectid = btrfs_root_dirid(&new_root->root_item);
5406         location->type = BTRFS_INODE_ITEM_KEY;
5407         location->offset = 0;
5408         err = 0;
5409 out:
5410         btrfs_free_path(path);
5411         return err;
5412 }
5413
5414 static void inode_tree_add(struct inode *inode)
5415 {
5416         struct btrfs_root *root = BTRFS_I(inode)->root;
5417         struct btrfs_inode *entry;
5418         struct rb_node **p;
5419         struct rb_node *parent;
5420         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5421         u64 ino = btrfs_ino(inode);
5422
5423         if (inode_unhashed(inode))
5424                 return;
5425         parent = NULL;
5426         spin_lock(&root->inode_lock);
5427         p = &root->inode_tree.rb_node;
5428         while (*p) {
5429                 parent = *p;
5430                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5431
5432                 if (ino < btrfs_ino(&entry->vfs_inode))
5433                         p = &parent->rb_left;
5434                 else if (ino > btrfs_ino(&entry->vfs_inode))
5435                         p = &parent->rb_right;
5436                 else {
5437                         WARN_ON(!(entry->vfs_inode.i_state &
5438                                   (I_WILL_FREE | I_FREEING)));
5439                         rb_replace_node(parent, new, &root->inode_tree);
5440                         RB_CLEAR_NODE(parent);
5441                         spin_unlock(&root->inode_lock);
5442                         return;
5443                 }
5444         }
5445         rb_link_node(new, parent, p);
5446         rb_insert_color(new, &root->inode_tree);
5447         spin_unlock(&root->inode_lock);
5448 }
5449
5450 static void inode_tree_del(struct inode *inode)
5451 {
5452         struct btrfs_root *root = BTRFS_I(inode)->root;
5453         int empty = 0;
5454
5455         spin_lock(&root->inode_lock);
5456         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5457                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5458                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5459                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5460         }
5461         spin_unlock(&root->inode_lock);
5462
5463         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5464                 synchronize_srcu(&root->fs_info->subvol_srcu);
5465                 spin_lock(&root->inode_lock);
5466                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5467                 spin_unlock(&root->inode_lock);
5468                 if (empty)
5469                         btrfs_add_dead_root(root);
5470         }
5471 }
5472
5473 void btrfs_invalidate_inodes(struct btrfs_root *root)
5474 {
5475         struct rb_node *node;
5476         struct rb_node *prev;
5477         struct btrfs_inode *entry;
5478         struct inode *inode;
5479         u64 objectid = 0;
5480
5481         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5482                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5483
5484         spin_lock(&root->inode_lock);
5485 again:
5486         node = root->inode_tree.rb_node;
5487         prev = NULL;
5488         while (node) {
5489                 prev = node;
5490                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5491
5492                 if (objectid < btrfs_ino(&entry->vfs_inode))
5493                         node = node->rb_left;
5494                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5495                         node = node->rb_right;
5496                 else
5497                         break;
5498         }
5499         if (!node) {
5500                 while (prev) {
5501                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5502                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5503                                 node = prev;
5504                                 break;
5505                         }
5506                         prev = rb_next(prev);
5507                 }
5508         }
5509         while (node) {
5510                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5511                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5512                 inode = igrab(&entry->vfs_inode);
5513                 if (inode) {
5514                         spin_unlock(&root->inode_lock);
5515                         if (atomic_read(&inode->i_count) > 1)
5516                                 d_prune_aliases(inode);
5517                         /*
5518                          * btrfs_drop_inode will have it removed from
5519                          * the inode cache when its usage count
5520                          * hits zero.
5521                          */
5522                         iput(inode);
5523                         cond_resched();
5524                         spin_lock(&root->inode_lock);
5525                         goto again;
5526                 }
5527
5528                 if (cond_resched_lock(&root->inode_lock))
5529                         goto again;
5530
5531                 node = rb_next(node);
5532         }
5533         spin_unlock(&root->inode_lock);
5534 }
5535
5536 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5537 {
5538         struct btrfs_iget_args *args = p;
5539         inode->i_ino = args->location->objectid;
5540         memcpy(&BTRFS_I(inode)->location, args->location,
5541                sizeof(*args->location));
5542         BTRFS_I(inode)->root = args->root;
5543         return 0;
5544 }
5545
5546 static int btrfs_find_actor(struct inode *inode, void *opaque)
5547 {
5548         struct btrfs_iget_args *args = opaque;
5549         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5550                 args->root == BTRFS_I(inode)->root;
5551 }
5552
5553 static struct inode *btrfs_iget_locked(struct super_block *s,
5554                                        struct btrfs_key *location,
5555                                        struct btrfs_root *root)
5556 {
5557         struct inode *inode;
5558         struct btrfs_iget_args args;
5559         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5560
5561         args.location = location;
5562         args.root = root;
5563
5564         inode = iget5_locked(s, hashval, btrfs_find_actor,
5565                              btrfs_init_locked_inode,
5566                              (void *)&args);
5567         return inode;
5568 }
5569
5570 /* Get an inode object given its location and corresponding root.
5571  * Returns in *is_new if the inode was read from disk
5572  */
5573 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5574                          struct btrfs_root *root, int *new)
5575 {
5576         struct inode *inode;
5577
5578         inode = btrfs_iget_locked(s, location, root);
5579         if (!inode)
5580                 return ERR_PTR(-ENOMEM);
5581
5582         if (inode->i_state & I_NEW) {
5583                 btrfs_read_locked_inode(inode);
5584                 if (!is_bad_inode(inode)) {
5585                         inode_tree_add(inode);
5586                         unlock_new_inode(inode);
5587                         if (new)
5588                                 *new = 1;
5589                 } else {
5590                         unlock_new_inode(inode);
5591                         iput(inode);
5592                         inode = ERR_PTR(-ESTALE);
5593                 }
5594         }
5595
5596         return inode;
5597 }
5598
5599 static struct inode *new_simple_dir(struct super_block *s,
5600                                     struct btrfs_key *key,
5601                                     struct btrfs_root *root)
5602 {
5603         struct inode *inode = new_inode(s);
5604
5605         if (!inode)
5606                 return ERR_PTR(-ENOMEM);
5607
5608         BTRFS_I(inode)->root = root;
5609         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5610         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5611
5612         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5613         inode->i_op = &btrfs_dir_ro_inode_operations;
5614         inode->i_fop = &simple_dir_operations;
5615         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5616         inode->i_mtime = CURRENT_TIME;
5617         inode->i_atime = inode->i_mtime;
5618         inode->i_ctime = inode->i_mtime;
5619         BTRFS_I(inode)->i_otime = inode->i_mtime;
5620
5621         return inode;
5622 }
5623
5624 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5625 {
5626         struct inode *inode;
5627         struct btrfs_root *root = BTRFS_I(dir)->root;
5628         struct btrfs_root *sub_root = root;
5629         struct btrfs_key location;
5630         int index;
5631         int ret = 0;
5632
5633         if (dentry->d_name.len > BTRFS_NAME_LEN)
5634                 return ERR_PTR(-ENAMETOOLONG);
5635
5636         ret = btrfs_inode_by_name(dir, dentry, &location);
5637         if (ret < 0)
5638                 return ERR_PTR(ret);
5639
5640         if (location.objectid == 0)
5641                 return ERR_PTR(-ENOENT);
5642
5643         if (location.type == BTRFS_INODE_ITEM_KEY) {
5644                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5645                 return inode;
5646         }
5647
5648         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5649
5650         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5651         ret = fixup_tree_root_location(root, dir, dentry,
5652                                        &location, &sub_root);
5653         if (ret < 0) {
5654                 if (ret != -ENOENT)
5655                         inode = ERR_PTR(ret);
5656                 else
5657                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5658         } else {
5659                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5660         }
5661         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5662
5663         if (!IS_ERR(inode) && root != sub_root) {
5664                 down_read(&root->fs_info->cleanup_work_sem);
5665                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5666                         ret = btrfs_orphan_cleanup(sub_root);
5667                 up_read(&root->fs_info->cleanup_work_sem);
5668                 if (ret) {
5669                         iput(inode);
5670                         inode = ERR_PTR(ret);
5671                 }
5672         }
5673
5674         return inode;
5675 }
5676
5677 static int btrfs_dentry_delete(const struct dentry *dentry)
5678 {
5679         struct btrfs_root *root;
5680         struct inode *inode = d_inode(dentry);
5681
5682         if (!inode && !IS_ROOT(dentry))
5683                 inode = d_inode(dentry->d_parent);
5684
5685         if (inode) {
5686                 root = BTRFS_I(inode)->root;
5687                 if (btrfs_root_refs(&root->root_item) == 0)
5688                         return 1;
5689
5690                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5691                         return 1;
5692         }
5693         return 0;
5694 }
5695
5696 static void btrfs_dentry_release(struct dentry *dentry)
5697 {
5698         kfree(dentry->d_fsdata);
5699 }
5700
5701 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5702                                    unsigned int flags)
5703 {
5704         struct inode *inode;
5705
5706         inode = btrfs_lookup_dentry(dir, dentry);
5707         if (IS_ERR(inode)) {
5708                 if (PTR_ERR(inode) == -ENOENT)
5709                         inode = NULL;
5710                 else
5711                         return ERR_CAST(inode);
5712         }
5713
5714         return d_splice_alias(inode, dentry);
5715 }
5716
5717 unsigned char btrfs_filetype_table[] = {
5718         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5719 };
5720
5721 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5722 {
5723         struct inode *inode = file_inode(file);
5724         struct btrfs_root *root = BTRFS_I(inode)->root;
5725         struct btrfs_item *item;
5726         struct btrfs_dir_item *di;
5727         struct btrfs_key key;
5728         struct btrfs_key found_key;
5729         struct btrfs_path *path;
5730         struct list_head ins_list;
5731         struct list_head del_list;
5732         int ret;
5733         struct extent_buffer *leaf;
5734         int slot;
5735         unsigned char d_type;
5736         int over = 0;
5737         u32 di_cur;
5738         u32 di_total;
5739         u32 di_len;
5740         int key_type = BTRFS_DIR_INDEX_KEY;
5741         char tmp_name[32];
5742         char *name_ptr;
5743         int name_len;
5744         int is_curr = 0;        /* ctx->pos points to the current index? */
5745
5746         /* FIXME, use a real flag for deciding about the key type */
5747         if (root->fs_info->tree_root == root)
5748                 key_type = BTRFS_DIR_ITEM_KEY;
5749
5750         if (!dir_emit_dots(file, ctx))
5751                 return 0;
5752
5753         path = btrfs_alloc_path();
5754         if (!path)
5755                 return -ENOMEM;
5756
5757         path->reada = 1;
5758
5759         if (key_type == BTRFS_DIR_INDEX_KEY) {
5760                 INIT_LIST_HEAD(&ins_list);
5761                 INIT_LIST_HEAD(&del_list);
5762                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5763         }
5764
5765         key.type = key_type;
5766         key.offset = ctx->pos;
5767         key.objectid = btrfs_ino(inode);
5768
5769         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5770         if (ret < 0)
5771                 goto err;
5772
5773         while (1) {
5774                 leaf = path->nodes[0];
5775                 slot = path->slots[0];
5776                 if (slot >= btrfs_header_nritems(leaf)) {
5777                         ret = btrfs_next_leaf(root, path);
5778                         if (ret < 0)
5779                                 goto err;
5780                         else if (ret > 0)
5781                                 break;
5782                         continue;
5783                 }
5784
5785                 item = btrfs_item_nr(slot);
5786                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5787
5788                 if (found_key.objectid != key.objectid)
5789                         break;
5790                 if (found_key.type != key_type)
5791                         break;
5792                 if (found_key.offset < ctx->pos)
5793                         goto next;
5794                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5795                     btrfs_should_delete_dir_index(&del_list,
5796                                                   found_key.offset))
5797                         goto next;
5798
5799                 ctx->pos = found_key.offset;
5800                 is_curr = 1;
5801
5802                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5803                 di_cur = 0;
5804                 di_total = btrfs_item_size(leaf, item);
5805
5806                 while (di_cur < di_total) {
5807                         struct btrfs_key location;
5808
5809                         if (verify_dir_item(root, leaf, di))
5810                                 break;
5811
5812                         name_len = btrfs_dir_name_len(leaf, di);
5813                         if (name_len <= sizeof(tmp_name)) {
5814                                 name_ptr = tmp_name;
5815                         } else {
5816                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5817                                 if (!name_ptr) {
5818                                         ret = -ENOMEM;
5819                                         goto err;
5820                                 }
5821                         }
5822                         read_extent_buffer(leaf, name_ptr,
5823                                            (unsigned long)(di + 1), name_len);
5824
5825                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5826                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5827
5828
5829                         /* is this a reference to our own snapshot? If so
5830                          * skip it.
5831                          *
5832                          * In contrast to old kernels, we insert the snapshot's
5833                          * dir item and dir index after it has been created, so
5834                          * we won't find a reference to our own snapshot. We
5835                          * still keep the following code for backward
5836                          * compatibility.
5837                          */
5838                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5839                             location.objectid == root->root_key.objectid) {
5840                                 over = 0;
5841                                 goto skip;
5842                         }
5843                         over = !dir_emit(ctx, name_ptr, name_len,
5844                                        location.objectid, d_type);
5845
5846 skip:
5847                         if (name_ptr != tmp_name)
5848                                 kfree(name_ptr);
5849
5850                         if (over)
5851                                 goto nopos;
5852                         di_len = btrfs_dir_name_len(leaf, di) +
5853                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5854                         di_cur += di_len;
5855                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5856                 }
5857 next:
5858                 path->slots[0]++;
5859         }
5860
5861         if (key_type == BTRFS_DIR_INDEX_KEY) {
5862                 if (is_curr)
5863                         ctx->pos++;
5864                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5865                 if (ret)
5866                         goto nopos;
5867         }
5868
5869         /* Reached end of directory/root. Bump pos past the last item. */
5870         ctx->pos++;
5871
5872         /*
5873          * Stop new entries from being returned after we return the last
5874          * entry.
5875          *
5876          * New directory entries are assigned a strictly increasing
5877          * offset.  This means that new entries created during readdir
5878          * are *guaranteed* to be seen in the future by that readdir.
5879          * This has broken buggy programs which operate on names as
5880          * they're returned by readdir.  Until we re-use freed offsets
5881          * we have this hack to stop new entries from being returned
5882          * under the assumption that they'll never reach this huge
5883          * offset.
5884          *
5885          * This is being careful not to overflow 32bit loff_t unless the
5886          * last entry requires it because doing so has broken 32bit apps
5887          * in the past.
5888          */
5889         if (key_type == BTRFS_DIR_INDEX_KEY) {
5890                 if (ctx->pos >= INT_MAX)
5891                         ctx->pos = LLONG_MAX;
5892                 else
5893                         ctx->pos = INT_MAX;
5894         }
5895 nopos:
5896         ret = 0;
5897 err:
5898         if (key_type == BTRFS_DIR_INDEX_KEY)
5899                 btrfs_put_delayed_items(&ins_list, &del_list);
5900         btrfs_free_path(path);
5901         return ret;
5902 }
5903
5904 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5905 {
5906         struct btrfs_root *root = BTRFS_I(inode)->root;
5907         struct btrfs_trans_handle *trans;
5908         int ret = 0;
5909         bool nolock = false;
5910
5911         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5912                 return 0;
5913
5914         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5915                 nolock = true;
5916
5917         if (wbc->sync_mode == WB_SYNC_ALL) {
5918                 if (nolock)
5919                         trans = btrfs_join_transaction_nolock(root);
5920                 else
5921                         trans = btrfs_join_transaction(root);
5922                 if (IS_ERR(trans))
5923                         return PTR_ERR(trans);
5924                 ret = btrfs_commit_transaction(trans, root);
5925         }
5926         return ret;
5927 }
5928
5929 /*
5930  * This is somewhat expensive, updating the tree every time the
5931  * inode changes.  But, it is most likely to find the inode in cache.
5932  * FIXME, needs more benchmarking...there are no reasons other than performance
5933  * to keep or drop this code.
5934  */
5935 static int btrfs_dirty_inode(struct inode *inode)
5936 {
5937         struct btrfs_root *root = BTRFS_I(inode)->root;
5938         struct btrfs_trans_handle *trans;
5939         int ret;
5940
5941         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5942                 return 0;
5943
5944         trans = btrfs_join_transaction(root);
5945         if (IS_ERR(trans))
5946                 return PTR_ERR(trans);
5947
5948         ret = btrfs_update_inode(trans, root, inode);
5949         if (ret && ret == -ENOSPC) {
5950                 /* whoops, lets try again with the full transaction */
5951                 btrfs_end_transaction(trans, root);
5952                 trans = btrfs_start_transaction(root, 1);
5953                 if (IS_ERR(trans))
5954                         return PTR_ERR(trans);
5955
5956                 ret = btrfs_update_inode(trans, root, inode);
5957         }
5958         btrfs_end_transaction(trans, root);
5959         if (BTRFS_I(inode)->delayed_node)
5960                 btrfs_balance_delayed_items(root);
5961
5962         return ret;
5963 }
5964
5965 /*
5966  * This is a copy of file_update_time.  We need this so we can return error on
5967  * ENOSPC for updating the inode in the case of file write and mmap writes.
5968  */
5969 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5970                              int flags)
5971 {
5972         struct btrfs_root *root = BTRFS_I(inode)->root;
5973
5974         if (btrfs_root_readonly(root))
5975                 return -EROFS;
5976
5977         if (flags & S_VERSION)
5978                 inode_inc_iversion(inode);
5979         if (flags & S_CTIME)
5980                 inode->i_ctime = *now;
5981         if (flags & S_MTIME)
5982                 inode->i_mtime = *now;
5983         if (flags & S_ATIME)
5984                 inode->i_atime = *now;
5985         return btrfs_dirty_inode(inode);
5986 }
5987
5988 /*
5989  * find the highest existing sequence number in a directory
5990  * and then set the in-memory index_cnt variable to reflect
5991  * free sequence numbers
5992  */
5993 static int btrfs_set_inode_index_count(struct inode *inode)
5994 {
5995         struct btrfs_root *root = BTRFS_I(inode)->root;
5996         struct btrfs_key key, found_key;
5997         struct btrfs_path *path;
5998         struct extent_buffer *leaf;
5999         int ret;
6000
6001         key.objectid = btrfs_ino(inode);
6002         key.type = BTRFS_DIR_INDEX_KEY;
6003         key.offset = (u64)-1;
6004
6005         path = btrfs_alloc_path();
6006         if (!path)
6007                 return -ENOMEM;
6008
6009         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6010         if (ret < 0)
6011                 goto out;
6012         /* FIXME: we should be able to handle this */
6013         if (ret == 0)
6014                 goto out;
6015         ret = 0;
6016
6017         /*
6018          * MAGIC NUMBER EXPLANATION:
6019          * since we search a directory based on f_pos we have to start at 2
6020          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6021          * else has to start at 2
6022          */
6023         if (path->slots[0] == 0) {
6024                 BTRFS_I(inode)->index_cnt = 2;
6025                 goto out;
6026         }
6027
6028         path->slots[0]--;
6029
6030         leaf = path->nodes[0];
6031         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6032
6033         if (found_key.objectid != btrfs_ino(inode) ||
6034             found_key.type != BTRFS_DIR_INDEX_KEY) {
6035                 BTRFS_I(inode)->index_cnt = 2;
6036                 goto out;
6037         }
6038
6039         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6040 out:
6041         btrfs_free_path(path);
6042         return ret;
6043 }
6044
6045 /*
6046  * helper to find a free sequence number in a given directory.  This current
6047  * code is very simple, later versions will do smarter things in the btree
6048  */
6049 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6050 {
6051         int ret = 0;
6052
6053         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6054                 ret = btrfs_inode_delayed_dir_index_count(dir);
6055                 if (ret) {
6056                         ret = btrfs_set_inode_index_count(dir);
6057                         if (ret)
6058                                 return ret;
6059                 }
6060         }
6061
6062         *index = BTRFS_I(dir)->index_cnt;
6063         BTRFS_I(dir)->index_cnt++;
6064
6065         return ret;
6066 }
6067
6068 static int btrfs_insert_inode_locked(struct inode *inode)
6069 {
6070         struct btrfs_iget_args args;
6071         args.location = &BTRFS_I(inode)->location;
6072         args.root = BTRFS_I(inode)->root;
6073
6074         return insert_inode_locked4(inode,
6075                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6076                    btrfs_find_actor, &args);
6077 }
6078
6079 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6080                                      struct btrfs_root *root,
6081                                      struct inode *dir,
6082                                      const char *name, int name_len,
6083                                      u64 ref_objectid, u64 objectid,
6084                                      umode_t mode, u64 *index)
6085 {
6086         struct inode *inode;
6087         struct btrfs_inode_item *inode_item;
6088         struct btrfs_key *location;
6089         struct btrfs_path *path;
6090         struct btrfs_inode_ref *ref;
6091         struct btrfs_key key[2];
6092         u32 sizes[2];
6093         int nitems = name ? 2 : 1;
6094         unsigned long ptr;
6095         int ret;
6096
6097         path = btrfs_alloc_path();
6098         if (!path)
6099                 return ERR_PTR(-ENOMEM);
6100
6101         inode = new_inode(root->fs_info->sb);
6102         if (!inode) {
6103                 btrfs_free_path(path);
6104                 return ERR_PTR(-ENOMEM);
6105         }
6106
6107         /*
6108          * O_TMPFILE, set link count to 0, so that after this point,
6109          * we fill in an inode item with the correct link count.
6110          */
6111         if (!name)
6112                 set_nlink(inode, 0);
6113
6114         /*
6115          * we have to initialize this early, so we can reclaim the inode
6116          * number if we fail afterwards in this function.
6117          */
6118         inode->i_ino = objectid;
6119
6120         if (dir && name) {
6121                 trace_btrfs_inode_request(dir);
6122
6123                 ret = btrfs_set_inode_index(dir, index);
6124                 if (ret) {
6125                         btrfs_free_path(path);
6126                         iput(inode);
6127                         return ERR_PTR(ret);
6128                 }
6129         } else if (dir) {
6130                 *index = 0;
6131         }
6132         /*
6133          * index_cnt is ignored for everything but a dir,
6134          * btrfs_get_inode_index_count has an explanation for the magic
6135          * number
6136          */
6137         BTRFS_I(inode)->index_cnt = 2;
6138         BTRFS_I(inode)->dir_index = *index;
6139         BTRFS_I(inode)->root = root;
6140         BTRFS_I(inode)->generation = trans->transid;
6141         inode->i_generation = BTRFS_I(inode)->generation;
6142
6143         /*
6144          * We could have gotten an inode number from somebody who was fsynced
6145          * and then removed in this same transaction, so let's just set full
6146          * sync since it will be a full sync anyway and this will blow away the
6147          * old info in the log.
6148          */
6149         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6150
6151         key[0].objectid = objectid;
6152         key[0].type = BTRFS_INODE_ITEM_KEY;
6153         key[0].offset = 0;
6154
6155         sizes[0] = sizeof(struct btrfs_inode_item);
6156
6157         if (name) {
6158                 /*
6159                  * Start new inodes with an inode_ref. This is slightly more
6160                  * efficient for small numbers of hard links since they will
6161                  * be packed into one item. Extended refs will kick in if we
6162                  * add more hard links than can fit in the ref item.
6163                  */
6164                 key[1].objectid = objectid;
6165                 key[1].type = BTRFS_INODE_REF_KEY;
6166                 key[1].offset = ref_objectid;
6167
6168                 sizes[1] = name_len + sizeof(*ref);
6169         }
6170
6171         location = &BTRFS_I(inode)->location;
6172         location->objectid = objectid;
6173         location->offset = 0;
6174         location->type = BTRFS_INODE_ITEM_KEY;
6175
6176         ret = btrfs_insert_inode_locked(inode);
6177         if (ret < 0)
6178                 goto fail;
6179
6180         path->leave_spinning = 1;
6181         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6182         if (ret != 0)
6183                 goto fail_unlock;
6184
6185         inode_init_owner(inode, dir, mode);
6186         inode_set_bytes(inode, 0);
6187
6188         inode->i_mtime = CURRENT_TIME;
6189         inode->i_atime = inode->i_mtime;
6190         inode->i_ctime = inode->i_mtime;
6191         BTRFS_I(inode)->i_otime = inode->i_mtime;
6192
6193         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6194                                   struct btrfs_inode_item);
6195         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6196                              sizeof(*inode_item));
6197         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6198
6199         if (name) {
6200                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6201                                      struct btrfs_inode_ref);
6202                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6203                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6204                 ptr = (unsigned long)(ref + 1);
6205                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6206         }
6207
6208         btrfs_mark_buffer_dirty(path->nodes[0]);
6209         btrfs_free_path(path);
6210
6211         btrfs_inherit_iflags(inode, dir);
6212
6213         if (S_ISREG(mode)) {
6214                 if (btrfs_test_opt(root, NODATASUM))
6215                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6216                 if (btrfs_test_opt(root, NODATACOW))
6217                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6218                                 BTRFS_INODE_NODATASUM;
6219         }
6220
6221         inode_tree_add(inode);
6222
6223         trace_btrfs_inode_new(inode);
6224         btrfs_set_inode_last_trans(trans, inode);
6225
6226         btrfs_update_root_times(trans, root);
6227
6228         ret = btrfs_inode_inherit_props(trans, inode, dir);
6229         if (ret)
6230                 btrfs_err(root->fs_info,
6231                           "error inheriting props for ino %llu (root %llu): %d",
6232                           btrfs_ino(inode), root->root_key.objectid, ret);
6233
6234         return inode;
6235
6236 fail_unlock:
6237         unlock_new_inode(inode);
6238 fail:
6239         if (dir && name)
6240                 BTRFS_I(dir)->index_cnt--;
6241         btrfs_free_path(path);
6242         iput(inode);
6243         return ERR_PTR(ret);
6244 }
6245
6246 static inline u8 btrfs_inode_type(struct inode *inode)
6247 {
6248         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6249 }
6250
6251 /*
6252  * utility function to add 'inode' into 'parent_inode' with
6253  * a give name and a given sequence number.
6254  * if 'add_backref' is true, also insert a backref from the
6255  * inode to the parent directory.
6256  */
6257 int btrfs_add_link(struct btrfs_trans_handle *trans,
6258                    struct inode *parent_inode, struct inode *inode,
6259                    const char *name, int name_len, int add_backref, u64 index)
6260 {
6261         int ret = 0;
6262         struct btrfs_key key;
6263         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6264         u64 ino = btrfs_ino(inode);
6265         u64 parent_ino = btrfs_ino(parent_inode);
6266
6267         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6268                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6269         } else {
6270                 key.objectid = ino;
6271                 key.type = BTRFS_INODE_ITEM_KEY;
6272                 key.offset = 0;
6273         }
6274
6275         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6276                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6277                                          key.objectid, root->root_key.objectid,
6278                                          parent_ino, index, name, name_len);
6279         } else if (add_backref) {
6280                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6281                                              parent_ino, index);
6282         }
6283
6284         /* Nothing to clean up yet */
6285         if (ret)
6286                 return ret;
6287
6288         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6289                                     parent_inode, &key,
6290                                     btrfs_inode_type(inode), index);
6291         if (ret == -EEXIST || ret == -EOVERFLOW)
6292                 goto fail_dir_item;
6293         else if (ret) {
6294                 btrfs_abort_transaction(trans, root, ret);
6295                 return ret;
6296         }
6297
6298         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6299                            name_len * 2);
6300         inode_inc_iversion(parent_inode);
6301         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
6302         ret = btrfs_update_inode(trans, root, parent_inode);
6303         if (ret)
6304                 btrfs_abort_transaction(trans, root, ret);
6305         return ret;
6306
6307 fail_dir_item:
6308         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6309                 u64 local_index;
6310                 int err;
6311                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6312                                  key.objectid, root->root_key.objectid,
6313                                  parent_ino, &local_index, name, name_len);
6314
6315         } else if (add_backref) {
6316                 u64 local_index;
6317                 int err;
6318
6319                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6320                                           ino, parent_ino, &local_index);
6321         }
6322         return ret;
6323 }
6324
6325 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6326                             struct inode *dir, struct dentry *dentry,
6327                             struct inode *inode, int backref, u64 index)
6328 {
6329         int err = btrfs_add_link(trans, dir, inode,
6330                                  dentry->d_name.name, dentry->d_name.len,
6331                                  backref, index);
6332         if (err > 0)
6333                 err = -EEXIST;
6334         return err;
6335 }
6336
6337 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6338                         umode_t mode, dev_t rdev)
6339 {
6340         struct btrfs_trans_handle *trans;
6341         struct btrfs_root *root = BTRFS_I(dir)->root;
6342         struct inode *inode = NULL;
6343         int err;
6344         int drop_inode = 0;
6345         u64 objectid;
6346         u64 index = 0;
6347
6348         /*
6349          * 2 for inode item and ref
6350          * 2 for dir items
6351          * 1 for xattr if selinux is on
6352          */
6353         trans = btrfs_start_transaction(root, 5);
6354         if (IS_ERR(trans))
6355                 return PTR_ERR(trans);
6356
6357         err = btrfs_find_free_ino(root, &objectid);
6358         if (err)
6359                 goto out_unlock;
6360
6361         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6362                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6363                                 mode, &index);
6364         if (IS_ERR(inode)) {
6365                 err = PTR_ERR(inode);
6366                 goto out_unlock;
6367         }
6368
6369         /*
6370         * If the active LSM wants to access the inode during
6371         * d_instantiate it needs these. Smack checks to see
6372         * if the filesystem supports xattrs by looking at the
6373         * ops vector.
6374         */
6375         inode->i_op = &btrfs_special_inode_operations;
6376         init_special_inode(inode, inode->i_mode, rdev);
6377
6378         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6379         if (err)
6380                 goto out_unlock_inode;
6381
6382         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6383         if (err) {
6384                 goto out_unlock_inode;
6385         } else {
6386                 btrfs_update_inode(trans, root, inode);
6387                 unlock_new_inode(inode);
6388                 d_instantiate(dentry, inode);
6389         }
6390
6391 out_unlock:
6392         btrfs_end_transaction(trans, root);
6393         btrfs_balance_delayed_items(root);
6394         btrfs_btree_balance_dirty(root);
6395         if (drop_inode) {
6396                 inode_dec_link_count(inode);
6397                 iput(inode);
6398         }
6399         return err;
6400
6401 out_unlock_inode:
6402         drop_inode = 1;
6403         unlock_new_inode(inode);
6404         goto out_unlock;
6405
6406 }
6407
6408 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6409                         umode_t mode, bool excl)
6410 {
6411         struct btrfs_trans_handle *trans;
6412         struct btrfs_root *root = BTRFS_I(dir)->root;
6413         struct inode *inode = NULL;
6414         int drop_inode_on_err = 0;
6415         int err;
6416         u64 objectid;
6417         u64 index = 0;
6418
6419         /*
6420          * 2 for inode item and ref
6421          * 2 for dir items
6422          * 1 for xattr if selinux is on
6423          */
6424         trans = btrfs_start_transaction(root, 5);
6425         if (IS_ERR(trans))
6426                 return PTR_ERR(trans);
6427
6428         err = btrfs_find_free_ino(root, &objectid);
6429         if (err)
6430                 goto out_unlock;
6431
6432         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6433                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6434                                 mode, &index);
6435         if (IS_ERR(inode)) {
6436                 err = PTR_ERR(inode);
6437                 goto out_unlock;
6438         }
6439         drop_inode_on_err = 1;
6440         /*
6441         * If the active LSM wants to access the inode during
6442         * d_instantiate it needs these. Smack checks to see
6443         * if the filesystem supports xattrs by looking at the
6444         * ops vector.
6445         */
6446         inode->i_fop = &btrfs_file_operations;
6447         inode->i_op = &btrfs_file_inode_operations;
6448         inode->i_mapping->a_ops = &btrfs_aops;
6449
6450         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6451         if (err)
6452                 goto out_unlock_inode;
6453
6454         err = btrfs_update_inode(trans, root, inode);
6455         if (err)
6456                 goto out_unlock_inode;
6457
6458         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6459         if (err)
6460                 goto out_unlock_inode;
6461
6462         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6463         unlock_new_inode(inode);
6464         d_instantiate(dentry, inode);
6465
6466 out_unlock:
6467         btrfs_end_transaction(trans, root);
6468         if (err && drop_inode_on_err) {
6469                 inode_dec_link_count(inode);
6470                 iput(inode);
6471         }
6472         btrfs_balance_delayed_items(root);
6473         btrfs_btree_balance_dirty(root);
6474         return err;
6475
6476 out_unlock_inode:
6477         unlock_new_inode(inode);
6478         goto out_unlock;
6479
6480 }
6481
6482 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6483                       struct dentry *dentry)
6484 {
6485         struct btrfs_trans_handle *trans;
6486         struct btrfs_root *root = BTRFS_I(dir)->root;
6487         struct inode *inode = d_inode(old_dentry);
6488         u64 index;
6489         int err;
6490         int drop_inode = 0;
6491
6492         /* do not allow sys_link's with other subvols of the same device */
6493         if (root->objectid != BTRFS_I(inode)->root->objectid)
6494                 return -EXDEV;
6495
6496         if (inode->i_nlink >= BTRFS_LINK_MAX)
6497                 return -EMLINK;
6498
6499         err = btrfs_set_inode_index(dir, &index);
6500         if (err)
6501                 goto fail;
6502
6503         /*
6504          * 2 items for inode and inode ref
6505          * 2 items for dir items
6506          * 1 item for parent inode
6507          */
6508         trans = btrfs_start_transaction(root, 5);
6509         if (IS_ERR(trans)) {
6510                 err = PTR_ERR(trans);
6511                 goto fail;
6512         }
6513
6514         /* There are several dir indexes for this inode, clear the cache. */
6515         BTRFS_I(inode)->dir_index = 0ULL;
6516         inc_nlink(inode);
6517         inode_inc_iversion(inode);
6518         inode->i_ctime = CURRENT_TIME;
6519         ihold(inode);
6520         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6521
6522         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6523
6524         if (err) {
6525                 drop_inode = 1;
6526         } else {
6527                 struct dentry *parent = dentry->d_parent;
6528                 err = btrfs_update_inode(trans, root, inode);
6529                 if (err)
6530                         goto fail;
6531                 if (inode->i_nlink == 1) {
6532                         /*
6533                          * If new hard link count is 1, it's a file created
6534                          * with open(2) O_TMPFILE flag.
6535                          */
6536                         err = btrfs_orphan_del(trans, inode);
6537                         if (err)
6538                                 goto fail;
6539                 }
6540                 d_instantiate(dentry, inode);
6541                 btrfs_log_new_name(trans, inode, NULL, parent);
6542         }
6543
6544         btrfs_end_transaction(trans, root);
6545         btrfs_balance_delayed_items(root);
6546 fail:
6547         if (drop_inode) {
6548                 inode_dec_link_count(inode);
6549                 iput(inode);
6550         }
6551         btrfs_btree_balance_dirty(root);
6552         return err;
6553 }
6554
6555 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6556 {
6557         struct inode *inode = NULL;
6558         struct btrfs_trans_handle *trans;
6559         struct btrfs_root *root = BTRFS_I(dir)->root;
6560         int err = 0;
6561         int drop_on_err = 0;
6562         u64 objectid = 0;
6563         u64 index = 0;
6564
6565         /*
6566          * 2 items for inode and ref
6567          * 2 items for dir items
6568          * 1 for xattr if selinux is on
6569          */
6570         trans = btrfs_start_transaction(root, 5);
6571         if (IS_ERR(trans))
6572                 return PTR_ERR(trans);
6573
6574         err = btrfs_find_free_ino(root, &objectid);
6575         if (err)
6576                 goto out_fail;
6577
6578         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6579                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6580                                 S_IFDIR | mode, &index);
6581         if (IS_ERR(inode)) {
6582                 err = PTR_ERR(inode);
6583                 goto out_fail;
6584         }
6585
6586         drop_on_err = 1;
6587         /* these must be set before we unlock the inode */
6588         inode->i_op = &btrfs_dir_inode_operations;
6589         inode->i_fop = &btrfs_dir_file_operations;
6590
6591         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6592         if (err)
6593                 goto out_fail_inode;
6594
6595         btrfs_i_size_write(inode, 0);
6596         err = btrfs_update_inode(trans, root, inode);
6597         if (err)
6598                 goto out_fail_inode;
6599
6600         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6601                              dentry->d_name.len, 0, index);
6602         if (err)
6603                 goto out_fail_inode;
6604
6605         d_instantiate(dentry, inode);
6606         /*
6607          * mkdir is special.  We're unlocking after we call d_instantiate
6608          * to avoid a race with nfsd calling d_instantiate.
6609          */
6610         unlock_new_inode(inode);
6611         drop_on_err = 0;
6612
6613 out_fail:
6614         btrfs_end_transaction(trans, root);
6615         if (drop_on_err) {
6616                 inode_dec_link_count(inode);
6617                 iput(inode);
6618         }
6619         btrfs_balance_delayed_items(root);
6620         btrfs_btree_balance_dirty(root);
6621         return err;
6622
6623 out_fail_inode:
6624         unlock_new_inode(inode);
6625         goto out_fail;
6626 }
6627
6628 /* Find next extent map of a given extent map, caller needs to ensure locks */
6629 static struct extent_map *next_extent_map(struct extent_map *em)
6630 {
6631         struct rb_node *next;
6632
6633         next = rb_next(&em->rb_node);
6634         if (!next)
6635                 return NULL;
6636         return container_of(next, struct extent_map, rb_node);
6637 }
6638
6639 static struct extent_map *prev_extent_map(struct extent_map *em)
6640 {
6641         struct rb_node *prev;
6642
6643         prev = rb_prev(&em->rb_node);
6644         if (!prev)
6645                 return NULL;
6646         return container_of(prev, struct extent_map, rb_node);
6647 }
6648
6649 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6650  * the existing extent is the nearest extent to map_start,
6651  * and an extent that you want to insert, deal with overlap and insert
6652  * the best fitted new extent into the tree.
6653  */
6654 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6655                                 struct extent_map *existing,
6656                                 struct extent_map *em,
6657                                 u64 map_start)
6658 {
6659         struct extent_map *prev;
6660         struct extent_map *next;
6661         u64 start;
6662         u64 end;
6663         u64 start_diff;
6664
6665         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6666
6667         if (existing->start > map_start) {
6668                 next = existing;
6669                 prev = prev_extent_map(next);
6670         } else {
6671                 prev = existing;
6672                 next = next_extent_map(prev);
6673         }
6674
6675         start = prev ? extent_map_end(prev) : em->start;
6676         start = max_t(u64, start, em->start);
6677         end = next ? next->start : extent_map_end(em);
6678         end = min_t(u64, end, extent_map_end(em));
6679         start_diff = start - em->start;
6680         em->start = start;
6681         em->len = end - start;
6682         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6683             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6684                 em->block_start += start_diff;
6685                 em->block_len -= start_diff;
6686         }
6687         return add_extent_mapping(em_tree, em, 0);
6688 }
6689
6690 static noinline int uncompress_inline(struct btrfs_path *path,
6691                                       struct inode *inode, struct page *page,
6692                                       size_t pg_offset, u64 extent_offset,
6693                                       struct btrfs_file_extent_item *item)
6694 {
6695         int ret;
6696         struct extent_buffer *leaf = path->nodes[0];
6697         char *tmp;
6698         size_t max_size;
6699         unsigned long inline_size;
6700         unsigned long ptr;
6701         int compress_type;
6702
6703         WARN_ON(pg_offset != 0);
6704         compress_type = btrfs_file_extent_compression(leaf, item);
6705         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6706         inline_size = btrfs_file_extent_inline_item_len(leaf,
6707                                         btrfs_item_nr(path->slots[0]));
6708         tmp = kmalloc(inline_size, GFP_NOFS);
6709         if (!tmp)
6710                 return -ENOMEM;
6711         ptr = btrfs_file_extent_inline_start(item);
6712
6713         read_extent_buffer(leaf, tmp, ptr, inline_size);
6714
6715         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
6716         ret = btrfs_decompress(compress_type, tmp, page,
6717                                extent_offset, inline_size, max_size);
6718         kfree(tmp);
6719         return ret;
6720 }
6721
6722 /*
6723  * a bit scary, this does extent mapping from logical file offset to the disk.
6724  * the ugly parts come from merging extents from the disk with the in-ram
6725  * representation.  This gets more complex because of the data=ordered code,
6726  * where the in-ram extents might be locked pending data=ordered completion.
6727  *
6728  * This also copies inline extents directly into the page.
6729  */
6730
6731 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6732                                     size_t pg_offset, u64 start, u64 len,
6733                                     int create)
6734 {
6735         int ret;
6736         int err = 0;
6737         u64 extent_start = 0;
6738         u64 extent_end = 0;
6739         u64 objectid = btrfs_ino(inode);
6740         u32 found_type;
6741         struct btrfs_path *path = NULL;
6742         struct btrfs_root *root = BTRFS_I(inode)->root;
6743         struct btrfs_file_extent_item *item;
6744         struct extent_buffer *leaf;
6745         struct btrfs_key found_key;
6746         struct extent_map *em = NULL;
6747         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6748         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6749         struct btrfs_trans_handle *trans = NULL;
6750         const bool new_inline = !page || create;
6751
6752 again:
6753         read_lock(&em_tree->lock);
6754         em = lookup_extent_mapping(em_tree, start, len);
6755         if (em)
6756                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6757         read_unlock(&em_tree->lock);
6758
6759         if (em) {
6760                 if (em->start > start || em->start + em->len <= start)
6761                         free_extent_map(em);
6762                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6763                         free_extent_map(em);
6764                 else
6765                         goto out;
6766         }
6767         em = alloc_extent_map();
6768         if (!em) {
6769                 err = -ENOMEM;
6770                 goto out;
6771         }
6772         em->bdev = root->fs_info->fs_devices->latest_bdev;
6773         em->start = EXTENT_MAP_HOLE;
6774         em->orig_start = EXTENT_MAP_HOLE;
6775         em->len = (u64)-1;
6776         em->block_len = (u64)-1;
6777
6778         if (!path) {
6779                 path = btrfs_alloc_path();
6780                 if (!path) {
6781                         err = -ENOMEM;
6782                         goto out;
6783                 }
6784                 /*
6785                  * Chances are we'll be called again, so go ahead and do
6786                  * readahead
6787                  */
6788                 path->reada = 1;
6789         }
6790
6791         ret = btrfs_lookup_file_extent(trans, root, path,
6792                                        objectid, start, trans != NULL);
6793         if (ret < 0) {
6794                 err = ret;
6795                 goto out;
6796         }
6797
6798         if (ret != 0) {
6799                 if (path->slots[0] == 0)
6800                         goto not_found;
6801                 path->slots[0]--;
6802         }
6803
6804         leaf = path->nodes[0];
6805         item = btrfs_item_ptr(leaf, path->slots[0],
6806                               struct btrfs_file_extent_item);
6807         /* are we inside the extent that was found? */
6808         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6809         found_type = found_key.type;
6810         if (found_key.objectid != objectid ||
6811             found_type != BTRFS_EXTENT_DATA_KEY) {
6812                 /*
6813                  * If we backup past the first extent we want to move forward
6814                  * and see if there is an extent in front of us, otherwise we'll
6815                  * say there is a hole for our whole search range which can
6816                  * cause problems.
6817                  */
6818                 extent_end = start;
6819                 goto next;
6820         }
6821
6822         found_type = btrfs_file_extent_type(leaf, item);
6823         extent_start = found_key.offset;
6824         if (found_type == BTRFS_FILE_EXTENT_REG ||
6825             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6826                 extent_end = extent_start +
6827                        btrfs_file_extent_num_bytes(leaf, item);
6828         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6829                 size_t size;
6830                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6831                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6832         }
6833 next:
6834         if (start >= extent_end) {
6835                 path->slots[0]++;
6836                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6837                         ret = btrfs_next_leaf(root, path);
6838                         if (ret < 0) {
6839                                 err = ret;
6840                                 goto out;
6841                         }
6842                         if (ret > 0)
6843                                 goto not_found;
6844                         leaf = path->nodes[0];
6845                 }
6846                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6847                 if (found_key.objectid != objectid ||
6848                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6849                         goto not_found;
6850                 if (start + len <= found_key.offset)
6851                         goto not_found;
6852                 if (start > found_key.offset)
6853                         goto next;
6854                 em->start = start;
6855                 em->orig_start = start;
6856                 em->len = found_key.offset - start;
6857                 goto not_found_em;
6858         }
6859
6860         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6861
6862         if (found_type == BTRFS_FILE_EXTENT_REG ||
6863             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6864                 goto insert;
6865         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6866                 unsigned long ptr;
6867                 char *map;
6868                 size_t size;
6869                 size_t extent_offset;
6870                 size_t copy_size;
6871
6872                 if (new_inline)
6873                         goto out;
6874
6875                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6876                 extent_offset = page_offset(page) + pg_offset - extent_start;
6877                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6878                                 size - extent_offset);
6879                 em->start = extent_start + extent_offset;
6880                 em->len = ALIGN(copy_size, root->sectorsize);
6881                 em->orig_block_len = em->len;
6882                 em->orig_start = em->start;
6883                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6884                 if (create == 0 && !PageUptodate(page)) {
6885                         if (btrfs_file_extent_compression(leaf, item) !=
6886                             BTRFS_COMPRESS_NONE) {
6887                                 ret = uncompress_inline(path, inode, page,
6888                                                         pg_offset,
6889                                                         extent_offset, item);
6890                                 if (ret) {
6891                                         err = ret;
6892                                         goto out;
6893                                 }
6894                         } else {
6895                                 map = kmap(page);
6896                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6897                                                    copy_size);
6898                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6899                                         memset(map + pg_offset + copy_size, 0,
6900                                                PAGE_CACHE_SIZE - pg_offset -
6901                                                copy_size);
6902                                 }
6903                                 kunmap(page);
6904                         }
6905                         flush_dcache_page(page);
6906                 } else if (create && PageUptodate(page)) {
6907                         BUG();
6908                         if (!trans) {
6909                                 kunmap(page);
6910                                 free_extent_map(em);
6911                                 em = NULL;
6912
6913                                 btrfs_release_path(path);
6914                                 trans = btrfs_join_transaction(root);
6915
6916                                 if (IS_ERR(trans))
6917                                         return ERR_CAST(trans);
6918                                 goto again;
6919                         }
6920                         map = kmap(page);
6921                         write_extent_buffer(leaf, map + pg_offset, ptr,
6922                                             copy_size);
6923                         kunmap(page);
6924                         btrfs_mark_buffer_dirty(leaf);
6925                 }
6926                 set_extent_uptodate(io_tree, em->start,
6927                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6928                 goto insert;
6929         }
6930 not_found:
6931         em->start = start;
6932         em->orig_start = start;
6933         em->len = len;
6934 not_found_em:
6935         em->block_start = EXTENT_MAP_HOLE;
6936         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6937 insert:
6938         btrfs_release_path(path);
6939         if (em->start > start || extent_map_end(em) <= start) {
6940                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6941                         em->start, em->len, start, len);
6942                 err = -EIO;
6943                 goto out;
6944         }
6945
6946         err = 0;
6947         write_lock(&em_tree->lock);
6948         ret = add_extent_mapping(em_tree, em, 0);
6949         /* it is possible that someone inserted the extent into the tree
6950          * while we had the lock dropped.  It is also possible that
6951          * an overlapping map exists in the tree
6952          */
6953         if (ret == -EEXIST) {
6954                 struct extent_map *existing;
6955
6956                 ret = 0;
6957
6958                 existing = search_extent_mapping(em_tree, start, len);
6959                 /*
6960                  * existing will always be non-NULL, since there must be
6961                  * extent causing the -EEXIST.
6962                  */
6963                 if (start >= extent_map_end(existing) ||
6964                     start <= existing->start) {
6965                         /*
6966                          * The existing extent map is the one nearest to
6967                          * the [start, start + len) range which overlaps
6968                          */
6969                         err = merge_extent_mapping(em_tree, existing,
6970                                                    em, start);
6971                         free_extent_map(existing);
6972                         if (err) {
6973                                 free_extent_map(em);
6974                                 em = NULL;
6975                         }
6976                 } else {
6977                         free_extent_map(em);
6978                         em = existing;
6979                         err = 0;
6980                 }
6981         }
6982         write_unlock(&em_tree->lock);
6983 out:
6984
6985         trace_btrfs_get_extent(root, em);
6986
6987         btrfs_free_path(path);
6988         if (trans) {
6989                 ret = btrfs_end_transaction(trans, root);
6990                 if (!err)
6991                         err = ret;
6992         }
6993         if (err) {
6994                 free_extent_map(em);
6995                 return ERR_PTR(err);
6996         }
6997         BUG_ON(!em); /* Error is always set */
6998         return em;
6999 }
7000
7001 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7002                                            size_t pg_offset, u64 start, u64 len,
7003                                            int create)
7004 {
7005         struct extent_map *em;
7006         struct extent_map *hole_em = NULL;
7007         u64 range_start = start;
7008         u64 end;
7009         u64 found;
7010         u64 found_end;
7011         int err = 0;
7012
7013         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7014         if (IS_ERR(em))
7015                 return em;
7016         if (em) {
7017                 /*
7018                  * if our em maps to
7019                  * -  a hole or
7020                  * -  a pre-alloc extent,
7021                  * there might actually be delalloc bytes behind it.
7022                  */
7023                 if (em->block_start != EXTENT_MAP_HOLE &&
7024                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7025                         return em;
7026                 else
7027                         hole_em = em;
7028         }
7029
7030         /* check to see if we've wrapped (len == -1 or similar) */
7031         end = start + len;
7032         if (end < start)
7033                 end = (u64)-1;
7034         else
7035                 end -= 1;
7036
7037         em = NULL;
7038
7039         /* ok, we didn't find anything, lets look for delalloc */
7040         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7041                                  end, len, EXTENT_DELALLOC, 1);
7042         found_end = range_start + found;
7043         if (found_end < range_start)
7044                 found_end = (u64)-1;
7045
7046         /*
7047          * we didn't find anything useful, return
7048          * the original results from get_extent()
7049          */
7050         if (range_start > end || found_end <= start) {
7051                 em = hole_em;
7052                 hole_em = NULL;
7053                 goto out;
7054         }
7055
7056         /* adjust the range_start to make sure it doesn't
7057          * go backwards from the start they passed in
7058          */
7059         range_start = max(start, range_start);
7060         found = found_end - range_start;
7061
7062         if (found > 0) {
7063                 u64 hole_start = start;
7064                 u64 hole_len = len;
7065
7066                 em = alloc_extent_map();
7067                 if (!em) {
7068                         err = -ENOMEM;
7069                         goto out;
7070                 }
7071                 /*
7072                  * when btrfs_get_extent can't find anything it
7073                  * returns one huge hole
7074                  *
7075                  * make sure what it found really fits our range, and
7076                  * adjust to make sure it is based on the start from
7077                  * the caller
7078                  */
7079                 if (hole_em) {
7080                         u64 calc_end = extent_map_end(hole_em);
7081
7082                         if (calc_end <= start || (hole_em->start > end)) {
7083                                 free_extent_map(hole_em);
7084                                 hole_em = NULL;
7085                         } else {
7086                                 hole_start = max(hole_em->start, start);
7087                                 hole_len = calc_end - hole_start;
7088                         }
7089                 }
7090                 em->bdev = NULL;
7091                 if (hole_em && range_start > hole_start) {
7092                         /* our hole starts before our delalloc, so we
7093                          * have to return just the parts of the hole
7094                          * that go until  the delalloc starts
7095                          */
7096                         em->len = min(hole_len,
7097                                       range_start - hole_start);
7098                         em->start = hole_start;
7099                         em->orig_start = hole_start;
7100                         /*
7101                          * don't adjust block start at all,
7102                          * it is fixed at EXTENT_MAP_HOLE
7103                          */
7104                         em->block_start = hole_em->block_start;
7105                         em->block_len = hole_len;
7106                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7107                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7108                 } else {
7109                         em->start = range_start;
7110                         em->len = found;
7111                         em->orig_start = range_start;
7112                         em->block_start = EXTENT_MAP_DELALLOC;
7113                         em->block_len = found;
7114                 }
7115         } else if (hole_em) {
7116                 return hole_em;
7117         }
7118 out:
7119
7120         free_extent_map(hole_em);
7121         if (err) {
7122                 free_extent_map(em);
7123                 return ERR_PTR(err);
7124         }
7125         return em;
7126 }
7127
7128 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7129                                                   u64 start, u64 len)
7130 {
7131         struct btrfs_root *root = BTRFS_I(inode)->root;
7132         struct extent_map *em;
7133         struct btrfs_key ins;
7134         u64 alloc_hint;
7135         int ret;
7136
7137         alloc_hint = get_extent_allocation_hint(inode, start, len);
7138         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7139                                    alloc_hint, &ins, 1, 1);
7140         if (ret)
7141                 return ERR_PTR(ret);
7142
7143         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
7144                               ins.offset, ins.offset, ins.offset, 0);
7145         if (IS_ERR(em)) {
7146                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7147                 return em;
7148         }
7149
7150         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
7151                                            ins.offset, ins.offset, 0);
7152         if (ret) {
7153                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7154                 free_extent_map(em);
7155                 return ERR_PTR(ret);
7156         }
7157
7158         return em;
7159 }
7160
7161 /*
7162  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7163  * block must be cow'd
7164  */
7165 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7166                               u64 *orig_start, u64 *orig_block_len,
7167                               u64 *ram_bytes)
7168 {
7169         struct btrfs_trans_handle *trans;
7170         struct btrfs_path *path;
7171         int ret;
7172         struct extent_buffer *leaf;
7173         struct btrfs_root *root = BTRFS_I(inode)->root;
7174         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7175         struct btrfs_file_extent_item *fi;
7176         struct btrfs_key key;
7177         u64 disk_bytenr;
7178         u64 backref_offset;
7179         u64 extent_end;
7180         u64 num_bytes;
7181         int slot;
7182         int found_type;
7183         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7184
7185         path = btrfs_alloc_path();
7186         if (!path)
7187                 return -ENOMEM;
7188
7189         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7190                                        offset, 0);
7191         if (ret < 0)
7192                 goto out;
7193
7194         slot = path->slots[0];
7195         if (ret == 1) {
7196                 if (slot == 0) {
7197                         /* can't find the item, must cow */
7198                         ret = 0;
7199                         goto out;
7200                 }
7201                 slot--;
7202         }
7203         ret = 0;
7204         leaf = path->nodes[0];
7205         btrfs_item_key_to_cpu(leaf, &key, slot);
7206         if (key.objectid != btrfs_ino(inode) ||
7207             key.type != BTRFS_EXTENT_DATA_KEY) {
7208                 /* not our file or wrong item type, must cow */
7209                 goto out;
7210         }
7211
7212         if (key.offset > offset) {
7213                 /* Wrong offset, must cow */
7214                 goto out;
7215         }
7216
7217         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7218         found_type = btrfs_file_extent_type(leaf, fi);
7219         if (found_type != BTRFS_FILE_EXTENT_REG &&
7220             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7221                 /* not a regular extent, must cow */
7222                 goto out;
7223         }
7224
7225         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7226                 goto out;
7227
7228         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7229         if (extent_end <= offset)
7230                 goto out;
7231
7232         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7233         if (disk_bytenr == 0)
7234                 goto out;
7235
7236         if (btrfs_file_extent_compression(leaf, fi) ||
7237             btrfs_file_extent_encryption(leaf, fi) ||
7238             btrfs_file_extent_other_encoding(leaf, fi))
7239                 goto out;
7240
7241         backref_offset = btrfs_file_extent_offset(leaf, fi);
7242
7243         if (orig_start) {
7244                 *orig_start = key.offset - backref_offset;
7245                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7246                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7247         }
7248
7249         if (btrfs_extent_readonly(root, disk_bytenr))
7250                 goto out;
7251
7252         num_bytes = min(offset + *len, extent_end) - offset;
7253         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7254                 u64 range_end;
7255
7256                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7257                 ret = test_range_bit(io_tree, offset, range_end,
7258                                      EXTENT_DELALLOC, 0, NULL);
7259                 if (ret) {
7260                         ret = -EAGAIN;
7261                         goto out;
7262                 }
7263         }
7264
7265         btrfs_release_path(path);
7266
7267         /*
7268          * look for other files referencing this extent, if we
7269          * find any we must cow
7270          */
7271         trans = btrfs_join_transaction(root);
7272         if (IS_ERR(trans)) {
7273                 ret = 0;
7274                 goto out;
7275         }
7276
7277         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7278                                     key.offset - backref_offset, disk_bytenr);
7279         btrfs_end_transaction(trans, root);
7280         if (ret) {
7281                 ret = 0;
7282                 goto out;
7283         }
7284
7285         /*
7286          * adjust disk_bytenr and num_bytes to cover just the bytes
7287          * in this extent we are about to write.  If there
7288          * are any csums in that range we have to cow in order
7289          * to keep the csums correct
7290          */
7291         disk_bytenr += backref_offset;
7292         disk_bytenr += offset - key.offset;
7293         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7294                                 goto out;
7295         /*
7296          * all of the above have passed, it is safe to overwrite this extent
7297          * without cow
7298          */
7299         *len = num_bytes;
7300         ret = 1;
7301 out:
7302         btrfs_free_path(path);
7303         return ret;
7304 }
7305
7306 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7307 {
7308         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7309         int found = false;
7310         void **pagep = NULL;
7311         struct page *page = NULL;
7312         int start_idx;
7313         int end_idx;
7314
7315         start_idx = start >> PAGE_CACHE_SHIFT;
7316
7317         /*
7318          * end is the last byte in the last page.  end == start is legal
7319          */
7320         end_idx = end >> PAGE_CACHE_SHIFT;
7321
7322         rcu_read_lock();
7323
7324         /* Most of the code in this while loop is lifted from
7325          * find_get_page.  It's been modified to begin searching from a
7326          * page and return just the first page found in that range.  If the
7327          * found idx is less than or equal to the end idx then we know that
7328          * a page exists.  If no pages are found or if those pages are
7329          * outside of the range then we're fine (yay!) */
7330         while (page == NULL &&
7331                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7332                 page = radix_tree_deref_slot(pagep);
7333                 if (unlikely(!page))
7334                         break;
7335
7336                 if (radix_tree_exception(page)) {
7337                         if (radix_tree_deref_retry(page)) {
7338                                 page = NULL;
7339                                 continue;
7340                         }
7341                         /*
7342                          * Otherwise, shmem/tmpfs must be storing a swap entry
7343                          * here as an exceptional entry: so return it without
7344                          * attempting to raise page count.
7345                          */
7346                         page = NULL;
7347                         break; /* TODO: Is this relevant for this use case? */
7348                 }
7349
7350                 if (!page_cache_get_speculative(page)) {
7351                         page = NULL;
7352                         continue;
7353                 }
7354
7355                 /*
7356                  * Has the page moved?
7357                  * This is part of the lockless pagecache protocol. See
7358                  * include/linux/pagemap.h for details.
7359                  */
7360                 if (unlikely(page != *pagep)) {
7361                         page_cache_release(page);
7362                         page = NULL;
7363                 }
7364         }
7365
7366         if (page) {
7367                 if (page->index <= end_idx)
7368                         found = true;
7369                 page_cache_release(page);
7370         }
7371
7372         rcu_read_unlock();
7373         return found;
7374 }
7375
7376 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7377                               struct extent_state **cached_state, int writing)
7378 {
7379         struct btrfs_ordered_extent *ordered;
7380         int ret = 0;
7381
7382         while (1) {
7383                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7384                                  0, cached_state);
7385                 /*
7386                  * We're concerned with the entire range that we're going to be
7387                  * doing DIO to, so we need to make sure theres no ordered
7388                  * extents in this range.
7389                  */
7390                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7391                                                      lockend - lockstart + 1);
7392
7393                 /*
7394                  * We need to make sure there are no buffered pages in this
7395                  * range either, we could have raced between the invalidate in
7396                  * generic_file_direct_write and locking the extent.  The
7397                  * invalidate needs to happen so that reads after a write do not
7398                  * get stale data.
7399                  */
7400                 if (!ordered &&
7401                     (!writing ||
7402                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7403                         break;
7404
7405                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7406                                      cached_state, GFP_NOFS);
7407
7408                 if (ordered) {
7409                         btrfs_start_ordered_extent(inode, ordered, 1);
7410                         btrfs_put_ordered_extent(ordered);
7411                 } else {
7412                         /* Screw you mmap */
7413                         ret = btrfs_fdatawrite_range(inode, lockstart, lockend);
7414                         if (ret)
7415                                 break;
7416                         ret = filemap_fdatawait_range(inode->i_mapping,
7417                                                       lockstart,
7418                                                       lockend);
7419                         if (ret)
7420                                 break;
7421
7422                         /*
7423                          * If we found a page that couldn't be invalidated just
7424                          * fall back to buffered.
7425                          */
7426                         ret = invalidate_inode_pages2_range(inode->i_mapping,
7427                                         lockstart >> PAGE_CACHE_SHIFT,
7428                                         lockend >> PAGE_CACHE_SHIFT);
7429                         if (ret)
7430                                 break;
7431                 }
7432
7433                 cond_resched();
7434         }
7435
7436         return ret;
7437 }
7438
7439 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7440                                            u64 len, u64 orig_start,
7441                                            u64 block_start, u64 block_len,
7442                                            u64 orig_block_len, u64 ram_bytes,
7443                                            int type)
7444 {
7445         struct extent_map_tree *em_tree;
7446         struct extent_map *em;
7447         struct btrfs_root *root = BTRFS_I(inode)->root;
7448         int ret;
7449
7450         em_tree = &BTRFS_I(inode)->extent_tree;
7451         em = alloc_extent_map();
7452         if (!em)
7453                 return ERR_PTR(-ENOMEM);
7454
7455         em->start = start;
7456         em->orig_start = orig_start;
7457         em->mod_start = start;
7458         em->mod_len = len;
7459         em->len = len;
7460         em->block_len = block_len;
7461         em->block_start = block_start;
7462         em->bdev = root->fs_info->fs_devices->latest_bdev;
7463         em->orig_block_len = orig_block_len;
7464         em->ram_bytes = ram_bytes;
7465         em->generation = -1;
7466         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7467         if (type == BTRFS_ORDERED_PREALLOC)
7468                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7469
7470         do {
7471                 btrfs_drop_extent_cache(inode, em->start,
7472                                 em->start + em->len - 1, 0);
7473                 write_lock(&em_tree->lock);
7474                 ret = add_extent_mapping(em_tree, em, 1);
7475                 write_unlock(&em_tree->lock);
7476         } while (ret == -EEXIST);
7477
7478         if (ret) {
7479                 free_extent_map(em);
7480                 return ERR_PTR(ret);
7481         }
7482
7483         return em;
7484 }
7485
7486 struct btrfs_dio_data {
7487         u64 outstanding_extents;
7488         u64 reserve;
7489 };
7490
7491 static void adjust_dio_outstanding_extents(struct inode *inode,
7492                                            struct btrfs_dio_data *dio_data,
7493                                            const u64 len)
7494 {
7495         unsigned num_extents;
7496
7497         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7498                                            BTRFS_MAX_EXTENT_SIZE);
7499         /*
7500          * If we have an outstanding_extents count still set then we're
7501          * within our reservation, otherwise we need to adjust our inode
7502          * counter appropriately.
7503          */
7504         if (dio_data->outstanding_extents) {
7505                 dio_data->outstanding_extents -= num_extents;
7506         } else {
7507                 spin_lock(&BTRFS_I(inode)->lock);
7508                 BTRFS_I(inode)->outstanding_extents += num_extents;
7509                 spin_unlock(&BTRFS_I(inode)->lock);
7510         }
7511 }
7512
7513 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7514                                    struct buffer_head *bh_result, int create)
7515 {
7516         struct extent_map *em;
7517         struct btrfs_root *root = BTRFS_I(inode)->root;
7518         struct extent_state *cached_state = NULL;
7519         struct btrfs_dio_data *dio_data = NULL;
7520         u64 start = iblock << inode->i_blkbits;
7521         u64 lockstart, lockend;
7522         u64 len = bh_result->b_size;
7523         int unlock_bits = EXTENT_LOCKED;
7524         int ret = 0;
7525
7526         if (create)
7527                 unlock_bits |= EXTENT_DIRTY;
7528         else
7529                 len = min_t(u64, len, root->sectorsize);
7530
7531         lockstart = start;
7532         lockend = start + len - 1;
7533
7534         if (current->journal_info) {
7535                 /*
7536                  * Need to pull our outstanding extents and set journal_info to NULL so
7537                  * that anything that needs to check if there's a transction doesn't get
7538                  * confused.
7539                  */
7540                 dio_data = current->journal_info;
7541                 current->journal_info = NULL;
7542         }
7543
7544         /*
7545          * If this errors out it's because we couldn't invalidate pagecache for
7546          * this range and we need to fallback to buffered.
7547          */
7548         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7549                                create)) {
7550                 ret = -ENOTBLK;
7551                 goto err;
7552         }
7553
7554         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7555         if (IS_ERR(em)) {
7556                 ret = PTR_ERR(em);
7557                 goto unlock_err;
7558         }
7559
7560         /*
7561          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7562          * io.  INLINE is special, and we could probably kludge it in here, but
7563          * it's still buffered so for safety lets just fall back to the generic
7564          * buffered path.
7565          *
7566          * For COMPRESSED we _have_ to read the entire extent in so we can
7567          * decompress it, so there will be buffering required no matter what we
7568          * do, so go ahead and fallback to buffered.
7569          *
7570          * We return -ENOTBLK because thats what makes DIO go ahead and go back
7571          * to buffered IO.  Don't blame me, this is the price we pay for using
7572          * the generic code.
7573          */
7574         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7575             em->block_start == EXTENT_MAP_INLINE) {
7576                 free_extent_map(em);
7577                 ret = -ENOTBLK;
7578                 goto unlock_err;
7579         }
7580
7581         /* Just a good old fashioned hole, return */
7582         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7583                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7584                 free_extent_map(em);
7585                 goto unlock_err;
7586         }
7587
7588         /*
7589          * We don't allocate a new extent in the following cases
7590          *
7591          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7592          * existing extent.
7593          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7594          * just use the extent.
7595          *
7596          */
7597         if (!create) {
7598                 len = min(len, em->len - (start - em->start));
7599                 lockstart = start + len;
7600                 goto unlock;
7601         }
7602
7603         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7604             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7605              em->block_start != EXTENT_MAP_HOLE)) {
7606                 int type;
7607                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7608
7609                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7610                         type = BTRFS_ORDERED_PREALLOC;
7611                 else
7612                         type = BTRFS_ORDERED_NOCOW;
7613                 len = min(len, em->len - (start - em->start));
7614                 block_start = em->block_start + (start - em->start);
7615
7616                 if (can_nocow_extent(inode, start, &len, &orig_start,
7617                                      &orig_block_len, &ram_bytes) == 1) {
7618                         if (type == BTRFS_ORDERED_PREALLOC) {
7619                                 free_extent_map(em);
7620                                 em = create_pinned_em(inode, start, len,
7621                                                        orig_start,
7622                                                        block_start, len,
7623                                                        orig_block_len,
7624                                                        ram_bytes, type);
7625                                 if (IS_ERR(em)) {
7626                                         ret = PTR_ERR(em);
7627                                         goto unlock_err;
7628                                 }
7629                         }
7630
7631                         ret = btrfs_add_ordered_extent_dio(inode, start,
7632                                            block_start, len, len, type);
7633                         if (ret) {
7634                                 free_extent_map(em);
7635                                 goto unlock_err;
7636                         }
7637                         goto unlock;
7638                 }
7639         }
7640
7641         /*
7642          * this will cow the extent, reset the len in case we changed
7643          * it above
7644          */
7645         len = bh_result->b_size;
7646         free_extent_map(em);
7647         em = btrfs_new_extent_direct(inode, start, len);
7648         if (IS_ERR(em)) {
7649                 ret = PTR_ERR(em);
7650                 goto unlock_err;
7651         }
7652         len = min(len, em->len - (start - em->start));
7653 unlock:
7654         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7655                 inode->i_blkbits;
7656         bh_result->b_size = len;
7657         bh_result->b_bdev = em->bdev;
7658         set_buffer_mapped(bh_result);
7659         if (create) {
7660                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7661                         set_buffer_new(bh_result);
7662
7663                 /*
7664                  * Need to update the i_size under the extent lock so buffered
7665                  * readers will get the updated i_size when we unlock.
7666                  */
7667                 if (start + len > i_size_read(inode))
7668                         i_size_write(inode, start + len);
7669
7670                 adjust_dio_outstanding_extents(inode, dio_data, len);
7671                 btrfs_free_reserved_data_space(inode, start, len);
7672                 WARN_ON(dio_data->reserve < len);
7673                 dio_data->reserve -= len;
7674                 current->journal_info = dio_data;
7675         }
7676
7677         /*
7678          * In the case of write we need to clear and unlock the entire range,
7679          * in the case of read we need to unlock only the end area that we
7680          * aren't using if there is any left over space.
7681          */
7682         if (lockstart < lockend) {
7683                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7684                                  lockend, unlock_bits, 1, 0,
7685                                  &cached_state, GFP_NOFS);
7686         } else {
7687                 free_extent_state(cached_state);
7688         }
7689
7690         free_extent_map(em);
7691
7692         return 0;
7693
7694 unlock_err:
7695         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7696                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7697 err:
7698         if (dio_data)
7699                 current->journal_info = dio_data;
7700         /*
7701          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7702          * write less data then expected, so that we don't underflow our inode's
7703          * outstanding extents counter.
7704          */
7705         if (create && dio_data)
7706                 adjust_dio_outstanding_extents(inode, dio_data, len);
7707
7708         return ret;
7709 }
7710
7711 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7712                                         int rw, int mirror_num)
7713 {
7714         struct btrfs_root *root = BTRFS_I(inode)->root;
7715         int ret;
7716
7717         BUG_ON(rw & REQ_WRITE);
7718
7719         bio_get(bio);
7720
7721         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7722                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7723         if (ret)
7724                 goto err;
7725
7726         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7727 err:
7728         bio_put(bio);
7729         return ret;
7730 }
7731
7732 static int btrfs_check_dio_repairable(struct inode *inode,
7733                                       struct bio *failed_bio,
7734                                       struct io_failure_record *failrec,
7735                                       int failed_mirror)
7736 {
7737         int num_copies;
7738
7739         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7740                                       failrec->logical, failrec->len);
7741         if (num_copies == 1) {
7742                 /*
7743                  * we only have a single copy of the data, so don't bother with
7744                  * all the retry and error correction code that follows. no
7745                  * matter what the error is, it is very likely to persist.
7746                  */
7747                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7748                          num_copies, failrec->this_mirror, failed_mirror);
7749                 return 0;
7750         }
7751
7752         failrec->failed_mirror = failed_mirror;
7753         failrec->this_mirror++;
7754         if (failrec->this_mirror == failed_mirror)
7755                 failrec->this_mirror++;
7756
7757         if (failrec->this_mirror > num_copies) {
7758                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7759                          num_copies, failrec->this_mirror, failed_mirror);
7760                 return 0;
7761         }
7762
7763         return 1;
7764 }
7765
7766 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7767                           struct page *page, u64 start, u64 end,
7768                           int failed_mirror, bio_end_io_t *repair_endio,
7769                           void *repair_arg)
7770 {
7771         struct io_failure_record *failrec;
7772         struct bio *bio;
7773         int isector;
7774         int read_mode;
7775         int ret;
7776
7777         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7778
7779         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7780         if (ret)
7781                 return ret;
7782
7783         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7784                                          failed_mirror);
7785         if (!ret) {
7786                 free_io_failure(inode, failrec);
7787                 return -EIO;
7788         }
7789
7790         if (failed_bio->bi_vcnt > 1)
7791                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7792         else
7793                 read_mode = READ_SYNC;
7794
7795         isector = start - btrfs_io_bio(failed_bio)->logical;
7796         isector >>= inode->i_sb->s_blocksize_bits;
7797         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7798                                       0, isector, repair_endio, repair_arg);
7799         if (!bio) {
7800                 free_io_failure(inode, failrec);
7801                 return -EIO;
7802         }
7803
7804         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7805                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7806                     read_mode, failrec->this_mirror, failrec->in_validation);
7807
7808         ret = submit_dio_repair_bio(inode, bio, read_mode,
7809                                     failrec->this_mirror);
7810         if (ret) {
7811                 free_io_failure(inode, failrec);
7812                 bio_put(bio);
7813         }
7814
7815         return ret;
7816 }
7817
7818 struct btrfs_retry_complete {
7819         struct completion done;
7820         struct inode *inode;
7821         u64 start;
7822         int uptodate;
7823 };
7824
7825 static void btrfs_retry_endio_nocsum(struct bio *bio)
7826 {
7827         struct btrfs_retry_complete *done = bio->bi_private;
7828         struct bio_vec *bvec;
7829         int i;
7830
7831         if (bio->bi_error)
7832                 goto end;
7833
7834         done->uptodate = 1;
7835         bio_for_each_segment_all(bvec, bio, i)
7836                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7837 end:
7838         complete(&done->done);
7839         bio_put(bio);
7840 }
7841
7842 static int __btrfs_correct_data_nocsum(struct inode *inode,
7843                                        struct btrfs_io_bio *io_bio)
7844 {
7845         struct bio_vec *bvec;
7846         struct btrfs_retry_complete done;
7847         u64 start;
7848         int i;
7849         int ret;
7850
7851         start = io_bio->logical;
7852         done.inode = inode;
7853
7854         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7855 try_again:
7856                 done.uptodate = 0;
7857                 done.start = start;
7858                 init_completion(&done.done);
7859
7860                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7861                                      start + bvec->bv_len - 1,
7862                                      io_bio->mirror_num,
7863                                      btrfs_retry_endio_nocsum, &done);
7864                 if (ret)
7865                         return ret;
7866
7867                 wait_for_completion(&done.done);
7868
7869                 if (!done.uptodate) {
7870                         /* We might have another mirror, so try again */
7871                         goto try_again;
7872                 }
7873
7874                 start += bvec->bv_len;
7875         }
7876
7877         return 0;
7878 }
7879
7880 static void btrfs_retry_endio(struct bio *bio)
7881 {
7882         struct btrfs_retry_complete *done = bio->bi_private;
7883         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7884         struct bio_vec *bvec;
7885         int uptodate;
7886         int ret;
7887         int i;
7888
7889         if (bio->bi_error)
7890                 goto end;
7891
7892         uptodate = 1;
7893         bio_for_each_segment_all(bvec, bio, i) {
7894                 ret = __readpage_endio_check(done->inode, io_bio, i,
7895                                              bvec->bv_page, 0,
7896                                              done->start, bvec->bv_len);
7897                 if (!ret)
7898                         clean_io_failure(done->inode, done->start,
7899                                          bvec->bv_page, 0);
7900                 else
7901                         uptodate = 0;
7902         }
7903
7904         done->uptodate = uptodate;
7905 end:
7906         complete(&done->done);
7907         bio_put(bio);
7908 }
7909
7910 static int __btrfs_subio_endio_read(struct inode *inode,
7911                                     struct btrfs_io_bio *io_bio, int err)
7912 {
7913         struct bio_vec *bvec;
7914         struct btrfs_retry_complete done;
7915         u64 start;
7916         u64 offset = 0;
7917         int i;
7918         int ret;
7919
7920         err = 0;
7921         start = io_bio->logical;
7922         done.inode = inode;
7923
7924         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7925                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7926                                              0, start, bvec->bv_len);
7927                 if (likely(!ret))
7928                         goto next;
7929 try_again:
7930                 done.uptodate = 0;
7931                 done.start = start;
7932                 init_completion(&done.done);
7933
7934                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7935                                      start + bvec->bv_len - 1,
7936                                      io_bio->mirror_num,
7937                                      btrfs_retry_endio, &done);
7938                 if (ret) {
7939                         err = ret;
7940                         goto next;
7941                 }
7942
7943                 wait_for_completion(&done.done);
7944
7945                 if (!done.uptodate) {
7946                         /* We might have another mirror, so try again */
7947                         goto try_again;
7948                 }
7949 next:
7950                 offset += bvec->bv_len;
7951                 start += bvec->bv_len;
7952         }
7953
7954         return err;
7955 }
7956
7957 static int btrfs_subio_endio_read(struct inode *inode,
7958                                   struct btrfs_io_bio *io_bio, int err)
7959 {
7960         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7961
7962         if (skip_csum) {
7963                 if (unlikely(err))
7964                         return __btrfs_correct_data_nocsum(inode, io_bio);
7965                 else
7966                         return 0;
7967         } else {
7968                 return __btrfs_subio_endio_read(inode, io_bio, err);
7969         }
7970 }
7971
7972 static void btrfs_endio_direct_read(struct bio *bio)
7973 {
7974         struct btrfs_dio_private *dip = bio->bi_private;
7975         struct inode *inode = dip->inode;
7976         struct bio *dio_bio;
7977         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7978         int err = bio->bi_error;
7979
7980         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
7981                 err = btrfs_subio_endio_read(inode, io_bio, err);
7982
7983         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
7984                       dip->logical_offset + dip->bytes - 1);
7985         dio_bio = dip->dio_bio;
7986
7987         kfree(dip);
7988
7989         dio_end_io(dio_bio, bio->bi_error);
7990
7991         if (io_bio->end_io)
7992                 io_bio->end_io(io_bio, err);
7993         bio_put(bio);
7994 }
7995
7996 static void btrfs_endio_direct_write(struct bio *bio)
7997 {
7998         struct btrfs_dio_private *dip = bio->bi_private;
7999         struct inode *inode = dip->inode;
8000         struct btrfs_root *root = BTRFS_I(inode)->root;
8001         struct btrfs_ordered_extent *ordered = NULL;
8002         u64 ordered_offset = dip->logical_offset;
8003         u64 ordered_bytes = dip->bytes;
8004         struct bio *dio_bio;
8005         int ret;
8006
8007 again:
8008         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8009                                                    &ordered_offset,
8010                                                    ordered_bytes,
8011                                                    !bio->bi_error);
8012         if (!ret)
8013                 goto out_test;
8014
8015         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8016                         finish_ordered_fn, NULL, NULL);
8017         btrfs_queue_work(root->fs_info->endio_write_workers,
8018                          &ordered->work);
8019 out_test:
8020         /*
8021          * our bio might span multiple ordered extents.  If we haven't
8022          * completed the accounting for the whole dio, go back and try again
8023          */
8024         if (ordered_offset < dip->logical_offset + dip->bytes) {
8025                 ordered_bytes = dip->logical_offset + dip->bytes -
8026                         ordered_offset;
8027                 ordered = NULL;
8028                 goto again;
8029         }
8030         dio_bio = dip->dio_bio;
8031
8032         kfree(dip);
8033
8034         dio_end_io(dio_bio, bio->bi_error);
8035         bio_put(bio);
8036 }
8037
8038 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8039                                     struct bio *bio, int mirror_num,
8040                                     unsigned long bio_flags, u64 offset)
8041 {
8042         int ret;
8043         struct btrfs_root *root = BTRFS_I(inode)->root;
8044         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8045         BUG_ON(ret); /* -ENOMEM */
8046         return 0;
8047 }
8048
8049 static void btrfs_end_dio_bio(struct bio *bio)
8050 {
8051         struct btrfs_dio_private *dip = bio->bi_private;
8052         int err = bio->bi_error;
8053
8054         if (err)
8055                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8056                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8057                            btrfs_ino(dip->inode), bio->bi_rw,
8058                            (unsigned long long)bio->bi_iter.bi_sector,
8059                            bio->bi_iter.bi_size, err);
8060
8061         if (dip->subio_endio)
8062                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8063
8064         if (err) {
8065                 dip->errors = 1;
8066
8067                 /*
8068                  * before atomic variable goto zero, we must make sure
8069                  * dip->errors is perceived to be set.
8070                  */
8071                 smp_mb__before_atomic();
8072         }
8073
8074         /* if there are more bios still pending for this dio, just exit */
8075         if (!atomic_dec_and_test(&dip->pending_bios))
8076                 goto out;
8077
8078         if (dip->errors) {
8079                 bio_io_error(dip->orig_bio);
8080         } else {
8081                 dip->dio_bio->bi_error = 0;
8082                 bio_endio(dip->orig_bio);
8083         }
8084 out:
8085         bio_put(bio);
8086 }
8087
8088 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8089                                        u64 first_sector, gfp_t gfp_flags)
8090 {
8091         struct bio *bio;
8092         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8093         if (bio)
8094                 bio_associate_current(bio);
8095         return bio;
8096 }
8097
8098 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8099                                                  struct inode *inode,
8100                                                  struct btrfs_dio_private *dip,
8101                                                  struct bio *bio,
8102                                                  u64 file_offset)
8103 {
8104         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8105         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8106         int ret;
8107
8108         /*
8109          * We load all the csum data we need when we submit
8110          * the first bio to reduce the csum tree search and
8111          * contention.
8112          */
8113         if (dip->logical_offset == file_offset) {
8114                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8115                                                 file_offset);
8116                 if (ret)
8117                         return ret;
8118         }
8119
8120         if (bio == dip->orig_bio)
8121                 return 0;
8122
8123         file_offset -= dip->logical_offset;
8124         file_offset >>= inode->i_sb->s_blocksize_bits;
8125         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8126
8127         return 0;
8128 }
8129
8130 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8131                                          int rw, u64 file_offset, int skip_sum,
8132                                          int async_submit)
8133 {
8134         struct btrfs_dio_private *dip = bio->bi_private;
8135         int write = rw & REQ_WRITE;
8136         struct btrfs_root *root = BTRFS_I(inode)->root;
8137         int ret;
8138
8139         if (async_submit)
8140                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8141
8142         bio_get(bio);
8143
8144         if (!write) {
8145                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8146                                 BTRFS_WQ_ENDIO_DATA);
8147                 if (ret)
8148                         goto err;
8149         }
8150
8151         if (skip_sum)
8152                 goto map;
8153
8154         if (write && async_submit) {
8155                 ret = btrfs_wq_submit_bio(root->fs_info,
8156                                    inode, rw, bio, 0, 0,
8157                                    file_offset,
8158                                    __btrfs_submit_bio_start_direct_io,
8159                                    __btrfs_submit_bio_done);
8160                 goto err;
8161         } else if (write) {
8162                 /*
8163                  * If we aren't doing async submit, calculate the csum of the
8164                  * bio now.
8165                  */
8166                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8167                 if (ret)
8168                         goto err;
8169         } else {
8170                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8171                                                      file_offset);
8172                 if (ret)
8173                         goto err;
8174         }
8175 map:
8176         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8177 err:
8178         bio_put(bio);
8179         return ret;
8180 }
8181
8182 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8183                                     int skip_sum)
8184 {
8185         struct inode *inode = dip->inode;
8186         struct btrfs_root *root = BTRFS_I(inode)->root;
8187         struct bio *bio;
8188         struct bio *orig_bio = dip->orig_bio;
8189         struct bio_vec *bvec = orig_bio->bi_io_vec;
8190         u64 start_sector = orig_bio->bi_iter.bi_sector;
8191         u64 file_offset = dip->logical_offset;
8192         u64 submit_len = 0;
8193         u64 map_length;
8194         int nr_pages = 0;
8195         int ret;
8196         int async_submit = 0;
8197
8198         map_length = orig_bio->bi_iter.bi_size;
8199         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8200                               &map_length, NULL, 0);
8201         if (ret)
8202                 return -EIO;
8203
8204         if (map_length >= orig_bio->bi_iter.bi_size) {
8205                 bio = orig_bio;
8206                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8207                 goto submit;
8208         }
8209
8210         /* async crcs make it difficult to collect full stripe writes. */
8211         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8212                 async_submit = 0;
8213         else
8214                 async_submit = 1;
8215
8216         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8217         if (!bio)
8218                 return -ENOMEM;
8219
8220         bio->bi_private = dip;
8221         bio->bi_end_io = btrfs_end_dio_bio;
8222         btrfs_io_bio(bio)->logical = file_offset;
8223         atomic_inc(&dip->pending_bios);
8224
8225         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8226                 if (map_length < submit_len + bvec->bv_len ||
8227                     bio_add_page(bio, bvec->bv_page, bvec->bv_len,
8228                                  bvec->bv_offset) < bvec->bv_len) {
8229                         /*
8230                          * inc the count before we submit the bio so
8231                          * we know the end IO handler won't happen before
8232                          * we inc the count. Otherwise, the dip might get freed
8233                          * before we're done setting it up
8234                          */
8235                         atomic_inc(&dip->pending_bios);
8236                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8237                                                      file_offset, skip_sum,
8238                                                      async_submit);
8239                         if (ret) {
8240                                 bio_put(bio);
8241                                 atomic_dec(&dip->pending_bios);
8242                                 goto out_err;
8243                         }
8244
8245                         start_sector += submit_len >> 9;
8246                         file_offset += submit_len;
8247
8248                         submit_len = 0;
8249                         nr_pages = 0;
8250
8251                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8252                                                   start_sector, GFP_NOFS);
8253                         if (!bio)
8254                                 goto out_err;
8255                         bio->bi_private = dip;
8256                         bio->bi_end_io = btrfs_end_dio_bio;
8257                         btrfs_io_bio(bio)->logical = file_offset;
8258
8259                         map_length = orig_bio->bi_iter.bi_size;
8260                         ret = btrfs_map_block(root->fs_info, rw,
8261                                               start_sector << 9,
8262                                               &map_length, NULL, 0);
8263                         if (ret) {
8264                                 bio_put(bio);
8265                                 goto out_err;
8266                         }
8267                 } else {
8268                         submit_len += bvec->bv_len;
8269                         nr_pages++;
8270                         bvec++;
8271                 }
8272         }
8273
8274 submit:
8275         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8276                                      async_submit);
8277         if (!ret)
8278                 return 0;
8279
8280         bio_put(bio);
8281 out_err:
8282         dip->errors = 1;
8283         /*
8284          * before atomic variable goto zero, we must
8285          * make sure dip->errors is perceived to be set.
8286          */
8287         smp_mb__before_atomic();
8288         if (atomic_dec_and_test(&dip->pending_bios))
8289                 bio_io_error(dip->orig_bio);
8290
8291         /* bio_end_io() will handle error, so we needn't return it */
8292         return 0;
8293 }
8294
8295 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8296                                 struct inode *inode, loff_t file_offset)
8297 {
8298         struct btrfs_dio_private *dip = NULL;
8299         struct bio *io_bio = NULL;
8300         struct btrfs_io_bio *btrfs_bio;
8301         int skip_sum;
8302         int write = rw & REQ_WRITE;
8303         int ret = 0;
8304
8305         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8306
8307         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8308         if (!io_bio) {
8309                 ret = -ENOMEM;
8310                 goto free_ordered;
8311         }
8312
8313         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8314         if (!dip) {
8315                 ret = -ENOMEM;
8316                 goto free_ordered;
8317         }
8318
8319         dip->private = dio_bio->bi_private;
8320         dip->inode = inode;
8321         dip->logical_offset = file_offset;
8322         dip->bytes = dio_bio->bi_iter.bi_size;
8323         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8324         io_bio->bi_private = dip;
8325         dip->orig_bio = io_bio;
8326         dip->dio_bio = dio_bio;
8327         atomic_set(&dip->pending_bios, 0);
8328         btrfs_bio = btrfs_io_bio(io_bio);
8329         btrfs_bio->logical = file_offset;
8330
8331         if (write) {
8332                 io_bio->bi_end_io = btrfs_endio_direct_write;
8333         } else {
8334                 io_bio->bi_end_io = btrfs_endio_direct_read;
8335                 dip->subio_endio = btrfs_subio_endio_read;
8336         }
8337
8338         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8339         if (!ret)
8340                 return;
8341
8342         if (btrfs_bio->end_io)
8343                 btrfs_bio->end_io(btrfs_bio, ret);
8344
8345 free_ordered:
8346         /*
8347          * If we arrived here it means either we failed to submit the dip
8348          * or we either failed to clone the dio_bio or failed to allocate the
8349          * dip. If we cloned the dio_bio and allocated the dip, we can just
8350          * call bio_endio against our io_bio so that we get proper resource
8351          * cleanup if we fail to submit the dip, otherwise, we must do the
8352          * same as btrfs_endio_direct_[write|read] because we can't call these
8353          * callbacks - they require an allocated dip and a clone of dio_bio.
8354          */
8355         if (io_bio && dip) {
8356                 io_bio->bi_error = -EIO;
8357                 bio_endio(io_bio);
8358                 /*
8359                  * The end io callbacks free our dip, do the final put on io_bio
8360                  * and all the cleanup and final put for dio_bio (through
8361                  * dio_end_io()).
8362                  */
8363                 dip = NULL;
8364                 io_bio = NULL;
8365         } else {
8366                 if (write) {
8367                         struct btrfs_ordered_extent *ordered;
8368
8369                         ordered = btrfs_lookup_ordered_extent(inode,
8370                                                               file_offset);
8371                         set_bit(BTRFS_ORDERED_IOERR, &ordered->flags);
8372                         /*
8373                          * Decrements our ref on the ordered extent and removes
8374                          * the ordered extent from the inode's ordered tree,
8375                          * doing all the proper resource cleanup such as for the
8376                          * reserved space and waking up any waiters for this
8377                          * ordered extent (through btrfs_remove_ordered_extent).
8378                          */
8379                         btrfs_finish_ordered_io(ordered);
8380                 } else {
8381                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8382                               file_offset + dio_bio->bi_iter.bi_size - 1);
8383                 }
8384                 dio_bio->bi_error = -EIO;
8385                 /*
8386                  * Releases and cleans up our dio_bio, no need to bio_put()
8387                  * nor bio_endio()/bio_io_error() against dio_bio.
8388                  */
8389                 dio_end_io(dio_bio, ret);
8390         }
8391         if (io_bio)
8392                 bio_put(io_bio);
8393         kfree(dip);
8394 }
8395
8396 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8397                         const struct iov_iter *iter, loff_t offset)
8398 {
8399         int seg;
8400         int i;
8401         unsigned blocksize_mask = root->sectorsize - 1;
8402         ssize_t retval = -EINVAL;
8403
8404         if (offset & blocksize_mask)
8405                 goto out;
8406
8407         if (iov_iter_alignment(iter) & blocksize_mask)
8408                 goto out;
8409
8410         /* If this is a write we don't need to check anymore */
8411         if (iov_iter_rw(iter) == WRITE)
8412                 return 0;
8413         /*
8414          * Check to make sure we don't have duplicate iov_base's in this
8415          * iovec, if so return EINVAL, otherwise we'll get csum errors
8416          * when reading back.
8417          */
8418         for (seg = 0; seg < iter->nr_segs; seg++) {
8419                 for (i = seg + 1; i < iter->nr_segs; i++) {
8420                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8421                                 goto out;
8422                 }
8423         }
8424         retval = 0;
8425 out:
8426         return retval;
8427 }
8428
8429 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
8430                                loff_t offset)
8431 {
8432         struct file *file = iocb->ki_filp;
8433         struct inode *inode = file->f_mapping->host;
8434         struct btrfs_root *root = BTRFS_I(inode)->root;
8435         struct btrfs_dio_data dio_data = { 0 };
8436         size_t count = 0;
8437         int flags = 0;
8438         bool wakeup = true;
8439         bool relock = false;
8440         ssize_t ret;
8441
8442         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8443                 return 0;
8444
8445         inode_dio_begin(inode);
8446         smp_mb__after_atomic();
8447
8448         /*
8449          * The generic stuff only does filemap_write_and_wait_range, which
8450          * isn't enough if we've written compressed pages to this area, so
8451          * we need to flush the dirty pages again to make absolutely sure
8452          * that any outstanding dirty pages are on disk.
8453          */
8454         count = iov_iter_count(iter);
8455         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8456                      &BTRFS_I(inode)->runtime_flags))
8457                 filemap_fdatawrite_range(inode->i_mapping, offset,
8458                                          offset + count - 1);
8459
8460         if (iov_iter_rw(iter) == WRITE) {
8461                 /*
8462                  * If the write DIO is beyond the EOF, we need update
8463                  * the isize, but it is protected by i_mutex. So we can
8464                  * not unlock the i_mutex at this case.
8465                  */
8466                 if (offset + count <= inode->i_size) {
8467                         mutex_unlock(&inode->i_mutex);
8468                         relock = true;
8469                 }
8470                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8471                 if (ret)
8472                         goto out;
8473                 dio_data.outstanding_extents = div64_u64(count +
8474                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8475                                                 BTRFS_MAX_EXTENT_SIZE);
8476
8477                 /*
8478                  * We need to know how many extents we reserved so that we can
8479                  * do the accounting properly if we go over the number we
8480                  * originally calculated.  Abuse current->journal_info for this.
8481                  */
8482                 dio_data.reserve = round_up(count, root->sectorsize);
8483                 current->journal_info = &dio_data;
8484         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8485                                      &BTRFS_I(inode)->runtime_flags)) {
8486                 inode_dio_end(inode);
8487                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8488                 wakeup = false;
8489         }
8490
8491         ret = __blockdev_direct_IO(iocb, inode,
8492                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8493                                    iter, offset, btrfs_get_blocks_direct, NULL,
8494                                    btrfs_submit_direct, flags);
8495         if (iov_iter_rw(iter) == WRITE) {
8496                 current->journal_info = NULL;
8497                 if (ret < 0 && ret != -EIOCBQUEUED) {
8498                         if (dio_data.reserve)
8499                                 btrfs_delalloc_release_space(inode, offset,
8500                                                              dio_data.reserve);
8501                 } else if (ret >= 0 && (size_t)ret < count)
8502                         btrfs_delalloc_release_space(inode, offset,
8503                                                      count - (size_t)ret);
8504         }
8505 out:
8506         if (wakeup)
8507                 inode_dio_end(inode);
8508         if (relock)
8509                 mutex_lock(&inode->i_mutex);
8510
8511         return ret;
8512 }
8513
8514 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8515
8516 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8517                 __u64 start, __u64 len)
8518 {
8519         int     ret;
8520
8521         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8522         if (ret)
8523                 return ret;
8524
8525         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8526 }
8527
8528 int btrfs_readpage(struct file *file, struct page *page)
8529 {
8530         struct extent_io_tree *tree;
8531         tree = &BTRFS_I(page->mapping->host)->io_tree;
8532         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8533 }
8534
8535 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8536 {
8537         struct extent_io_tree *tree;
8538
8539
8540         if (current->flags & PF_MEMALLOC) {
8541                 redirty_page_for_writepage(wbc, page);
8542                 unlock_page(page);
8543                 return 0;
8544         }
8545         tree = &BTRFS_I(page->mapping->host)->io_tree;
8546         return extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8547 }
8548
8549 static int btrfs_writepages(struct address_space *mapping,
8550                             struct writeback_control *wbc)
8551 {
8552         struct extent_io_tree *tree;
8553
8554         tree = &BTRFS_I(mapping->host)->io_tree;
8555         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8556 }
8557
8558 static int
8559 btrfs_readpages(struct file *file, struct address_space *mapping,
8560                 struct list_head *pages, unsigned nr_pages)
8561 {
8562         struct extent_io_tree *tree;
8563         tree = &BTRFS_I(mapping->host)->io_tree;
8564         return extent_readpages(tree, mapping, pages, nr_pages,
8565                                 btrfs_get_extent);
8566 }
8567 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8568 {
8569         struct extent_io_tree *tree;
8570         struct extent_map_tree *map;
8571         int ret;
8572
8573         tree = &BTRFS_I(page->mapping->host)->io_tree;
8574         map = &BTRFS_I(page->mapping->host)->extent_tree;
8575         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8576         if (ret == 1) {
8577                 ClearPagePrivate(page);
8578                 set_page_private(page, 0);
8579                 page_cache_release(page);
8580         }
8581         return ret;
8582 }
8583
8584 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8585 {
8586         if (PageWriteback(page) || PageDirty(page))
8587                 return 0;
8588         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8589 }
8590
8591 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8592                                  unsigned int length)
8593 {
8594         struct inode *inode = page->mapping->host;
8595         struct extent_io_tree *tree;
8596         struct btrfs_ordered_extent *ordered;
8597         struct extent_state *cached_state = NULL;
8598         u64 page_start = page_offset(page);
8599         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
8600         int inode_evicting = inode->i_state & I_FREEING;
8601
8602         /*
8603          * we have the page locked, so new writeback can't start,
8604          * and the dirty bit won't be cleared while we are here.
8605          *
8606          * Wait for IO on this page so that we can safely clear
8607          * the PagePrivate2 bit and do ordered accounting
8608          */
8609         wait_on_page_writeback(page);
8610
8611         tree = &BTRFS_I(inode)->io_tree;
8612         if (offset) {
8613                 btrfs_releasepage(page, GFP_NOFS);
8614                 return;
8615         }
8616
8617         if (!inode_evicting)
8618                 lock_extent_bits(tree, page_start, page_end, 0, &cached_state);
8619         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8620         if (ordered) {
8621                 /*
8622                  * IO on this page will never be started, so we need
8623                  * to account for any ordered extents now
8624                  */
8625                 if (!inode_evicting)
8626                         clear_extent_bit(tree, page_start, page_end,
8627                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8628                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8629                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8630                                          GFP_NOFS);
8631                 /*
8632                  * whoever cleared the private bit is responsible
8633                  * for the finish_ordered_io
8634                  */
8635                 if (TestClearPagePrivate2(page)) {
8636                         struct btrfs_ordered_inode_tree *tree;
8637                         u64 new_len;
8638
8639                         tree = &BTRFS_I(inode)->ordered_tree;
8640
8641                         spin_lock_irq(&tree->lock);
8642                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8643                         new_len = page_start - ordered->file_offset;
8644                         if (new_len < ordered->truncated_len)
8645                                 ordered->truncated_len = new_len;
8646                         spin_unlock_irq(&tree->lock);
8647
8648                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8649                                                            page_start,
8650                                                            PAGE_CACHE_SIZE, 1))
8651                                 btrfs_finish_ordered_io(ordered);
8652                 }
8653                 btrfs_put_ordered_extent(ordered);
8654                 if (!inode_evicting) {
8655                         cached_state = NULL;
8656                         lock_extent_bits(tree, page_start, page_end, 0,
8657                                          &cached_state);
8658                 }
8659         }
8660
8661         /*
8662          * Qgroup reserved space handler
8663          * Page here will be either
8664          * 1) Already written to disk
8665          *    In this case, its reserved space is released from data rsv map
8666          *    and will be freed by delayed_ref handler finally.
8667          *    So even we call qgroup_free_data(), it won't decrease reserved
8668          *    space.
8669          * 2) Not written to disk
8670          *    This means the reserved space should be freed here.
8671          */
8672         btrfs_qgroup_free_data(inode, page_start, PAGE_CACHE_SIZE);
8673         if (!inode_evicting) {
8674                 clear_extent_bit(tree, page_start, page_end,
8675                                  EXTENT_LOCKED | EXTENT_DIRTY |
8676                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8677                                  EXTENT_DEFRAG, 1, 1,
8678                                  &cached_state, GFP_NOFS);
8679
8680                 __btrfs_releasepage(page, GFP_NOFS);
8681         }
8682
8683         ClearPageChecked(page);
8684         if (PagePrivate(page)) {
8685                 ClearPagePrivate(page);
8686                 set_page_private(page, 0);
8687                 page_cache_release(page);
8688         }
8689 }
8690
8691 /*
8692  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8693  * called from a page fault handler when a page is first dirtied. Hence we must
8694  * be careful to check for EOF conditions here. We set the page up correctly
8695  * for a written page which means we get ENOSPC checking when writing into
8696  * holes and correct delalloc and unwritten extent mapping on filesystems that
8697  * support these features.
8698  *
8699  * We are not allowed to take the i_mutex here so we have to play games to
8700  * protect against truncate races as the page could now be beyond EOF.  Because
8701  * vmtruncate() writes the inode size before removing pages, once we have the
8702  * page lock we can determine safely if the page is beyond EOF. If it is not
8703  * beyond EOF, then the page is guaranteed safe against truncation until we
8704  * unlock the page.
8705  */
8706 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8707 {
8708         struct page *page = vmf->page;
8709         struct inode *inode = file_inode(vma->vm_file);
8710         struct btrfs_root *root = BTRFS_I(inode)->root;
8711         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8712         struct btrfs_ordered_extent *ordered;
8713         struct extent_state *cached_state = NULL;
8714         char *kaddr;
8715         unsigned long zero_start;
8716         loff_t size;
8717         int ret;
8718         int reserved = 0;
8719         u64 page_start;
8720         u64 page_end;
8721
8722         sb_start_pagefault(inode->i_sb);
8723         page_start = page_offset(page);
8724         page_end = page_start + PAGE_CACHE_SIZE - 1;
8725
8726         ret = btrfs_delalloc_reserve_space(inode, page_start,
8727                                            PAGE_CACHE_SIZE);
8728         if (!ret) {
8729                 ret = file_update_time(vma->vm_file);
8730                 reserved = 1;
8731         }
8732         if (ret) {
8733                 if (ret == -ENOMEM)
8734                         ret = VM_FAULT_OOM;
8735                 else /* -ENOSPC, -EIO, etc */
8736                         ret = VM_FAULT_SIGBUS;
8737                 if (reserved)
8738                         goto out;
8739                 goto out_noreserve;
8740         }
8741
8742         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8743 again:
8744         lock_page(page);
8745         size = i_size_read(inode);
8746
8747         if ((page->mapping != inode->i_mapping) ||
8748             (page_start >= size)) {
8749                 /* page got truncated out from underneath us */
8750                 goto out_unlock;
8751         }
8752         wait_on_page_writeback(page);
8753
8754         lock_extent_bits(io_tree, page_start, page_end, 0, &cached_state);
8755         set_page_extent_mapped(page);
8756
8757         /*
8758          * we can't set the delalloc bits if there are pending ordered
8759          * extents.  Drop our locks and wait for them to finish
8760          */
8761         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8762         if (ordered) {
8763                 unlock_extent_cached(io_tree, page_start, page_end,
8764                                      &cached_state, GFP_NOFS);
8765                 unlock_page(page);
8766                 btrfs_start_ordered_extent(inode, ordered, 1);
8767                 btrfs_put_ordered_extent(ordered);
8768                 goto again;
8769         }
8770
8771         /*
8772          * XXX - page_mkwrite gets called every time the page is dirtied, even
8773          * if it was already dirty, so for space accounting reasons we need to
8774          * clear any delalloc bits for the range we are fixing to save.  There
8775          * is probably a better way to do this, but for now keep consistent with
8776          * prepare_pages in the normal write path.
8777          */
8778         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
8779                           EXTENT_DIRTY | EXTENT_DELALLOC |
8780                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8781                           0, 0, &cached_state, GFP_NOFS);
8782
8783         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
8784                                         &cached_state);
8785         if (ret) {
8786                 unlock_extent_cached(io_tree, page_start, page_end,
8787                                      &cached_state, GFP_NOFS);
8788                 ret = VM_FAULT_SIGBUS;
8789                 goto out_unlock;
8790         }
8791         ret = 0;
8792
8793         /* page is wholly or partially inside EOF */
8794         if (page_start + PAGE_CACHE_SIZE > size)
8795                 zero_start = size & ~PAGE_CACHE_MASK;
8796         else
8797                 zero_start = PAGE_CACHE_SIZE;
8798
8799         if (zero_start != PAGE_CACHE_SIZE) {
8800                 kaddr = kmap(page);
8801                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
8802                 flush_dcache_page(page);
8803                 kunmap(page);
8804         }
8805         ClearPageChecked(page);
8806         set_page_dirty(page);
8807         SetPageUptodate(page);
8808
8809         BTRFS_I(inode)->last_trans = root->fs_info->generation;
8810         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8811         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8812
8813         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
8814
8815 out_unlock:
8816         if (!ret) {
8817                 sb_end_pagefault(inode->i_sb);
8818                 return VM_FAULT_LOCKED;
8819         }
8820         unlock_page(page);
8821 out:
8822         btrfs_delalloc_release_space(inode, page_start, PAGE_CACHE_SIZE);
8823 out_noreserve:
8824         sb_end_pagefault(inode->i_sb);
8825         return ret;
8826 }
8827
8828 static int btrfs_truncate(struct inode *inode)
8829 {
8830         struct btrfs_root *root = BTRFS_I(inode)->root;
8831         struct btrfs_block_rsv *rsv;
8832         int ret = 0;
8833         int err = 0;
8834         struct btrfs_trans_handle *trans;
8835         u64 mask = root->sectorsize - 1;
8836         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
8837
8838         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8839                                        (u64)-1);
8840         if (ret)
8841                 return ret;
8842
8843         /*
8844          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
8845          * 3 things going on here
8846          *
8847          * 1) We need to reserve space for our orphan item and the space to
8848          * delete our orphan item.  Lord knows we don't want to have a dangling
8849          * orphan item because we didn't reserve space to remove it.
8850          *
8851          * 2) We need to reserve space to update our inode.
8852          *
8853          * 3) We need to have something to cache all the space that is going to
8854          * be free'd up by the truncate operation, but also have some slack
8855          * space reserved in case it uses space during the truncate (thank you
8856          * very much snapshotting).
8857          *
8858          * And we need these to all be seperate.  The fact is we can use alot of
8859          * space doing the truncate, and we have no earthly idea how much space
8860          * we will use, so we need the truncate reservation to be seperate so it
8861          * doesn't end up using space reserved for updating the inode or
8862          * removing the orphan item.  We also need to be able to stop the
8863          * transaction and start a new one, which means we need to be able to
8864          * update the inode several times, and we have no idea of knowing how
8865          * many times that will be, so we can't just reserve 1 item for the
8866          * entirety of the opration, so that has to be done seperately as well.
8867          * Then there is the orphan item, which does indeed need to be held on
8868          * to for the whole operation, and we need nobody to touch this reserved
8869          * space except the orphan code.
8870          *
8871          * So that leaves us with
8872          *
8873          * 1) root->orphan_block_rsv - for the orphan deletion.
8874          * 2) rsv - for the truncate reservation, which we will steal from the
8875          * transaction reservation.
8876          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
8877          * updating the inode.
8878          */
8879         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
8880         if (!rsv)
8881                 return -ENOMEM;
8882         rsv->size = min_size;
8883         rsv->failfast = 1;
8884
8885         /*
8886          * 1 for the truncate slack space
8887          * 1 for updating the inode.
8888          */
8889         trans = btrfs_start_transaction(root, 2);
8890         if (IS_ERR(trans)) {
8891                 err = PTR_ERR(trans);
8892                 goto out;
8893         }
8894
8895         /* Migrate the slack space for the truncate to our reserve */
8896         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
8897                                       min_size);
8898         BUG_ON(ret);
8899
8900         /*
8901          * So if we truncate and then write and fsync we normally would just
8902          * write the extents that changed, which is a problem if we need to
8903          * first truncate that entire inode.  So set this flag so we write out
8904          * all of the extents in the inode to the sync log so we're completely
8905          * safe.
8906          */
8907         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8908         trans->block_rsv = rsv;
8909
8910         while (1) {
8911                 ret = btrfs_truncate_inode_items(trans, root, inode,
8912                                                  inode->i_size,
8913                                                  BTRFS_EXTENT_DATA_KEY);
8914                 if (ret != -ENOSPC && ret != -EAGAIN) {
8915                         err = ret;
8916                         break;
8917                 }
8918
8919                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8920                 ret = btrfs_update_inode(trans, root, inode);
8921                 if (ret) {
8922                         err = ret;
8923                         break;
8924                 }
8925
8926                 btrfs_end_transaction(trans, root);
8927                 btrfs_btree_balance_dirty(root);
8928
8929                 trans = btrfs_start_transaction(root, 2);
8930                 if (IS_ERR(trans)) {
8931                         ret = err = PTR_ERR(trans);
8932                         trans = NULL;
8933                         break;
8934                 }
8935
8936                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
8937                                               rsv, min_size);
8938                 BUG_ON(ret);    /* shouldn't happen */
8939                 trans->block_rsv = rsv;
8940         }
8941
8942         if (ret == 0 && inode->i_nlink > 0) {
8943                 trans->block_rsv = root->orphan_block_rsv;
8944                 ret = btrfs_orphan_del(trans, inode);
8945                 if (ret)
8946                         err = ret;
8947         }
8948
8949         if (trans) {
8950                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8951                 ret = btrfs_update_inode(trans, root, inode);
8952                 if (ret && !err)
8953                         err = ret;
8954
8955                 ret = btrfs_end_transaction(trans, root);
8956                 btrfs_btree_balance_dirty(root);
8957         }
8958
8959 out:
8960         btrfs_free_block_rsv(root, rsv);
8961
8962         if (ret && !err)
8963                 err = ret;
8964
8965         return err;
8966 }
8967
8968 /*
8969  * create a new subvolume directory/inode (helper for the ioctl).
8970  */
8971 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
8972                              struct btrfs_root *new_root,
8973                              struct btrfs_root *parent_root,
8974                              u64 new_dirid)
8975 {
8976         struct inode *inode;
8977         int err;
8978         u64 index = 0;
8979
8980         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
8981                                 new_dirid, new_dirid,
8982                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
8983                                 &index);
8984         if (IS_ERR(inode))
8985                 return PTR_ERR(inode);
8986         inode->i_op = &btrfs_dir_inode_operations;
8987         inode->i_fop = &btrfs_dir_file_operations;
8988
8989         set_nlink(inode, 1);
8990         btrfs_i_size_write(inode, 0);
8991         unlock_new_inode(inode);
8992
8993         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
8994         if (err)
8995                 btrfs_err(new_root->fs_info,
8996                           "error inheriting subvolume %llu properties: %d",
8997                           new_root->root_key.objectid, err);
8998
8999         err = btrfs_update_inode(trans, new_root, inode);
9000
9001         iput(inode);
9002         return err;
9003 }
9004
9005 struct inode *btrfs_alloc_inode(struct super_block *sb)
9006 {
9007         struct btrfs_inode *ei;
9008         struct inode *inode;
9009
9010         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9011         if (!ei)
9012                 return NULL;
9013
9014         ei->root = NULL;
9015         ei->generation = 0;
9016         ei->last_trans = 0;
9017         ei->last_sub_trans = 0;
9018         ei->logged_trans = 0;
9019         ei->delalloc_bytes = 0;
9020         ei->defrag_bytes = 0;
9021         ei->disk_i_size = 0;
9022         ei->flags = 0;
9023         ei->csum_bytes = 0;
9024         ei->index_cnt = (u64)-1;
9025         ei->dir_index = 0;
9026         ei->last_unlink_trans = 0;
9027         ei->last_log_commit = 0;
9028
9029         spin_lock_init(&ei->lock);
9030         ei->outstanding_extents = 0;
9031         ei->reserved_extents = 0;
9032
9033         ei->runtime_flags = 0;
9034         ei->force_compress = BTRFS_COMPRESS_NONE;
9035
9036         ei->delayed_node = NULL;
9037
9038         ei->i_otime.tv_sec = 0;
9039         ei->i_otime.tv_nsec = 0;
9040
9041         inode = &ei->vfs_inode;
9042         extent_map_tree_init(&ei->extent_tree);
9043         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9044         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9045         ei->io_tree.track_uptodate = 1;
9046         ei->io_failure_tree.track_uptodate = 1;
9047         atomic_set(&ei->sync_writers, 0);
9048         mutex_init(&ei->log_mutex);
9049         mutex_init(&ei->delalloc_mutex);
9050         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9051         INIT_LIST_HEAD(&ei->delalloc_inodes);
9052         RB_CLEAR_NODE(&ei->rb_node);
9053
9054         return inode;
9055 }
9056
9057 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9058 void btrfs_test_destroy_inode(struct inode *inode)
9059 {
9060         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9061         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9062 }
9063 #endif
9064
9065 static void btrfs_i_callback(struct rcu_head *head)
9066 {
9067         struct inode *inode = container_of(head, struct inode, i_rcu);
9068         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9069 }
9070
9071 void btrfs_destroy_inode(struct inode *inode)
9072 {
9073         struct btrfs_ordered_extent *ordered;
9074         struct btrfs_root *root = BTRFS_I(inode)->root;
9075
9076         WARN_ON(!hlist_empty(&inode->i_dentry));
9077         WARN_ON(inode->i_data.nrpages);
9078         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9079         WARN_ON(BTRFS_I(inode)->reserved_extents);
9080         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9081         WARN_ON(BTRFS_I(inode)->csum_bytes);
9082         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9083
9084         /*
9085          * This can happen where we create an inode, but somebody else also
9086          * created the same inode and we need to destroy the one we already
9087          * created.
9088          */
9089         if (!root)
9090                 goto free;
9091
9092         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9093                      &BTRFS_I(inode)->runtime_flags)) {
9094                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9095                         btrfs_ino(inode));
9096                 atomic_dec(&root->orphan_inodes);
9097         }
9098
9099         while (1) {
9100                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9101                 if (!ordered)
9102                         break;
9103                 else {
9104                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9105                                 ordered->file_offset, ordered->len);
9106                         btrfs_remove_ordered_extent(inode, ordered);
9107                         btrfs_put_ordered_extent(ordered);
9108                         btrfs_put_ordered_extent(ordered);
9109                 }
9110         }
9111         btrfs_qgroup_check_reserved_leak(inode);
9112         inode_tree_del(inode);
9113         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9114 free:
9115         call_rcu(&inode->i_rcu, btrfs_i_callback);
9116 }
9117
9118 int btrfs_drop_inode(struct inode *inode)
9119 {
9120         struct btrfs_root *root = BTRFS_I(inode)->root;
9121
9122         if (root == NULL)
9123                 return 1;
9124
9125         /* the snap/subvol tree is on deleting */
9126         if (btrfs_root_refs(&root->root_item) == 0)
9127                 return 1;
9128         else
9129                 return generic_drop_inode(inode);
9130 }
9131
9132 static void init_once(void *foo)
9133 {
9134         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9135
9136         inode_init_once(&ei->vfs_inode);
9137 }
9138
9139 void btrfs_destroy_cachep(void)
9140 {
9141         /*
9142          * Make sure all delayed rcu free inodes are flushed before we
9143          * destroy cache.
9144          */
9145         rcu_barrier();
9146         if (btrfs_inode_cachep)
9147                 kmem_cache_destroy(btrfs_inode_cachep);
9148         if (btrfs_trans_handle_cachep)
9149                 kmem_cache_destroy(btrfs_trans_handle_cachep);
9150         if (btrfs_transaction_cachep)
9151                 kmem_cache_destroy(btrfs_transaction_cachep);
9152         if (btrfs_path_cachep)
9153                 kmem_cache_destroy(btrfs_path_cachep);
9154         if (btrfs_free_space_cachep)
9155                 kmem_cache_destroy(btrfs_free_space_cachep);
9156         if (btrfs_delalloc_work_cachep)
9157                 kmem_cache_destroy(btrfs_delalloc_work_cachep);
9158 }
9159
9160 int btrfs_init_cachep(void)
9161 {
9162         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9163                         sizeof(struct btrfs_inode), 0,
9164                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9165                         init_once);
9166         if (!btrfs_inode_cachep)
9167                 goto fail;
9168
9169         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9170                         sizeof(struct btrfs_trans_handle), 0,
9171                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9172         if (!btrfs_trans_handle_cachep)
9173                 goto fail;
9174
9175         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9176                         sizeof(struct btrfs_transaction), 0,
9177                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9178         if (!btrfs_transaction_cachep)
9179                 goto fail;
9180
9181         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9182                         sizeof(struct btrfs_path), 0,
9183                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9184         if (!btrfs_path_cachep)
9185                 goto fail;
9186
9187         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9188                         sizeof(struct btrfs_free_space), 0,
9189                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9190         if (!btrfs_free_space_cachep)
9191                 goto fail;
9192
9193         btrfs_delalloc_work_cachep = kmem_cache_create("btrfs_delalloc_work",
9194                         sizeof(struct btrfs_delalloc_work), 0,
9195                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
9196                         NULL);
9197         if (!btrfs_delalloc_work_cachep)
9198                 goto fail;
9199
9200         return 0;
9201 fail:
9202         btrfs_destroy_cachep();
9203         return -ENOMEM;
9204 }
9205
9206 static int btrfs_getattr(struct vfsmount *mnt,
9207                          struct dentry *dentry, struct kstat *stat)
9208 {
9209         u64 delalloc_bytes;
9210         struct inode *inode = d_inode(dentry);
9211         u32 blocksize = inode->i_sb->s_blocksize;
9212
9213         generic_fillattr(inode, stat);
9214         stat->dev = BTRFS_I(inode)->root->anon_dev;
9215         stat->blksize = PAGE_CACHE_SIZE;
9216
9217         spin_lock(&BTRFS_I(inode)->lock);
9218         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9219         spin_unlock(&BTRFS_I(inode)->lock);
9220         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9221                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9222         return 0;
9223 }
9224
9225 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9226                            struct inode *new_dir, struct dentry *new_dentry)
9227 {
9228         struct btrfs_trans_handle *trans;
9229         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9230         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9231         struct inode *new_inode = d_inode(new_dentry);
9232         struct inode *old_inode = d_inode(old_dentry);
9233         struct timespec ctime = CURRENT_TIME;
9234         u64 index = 0;
9235         u64 root_objectid;
9236         int ret;
9237         u64 old_ino = btrfs_ino(old_inode);
9238
9239         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9240                 return -EPERM;
9241
9242         /* we only allow rename subvolume link between subvolumes */
9243         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9244                 return -EXDEV;
9245
9246         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9247             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9248                 return -ENOTEMPTY;
9249
9250         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9251             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9252                 return -ENOTEMPTY;
9253
9254
9255         /* check for collisions, even if the  name isn't there */
9256         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9257                              new_dentry->d_name.name,
9258                              new_dentry->d_name.len);
9259
9260         if (ret) {
9261                 if (ret == -EEXIST) {
9262                         /* we shouldn't get
9263                          * eexist without a new_inode */
9264                         if (WARN_ON(!new_inode)) {
9265                                 return ret;
9266                         }
9267                 } else {
9268                         /* maybe -EOVERFLOW */
9269                         return ret;
9270                 }
9271         }
9272         ret = 0;
9273
9274         /*
9275          * we're using rename to replace one file with another.  Start IO on it
9276          * now so  we don't add too much work to the end of the transaction
9277          */
9278         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9279                 filemap_flush(old_inode->i_mapping);
9280
9281         /* close the racy window with snapshot create/destroy ioctl */
9282         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9283                 down_read(&root->fs_info->subvol_sem);
9284         /*
9285          * We want to reserve the absolute worst case amount of items.  So if
9286          * both inodes are subvols and we need to unlink them then that would
9287          * require 4 item modifications, but if they are both normal inodes it
9288          * would require 5 item modifications, so we'll assume their normal
9289          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9290          * should cover the worst case number of items we'll modify.
9291          */
9292         trans = btrfs_start_transaction(root, 11);
9293         if (IS_ERR(trans)) {
9294                 ret = PTR_ERR(trans);
9295                 goto out_notrans;
9296         }
9297
9298         if (dest != root)
9299                 btrfs_record_root_in_trans(trans, dest);
9300
9301         ret = btrfs_set_inode_index(new_dir, &index);
9302         if (ret)
9303                 goto out_fail;
9304
9305         BTRFS_I(old_inode)->dir_index = 0ULL;
9306         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9307                 /* force full log commit if subvolume involved. */
9308                 btrfs_set_log_full_commit(root->fs_info, trans);
9309         } else {
9310                 ret = btrfs_insert_inode_ref(trans, dest,
9311                                              new_dentry->d_name.name,
9312                                              new_dentry->d_name.len,
9313                                              old_ino,
9314                                              btrfs_ino(new_dir), index);
9315                 if (ret)
9316                         goto out_fail;
9317                 /*
9318                  * this is an ugly little race, but the rename is required
9319                  * to make sure that if we crash, the inode is either at the
9320                  * old name or the new one.  pinning the log transaction lets
9321                  * us make sure we don't allow a log commit to come in after
9322                  * we unlink the name but before we add the new name back in.
9323                  */
9324                 btrfs_pin_log_trans(root);
9325         }
9326
9327         inode_inc_iversion(old_dir);
9328         inode_inc_iversion(new_dir);
9329         inode_inc_iversion(old_inode);
9330         old_dir->i_ctime = old_dir->i_mtime = ctime;
9331         new_dir->i_ctime = new_dir->i_mtime = ctime;
9332         old_inode->i_ctime = ctime;
9333
9334         if (old_dentry->d_parent != new_dentry->d_parent)
9335                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9336
9337         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9338                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9339                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9340                                         old_dentry->d_name.name,
9341                                         old_dentry->d_name.len);
9342         } else {
9343                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9344                                         d_inode(old_dentry),
9345                                         old_dentry->d_name.name,
9346                                         old_dentry->d_name.len);
9347                 if (!ret)
9348                         ret = btrfs_update_inode(trans, root, old_inode);
9349         }
9350         if (ret) {
9351                 btrfs_abort_transaction(trans, root, ret);
9352                 goto out_fail;
9353         }
9354
9355         if (new_inode) {
9356                 inode_inc_iversion(new_inode);
9357                 new_inode->i_ctime = CURRENT_TIME;
9358                 if (unlikely(btrfs_ino(new_inode) ==
9359                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9360                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9361                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9362                                                 root_objectid,
9363                                                 new_dentry->d_name.name,
9364                                                 new_dentry->d_name.len);
9365                         BUG_ON(new_inode->i_nlink == 0);
9366                 } else {
9367                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9368                                                  d_inode(new_dentry),
9369                                                  new_dentry->d_name.name,
9370                                                  new_dentry->d_name.len);
9371                 }
9372                 if (!ret && new_inode->i_nlink == 0)
9373                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9374                 if (ret) {
9375                         btrfs_abort_transaction(trans, root, ret);
9376                         goto out_fail;
9377                 }
9378         }
9379
9380         ret = btrfs_add_link(trans, new_dir, old_inode,
9381                              new_dentry->d_name.name,
9382                              new_dentry->d_name.len, 0, index);
9383         if (ret) {
9384                 btrfs_abort_transaction(trans, root, ret);
9385                 goto out_fail;
9386         }
9387
9388         if (old_inode->i_nlink == 1)
9389                 BTRFS_I(old_inode)->dir_index = index;
9390
9391         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
9392                 struct dentry *parent = new_dentry->d_parent;
9393                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9394                 btrfs_end_log_trans(root);
9395         }
9396 out_fail:
9397         btrfs_end_transaction(trans, root);
9398 out_notrans:
9399         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9400                 up_read(&root->fs_info->subvol_sem);
9401
9402         return ret;
9403 }
9404
9405 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9406                          struct inode *new_dir, struct dentry *new_dentry,
9407                          unsigned int flags)
9408 {
9409         if (flags & ~RENAME_NOREPLACE)
9410                 return -EINVAL;
9411
9412         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry);
9413 }
9414
9415 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9416 {
9417         struct btrfs_delalloc_work *delalloc_work;
9418         struct inode *inode;
9419
9420         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9421                                      work);
9422         inode = delalloc_work->inode;
9423         if (delalloc_work->wait) {
9424                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
9425         } else {
9426                 filemap_flush(inode->i_mapping);
9427                 if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9428                              &BTRFS_I(inode)->runtime_flags))
9429                         filemap_flush(inode->i_mapping);
9430         }
9431
9432         if (delalloc_work->delay_iput)
9433                 btrfs_add_delayed_iput(inode);
9434         else
9435                 iput(inode);
9436         complete(&delalloc_work->completion);
9437 }
9438
9439 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9440                                                     int wait, int delay_iput)
9441 {
9442         struct btrfs_delalloc_work *work;
9443
9444         work = kmem_cache_zalloc(btrfs_delalloc_work_cachep, GFP_NOFS);
9445         if (!work)
9446                 return NULL;
9447
9448         init_completion(&work->completion);
9449         INIT_LIST_HEAD(&work->list);
9450         work->inode = inode;
9451         work->wait = wait;
9452         work->delay_iput = delay_iput;
9453         WARN_ON_ONCE(!inode);
9454         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9455                         btrfs_run_delalloc_work, NULL, NULL);
9456
9457         return work;
9458 }
9459
9460 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9461 {
9462         wait_for_completion(&work->completion);
9463         kmem_cache_free(btrfs_delalloc_work_cachep, work);
9464 }
9465
9466 /*
9467  * some fairly slow code that needs optimization. This walks the list
9468  * of all the inodes with pending delalloc and forces them to disk.
9469  */
9470 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9471                                    int nr)
9472 {
9473         struct btrfs_inode *binode;
9474         struct inode *inode;
9475         struct btrfs_delalloc_work *work, *next;
9476         struct list_head works;
9477         struct list_head splice;
9478         int ret = 0;
9479
9480         INIT_LIST_HEAD(&works);
9481         INIT_LIST_HEAD(&splice);
9482
9483         mutex_lock(&root->delalloc_mutex);
9484         spin_lock(&root->delalloc_lock);
9485         list_splice_init(&root->delalloc_inodes, &splice);
9486         while (!list_empty(&splice)) {
9487                 binode = list_entry(splice.next, struct btrfs_inode,
9488                                     delalloc_inodes);
9489
9490                 list_move_tail(&binode->delalloc_inodes,
9491                                &root->delalloc_inodes);
9492                 inode = igrab(&binode->vfs_inode);
9493                 if (!inode) {
9494                         cond_resched_lock(&root->delalloc_lock);
9495                         continue;
9496                 }
9497                 spin_unlock(&root->delalloc_lock);
9498
9499                 work = btrfs_alloc_delalloc_work(inode, 0, delay_iput);
9500                 if (!work) {
9501                         if (delay_iput)
9502                                 btrfs_add_delayed_iput(inode);
9503                         else
9504                                 iput(inode);
9505                         ret = -ENOMEM;
9506                         goto out;
9507                 }
9508                 list_add_tail(&work->list, &works);
9509                 btrfs_queue_work(root->fs_info->flush_workers,
9510                                  &work->work);
9511                 ret++;
9512                 if (nr != -1 && ret >= nr)
9513                         goto out;
9514                 cond_resched();
9515                 spin_lock(&root->delalloc_lock);
9516         }
9517         spin_unlock(&root->delalloc_lock);
9518
9519 out:
9520         list_for_each_entry_safe(work, next, &works, list) {
9521                 list_del_init(&work->list);
9522                 btrfs_wait_and_free_delalloc_work(work);
9523         }
9524
9525         if (!list_empty_careful(&splice)) {
9526                 spin_lock(&root->delalloc_lock);
9527                 list_splice_tail(&splice, &root->delalloc_inodes);
9528                 spin_unlock(&root->delalloc_lock);
9529         }
9530         mutex_unlock(&root->delalloc_mutex);
9531         return ret;
9532 }
9533
9534 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
9535 {
9536         int ret;
9537
9538         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
9539                 return -EROFS;
9540
9541         ret = __start_delalloc_inodes(root, delay_iput, -1);
9542         if (ret > 0)
9543                 ret = 0;
9544         /*
9545          * the filemap_flush will queue IO into the worker threads, but
9546          * we have to make sure the IO is actually started and that
9547          * ordered extents get created before we return
9548          */
9549         atomic_inc(&root->fs_info->async_submit_draining);
9550         while (atomic_read(&root->fs_info->nr_async_submits) ||
9551               atomic_read(&root->fs_info->async_delalloc_pages)) {
9552                 wait_event(root->fs_info->async_submit_wait,
9553                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
9554                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
9555         }
9556         atomic_dec(&root->fs_info->async_submit_draining);
9557         return ret;
9558 }
9559
9560 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
9561                                int nr)
9562 {
9563         struct btrfs_root *root;
9564         struct list_head splice;
9565         int ret;
9566
9567         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9568                 return -EROFS;
9569
9570         INIT_LIST_HEAD(&splice);
9571
9572         mutex_lock(&fs_info->delalloc_root_mutex);
9573         spin_lock(&fs_info->delalloc_root_lock);
9574         list_splice_init(&fs_info->delalloc_roots, &splice);
9575         while (!list_empty(&splice) && nr) {
9576                 root = list_first_entry(&splice, struct btrfs_root,
9577                                         delalloc_root);
9578                 root = btrfs_grab_fs_root(root);
9579                 BUG_ON(!root);
9580                 list_move_tail(&root->delalloc_root,
9581                                &fs_info->delalloc_roots);
9582                 spin_unlock(&fs_info->delalloc_root_lock);
9583
9584                 ret = __start_delalloc_inodes(root, delay_iput, nr);
9585                 btrfs_put_fs_root(root);
9586                 if (ret < 0)
9587                         goto out;
9588
9589                 if (nr != -1) {
9590                         nr -= ret;
9591                         WARN_ON(nr < 0);
9592                 }
9593                 spin_lock(&fs_info->delalloc_root_lock);
9594         }
9595         spin_unlock(&fs_info->delalloc_root_lock);
9596
9597         ret = 0;
9598         atomic_inc(&fs_info->async_submit_draining);
9599         while (atomic_read(&fs_info->nr_async_submits) ||
9600               atomic_read(&fs_info->async_delalloc_pages)) {
9601                 wait_event(fs_info->async_submit_wait,
9602                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
9603                     atomic_read(&fs_info->async_delalloc_pages) == 0));
9604         }
9605         atomic_dec(&fs_info->async_submit_draining);
9606 out:
9607         if (!list_empty_careful(&splice)) {
9608                 spin_lock(&fs_info->delalloc_root_lock);
9609                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9610                 spin_unlock(&fs_info->delalloc_root_lock);
9611         }
9612         mutex_unlock(&fs_info->delalloc_root_mutex);
9613         return ret;
9614 }
9615
9616 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9617                          const char *symname)
9618 {
9619         struct btrfs_trans_handle *trans;
9620         struct btrfs_root *root = BTRFS_I(dir)->root;
9621         struct btrfs_path *path;
9622         struct btrfs_key key;
9623         struct inode *inode = NULL;
9624         int err;
9625         int drop_inode = 0;
9626         u64 objectid;
9627         u64 index = 0;
9628         int name_len;
9629         int datasize;
9630         unsigned long ptr;
9631         struct btrfs_file_extent_item *ei;
9632         struct extent_buffer *leaf;
9633
9634         name_len = strlen(symname);
9635         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
9636                 return -ENAMETOOLONG;
9637
9638         /*
9639          * 2 items for inode item and ref
9640          * 2 items for dir items
9641          * 1 item for xattr if selinux is on
9642          */
9643         trans = btrfs_start_transaction(root, 5);
9644         if (IS_ERR(trans))
9645                 return PTR_ERR(trans);
9646
9647         err = btrfs_find_free_ino(root, &objectid);
9648         if (err)
9649                 goto out_unlock;
9650
9651         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9652                                 dentry->d_name.len, btrfs_ino(dir), objectid,
9653                                 S_IFLNK|S_IRWXUGO, &index);
9654         if (IS_ERR(inode)) {
9655                 err = PTR_ERR(inode);
9656                 goto out_unlock;
9657         }
9658
9659         /*
9660         * If the active LSM wants to access the inode during
9661         * d_instantiate it needs these. Smack checks to see
9662         * if the filesystem supports xattrs by looking at the
9663         * ops vector.
9664         */
9665         inode->i_fop = &btrfs_file_operations;
9666         inode->i_op = &btrfs_file_inode_operations;
9667         inode->i_mapping->a_ops = &btrfs_aops;
9668         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9669
9670         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9671         if (err)
9672                 goto out_unlock_inode;
9673
9674         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
9675         if (err)
9676                 goto out_unlock_inode;
9677
9678         path = btrfs_alloc_path();
9679         if (!path) {
9680                 err = -ENOMEM;
9681                 goto out_unlock_inode;
9682         }
9683         key.objectid = btrfs_ino(inode);
9684         key.offset = 0;
9685         key.type = BTRFS_EXTENT_DATA_KEY;
9686         datasize = btrfs_file_extent_calc_inline_size(name_len);
9687         err = btrfs_insert_empty_item(trans, root, path, &key,
9688                                       datasize);
9689         if (err) {
9690                 btrfs_free_path(path);
9691                 goto out_unlock_inode;
9692         }
9693         leaf = path->nodes[0];
9694         ei = btrfs_item_ptr(leaf, path->slots[0],
9695                             struct btrfs_file_extent_item);
9696         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9697         btrfs_set_file_extent_type(leaf, ei,
9698                                    BTRFS_FILE_EXTENT_INLINE);
9699         btrfs_set_file_extent_encryption(leaf, ei, 0);
9700         btrfs_set_file_extent_compression(leaf, ei, 0);
9701         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9702         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9703
9704         ptr = btrfs_file_extent_inline_start(ei);
9705         write_extent_buffer(leaf, symname, ptr, name_len);
9706         btrfs_mark_buffer_dirty(leaf);
9707         btrfs_free_path(path);
9708
9709         inode->i_op = &btrfs_symlink_inode_operations;
9710         inode_nohighmem(inode);
9711         inode->i_mapping->a_ops = &btrfs_symlink_aops;
9712         inode_set_bytes(inode, name_len);
9713         btrfs_i_size_write(inode, name_len);
9714         err = btrfs_update_inode(trans, root, inode);
9715         if (err) {
9716                 drop_inode = 1;
9717                 goto out_unlock_inode;
9718         }
9719
9720         unlock_new_inode(inode);
9721         d_instantiate(dentry, inode);
9722
9723 out_unlock:
9724         btrfs_end_transaction(trans, root);
9725         if (drop_inode) {
9726                 inode_dec_link_count(inode);
9727                 iput(inode);
9728         }
9729         btrfs_btree_balance_dirty(root);
9730         return err;
9731
9732 out_unlock_inode:
9733         drop_inode = 1;
9734         unlock_new_inode(inode);
9735         goto out_unlock;
9736 }
9737
9738 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9739                                        u64 start, u64 num_bytes, u64 min_size,
9740                                        loff_t actual_len, u64 *alloc_hint,
9741                                        struct btrfs_trans_handle *trans)
9742 {
9743         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9744         struct extent_map *em;
9745         struct btrfs_root *root = BTRFS_I(inode)->root;
9746         struct btrfs_key ins;
9747         u64 cur_offset = start;
9748         u64 i_size;
9749         u64 cur_bytes;
9750         u64 last_alloc = (u64)-1;
9751         int ret = 0;
9752         bool own_trans = true;
9753
9754         if (trans)
9755                 own_trans = false;
9756         while (num_bytes > 0) {
9757                 if (own_trans) {
9758                         trans = btrfs_start_transaction(root, 3);
9759                         if (IS_ERR(trans)) {
9760                                 ret = PTR_ERR(trans);
9761                                 break;
9762                         }
9763                 }
9764
9765                 cur_bytes = min(num_bytes, 256ULL * 1024 * 1024);
9766                 cur_bytes = max(cur_bytes, min_size);
9767                 /*
9768                  * If we are severely fragmented we could end up with really
9769                  * small allocations, so if the allocator is returning small
9770                  * chunks lets make its job easier by only searching for those
9771                  * sized chunks.
9772                  */
9773                 cur_bytes = min(cur_bytes, last_alloc);
9774                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
9775                                            *alloc_hint, &ins, 1, 0);
9776                 if (ret) {
9777                         if (own_trans)
9778                                 btrfs_end_transaction(trans, root);
9779                         break;
9780                 }
9781
9782                 last_alloc = ins.offset;
9783                 ret = insert_reserved_file_extent(trans, inode,
9784                                                   cur_offset, ins.objectid,
9785                                                   ins.offset, ins.offset,
9786                                                   ins.offset, 0, 0, 0,
9787                                                   BTRFS_FILE_EXTENT_PREALLOC);
9788                 if (ret) {
9789                         btrfs_free_reserved_extent(root, ins.objectid,
9790                                                    ins.offset, 0);
9791                         btrfs_abort_transaction(trans, root, ret);
9792                         if (own_trans)
9793                                 btrfs_end_transaction(trans, root);
9794                         break;
9795                 }
9796
9797                 btrfs_drop_extent_cache(inode, cur_offset,
9798                                         cur_offset + ins.offset -1, 0);
9799
9800                 em = alloc_extent_map();
9801                 if (!em) {
9802                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9803                                 &BTRFS_I(inode)->runtime_flags);
9804                         goto next;
9805                 }
9806
9807                 em->start = cur_offset;
9808                 em->orig_start = cur_offset;
9809                 em->len = ins.offset;
9810                 em->block_start = ins.objectid;
9811                 em->block_len = ins.offset;
9812                 em->orig_block_len = ins.offset;
9813                 em->ram_bytes = ins.offset;
9814                 em->bdev = root->fs_info->fs_devices->latest_bdev;
9815                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9816                 em->generation = trans->transid;
9817
9818                 while (1) {
9819                         write_lock(&em_tree->lock);
9820                         ret = add_extent_mapping(em_tree, em, 1);
9821                         write_unlock(&em_tree->lock);
9822                         if (ret != -EEXIST)
9823                                 break;
9824                         btrfs_drop_extent_cache(inode, cur_offset,
9825                                                 cur_offset + ins.offset - 1,
9826                                                 0);
9827                 }
9828                 free_extent_map(em);
9829 next:
9830                 num_bytes -= ins.offset;
9831                 cur_offset += ins.offset;
9832                 *alloc_hint = ins.objectid + ins.offset;
9833
9834                 inode_inc_iversion(inode);
9835                 inode->i_ctime = CURRENT_TIME;
9836                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9837                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9838                     (actual_len > inode->i_size) &&
9839                     (cur_offset > inode->i_size)) {
9840                         if (cur_offset > actual_len)
9841                                 i_size = actual_len;
9842                         else
9843                                 i_size = cur_offset;
9844                         i_size_write(inode, i_size);
9845                         btrfs_ordered_update_i_size(inode, i_size, NULL);
9846                 }
9847
9848                 ret = btrfs_update_inode(trans, root, inode);
9849
9850                 if (ret) {
9851                         btrfs_abort_transaction(trans, root, ret);
9852                         if (own_trans)
9853                                 btrfs_end_transaction(trans, root);
9854                         break;
9855                 }
9856
9857                 if (own_trans)
9858                         btrfs_end_transaction(trans, root);
9859         }
9860         return ret;
9861 }
9862
9863 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9864                               u64 start, u64 num_bytes, u64 min_size,
9865                               loff_t actual_len, u64 *alloc_hint)
9866 {
9867         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9868                                            min_size, actual_len, alloc_hint,
9869                                            NULL);
9870 }
9871
9872 int btrfs_prealloc_file_range_trans(struct inode *inode,
9873                                     struct btrfs_trans_handle *trans, int mode,
9874                                     u64 start, u64 num_bytes, u64 min_size,
9875                                     loff_t actual_len, u64 *alloc_hint)
9876 {
9877         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9878                                            min_size, actual_len, alloc_hint, trans);
9879 }
9880
9881 static int btrfs_set_page_dirty(struct page *page)
9882 {
9883         return __set_page_dirty_nobuffers(page);
9884 }
9885
9886 static int btrfs_permission(struct inode *inode, int mask)
9887 {
9888         struct btrfs_root *root = BTRFS_I(inode)->root;
9889         umode_t mode = inode->i_mode;
9890
9891         if (mask & MAY_WRITE &&
9892             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
9893                 if (btrfs_root_readonly(root))
9894                         return -EROFS;
9895                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
9896                         return -EACCES;
9897         }
9898         return generic_permission(inode, mask);
9899 }
9900
9901 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
9902 {
9903         struct btrfs_trans_handle *trans;
9904         struct btrfs_root *root = BTRFS_I(dir)->root;
9905         struct inode *inode = NULL;
9906         u64 objectid;
9907         u64 index;
9908         int ret = 0;
9909
9910         /*
9911          * 5 units required for adding orphan entry
9912          */
9913         trans = btrfs_start_transaction(root, 5);
9914         if (IS_ERR(trans))
9915                 return PTR_ERR(trans);
9916
9917         ret = btrfs_find_free_ino(root, &objectid);
9918         if (ret)
9919                 goto out;
9920
9921         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
9922                                 btrfs_ino(dir), objectid, mode, &index);
9923         if (IS_ERR(inode)) {
9924                 ret = PTR_ERR(inode);
9925                 inode = NULL;
9926                 goto out;
9927         }
9928
9929         inode->i_fop = &btrfs_file_operations;
9930         inode->i_op = &btrfs_file_inode_operations;
9931
9932         inode->i_mapping->a_ops = &btrfs_aops;
9933         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9934
9935         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
9936         if (ret)
9937                 goto out_inode;
9938
9939         ret = btrfs_update_inode(trans, root, inode);
9940         if (ret)
9941                 goto out_inode;
9942         ret = btrfs_orphan_add(trans, inode);
9943         if (ret)
9944                 goto out_inode;
9945
9946         /*
9947          * We set number of links to 0 in btrfs_new_inode(), and here we set
9948          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
9949          * through:
9950          *
9951          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
9952          */
9953         set_nlink(inode, 1);
9954         unlock_new_inode(inode);
9955         d_tmpfile(dentry, inode);
9956         mark_inode_dirty(inode);
9957
9958 out:
9959         btrfs_end_transaction(trans, root);
9960         if (ret)
9961                 iput(inode);
9962         btrfs_balance_delayed_items(root);
9963         btrfs_btree_balance_dirty(root);
9964         return ret;
9965
9966 out_inode:
9967         unlock_new_inode(inode);
9968         goto out;
9969
9970 }
9971
9972 /* Inspired by filemap_check_errors() */
9973 int btrfs_inode_check_errors(struct inode *inode)
9974 {
9975         int ret = 0;
9976
9977         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
9978             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
9979                 ret = -ENOSPC;
9980         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
9981             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
9982                 ret = -EIO;
9983
9984         return ret;
9985 }
9986
9987 static const struct inode_operations btrfs_dir_inode_operations = {
9988         .getattr        = btrfs_getattr,
9989         .lookup         = btrfs_lookup,
9990         .create         = btrfs_create,
9991         .unlink         = btrfs_unlink,
9992         .link           = btrfs_link,
9993         .mkdir          = btrfs_mkdir,
9994         .rmdir          = btrfs_rmdir,
9995         .rename2        = btrfs_rename2,
9996         .symlink        = btrfs_symlink,
9997         .setattr        = btrfs_setattr,
9998         .mknod          = btrfs_mknod,
9999         .setxattr       = btrfs_setxattr,
10000         .getxattr       = generic_getxattr,
10001         .listxattr      = btrfs_listxattr,
10002         .removexattr    = btrfs_removexattr,
10003         .permission     = btrfs_permission,
10004         .get_acl        = btrfs_get_acl,
10005         .set_acl        = btrfs_set_acl,
10006         .update_time    = btrfs_update_time,
10007         .tmpfile        = btrfs_tmpfile,
10008 };
10009 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10010         .lookup         = btrfs_lookup,
10011         .permission     = btrfs_permission,
10012         .get_acl        = btrfs_get_acl,
10013         .set_acl        = btrfs_set_acl,
10014         .update_time    = btrfs_update_time,
10015 };
10016
10017 static const struct file_operations btrfs_dir_file_operations = {
10018         .llseek         = generic_file_llseek,
10019         .read           = generic_read_dir,
10020         .iterate        = btrfs_real_readdir,
10021         .unlocked_ioctl = btrfs_ioctl,
10022 #ifdef CONFIG_COMPAT
10023         .compat_ioctl   = btrfs_ioctl,
10024 #endif
10025         .release        = btrfs_release_file,
10026         .fsync          = btrfs_sync_file,
10027 };
10028
10029 static struct extent_io_ops btrfs_extent_io_ops = {
10030         .fill_delalloc = run_delalloc_range,
10031         .submit_bio_hook = btrfs_submit_bio_hook,
10032         .merge_bio_hook = btrfs_merge_bio_hook,
10033         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10034         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10035         .writepage_start_hook = btrfs_writepage_start_hook,
10036         .set_bit_hook = btrfs_set_bit_hook,
10037         .clear_bit_hook = btrfs_clear_bit_hook,
10038         .merge_extent_hook = btrfs_merge_extent_hook,
10039         .split_extent_hook = btrfs_split_extent_hook,
10040 };
10041
10042 /*
10043  * btrfs doesn't support the bmap operation because swapfiles
10044  * use bmap to make a mapping of extents in the file.  They assume
10045  * these extents won't change over the life of the file and they
10046  * use the bmap result to do IO directly to the drive.
10047  *
10048  * the btrfs bmap call would return logical addresses that aren't
10049  * suitable for IO and they also will change frequently as COW
10050  * operations happen.  So, swapfile + btrfs == corruption.
10051  *
10052  * For now we're avoiding this by dropping bmap.
10053  */
10054 static const struct address_space_operations btrfs_aops = {
10055         .readpage       = btrfs_readpage,
10056         .writepage      = btrfs_writepage,
10057         .writepages     = btrfs_writepages,
10058         .readpages      = btrfs_readpages,
10059         .direct_IO      = btrfs_direct_IO,
10060         .invalidatepage = btrfs_invalidatepage,
10061         .releasepage    = btrfs_releasepage,
10062         .set_page_dirty = btrfs_set_page_dirty,
10063         .error_remove_page = generic_error_remove_page,
10064 };
10065
10066 static const struct address_space_operations btrfs_symlink_aops = {
10067         .readpage       = btrfs_readpage,
10068         .writepage      = btrfs_writepage,
10069         .invalidatepage = btrfs_invalidatepage,
10070         .releasepage    = btrfs_releasepage,
10071 };
10072
10073 static const struct inode_operations btrfs_file_inode_operations = {
10074         .getattr        = btrfs_getattr,
10075         .setattr        = btrfs_setattr,
10076         .setxattr       = btrfs_setxattr,
10077         .getxattr       = generic_getxattr,
10078         .listxattr      = btrfs_listxattr,
10079         .removexattr    = btrfs_removexattr,
10080         .permission     = btrfs_permission,
10081         .fiemap         = btrfs_fiemap,
10082         .get_acl        = btrfs_get_acl,
10083         .set_acl        = btrfs_set_acl,
10084         .update_time    = btrfs_update_time,
10085 };
10086 static const struct inode_operations btrfs_special_inode_operations = {
10087         .getattr        = btrfs_getattr,
10088         .setattr        = btrfs_setattr,
10089         .permission     = btrfs_permission,
10090         .setxattr       = btrfs_setxattr,
10091         .getxattr       = generic_getxattr,
10092         .listxattr      = btrfs_listxattr,
10093         .removexattr    = btrfs_removexattr,
10094         .get_acl        = btrfs_get_acl,
10095         .set_acl        = btrfs_set_acl,
10096         .update_time    = btrfs_update_time,
10097 };
10098 static const struct inode_operations btrfs_symlink_inode_operations = {
10099         .readlink       = generic_readlink,
10100         .get_link       = page_get_link,
10101         .getattr        = btrfs_getattr,
10102         .setattr        = btrfs_setattr,
10103         .permission     = btrfs_permission,
10104         .setxattr       = btrfs_setxattr,
10105         .getxattr       = generic_getxattr,
10106         .listxattr      = btrfs_listxattr,
10107         .removexattr    = btrfs_removexattr,
10108         .update_time    = btrfs_update_time,
10109 };
10110
10111 const struct dentry_operations btrfs_dentry_operations = {
10112         .d_delete       = btrfs_dentry_delete,
10113         .d_release      = btrfs_dentry_release,
10114 };