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 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74 };
  75
  76 static const struct inode_operations btrfs_dir_inode_operations;
  77 static const struct inode_operations btrfs_symlink_inode_operations;
  78 static const struct inode_operations btrfs_dir_ro_inode_operations;
  79 static const struct inode_operations btrfs_special_inode_operations;
  80 static const struct inode_operations btrfs_file_inode_operations;
  81 static const struct address_space_operations btrfs_aops;
  82 static const struct address_space_operations btrfs_symlink_aops;
  83 static const struct file_operations btrfs_dir_file_operations;
  84 static const struct extent_io_ops btrfs_extent_io_ops;
  85
  86 static struct kmem_cache *btrfs_inode_cachep;
  87 struct kmem_cache *btrfs_trans_handle_cachep;
  88 struct kmem_cache *btrfs_transaction_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, int *page_started,
 109                                    unsigned long *nr_written, int unlock);
 110 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 111                                            u64 len, u64 orig_start,
 112                                            u64 block_start, u64 block_len,
 113                                            u64 orig_block_len, u64 ram_bytes,
 114                                            int type);
 115
 116 static int btrfs_dirty_inode(struct inode *inode);
 117
 118 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 119 void btrfs_test_inode_set_ops(struct inode *inode)
 120 {
 121         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 122 }
 123 #endif
 124
 125 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 126                                      struct inode *inode,  struct inode *dir,
 127                                      const struct qstr *qstr)
 128 {
 129         int err;
 130
 131         err = btrfs_init_acl(trans, inode, dir);
 132         if (!err)
 133                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 134         return err;
 135 }
 136
 137 /*
 138  * this does all the hard work for inserting an inline extent into
 139  * the btree.  The caller should have done a btrfs_drop_extents so that
 140  * no overlapping inline items exist in the btree
 141  */
 142 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 143                                 struct btrfs_path *path, int extent_inserted,
 144                                 struct btrfs_root *root, struct inode *inode,
 145                                 u64 start, size_t size, size_t compressed_size,
 146                                 int compress_type,
 147                                 struct page **compressed_pages)
 148 {
 149         struct extent_buffer *leaf;
 150         struct page *page = NULL;
 151         char *kaddr;
 152         unsigned long ptr;
 153         struct btrfs_file_extent_item *ei;
 154         int err = 0;
 155         int ret;
 156         size_t cur_size = size;
 157         unsigned long offset;
 158
 159         if (compressed_size && compressed_pages)
 160                 cur_size = compressed_size;
 161
 162         inode_add_bytes(inode, size);
 163
 164         if (!extent_inserted) {
 165                 struct btrfs_key key;
 166                 size_t datasize;
 167
 168                 key.objectid = btrfs_ino(inode);
 169                 key.offset = start;
 170                 key.type = BTRFS_EXTENT_DATA_KEY;
 171
 172                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 173                 path->leave_spinning = 1;
 174                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                               datasize);
 176                 if (ret) {
 177                         err = ret;
 178                         goto fail;
 179                 }
 180         }
 181         leaf = path->nodes[0];
 182         ei = btrfs_item_ptr(leaf, path->slots[0],
 183                             struct btrfs_file_extent_item);
 184         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 185         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 186         btrfs_set_file_extent_encryption(leaf, ei, 0);
 187         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 188         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 189         ptr = btrfs_file_extent_inline_start(ei);
 190
 191         if (compress_type != BTRFS_COMPRESS_NONE) {
 192                 struct page *cpage;
 193                 int i = 0;
 194                 while (compressed_size > 0) {
 195                         cpage = compressed_pages[i];
 196                         cur_size = min_t(unsigned long, compressed_size,
 197                                        PAGE_SIZE);
 198
 199                         kaddr = kmap_atomic(cpage);
 200                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 201                         kunmap_atomic(kaddr);
 202
 203                         i++;
 204                         ptr += cur_size;
 205                         compressed_size -= cur_size;
 206                 }
 207                 btrfs_set_file_extent_compression(leaf, ei,
 208                                                   compress_type);
 209         } else {
 210                 page = find_get_page(inode->i_mapping,
 211                                      start >> PAGE_SHIFT);
 212                 btrfs_set_file_extent_compression(leaf, ei, 0);
 213                 kaddr = kmap_atomic(page);
 214                 offset = start & (PAGE_SIZE - 1);
 215                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 216                 kunmap_atomic(kaddr);
 217                 put_page(page);
 218         }
 219         btrfs_mark_buffer_dirty(leaf);
 220         btrfs_release_path(path);
 221
 222         /*
 223          * we're an inline extent, so nobody can
 224          * extend the file past i_size without locking
 225          * a page we already have locked.
 226          *
 227          * We must do any isize and inode updates
 228          * before we unlock the pages.  Otherwise we
 229          * could end up racing with unlink.
 230          */
 231         BTRFS_I(inode)->disk_i_size = inode->i_size;
 232         ret = btrfs_update_inode(trans, root, inode);
 233
 234         return ret;
 235 fail:
 236         return err;
 237 }
 238
 239
 240 /*
 241  * conditionally insert an inline extent into the file.  This
 242  * does the checks required to make sure the data is small enough
 243  * to fit as an inline extent.
 244  */
 245 static noinline int cow_file_range_inline(struct btrfs_root *root,
 246                                           struct inode *inode, u64 start,
 247                                           u64 end, size_t compressed_size,
 248                                           int compress_type,
 249                                           struct page **compressed_pages)
 250 {
 251         struct btrfs_trans_handle *trans;
 252         u64 isize = i_size_read(inode);
 253         u64 actual_end = min(end + 1, isize);
 254         u64 inline_len = actual_end - start;
 255         u64 aligned_end = ALIGN(end, root->sectorsize);
 256         u64 data_len = inline_len;
 257         int ret;
 258         struct btrfs_path *path;
 259         int extent_inserted = 0;
 260         u32 extent_item_size;
 261
 262         if (compressed_size)
 263                 data_len = compressed_size;
 264
 265         if (start > 0 ||
 266             actual_end > root->sectorsize ||
 267             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 268             (!compressed_size &&
 269             (actual_end & (root->sectorsize - 1)) == 0) ||
 270             end + 1 < isize ||
 271             data_len > root->fs_info->max_inline) {
 272                 return 1;
 273         }
 274
 275         path = btrfs_alloc_path();
 276         if (!path)
 277                 return -ENOMEM;
 278
 279         trans = btrfs_join_transaction(root);
 280         if (IS_ERR(trans)) {
 281                 btrfs_free_path(path);
 282                 return PTR_ERR(trans);
 283         }
 284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 285
 286         if (compressed_size && compressed_pages)
 287                 extent_item_size = btrfs_file_extent_calc_inline_size(
 288                    compressed_size);
 289         else
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                     inline_len);
 292
 293         ret = __btrfs_drop_extents(trans, root, inode, path,
 294                                    start, aligned_end, NULL,
 295                                    1, 1, extent_item_size, &extent_inserted);
 296         if (ret) {
 297                 btrfs_abort_transaction(trans, root, ret);
 298                 goto out;
 299         }
 300
 301         if (isize > actual_end)
 302                 inline_len = min_t(u64, isize, actual_end);
 303         ret = insert_inline_extent(trans, path, extent_inserted,
 304                                    root, inode, start,
 305                                    inline_len, compressed_size,
 306                                    compress_type, compressed_pages);
 307         if (ret && ret != -ENOSPC) {
 308                 btrfs_abort_transaction(trans, root, ret);
 309                 goto out;
 310         } else if (ret == -ENOSPC) {
 311                 ret = 1;
 312                 goto out;
 313         }
 314
 315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 316         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 317         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 318 out:
 319         /*
 320          * Don't forget to free the reserved space, as for inlined extent
 321          * it won't count as data extent, free them directly here.
 322          * And at reserve time, it's always aligned to page size, so
 323          * just free one page here.
 324          */
 325         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 326         btrfs_free_path(path);
 327         btrfs_end_transaction(trans, root);
 328         return ret;
 329 }
 330
 331 struct async_extent {
 332         u64 start;
 333         u64 ram_size;
 334         u64 compressed_size;
 335         struct page **pages;
 336         unsigned long nr_pages;
 337         int compress_type;
 338         struct list_head list;
 339 };
 340
 341 struct async_cow {
 342         struct inode *inode;
 343         struct btrfs_root *root;
 344         struct page *locked_page;
 345         u64 start;
 346         u64 end;
 347         struct list_head extents;
 348         struct btrfs_work work;
 349 };
 350
 351 static noinline int add_async_extent(struct async_cow *cow,
 352                                      u64 start, u64 ram_size,
 353                                      u64 compressed_size,
 354                                      struct page **pages,
 355                                      unsigned long nr_pages,
 356                                      int compress_type)
 357 {
 358         struct async_extent *async_extent;
 359
 360         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 361         BUG_ON(!async_extent); /* -ENOMEM */
 362         async_extent->start = start;
 363         async_extent->ram_size = ram_size;
 364         async_extent->compressed_size = compressed_size;
 365         async_extent->pages = pages;
 366         async_extent->nr_pages = nr_pages;
 367         async_extent->compress_type = compress_type;
 368         list_add_tail(&async_extent->list, &cow->extents);
 369         return 0;
 370 }
 371
 372 static inline int inode_need_compress(struct inode *inode)
 373 {
 374         struct btrfs_root *root = BTRFS_I(inode)->root;
 375
 376         /* force compress */
 377         if (btrfs_test_opt(root, FORCE_COMPRESS))
 378                 return 1;
 379         /* bad compression ratios */
 380         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 381                 return 0;
 382         if (btrfs_test_opt(root, COMPRESS) ||
 383             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 384             BTRFS_I(inode)->force_compress)
 385                 return 1;
 386         return 0;
 387 }
 388
 389 /*
 390  * we create compressed extents in two phases.  The first
 391  * phase compresses a range of pages that have already been
 392  * locked (both pages and state bits are locked).
 393  *
 394  * This is done inside an ordered work queue, and the compression
 395  * is spread across many cpus.  The actual IO submission is step
 396  * two, and the ordered work queue takes care of making sure that
 397  * happens in the same order things were put onto the queue by
 398  * writepages and friends.
 399  *
 400  * If this code finds it can't get good compression, it puts an
 401  * entry onto the work queue to write the uncompressed bytes.  This
 402  * makes sure that both compressed inodes and uncompressed inodes
 403  * are written in the same order that the flusher thread sent them
 404  * down.
 405  */
 406 static noinline void compress_file_range(struct inode *inode,
 407                                         struct page *locked_page,
 408                                         u64 start, u64 end,
 409                                         struct async_cow *async_cow,
 410                                         int *num_added)
 411 {
 412         struct btrfs_root *root = BTRFS_I(inode)->root;
 413         u64 num_bytes;
 414         u64 blocksize = root->sectorsize;
 415         u64 actual_end;
 416         u64 isize = i_size_read(inode);
 417         int ret = 0;
 418         struct page **pages = NULL;
 419         unsigned long nr_pages;
 420         unsigned long nr_pages_ret = 0;
 421         unsigned long total_compressed = 0;
 422         unsigned long total_in = 0;
 423         unsigned long max_compressed = SZ_128K;
 424         unsigned long max_uncompressed = SZ_128K;
 425         int i;
 426         int will_compress;
 427         int compress_type = root->fs_info->compress_type;
 428         int redirty = 0;
 429
 430         /* if this is a small write inside eof, kick off a defrag */
 431         if ((end - start + 1) < SZ_16K &&
 432             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434
 435         actual_end = min_t(u64, isize, end + 1);
 436 again:
 437         will_compress = 0;
 438         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 439         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_SIZE);
 440
 441         /*
 442          * we don't want to send crud past the end of i_size through
 443          * compression, that's just a waste of CPU time.  So, if the
 444          * end of the file is before the start of our current
 445          * requested range of bytes, we bail out to the uncompressed
 446          * cleanup code that can deal with all of this.
 447          *
 448          * It isn't really the fastest way to fix things, but this is a
 449          * very uncommon corner.
 450          */
 451         if (actual_end <= start)
 452                 goto cleanup_and_bail_uncompressed;
 453
 454         total_compressed = actual_end - start;
 455
 456         /*
 457          * skip compression for a small file range(<=blocksize) that
 458          * isn't an inline extent, since it doesn't save disk space at all.
 459          */
 460         if (total_compressed <= blocksize &&
 461            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 462                 goto cleanup_and_bail_uncompressed;
 463
 464         /* we want to make sure that amount of ram required to uncompress
 465          * an extent is reasonable, so we limit the total size in ram
 466          * of a compressed extent to 128k.  This is a crucial number
 467          * because it also controls how easily we can spread reads across
 468          * cpus for decompression.
 469          *
 470          * We also want to make sure the amount of IO required to do
 471          * a random read is reasonably small, so we limit the size of
 472          * a compressed extent to 128k.
 473          */
 474         total_compressed = min(total_compressed, max_uncompressed);
 475         num_bytes = ALIGN(end - start + 1, blocksize);
 476         num_bytes = max(blocksize,  num_bytes);
 477         total_in = 0;
 478         ret = 0;
 479
 480         /*
 481          * we do compression for mount -o compress and when the
 482          * inode has not been flagged as nocompress.  This flag can
 483          * change at any time if we discover bad compression ratios.
 484          */
 485         if (inode_need_compress(inode)) {
 486                 WARN_ON(pages);
 487                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 488                 if (!pages) {
 489                         /* just bail out to the uncompressed code */
 490                         goto cont;
 491                 }
 492
 493                 if (BTRFS_I(inode)->force_compress)
 494                         compress_type = BTRFS_I(inode)->force_compress;
 495
 496                 /*
 497                  * we need to call clear_page_dirty_for_io on each
 498                  * page in the range.  Otherwise applications with the file
 499                  * mmap'd can wander in and change the page contents while
 500                  * we are compressing them.
 501                  *
 502                  * If the compression fails for any reason, we set the pages
 503                  * dirty again later on.
 504                  */
 505                 extent_range_clear_dirty_for_io(inode, start, end);
 506                 redirty = 1;
 507                 ret = btrfs_compress_pages(compress_type,
 508                                            inode->i_mapping, start,
 509                                            total_compressed, pages,
 510                                            nr_pages, &nr_pages_ret,
 511                                            &total_in,
 512                                            &total_compressed,
 513                                            max_compressed);
 514
 515                 if (!ret) {
 516                         unsigned long offset = total_compressed &
 517                                 (PAGE_SIZE - 1);
 518                         struct page *page = pages[nr_pages_ret - 1];
 519                         char *kaddr;
 520
 521                         /* zero the tail end of the last page, we might be
 522                          * sending it down to disk
 523                          */
 524                         if (offset) {
 525                                 kaddr = kmap_atomic(page);
 526                                 memset(kaddr + offset, 0,
 527                                        PAGE_SIZE - offset);
 528                                 kunmap_atomic(kaddr);
 529                         }
 530                         will_compress = 1;
 531                 }
 532         }
 533 cont:
 534         if (start == 0) {
 535                 /* lets try to make an inline extent */
 536                 if (ret || total_in < (actual_end - start)) {
 537                         /* we didn't compress the entire range, try
 538                          * to make an uncompressed inline extent.
 539                          */
 540                         ret = cow_file_range_inline(root, inode, start, end,
 541                                                     0, 0, NULL);
 542                 } else {
 543                         /* try making a compressed inline extent */
 544                         ret = cow_file_range_inline(root, inode, start, end,
 545                                                     total_compressed,
 546                                                     compress_type, pages);
 547                 }
 548                 if (ret <= 0) {
 549                         unsigned long clear_flags = EXTENT_DELALLOC |
 550                                 EXTENT_DEFRAG;
 551                         unsigned long page_error_op;
 552
 553                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 554                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 555
 556                         /*
 557                          * inline extent creation worked or returned error,
 558                          * we don't need to create any more async work items.
 559                          * Unlock and free up our temp pages.
 560                          */
 561                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 562                                                      clear_flags, PAGE_UNLOCK |
 563                                                      PAGE_CLEAR_DIRTY |
 564                                                      PAGE_SET_WRITEBACK |
 565                                                      page_error_op |
 566                                                      PAGE_END_WRITEBACK);
 567                         goto free_pages_out;
 568                 }
 569         }
 570
 571         if (will_compress) {
 572                 /*
 573                  * we aren't doing an inline extent round the compressed size
 574                  * up to a block size boundary so the allocator does sane
 575                  * things
 576                  */
 577                 total_compressed = ALIGN(total_compressed, blocksize);
 578
 579                 /*
 580                  * one last check to make sure the compression is really a
 581                  * win, compare the page count read with the blocks on disk
 582                  */
 583                 total_in = ALIGN(total_in, PAGE_SIZE);
 584                 if (total_compressed >= total_in) {
 585                         will_compress = 0;
 586                 } else {
 587                         num_bytes = total_in;
 588                 }
 589         }
 590         if (!will_compress && pages) {
 591                 /*
 592                  * the compression code ran but failed to make things smaller,
 593                  * free any pages it allocated and our page pointer array
 594                  */
 595                 for (i = 0; i < nr_pages_ret; i++) {
 596                         WARN_ON(pages[i]->mapping);
 597                         put_page(pages[i]);
 598                 }
 599                 kfree(pages);
 600                 pages = NULL;
 601                 total_compressed = 0;
 602                 nr_pages_ret = 0;
 603
 604                 /* flag the file so we don't compress in the future */
 605                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 606                     !(BTRFS_I(inode)->force_compress)) {
 607                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 608                 }
 609         }
 610         if (will_compress) {
 611                 *num_added += 1;
 612
 613                 /* the async work queues will take care of doing actual
 614                  * allocation on disk for these compressed pages,
 615                  * and will submit them to the elevator.
 616                  */
 617                 add_async_extent(async_cow, start, num_bytes,
 618                                  total_compressed, pages, nr_pages_ret,
 619                                  compress_type);
 620
 621                 if (start + num_bytes < end) {
 622                         start += num_bytes;
 623                         pages = NULL;
 624                         cond_resched();
 625                         goto again;
 626                 }
 627         } else {
 628 cleanup_and_bail_uncompressed:
 629                 /*
 630                  * No compression, but we still need to write the pages in
 631                  * the file we've been given so far.  redirty the locked
 632                  * page if it corresponds to our extent and set things up
 633                  * for the async work queue to run cow_file_range to do
 634                  * the normal delalloc dance
 635                  */
 636                 if (page_offset(locked_page) >= start &&
 637                     page_offset(locked_page) <= end) {
 638                         __set_page_dirty_nobuffers(locked_page);
 639                         /* unlocked later on in the async handlers */
 640                 }
 641                 if (redirty)
 642                         extent_range_redirty_for_io(inode, start, end);
 643                 add_async_extent(async_cow, start, end - start + 1,
 644                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 645                 *num_added += 1;
 646         }
 647
 648         return;
 649
 650 free_pages_out:
 651         for (i = 0; i < nr_pages_ret; i++) {
 652                 WARN_ON(pages[i]->mapping);
 653                 put_page(pages[i]);
 654         }
 655         kfree(pages);
 656 }
 657
 658 static void free_async_extent_pages(struct async_extent *async_extent)
 659 {
 660         int i;
 661
 662         if (!async_extent->pages)
 663                 return;
 664
 665         for (i = 0; i < async_extent->nr_pages; i++) {
 666                 WARN_ON(async_extent->pages[i]->mapping);
 667                 put_page(async_extent->pages[i]);
 668         }
 669         kfree(async_extent->pages);
 670         async_extent->nr_pages = 0;
 671         async_extent->pages = NULL;
 672 }
 673
 674 /*
 675  * phase two of compressed writeback.  This is the ordered portion
 676  * of the code, which only gets called in the order the work was
 677  * queued.  We walk all the async extents created by compress_file_range
 678  * and send them down to the disk.
 679  */
 680 static noinline void submit_compressed_extents(struct inode *inode,
 681                                               struct async_cow *async_cow)
 682 {
 683         struct async_extent *async_extent;
 684         u64 alloc_hint = 0;
 685         struct btrfs_key ins;
 686         struct extent_map *em;
 687         struct btrfs_root *root = BTRFS_I(inode)->root;
 688         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 689         struct extent_io_tree *io_tree;
 690         int ret = 0;
 691
 692 again:
 693         while (!list_empty(&async_cow->extents)) {
 694                 async_extent = list_entry(async_cow->extents.next,
 695                                           struct async_extent, list);
 696                 list_del(&async_extent->list);
 697
 698                 io_tree = &BTRFS_I(inode)->io_tree;
 699
 700 retry:
 701                 /* did the compression code fall back to uncompressed IO? */
 702                 if (!async_extent->pages) {
 703                         int page_started = 0;
 704                         unsigned long nr_written = 0;
 705
 706                         lock_extent(io_tree, async_extent->start,
 707                                          async_extent->start +
 708                                          async_extent->ram_size - 1);
 709
 710                         /* allocate blocks */
 711                         ret = cow_file_range(inode, async_cow->locked_page,
 712                                              async_extent->start,
 713                                              async_extent->start +
 714                                              async_extent->ram_size - 1,
 715                                              &page_started, &nr_written, 0);
 716
 717                         /* JDM XXX */
 718
 719                         /*
 720                          * if page_started, cow_file_range inserted an
 721                          * inline extent and took care of all the unlocking
 722                          * and IO for us.  Otherwise, we need to submit
 723                          * all those pages down to the drive.
 724                          */
 725                         if (!page_started && !ret)
 726                                 extent_write_locked_range(io_tree,
 727                                                   inode, async_extent->start,
 728                                                   async_extent->start +
 729                                                   async_extent->ram_size - 1,
 730                                                   btrfs_get_extent,
 731                                                   WB_SYNC_ALL);
 732                         else if (ret)
 733                                 unlock_page(async_cow->locked_page);
 734                         kfree(async_extent);
 735                         cond_resched();
 736                         continue;
 737                 }
 738
 739                 lock_extent(io_tree, async_extent->start,
 740                             async_extent->start + async_extent->ram_size - 1);
 741
 742                 ret = btrfs_reserve_extent(root,
 743                                            async_extent->compressed_size,
 744                                            async_extent->compressed_size,
 745                                            0, alloc_hint, &ins, 1, 1);
 746                 if (ret) {
 747                         free_async_extent_pages(async_extent);
 748
 749                         if (ret == -ENOSPC) {
 750                                 unlock_extent(io_tree, async_extent->start,
 751                                               async_extent->start +
 752                                               async_extent->ram_size - 1);
 753
 754                                 /*
 755                                  * we need to redirty the pages if we decide to
 756                                  * fallback to uncompressed IO, otherwise we
 757                                  * will not submit these pages down to lower
 758                                  * layers.
 759                                  */
 760                                 extent_range_redirty_for_io(inode,
 761                                                 async_extent->start,
 762                                                 async_extent->start +
 763                                                 async_extent->ram_size - 1);
 764
 765                                 goto retry;
 766                         }
 767                         goto out_free;
 768                 }
 769                 /*
 770                  * here we're doing allocation and writeback of the
 771                  * compressed pages
 772                  */
 773                 btrfs_drop_extent_cache(inode, async_extent->start,
 774                                         async_extent->start +
 775                                         async_extent->ram_size - 1, 0);
 776
 777                 em = alloc_extent_map();
 778                 if (!em) {
 779                         ret = -ENOMEM;
 780                         goto out_free_reserve;
 781                 }
 782                 em->start = async_extent->start;
 783                 em->len = async_extent->ram_size;
 784                 em->orig_start = em->start;
 785                 em->mod_start = em->start;
 786                 em->mod_len = em->len;
 787
 788                 em->block_start = ins.objectid;
 789                 em->block_len = ins.offset;
 790                 em->orig_block_len = ins.offset;
 791                 em->ram_bytes = async_extent->ram_size;
 792                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 793                 em->compress_type = async_extent->compress_type;
 794                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 795                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 796                 em->generation = -1;
 797
 798                 while (1) {
 799                         write_lock(&em_tree->lock);
 800                         ret = add_extent_mapping(em_tree, em, 1);
 801                         write_unlock(&em_tree->lock);
 802                         if (ret != -EEXIST) {
 803                                 free_extent_map(em);
 804                                 break;
 805                         }
 806                         btrfs_drop_extent_cache(inode, async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1, 0);
 809                 }
 810
 811                 if (ret)
 812                         goto out_free_reserve;
 813
 814                 ret = btrfs_add_ordered_extent_compress(inode,
 815                                                 async_extent->start,
 816                                                 ins.objectid,
 817                                                 async_extent->ram_size,
 818                                                 ins.offset,
 819                                                 BTRFS_ORDERED_COMPRESSED,
 820                                                 async_extent->compress_type);
 821                 if (ret) {
 822                         btrfs_drop_extent_cache(inode, async_extent->start,
 823                                                 async_extent->start +
 824                                                 async_extent->ram_size - 1, 0);
 825                         goto out_free_reserve;
 826                 }
 827                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 828
 829                 /*
 830                  * clear dirty, set writeback and unlock the pages.
 831                  */
 832                 extent_clear_unlock_delalloc(inode, async_extent->start,
 833                                 async_extent->start +
 834                                 async_extent->ram_size - 1,
 835                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 836                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 837                                 PAGE_SET_WRITEBACK);
 838                 ret = btrfs_submit_compressed_write(inode,
 839                                     async_extent->start,
 840                                     async_extent->ram_size,
 841                                     ins.objectid,
 842                                     ins.offset, async_extent->pages,
 843                                     async_extent->nr_pages);
 844                 if (ret) {
 845                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 846                         struct page *p = async_extent->pages[0];
 847                         const u64 start = async_extent->start;
 848                         const u64 end = start + async_extent->ram_size - 1;
 849
 850                         p->mapping = inode->i_mapping;
 851                         tree->ops->writepage_end_io_hook(p, start, end,
 852                                                          NULL, 0);
 853                         p->mapping = NULL;
 854                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 855                                                      PAGE_END_WRITEBACK |
 856                                                      PAGE_SET_ERROR);
 857                         free_async_extent_pages(async_extent);
 858                 }
 859                 alloc_hint = ins.objectid + ins.offset;
 860                 kfree(async_extent);
 861                 cond_resched();
 862         }
 863         return;
 864 out_free_reserve:
 865         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 866         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 867 out_free:
 868         extent_clear_unlock_delalloc(inode, async_extent->start,
 869                                      async_extent->start +
 870                                      async_extent->ram_size - 1,
 871                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 872                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 873                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 874                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 875                                      PAGE_SET_ERROR);
 876         free_async_extent_pages(async_extent);
 877         kfree(async_extent);
 878         goto again;
 879 }
 880
 881 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 882                                       u64 num_bytes)
 883 {
 884         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 885         struct extent_map *em;
 886         u64 alloc_hint = 0;
 887
 888         read_lock(&em_tree->lock);
 889         em = search_extent_mapping(em_tree, start, num_bytes);
 890         if (em) {
 891                 /*
 892                  * if block start isn't an actual block number then find the
 893                  * first block in this inode and use that as a hint.  If that
 894                  * block is also bogus then just don't worry about it.
 895                  */
 896                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 897                         free_extent_map(em);
 898                         em = search_extent_mapping(em_tree, 0, 0);
 899                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 900                                 alloc_hint = em->block_start;
 901                         if (em)
 902                                 free_extent_map(em);
 903                 } else {
 904                         alloc_hint = em->block_start;
 905                         free_extent_map(em);
 906                 }
 907         }
 908         read_unlock(&em_tree->lock);
 909
 910         return alloc_hint;
 911 }
 912
 913 /*
 914  * when extent_io.c finds a delayed allocation range in the file,
 915  * the call backs end up in this code.  The basic idea is to
 916  * allocate extents on disk for the range, and create ordered data structs
 917  * in ram to track those extents.
 918  *
 919  * locked_page is the page that writepage had locked already.  We use
 920  * it to make sure we don't do extra locks or unlocks.
 921  *
 922  * *page_started is set to one if we unlock locked_page and do everything
 923  * required to start IO on it.  It may be clean and already done with
 924  * IO when we return.
 925  */
 926 static noinline int cow_file_range(struct inode *inode,
 927                                    struct page *locked_page,
 928                                    u64 start, u64 end, int *page_started,
 929                                    unsigned long *nr_written,
 930                                    int unlock)
 931 {
 932         struct btrfs_root *root = BTRFS_I(inode)->root;
 933         u64 alloc_hint = 0;
 934         u64 num_bytes;
 935         unsigned long ram_size;
 936         u64 disk_num_bytes;
 937         u64 cur_alloc_size;
 938         u64 blocksize = root->sectorsize;
 939         struct btrfs_key ins;
 940         struct extent_map *em;
 941         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 942         int ret = 0;
 943
 944         if (btrfs_is_free_space_inode(inode)) {
 945                 WARN_ON_ONCE(1);
 946                 ret = -EINVAL;
 947                 goto out_unlock;
 948         }
 949
 950         num_bytes = ALIGN(end - start + 1, blocksize);
 951         num_bytes = max(blocksize,  num_bytes);
 952         disk_num_bytes = num_bytes;
 953
 954         /* if this is a small write inside eof, kick off defrag */
 955         if (num_bytes < SZ_64K &&
 956             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 957                 btrfs_add_inode_defrag(NULL, inode);
 958
 959         if (start == 0) {
 960                 /* lets try to make an inline extent */
 961                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 962                                             NULL);
 963                 if (ret == 0) {
 964                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 965                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 966                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 967                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 968                                      PAGE_END_WRITEBACK);
 969
 970                         *nr_written = *nr_written +
 971                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 972                         *page_started = 1;
 973                         goto out;
 974                 } else if (ret < 0) {
 975                         goto out_unlock;
 976                 }
 977         }
 978
 979         BUG_ON(disk_num_bytes >
 980                btrfs_super_total_bytes(root->fs_info->super_copy));
 981
 982         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 983         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 984
 985         while (disk_num_bytes > 0) {
 986                 unsigned long op;
 987
 988                 cur_alloc_size = disk_num_bytes;
 989                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 990                                            root->sectorsize, 0, alloc_hint,
 991                                            &ins, 1, 1);
 992                 if (ret < 0)
 993                         goto out_unlock;
 994
 995                 em = alloc_extent_map();
 996                 if (!em) {
 997                         ret = -ENOMEM;
 998                         goto out_reserve;
 999                 }
1000                 em->start = start;
1001                 em->orig_start = em->start;
1002                 ram_size = ins.offset;
1003                 em->len = ins.offset;
1004                 em->mod_start = em->start;
1005                 em->mod_len = em->len;
1006
1007                 em->block_start = ins.objectid;
1008                 em->block_len = ins.offset;
1009                 em->orig_block_len = ins.offset;
1010                 em->ram_bytes = ram_size;
1011                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1012                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1013                 em->generation = -1;
1014
1015                 while (1) {
1016                         write_lock(&em_tree->lock);
1017                         ret = add_extent_mapping(em_tree, em, 1);
1018                         write_unlock(&em_tree->lock);
1019                         if (ret != -EEXIST) {
1020                                 free_extent_map(em);
1021                                 break;
1022                         }
1023                         btrfs_drop_extent_cache(inode, start,
1024                                                 start + ram_size - 1, 0);
1025                 }
1026                 if (ret)
1027                         goto out_reserve;
1028
1029                 cur_alloc_size = ins.offset;
1030                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1031                                                ram_size, cur_alloc_size, 0);
1032                 if (ret)
1033                         goto out_drop_extent_cache;
1034
1035                 if (root->root_key.objectid ==
1036                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1037                         ret = btrfs_reloc_clone_csums(inode, start,
1038                                                       cur_alloc_size);
1039                         if (ret)
1040                                 goto out_drop_extent_cache;
1041                 }
1042
1043                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1044
1045                 if (disk_num_bytes < cur_alloc_size)
1046                         break;
1047
1048                 /* we're not doing compressed IO, don't unlock the first
1049                  * page (which the caller expects to stay locked), don't
1050                  * clear any dirty bits and don't set any writeback bits
1051                  *
1052                  * Do set the Private2 bit so we know this page was properly
1053                  * setup for writepage
1054                  */
1055                 op = unlock ? PAGE_UNLOCK : 0;
1056                 op |= PAGE_SET_PRIVATE2;
1057
1058                 extent_clear_unlock_delalloc(inode, start,
1059                                              start + ram_size - 1, locked_page,
1060                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1061                                              op);
1062                 disk_num_bytes -= cur_alloc_size;
1063                 num_bytes -= cur_alloc_size;
1064                 alloc_hint = ins.objectid + ins.offset;
1065                 start += cur_alloc_size;
1066         }
1067 out:
1068         return ret;
1069
1070 out_drop_extent_cache:
1071         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1072 out_reserve:
1073         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1074         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1075 out_unlock:
1076         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1077                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1078                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1079                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1080                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1081         goto out;
1082 }
1083
1084 /*
1085  * work queue call back to started compression on a file and pages
1086  */
1087 static noinline void async_cow_start(struct btrfs_work *work)
1088 {
1089         struct async_cow *async_cow;
1090         int num_added = 0;
1091         async_cow = container_of(work, struct async_cow, work);
1092
1093         compress_file_range(async_cow->inode, async_cow->locked_page,
1094                             async_cow->start, async_cow->end, async_cow,
1095                             &num_added);
1096         if (num_added == 0) {
1097                 btrfs_add_delayed_iput(async_cow->inode);
1098                 async_cow->inode = NULL;
1099         }
1100 }
1101
1102 /*
1103  * work queue call back to submit previously compressed pages
1104  */
1105 static noinline void async_cow_submit(struct btrfs_work *work)
1106 {
1107         struct async_cow *async_cow;
1108         struct btrfs_root *root;
1109         unsigned long nr_pages;
1110
1111         async_cow = container_of(work, struct async_cow, work);
1112
1113         root = async_cow->root;
1114         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1115                 PAGE_SHIFT;
1116
1117         /*
1118          * atomic_sub_return implies a barrier for waitqueue_active
1119          */
1120         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1121             5 * SZ_1M &&
1122             waitqueue_active(&root->fs_info->async_submit_wait))
1123                 wake_up(&root->fs_info->async_submit_wait);
1124
1125         if (async_cow->inode)
1126                 submit_compressed_extents(async_cow->inode, async_cow);
1127 }
1128
1129 static noinline void async_cow_free(struct btrfs_work *work)
1130 {
1131         struct async_cow *async_cow;
1132         async_cow = container_of(work, struct async_cow, work);
1133         if (async_cow->inode)
1134                 btrfs_add_delayed_iput(async_cow->inode);
1135         kfree(async_cow);
1136 }
1137
1138 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1139                                 u64 start, u64 end, int *page_started,
1140                                 unsigned long *nr_written)
1141 {
1142         struct async_cow *async_cow;
1143         struct btrfs_root *root = BTRFS_I(inode)->root;
1144         unsigned long nr_pages;
1145         u64 cur_end;
1146         int limit = 10 * SZ_1M;
1147
1148         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1149                          1, 0, NULL, GFP_NOFS);
1150         while (start < end) {
1151                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1152                 BUG_ON(!async_cow); /* -ENOMEM */
1153                 async_cow->inode = igrab(inode);
1154                 async_cow->root = root;
1155                 async_cow->locked_page = locked_page;
1156                 async_cow->start = start;
1157
1158                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1159                     !btrfs_test_opt(root, FORCE_COMPRESS))
1160                         cur_end = end;
1161                 else
1162                         cur_end = min(end, start + SZ_512K - 1);
1163
1164                 async_cow->end = cur_end;
1165                 INIT_LIST_HEAD(&async_cow->extents);
1166
1167                 btrfs_init_work(&async_cow->work,
1168                                 btrfs_delalloc_helper,
1169                                 async_cow_start, async_cow_submit,
1170                                 async_cow_free);
1171
1172                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1173                         PAGE_SHIFT;
1174                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1175
1176                 btrfs_queue_work(root->fs_info->delalloc_workers,
1177                                  &async_cow->work);
1178
1179                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1180                         wait_event(root->fs_info->async_submit_wait,
1181                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1182                             limit));
1183                 }
1184
1185                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1186                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1187                         wait_event(root->fs_info->async_submit_wait,
1188                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1189                            0));
1190                 }
1191
1192                 *nr_written += nr_pages;
1193                 start = cur_end + 1;
1194         }
1195         *page_started = 1;
1196         return 0;
1197 }
1198
1199 static noinline int csum_exist_in_range(struct btrfs_root *root,
1200                                         u64 bytenr, u64 num_bytes)
1201 {
1202         int ret;
1203         struct btrfs_ordered_sum *sums;
1204         LIST_HEAD(list);
1205
1206         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1207                                        bytenr + num_bytes - 1, &list, 0);
1208         if (ret == 0 && list_empty(&list))
1209                 return 0;
1210
1211         while (!list_empty(&list)) {
1212                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1213                 list_del(&sums->list);
1214                 kfree(sums);
1215         }
1216         return 1;
1217 }
1218
1219 /*
1220  * when nowcow writeback call back.  This checks for snapshots or COW copies
1221  * of the extents that exist in the file, and COWs the file as required.
1222  *
1223  * If no cow copies or snapshots exist, we write directly to the existing
1224  * blocks on disk
1225  */
1226 static noinline int run_delalloc_nocow(struct inode *inode,
1227                                        struct page *locked_page,
1228                               u64 start, u64 end, int *page_started, int force,
1229                               unsigned long *nr_written)
1230 {
1231         struct btrfs_root *root = BTRFS_I(inode)->root;
1232         struct btrfs_trans_handle *trans;
1233         struct extent_buffer *leaf;
1234         struct btrfs_path *path;
1235         struct btrfs_file_extent_item *fi;
1236         struct btrfs_key found_key;
1237         u64 cow_start;
1238         u64 cur_offset;
1239         u64 extent_end;
1240         u64 extent_offset;
1241         u64 disk_bytenr;
1242         u64 num_bytes;
1243         u64 disk_num_bytes;
1244         u64 ram_bytes;
1245         int extent_type;
1246         int ret, err;
1247         int type;
1248         int nocow;
1249         int check_prev = 1;
1250         bool nolock;
1251         u64 ino = btrfs_ino(inode);
1252
1253         path = btrfs_alloc_path();
1254         if (!path) {
1255                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1256                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1257                                              EXTENT_DO_ACCOUNTING |
1258                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1259                                              PAGE_CLEAR_DIRTY |
1260                                              PAGE_SET_WRITEBACK |
1261                                              PAGE_END_WRITEBACK);
1262                 return -ENOMEM;
1263         }
1264
1265         nolock = btrfs_is_free_space_inode(inode);
1266
1267         if (nolock)
1268                 trans = btrfs_join_transaction_nolock(root);
1269         else
1270                 trans = btrfs_join_transaction(root);
1271
1272         if (IS_ERR(trans)) {
1273                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1274                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1275                                              EXTENT_DO_ACCOUNTING |
1276                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1277                                              PAGE_CLEAR_DIRTY |
1278                                              PAGE_SET_WRITEBACK |
1279                                              PAGE_END_WRITEBACK);
1280                 btrfs_free_path(path);
1281                 return PTR_ERR(trans);
1282         }
1283
1284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1285
1286         cow_start = (u64)-1;
1287         cur_offset = start;
1288         while (1) {
1289                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1290                                                cur_offset, 0);
1291                 if (ret < 0)
1292                         goto error;
1293                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1294                         leaf = path->nodes[0];
1295                         btrfs_item_key_to_cpu(leaf, &found_key,
1296                                               path->slots[0] - 1);
1297                         if (found_key.objectid == ino &&
1298                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1299                                 path->slots[0]--;
1300                 }
1301                 check_prev = 0;
1302 next_slot:
1303                 leaf = path->nodes[0];
1304                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1305                         ret = btrfs_next_leaf(root, path);
1306                         if (ret < 0)
1307                                 goto error;
1308                         if (ret > 0)
1309                                 break;
1310                         leaf = path->nodes[0];
1311                 }
1312
1313                 nocow = 0;
1314                 disk_bytenr = 0;
1315                 num_bytes = 0;
1316                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1317
1318                 if (found_key.objectid > ino)
1319                         break;
1320                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1321                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1322                         path->slots[0]++;
1323                         goto next_slot;
1324                 }
1325                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1326                     found_key.offset > end)
1327                         break;
1328
1329                 if (found_key.offset > cur_offset) {
1330                         extent_end = found_key.offset;
1331                         extent_type = 0;
1332                         goto out_check;
1333                 }
1334
1335                 fi = btrfs_item_ptr(leaf, path->slots[0],
1336                                     struct btrfs_file_extent_item);
1337                 extent_type = btrfs_file_extent_type(leaf, fi);
1338
1339                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1340                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1341                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1342                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1343                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1344                         extent_end = found_key.offset +
1345                                 btrfs_file_extent_num_bytes(leaf, fi);
1346                         disk_num_bytes =
1347                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1348                         if (extent_end <= start) {
1349                                 path->slots[0]++;
1350                                 goto next_slot;
1351                         }
1352                         if (disk_bytenr == 0)
1353                                 goto out_check;
1354                         if (btrfs_file_extent_compression(leaf, fi) ||
1355                             btrfs_file_extent_encryption(leaf, fi) ||
1356                             btrfs_file_extent_other_encoding(leaf, fi))
1357                                 goto out_check;
1358                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1359                                 goto out_check;
1360                         if (btrfs_extent_readonly(root, disk_bytenr))
1361                                 goto out_check;
1362                         if (btrfs_cross_ref_exist(trans, root, ino,
1363                                                   found_key.offset -
1364                                                   extent_offset, disk_bytenr))
1365                                 goto out_check;
1366                         disk_bytenr += extent_offset;
1367                         disk_bytenr += cur_offset - found_key.offset;
1368                         num_bytes = min(end + 1, extent_end) - cur_offset;
1369                         /*
1370                          * if there are pending snapshots for this root,
1371                          * we fall into common COW way.
1372                          */
1373                         if (!nolock) {
1374                                 err = btrfs_start_write_no_snapshoting(root);
1375                                 if (!err)
1376                                         goto out_check;
1377                         }
1378                         /*
1379                          * force cow if csum exists in the range.
1380                          * this ensure that csum for a given extent are
1381                          * either valid or do not exist.
1382                          */
1383                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1384                                 goto out_check;
1385                         if (!btrfs_inc_nocow_writers(root->fs_info,
1386                                                      disk_bytenr))
1387                                 goto out_check;
1388                         nocow = 1;
1389                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1390                         extent_end = found_key.offset +
1391                                 btrfs_file_extent_inline_len(leaf,
1392                                                      path->slots[0], fi);
1393                         extent_end = ALIGN(extent_end, root->sectorsize);
1394                 } else {
1395                         BUG_ON(1);
1396                 }
1397 out_check:
1398                 if (extent_end <= start) {
1399                         path->slots[0]++;
1400                         if (!nolock && nocow)
1401                                 btrfs_end_write_no_snapshoting(root);
1402                         if (nocow)
1403                                 btrfs_dec_nocow_writers(root->fs_info,
1404                                                         disk_bytenr);
1405                         goto next_slot;
1406                 }
1407                 if (!nocow) {
1408                         if (cow_start == (u64)-1)
1409                                 cow_start = cur_offset;
1410                         cur_offset = extent_end;
1411                         if (cur_offset > end)
1412                                 break;
1413                         path->slots[0]++;
1414                         goto next_slot;
1415                 }
1416
1417                 btrfs_release_path(path);
1418                 if (cow_start != (u64)-1) {
1419                         ret = cow_file_range(inode, locked_page,
1420                                              cow_start, found_key.offset - 1,
1421                                              page_started, nr_written, 1);
1422                         if (ret) {
1423                                 if (!nolock && nocow)
1424                                         btrfs_end_write_no_snapshoting(root);
1425                                 if (nocow)
1426                                         btrfs_dec_nocow_writers(root->fs_info,
1427                                                                 disk_bytenr);
1428                                 goto error;
1429                         }
1430                         cow_start = (u64)-1;
1431                 }
1432
1433                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1434                         struct extent_map *em;
1435                         struct extent_map_tree *em_tree;
1436                         em_tree = &BTRFS_I(inode)->extent_tree;
1437                         em = alloc_extent_map();
1438                         BUG_ON(!em); /* -ENOMEM */
1439                         em->start = cur_offset;
1440                         em->orig_start = found_key.offset - extent_offset;
1441                         em->len = num_bytes;
1442                         em->block_len = num_bytes;
1443                         em->block_start = disk_bytenr;
1444                         em->orig_block_len = disk_num_bytes;
1445                         em->ram_bytes = ram_bytes;
1446                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1447                         em->mod_start = em->start;
1448                         em->mod_len = em->len;
1449                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1450                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1451                         em->generation = -1;
1452                         while (1) {
1453                                 write_lock(&em_tree->lock);
1454                                 ret = add_extent_mapping(em_tree, em, 1);
1455                                 write_unlock(&em_tree->lock);
1456                                 if (ret != -EEXIST) {
1457                                         free_extent_map(em);
1458                                         break;
1459                                 }
1460                                 btrfs_drop_extent_cache(inode, em->start,
1461                                                 em->start + em->len - 1, 0);
1462                         }
1463                         type = BTRFS_ORDERED_PREALLOC;
1464                 } else {
1465                         type = BTRFS_ORDERED_NOCOW;
1466                 }
1467
1468                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1469                                                num_bytes, num_bytes, type);
1470                 if (nocow)
1471                         btrfs_dec_nocow_writers(root->fs_info, disk_bytenr);
1472                 BUG_ON(ret); /* -ENOMEM */
1473
1474                 if (root->root_key.objectid ==
1475                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1476                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1477                                                       num_bytes);
1478                         if (ret) {
1479                                 if (!nolock && nocow)
1480                                         btrfs_end_write_no_snapshoting(root);
1481                                 goto error;
1482                         }
1483                 }
1484
1485                 extent_clear_unlock_delalloc(inode, cur_offset,
1486                                              cur_offset + num_bytes - 1,
1487                                              locked_page, EXTENT_LOCKED |
1488                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1489                                              PAGE_SET_PRIVATE2);
1490                 if (!nolock && nocow)
1491                         btrfs_end_write_no_snapshoting(root);
1492                 cur_offset = extent_end;
1493                 if (cur_offset > end)
1494                         break;
1495         }
1496         btrfs_release_path(path);
1497
1498         if (cur_offset <= end && cow_start == (u64)-1) {
1499                 cow_start = cur_offset;
1500                 cur_offset = end;
1501         }
1502
1503         if (cow_start != (u64)-1) {
1504                 ret = cow_file_range(inode, locked_page, cow_start, end,
1505                                      page_started, nr_written, 1);
1506                 if (ret)
1507                         goto error;
1508         }
1509
1510 error:
1511         err = btrfs_end_transaction(trans, root);
1512         if (!ret)
1513                 ret = err;
1514
1515         if (ret && cur_offset < end)
1516                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1517                                              locked_page, EXTENT_LOCKED |
1518                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1519                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1520                                              PAGE_CLEAR_DIRTY |
1521                                              PAGE_SET_WRITEBACK |
1522                                              PAGE_END_WRITEBACK);
1523         btrfs_free_path(path);
1524         return ret;
1525 }
1526
1527 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1528 {
1529
1530         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1531             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1532                 return 0;
1533
1534         /*
1535          * @defrag_bytes is a hint value, no spinlock held here,
1536          * if is not zero, it means the file is defragging.
1537          * Force cow if given extent needs to be defragged.
1538          */
1539         if (BTRFS_I(inode)->defrag_bytes &&
1540             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1541                            EXTENT_DEFRAG, 0, NULL))
1542                 return 1;
1543
1544         return 0;
1545 }
1546
1547 /*
1548  * extent_io.c call back to do delayed allocation processing
1549  */
1550 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1551                               u64 start, u64 end, int *page_started,
1552                               unsigned long *nr_written)
1553 {
1554         int ret;
1555         int force_cow = need_force_cow(inode, start, end);
1556
1557         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1558                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1559                                          page_started, 1, nr_written);
1560         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1561                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1562                                          page_started, 0, nr_written);
1563         } else if (!inode_need_compress(inode)) {
1564                 ret = cow_file_range(inode, locked_page, start, end,
1565                                       page_started, nr_written, 1);
1566         } else {
1567                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1568                         &BTRFS_I(inode)->runtime_flags);
1569                 ret = cow_file_range_async(inode, locked_page, start, end,
1570                                            page_started, nr_written);
1571         }
1572         return ret;
1573 }
1574
1575 static void btrfs_split_extent_hook(struct inode *inode,
1576                                     struct extent_state *orig, u64 split)
1577 {
1578         u64 size;
1579
1580         /* not delalloc, ignore it */
1581         if (!(orig->state & EXTENT_DELALLOC))
1582                 return;
1583
1584         size = orig->end - orig->start + 1;
1585         if (size > BTRFS_MAX_EXTENT_SIZE) {
1586                 u64 num_extents;
1587                 u64 new_size;
1588
1589                 /*
1590                  * See the explanation in btrfs_merge_extent_hook, the same
1591                  * applies here, just in reverse.
1592                  */
1593                 new_size = orig->end - split + 1;
1594                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1595                                         BTRFS_MAX_EXTENT_SIZE);
1596                 new_size = split - orig->start;
1597                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1598                                         BTRFS_MAX_EXTENT_SIZE);
1599                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1600                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1601                         return;
1602         }
1603
1604         spin_lock(&BTRFS_I(inode)->lock);
1605         BTRFS_I(inode)->outstanding_extents++;
1606         spin_unlock(&BTRFS_I(inode)->lock);
1607 }
1608
1609 /*
1610  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1611  * extents so we can keep track of new extents that are just merged onto old
1612  * extents, such as when we are doing sequential writes, so we can properly
1613  * account for the metadata space we'll need.
1614  */
1615 static void btrfs_merge_extent_hook(struct inode *inode,
1616                                     struct extent_state *new,
1617                                     struct extent_state *other)
1618 {
1619         u64 new_size, old_size;
1620         u64 num_extents;
1621
1622         /* not delalloc, ignore it */
1623         if (!(other->state & EXTENT_DELALLOC))
1624                 return;
1625
1626         if (new->start > other->start)
1627                 new_size = new->end - other->start + 1;
1628         else
1629                 new_size = other->end - new->start + 1;
1630
1631         /* we're not bigger than the max, unreserve the space and go */
1632         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1633                 spin_lock(&BTRFS_I(inode)->lock);
1634                 BTRFS_I(inode)->outstanding_extents--;
1635                 spin_unlock(&BTRFS_I(inode)->lock);
1636                 return;
1637         }
1638
1639         /*
1640          * We have to add up either side to figure out how many extents were
1641          * accounted for before we merged into one big extent.  If the number of
1642          * extents we accounted for is <= the amount we need for the new range
1643          * then we can return, otherwise drop.  Think of it like this
1644          *
1645          * [ 4k][MAX_SIZE]
1646          *
1647          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1648          * need 2 outstanding extents, on one side we have 1 and the other side
1649          * we have 1 so they are == and we can return.  But in this case
1650          *
1651          * [MAX_SIZE+4k][MAX_SIZE+4k]
1652          *
1653          * Each range on their own accounts for 2 extents, but merged together
1654          * they are only 3 extents worth of accounting, so we need to drop in
1655          * this case.
1656          */
1657         old_size = other->end - other->start + 1;
1658         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1659                                 BTRFS_MAX_EXTENT_SIZE);
1660         old_size = new->end - new->start + 1;
1661         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1662                                  BTRFS_MAX_EXTENT_SIZE);
1663
1664         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1665                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1666                 return;
1667
1668         spin_lock(&BTRFS_I(inode)->lock);
1669         BTRFS_I(inode)->outstanding_extents--;
1670         spin_unlock(&BTRFS_I(inode)->lock);
1671 }
1672
1673 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1674                                       struct inode *inode)
1675 {
1676         spin_lock(&root->delalloc_lock);
1677         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1678                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1679                               &root->delalloc_inodes);
1680                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1681                         &BTRFS_I(inode)->runtime_flags);
1682                 root->nr_delalloc_inodes++;
1683                 if (root->nr_delalloc_inodes == 1) {
1684                         spin_lock(&root->fs_info->delalloc_root_lock);
1685                         BUG_ON(!list_empty(&root->delalloc_root));
1686                         list_add_tail(&root->delalloc_root,
1687                                       &root->fs_info->delalloc_roots);
1688                         spin_unlock(&root->fs_info->delalloc_root_lock);
1689                 }
1690         }
1691         spin_unlock(&root->delalloc_lock);
1692 }
1693
1694 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1695                                      struct inode *inode)
1696 {
1697         spin_lock(&root->delalloc_lock);
1698         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1699                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1700                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1701                           &BTRFS_I(inode)->runtime_flags);
1702                 root->nr_delalloc_inodes--;
1703                 if (!root->nr_delalloc_inodes) {
1704                         spin_lock(&root->fs_info->delalloc_root_lock);
1705                         BUG_ON(list_empty(&root->delalloc_root));
1706                         list_del_init(&root->delalloc_root);
1707                         spin_unlock(&root->fs_info->delalloc_root_lock);
1708                 }
1709         }
1710         spin_unlock(&root->delalloc_lock);
1711 }
1712
1713 /*
1714  * extent_io.c set_bit_hook, used to track delayed allocation
1715  * bytes in this file, and to maintain the list of inodes that
1716  * have pending delalloc work to be done.
1717  */
1718 static void btrfs_set_bit_hook(struct inode *inode,
1719                                struct extent_state *state, unsigned *bits)
1720 {
1721
1722         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1723                 WARN_ON(1);
1724         /*
1725          * set_bit and clear bit hooks normally require _irqsave/restore
1726          * but in this case, we are only testing for the DELALLOC
1727          * bit, which is only set or cleared with irqs on
1728          */
1729         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1730                 struct btrfs_root *root = BTRFS_I(inode)->root;
1731                 u64 len = state->end + 1 - state->start;
1732                 bool do_list = !btrfs_is_free_space_inode(inode);
1733
1734                 if (*bits & EXTENT_FIRST_DELALLOC) {
1735                         *bits &= ~EXTENT_FIRST_DELALLOC;
1736                 } else {
1737                         spin_lock(&BTRFS_I(inode)->lock);
1738                         BTRFS_I(inode)->outstanding_extents++;
1739                         spin_unlock(&BTRFS_I(inode)->lock);
1740                 }
1741
1742                 /* For sanity tests */
1743                 if (btrfs_test_is_dummy_root(root))
1744                         return;
1745
1746                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1747                                      root->fs_info->delalloc_batch);
1748                 spin_lock(&BTRFS_I(inode)->lock);
1749                 BTRFS_I(inode)->delalloc_bytes += len;
1750                 if (*bits & EXTENT_DEFRAG)
1751                         BTRFS_I(inode)->defrag_bytes += len;
1752                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1753                                          &BTRFS_I(inode)->runtime_flags))
1754                         btrfs_add_delalloc_inodes(root, inode);
1755                 spin_unlock(&BTRFS_I(inode)->lock);
1756         }
1757 }
1758
1759 /*
1760  * extent_io.c clear_bit_hook, see set_bit_hook for why
1761  */
1762 static void btrfs_clear_bit_hook(struct inode *inode,
1763                                  struct extent_state *state,
1764                                  unsigned *bits)
1765 {
1766         u64 len = state->end + 1 - state->start;
1767         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1768                                     BTRFS_MAX_EXTENT_SIZE);
1769
1770         spin_lock(&BTRFS_I(inode)->lock);
1771         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1772                 BTRFS_I(inode)->defrag_bytes -= len;
1773         spin_unlock(&BTRFS_I(inode)->lock);
1774
1775         /*
1776          * set_bit and clear bit hooks normally require _irqsave/restore
1777          * but in this case, we are only testing for the DELALLOC
1778          * bit, which is only set or cleared with irqs on
1779          */
1780         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1781                 struct btrfs_root *root = BTRFS_I(inode)->root;
1782                 bool do_list = !btrfs_is_free_space_inode(inode);
1783
1784                 if (*bits & EXTENT_FIRST_DELALLOC) {
1785                         *bits &= ~EXTENT_FIRST_DELALLOC;
1786                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1787                         spin_lock(&BTRFS_I(inode)->lock);
1788                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1789                         spin_unlock(&BTRFS_I(inode)->lock);
1790                 }
1791
1792                 /*
1793                  * We don't reserve metadata space for space cache inodes so we
1794                  * don't need to call dellalloc_release_metadata if there is an
1795                  * error.
1796                  */
1797                 if (*bits & EXTENT_DO_ACCOUNTING &&
1798                     root != root->fs_info->tree_root)
1799                         btrfs_delalloc_release_metadata(inode, len);
1800
1801                 /* For sanity tests. */
1802                 if (btrfs_test_is_dummy_root(root))
1803                         return;
1804
1805                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1806                     && do_list && !(state->state & EXTENT_NORESERVE))
1807                         btrfs_free_reserved_data_space_noquota(inode,
1808                                         state->start, len);
1809
1810                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1811                                      root->fs_info->delalloc_batch);
1812                 spin_lock(&BTRFS_I(inode)->lock);
1813                 BTRFS_I(inode)->delalloc_bytes -= len;
1814                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1815                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1816                              &BTRFS_I(inode)->runtime_flags))
1817                         btrfs_del_delalloc_inode(root, inode);
1818                 spin_unlock(&BTRFS_I(inode)->lock);
1819         }
1820 }
1821
1822 /*
1823  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1824  * we don't create bios that span stripes or chunks
1825  */
1826 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1827                          size_t size, struct bio *bio,
1828                          unsigned long bio_flags)
1829 {
1830         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1831         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1832         u64 length = 0;
1833         u64 map_length;
1834         int ret;
1835
1836         if (bio_flags & EXTENT_BIO_COMPRESSED)
1837                 return 0;
1838
1839         length = bio->bi_iter.bi_size;
1840         map_length = length;
1841         ret = btrfs_map_block(root->fs_info, rw, logical,
1842                               &map_length, NULL, 0);
1843         /* Will always return 0 with map_multi == NULL */
1844         BUG_ON(ret < 0);
1845         if (map_length < length + size)
1846                 return 1;
1847         return 0;
1848 }
1849
1850 /*
1851  * in order to insert checksums into the metadata in large chunks,
1852  * we wait until bio submission time.   All the pages in the bio are
1853  * checksummed and sums are attached onto the ordered extent record.
1854  *
1855  * At IO completion time the cums attached on the ordered extent record
1856  * are inserted into the btree
1857  */
1858 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1859                                     struct bio *bio, int mirror_num,
1860                                     unsigned long bio_flags,
1861                                     u64 bio_offset)
1862 {
1863         struct btrfs_root *root = BTRFS_I(inode)->root;
1864         int ret = 0;
1865
1866         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1867         BUG_ON(ret); /* -ENOMEM */
1868         return 0;
1869 }
1870
1871 /*
1872  * in order to insert checksums into the metadata in large chunks,
1873  * we wait until bio submission time.   All the pages in the bio are
1874  * checksummed and sums are attached onto the ordered extent record.
1875  *
1876  * At IO completion time the cums attached on the ordered extent record
1877  * are inserted into the btree
1878  */
1879 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1880                           int mirror_num, unsigned long bio_flags,
1881                           u64 bio_offset)
1882 {
1883         struct btrfs_root *root = BTRFS_I(inode)->root;
1884         int ret;
1885
1886         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1887         if (ret) {
1888                 bio->bi_error = ret;
1889                 bio_endio(bio);
1890         }
1891         return ret;
1892 }
1893
1894 /*
1895  * extent_io.c submission hook. This does the right thing for csum calculation
1896  * on write, or reading the csums from the tree before a read
1897  */
1898 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1899                           int mirror_num, unsigned long bio_flags,
1900                           u64 bio_offset)
1901 {
1902         struct btrfs_root *root = BTRFS_I(inode)->root;
1903         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1904         int ret = 0;
1905         int skip_sum;
1906         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1907
1908         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1909
1910         if (btrfs_is_free_space_inode(inode))
1911                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1912
1913         if (!(rw & REQ_WRITE)) {
1914                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1915                 if (ret)
1916                         goto out;
1917
1918                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1919                         ret = btrfs_submit_compressed_read(inode, bio,
1920                                                            mirror_num,
1921                                                            bio_flags);
1922                         goto out;
1923                 } else if (!skip_sum) {
1924                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1925                         if (ret)
1926                                 goto out;
1927                 }
1928                 goto mapit;
1929         } else if (async && !skip_sum) {
1930                 /* csum items have already been cloned */
1931                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1932                         goto mapit;
1933                 /* we're doing a write, do the async checksumming */
1934                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1935                                    inode, rw, bio, mirror_num,
1936                                    bio_flags, bio_offset,
1937                                    __btrfs_submit_bio_start,
1938                                    __btrfs_submit_bio_done);
1939                 goto out;
1940         } else if (!skip_sum) {
1941                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1942                 if (ret)
1943                         goto out;
1944         }
1945
1946 mapit:
1947         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1948
1949 out:
1950         if (ret < 0) {
1951                 bio->bi_error = ret;
1952                 bio_endio(bio);
1953         }
1954         return ret;
1955 }
1956
1957 /*
1958  * given a list of ordered sums record them in the inode.  This happens
1959  * at IO completion time based on sums calculated at bio submission time.
1960  */
1961 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1962                              struct inode *inode, u64 file_offset,
1963                              struct list_head *list)
1964 {
1965         struct btrfs_ordered_sum *sum;
1966
1967         list_for_each_entry(sum, list, list) {
1968                 trans->adding_csums = 1;
1969                 btrfs_csum_file_blocks(trans,
1970                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1971                 trans->adding_csums = 0;
1972         }
1973         return 0;
1974 }
1975
1976 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1977                               struct extent_state **cached_state)
1978 {
1979         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
1980         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1981                                    cached_state);
1982 }
1983
1984 /* see btrfs_writepage_start_hook for details on why this is required */
1985 struct btrfs_writepage_fixup {
1986         struct page *page;
1987         struct btrfs_work work;
1988 };
1989
1990 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1991 {
1992         struct btrfs_writepage_fixup *fixup;
1993         struct btrfs_ordered_extent *ordered;
1994         struct extent_state *cached_state = NULL;
1995         struct page *page;
1996         struct inode *inode;
1997         u64 page_start;
1998         u64 page_end;
1999         int ret;
2000
2001         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2002         page = fixup->page;
2003 again:
2004         lock_page(page);
2005         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2006                 ClearPageChecked(page);
2007                 goto out_page;
2008         }
2009
2010         inode = page->mapping->host;
2011         page_start = page_offset(page);
2012         page_end = page_offset(page) + PAGE_SIZE - 1;
2013
2014         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2015                          &cached_state);
2016
2017         /* already ordered? We're done */
2018         if (PagePrivate2(page))
2019                 goto out;
2020
2021         ordered = btrfs_lookup_ordered_range(inode, page_start,
2022                                         PAGE_SIZE);
2023         if (ordered) {
2024                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2025                                      page_end, &cached_state, GFP_NOFS);
2026                 unlock_page(page);
2027                 btrfs_start_ordered_extent(inode, ordered, 1);
2028                 btrfs_put_ordered_extent(ordered);
2029                 goto again;
2030         }
2031
2032         ret = btrfs_delalloc_reserve_space(inode, page_start,
2033                                            PAGE_SIZE);
2034         if (ret) {
2035                 mapping_set_error(page->mapping, ret);
2036                 end_extent_writepage(page, ret, page_start, page_end);
2037                 ClearPageChecked(page);
2038                 goto out;
2039          }
2040
2041         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2042         ClearPageChecked(page);
2043         set_page_dirty(page);
2044 out:
2045         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2046                              &cached_state, GFP_NOFS);
2047 out_page:
2048         unlock_page(page);
2049         put_page(page);
2050         kfree(fixup);
2051 }
2052
2053 /*
2054  * There are a few paths in the higher layers of the kernel that directly
2055  * set the page dirty bit without asking the filesystem if it is a
2056  * good idea.  This causes problems because we want to make sure COW
2057  * properly happens and the data=ordered rules are followed.
2058  *
2059  * In our case any range that doesn't have the ORDERED bit set
2060  * hasn't been properly setup for IO.  We kick off an async process
2061  * to fix it up.  The async helper will wait for ordered extents, set
2062  * the delalloc bit and make it safe to write the page.
2063  */
2064 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2065 {
2066         struct inode *inode = page->mapping->host;
2067         struct btrfs_writepage_fixup *fixup;
2068         struct btrfs_root *root = BTRFS_I(inode)->root;
2069
2070         /* this page is properly in the ordered list */
2071         if (TestClearPagePrivate2(page))
2072                 return 0;
2073
2074         if (PageChecked(page))
2075                 return -EAGAIN;
2076
2077         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2078         if (!fixup)
2079                 return -EAGAIN;
2080
2081         SetPageChecked(page);
2082         get_page(page);
2083         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2084                         btrfs_writepage_fixup_worker, NULL, NULL);
2085         fixup->page = page;
2086         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2087         return -EBUSY;
2088 }
2089
2090 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2091                                        struct inode *inode, u64 file_pos,
2092                                        u64 disk_bytenr, u64 disk_num_bytes,
2093                                        u64 num_bytes, u64 ram_bytes,
2094                                        u8 compression, u8 encryption,
2095                                        u16 other_encoding, int extent_type)
2096 {
2097         struct btrfs_root *root = BTRFS_I(inode)->root;
2098         struct btrfs_file_extent_item *fi;
2099         struct btrfs_path *path;
2100         struct extent_buffer *leaf;
2101         struct btrfs_key ins;
2102         int extent_inserted = 0;
2103         int ret;
2104
2105         path = btrfs_alloc_path();
2106         if (!path)
2107                 return -ENOMEM;
2108
2109         /*
2110          * we may be replacing one extent in the tree with another.
2111          * The new extent is pinned in the extent map, and we don't want
2112          * to drop it from the cache until it is completely in the btree.
2113          *
2114          * So, tell btrfs_drop_extents to leave this extent in the cache.
2115          * the caller is expected to unpin it and allow it to be merged
2116          * with the others.
2117          */
2118         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2119                                    file_pos + num_bytes, NULL, 0,
2120                                    1, sizeof(*fi), &extent_inserted);
2121         if (ret)
2122                 goto out;
2123
2124         if (!extent_inserted) {
2125                 ins.objectid = btrfs_ino(inode);
2126                 ins.offset = file_pos;
2127                 ins.type = BTRFS_EXTENT_DATA_KEY;
2128
2129                 path->leave_spinning = 1;
2130                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2131                                               sizeof(*fi));
2132                 if (ret)
2133                         goto out;
2134         }
2135         leaf = path->nodes[0];
2136         fi = btrfs_item_ptr(leaf, path->slots[0],
2137                             struct btrfs_file_extent_item);
2138         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2139         btrfs_set_file_extent_type(leaf, fi, extent_type);
2140         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2141         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2142         btrfs_set_file_extent_offset(leaf, fi, 0);
2143         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2144         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2145         btrfs_set_file_extent_compression(leaf, fi, compression);
2146         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2147         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2148
2149         btrfs_mark_buffer_dirty(leaf);
2150         btrfs_release_path(path);
2151
2152         inode_add_bytes(inode, num_bytes);
2153
2154         ins.objectid = disk_bytenr;
2155         ins.offset = disk_num_bytes;
2156         ins.type = BTRFS_EXTENT_ITEM_KEY;
2157         ret = btrfs_alloc_reserved_file_extent(trans, root,
2158                                         root->root_key.objectid,
2159                                         btrfs_ino(inode), file_pos,
2160                                         ram_bytes, &ins);
2161         /*
2162          * Release the reserved range from inode dirty range map, as it is
2163          * already moved into delayed_ref_head
2164          */
2165         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2166 out:
2167         btrfs_free_path(path);
2168
2169         return ret;
2170 }
2171
2172 /* snapshot-aware defrag */
2173 struct sa_defrag_extent_backref {
2174         struct rb_node node;
2175         struct old_sa_defrag_extent *old;
2176         u64 root_id;
2177         u64 inum;
2178         u64 file_pos;
2179         u64 extent_offset;
2180         u64 num_bytes;
2181         u64 generation;
2182 };
2183
2184 struct old_sa_defrag_extent {
2185         struct list_head list;
2186         struct new_sa_defrag_extent *new;
2187
2188         u64 extent_offset;
2189         u64 bytenr;
2190         u64 offset;
2191         u64 len;
2192         int count;
2193 };
2194
2195 struct new_sa_defrag_extent {
2196         struct rb_root root;
2197         struct list_head head;
2198         struct btrfs_path *path;
2199         struct inode *inode;
2200         u64 file_pos;
2201         u64 len;
2202         u64 bytenr;
2203         u64 disk_len;
2204         u8 compress_type;
2205 };
2206
2207 static int backref_comp(struct sa_defrag_extent_backref *b1,
2208                         struct sa_defrag_extent_backref *b2)
2209 {
2210         if (b1->root_id < b2->root_id)
2211                 return -1;
2212         else if (b1->root_id > b2->root_id)
2213                 return 1;
2214
2215         if (b1->inum < b2->inum)
2216                 return -1;
2217         else if (b1->inum > b2->inum)
2218                 return 1;
2219
2220         if (b1->file_pos < b2->file_pos)
2221                 return -1;
2222         else if (b1->file_pos > b2->file_pos)
2223                 return 1;
2224
2225         /*
2226          * [------------------------------] ===> (a range of space)
2227          *     |<--->|   |<---->| =============> (fs/file tree A)
2228          * |<---------------------------->| ===> (fs/file tree B)
2229          *
2230          * A range of space can refer to two file extents in one tree while
2231          * refer to only one file extent in another tree.
2232          *
2233          * So we may process a disk offset more than one time(two extents in A)
2234          * and locate at the same extent(one extent in B), then insert two same
2235          * backrefs(both refer to the extent in B).
2236          */
2237         return 0;
2238 }
2239
2240 static void backref_insert(struct rb_root *root,
2241                            struct sa_defrag_extent_backref *backref)
2242 {
2243         struct rb_node **p = &root->rb_node;
2244         struct rb_node *parent = NULL;
2245         struct sa_defrag_extent_backref *entry;
2246         int ret;
2247
2248         while (*p) {
2249                 parent = *p;
2250                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2251
2252                 ret = backref_comp(backref, entry);
2253                 if (ret < 0)
2254                         p = &(*p)->rb_left;
2255                 else
2256                         p = &(*p)->rb_right;
2257         }
2258
2259         rb_link_node(&backref->node, parent, p);
2260         rb_insert_color(&backref->node, root);
2261 }
2262
2263 /*
2264  * Note the backref might has changed, and in this case we just return 0.
2265  */
2266 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2267                                        void *ctx)
2268 {
2269         struct btrfs_file_extent_item *extent;
2270         struct btrfs_fs_info *fs_info;
2271         struct old_sa_defrag_extent *old = ctx;
2272         struct new_sa_defrag_extent *new = old->new;
2273         struct btrfs_path *path = new->path;
2274         struct btrfs_key key;
2275         struct btrfs_root *root;
2276         struct sa_defrag_extent_backref *backref;
2277         struct extent_buffer *leaf;
2278         struct inode *inode = new->inode;
2279         int slot;
2280         int ret;
2281         u64 extent_offset;
2282         u64 num_bytes;
2283
2284         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2285             inum == btrfs_ino(inode))
2286                 return 0;
2287
2288         key.objectid = root_id;
2289         key.type = BTRFS_ROOT_ITEM_KEY;
2290         key.offset = (u64)-1;
2291
2292         fs_info = BTRFS_I(inode)->root->fs_info;
2293         root = btrfs_read_fs_root_no_name(fs_info, &key);
2294         if (IS_ERR(root)) {
2295                 if (PTR_ERR(root) == -ENOENT)
2296                         return 0;
2297                 WARN_ON(1);
2298                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2299                          inum, offset, root_id);
2300                 return PTR_ERR(root);
2301         }
2302
2303         key.objectid = inum;
2304         key.type = BTRFS_EXTENT_DATA_KEY;
2305         if (offset > (u64)-1 << 32)
2306                 key.offset = 0;
2307         else
2308                 key.offset = offset;
2309
2310         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2311         if (WARN_ON(ret < 0))
2312                 return ret;
2313         ret = 0;
2314
2315         while (1) {
2316                 cond_resched();
2317
2318                 leaf = path->nodes[0];
2319                 slot = path->slots[0];
2320
2321                 if (slot >= btrfs_header_nritems(leaf)) {
2322                         ret = btrfs_next_leaf(root, path);
2323                         if (ret < 0) {
2324                                 goto out;
2325                         } else if (ret > 0) {
2326                                 ret = 0;
2327                                 goto out;
2328                         }
2329                         continue;
2330                 }
2331
2332                 path->slots[0]++;
2333
2334                 btrfs_item_key_to_cpu(leaf, &key, slot);
2335
2336                 if (key.objectid > inum)
2337                         goto out;
2338
2339                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2340                         continue;
2341
2342                 extent = btrfs_item_ptr(leaf, slot,
2343                                         struct btrfs_file_extent_item);
2344
2345                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2346                         continue;
2347
2348                 /*
2349                  * 'offset' refers to the exact key.offset,
2350                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2351                  * (key.offset - extent_offset).
2352                  */
2353                 if (key.offset != offset)
2354                         continue;
2355
2356                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2357                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2358
2359                 if (extent_offset >= old->extent_offset + old->offset +
2360                     old->len || extent_offset + num_bytes <=
2361                     old->extent_offset + old->offset)
2362                         continue;
2363                 break;
2364         }
2365
2366         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2367         if (!backref) {
2368                 ret = -ENOENT;
2369                 goto out;
2370         }
2371
2372         backref->root_id = root_id;
2373         backref->inum = inum;
2374         backref->file_pos = offset;
2375         backref->num_bytes = num_bytes;
2376         backref->extent_offset = extent_offset;
2377         backref->generation = btrfs_file_extent_generation(leaf, extent);
2378         backref->old = old;
2379         backref_insert(&new->root, backref);
2380         old->count++;
2381 out:
2382         btrfs_release_path(path);
2383         WARN_ON(ret);
2384         return ret;
2385 }
2386
2387 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2388                                    struct new_sa_defrag_extent *new)
2389 {
2390         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2391         struct old_sa_defrag_extent *old, *tmp;
2392         int ret;
2393
2394         new->path = path;
2395
2396         list_for_each_entry_safe(old, tmp, &new->head, list) {
2397                 ret = iterate_inodes_from_logical(old->bytenr +
2398                                                   old->extent_offset, fs_info,
2399                                                   path, record_one_backref,
2400                                                   old);
2401                 if (ret < 0 && ret != -ENOENT)
2402                         return false;
2403
2404                 /* no backref to be processed for this extent */
2405                 if (!old->count) {
2406                         list_del(&old->list);
2407                         kfree(old);
2408                 }
2409         }
2410
2411         if (list_empty(&new->head))
2412                 return false;
2413
2414         return true;
2415 }
2416
2417 static int relink_is_mergable(struct extent_buffer *leaf,
2418                               struct btrfs_file_extent_item *fi,
2419                               struct new_sa_defrag_extent *new)
2420 {
2421         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2422                 return 0;
2423
2424         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2425                 return 0;
2426
2427         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2428                 return 0;
2429
2430         if (btrfs_file_extent_encryption(leaf, fi) ||
2431             btrfs_file_extent_other_encoding(leaf, fi))
2432                 return 0;
2433
2434         return 1;
2435 }
2436
2437 /*
2438  * Note the backref might has changed, and in this case we just return 0.
2439  */
2440 static noinline int relink_extent_backref(struct btrfs_path *path,
2441                                  struct sa_defrag_extent_backref *prev,
2442                                  struct sa_defrag_extent_backref *backref)
2443 {
2444         struct btrfs_file_extent_item *extent;
2445         struct btrfs_file_extent_item *item;
2446         struct btrfs_ordered_extent *ordered;
2447         struct btrfs_trans_handle *trans;
2448         struct btrfs_fs_info *fs_info;
2449         struct btrfs_root *root;
2450         struct btrfs_key key;
2451         struct extent_buffer *leaf;
2452         struct old_sa_defrag_extent *old = backref->old;
2453         struct new_sa_defrag_extent *new = old->new;
2454         struct inode *src_inode = new->inode;
2455         struct inode *inode;
2456         struct extent_state *cached = NULL;
2457         int ret = 0;
2458         u64 start;
2459         u64 len;
2460         u64 lock_start;
2461         u64 lock_end;
2462         bool merge = false;
2463         int index;
2464
2465         if (prev && prev->root_id == backref->root_id &&
2466             prev->inum == backref->inum &&
2467             prev->file_pos + prev->num_bytes == backref->file_pos)
2468                 merge = true;
2469
2470         /* step 1: get root */
2471         key.objectid = backref->root_id;
2472         key.type = BTRFS_ROOT_ITEM_KEY;
2473         key.offset = (u64)-1;
2474
2475         fs_info = BTRFS_I(src_inode)->root->fs_info;
2476         index = srcu_read_lock(&fs_info->subvol_srcu);
2477
2478         root = btrfs_read_fs_root_no_name(fs_info, &key);
2479         if (IS_ERR(root)) {
2480                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2481                 if (PTR_ERR(root) == -ENOENT)
2482                         return 0;
2483                 return PTR_ERR(root);
2484         }
2485
2486         if (btrfs_root_readonly(root)) {
2487                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2488                 return 0;
2489         }
2490
2491         /* step 2: get inode */
2492         key.objectid = backref->inum;
2493         key.type = BTRFS_INODE_ITEM_KEY;
2494         key.offset = 0;
2495
2496         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2497         if (IS_ERR(inode)) {
2498                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2499                 return 0;
2500         }
2501
2502         srcu_read_unlock(&fs_info->subvol_srcu, index);
2503
2504         /* step 3: relink backref */
2505         lock_start = backref->file_pos;
2506         lock_end = backref->file_pos + backref->num_bytes - 1;
2507         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2508                          &cached);
2509
2510         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2511         if (ordered) {
2512                 btrfs_put_ordered_extent(ordered);
2513                 goto out_unlock;
2514         }
2515
2516         trans = btrfs_join_transaction(root);
2517         if (IS_ERR(trans)) {
2518                 ret = PTR_ERR(trans);
2519                 goto out_unlock;
2520         }
2521
2522         key.objectid = backref->inum;
2523         key.type = BTRFS_EXTENT_DATA_KEY;
2524         key.offset = backref->file_pos;
2525
2526         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2527         if (ret < 0) {
2528                 goto out_free_path;
2529         } else if (ret > 0) {
2530                 ret = 0;
2531                 goto out_free_path;
2532         }
2533
2534         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2535                                 struct btrfs_file_extent_item);
2536
2537         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2538             backref->generation)
2539                 goto out_free_path;
2540
2541         btrfs_release_path(path);
2542
2543         start = backref->file_pos;
2544         if (backref->extent_offset < old->extent_offset + old->offset)
2545                 start += old->extent_offset + old->offset -
2546                          backref->extent_offset;
2547
2548         len = min(backref->extent_offset + backref->num_bytes,
2549                   old->extent_offset + old->offset + old->len);
2550         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2551
2552         ret = btrfs_drop_extents(trans, root, inode, start,
2553                                  start + len, 1);
2554         if (ret)
2555                 goto out_free_path;
2556 again:
2557         key.objectid = btrfs_ino(inode);
2558         key.type = BTRFS_EXTENT_DATA_KEY;
2559         key.offset = start;
2560
2561         path->leave_spinning = 1;
2562         if (merge) {
2563                 struct btrfs_file_extent_item *fi;
2564                 u64 extent_len;
2565                 struct btrfs_key found_key;
2566
2567                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2568                 if (ret < 0)
2569                         goto out_free_path;
2570
2571                 path->slots[0]--;
2572                 leaf = path->nodes[0];
2573                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2574
2575                 fi = btrfs_item_ptr(leaf, path->slots[0],
2576                                     struct btrfs_file_extent_item);
2577                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2578
2579                 if (extent_len + found_key.offset == start &&
2580                     relink_is_mergable(leaf, fi, new)) {
2581                         btrfs_set_file_extent_num_bytes(leaf, fi,
2582                                                         extent_len + len);
2583                         btrfs_mark_buffer_dirty(leaf);
2584                         inode_add_bytes(inode, len);
2585
2586                         ret = 1;
2587                         goto out_free_path;
2588                 } else {
2589                         merge = false;
2590                         btrfs_release_path(path);
2591                         goto again;
2592                 }
2593         }
2594
2595         ret = btrfs_insert_empty_item(trans, root, path, &key,
2596                                         sizeof(*extent));
2597         if (ret) {
2598                 btrfs_abort_transaction(trans, root, ret);
2599                 goto out_free_path;
2600         }
2601
2602         leaf = path->nodes[0];
2603         item = btrfs_item_ptr(leaf, path->slots[0],
2604                                 struct btrfs_file_extent_item);
2605         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2606         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2607         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2608         btrfs_set_file_extent_num_bytes(leaf, item, len);
2609         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2610         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2611         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2612         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2613         btrfs_set_file_extent_encryption(leaf, item, 0);
2614         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2615
2616         btrfs_mark_buffer_dirty(leaf);
2617         inode_add_bytes(inode, len);
2618         btrfs_release_path(path);
2619
2620         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2621                         new->disk_len, 0,
2622                         backref->root_id, backref->inum,
2623                         new->file_pos); /* start - extent_offset */
2624         if (ret) {
2625                 btrfs_abort_transaction(trans, root, ret);
2626                 goto out_free_path;
2627         }
2628
2629         ret = 1;
2630 out_free_path:
2631         btrfs_release_path(path);
2632         path->leave_spinning = 0;
2633         btrfs_end_transaction(trans, root);
2634 out_unlock:
2635         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2636                              &cached, GFP_NOFS);
2637         iput(inode);
2638         return ret;
2639 }
2640
2641 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2642 {
2643         struct old_sa_defrag_extent *old, *tmp;
2644
2645         if (!new)
2646                 return;
2647
2648         list_for_each_entry_safe(old, tmp, &new->head, list) {
2649                 kfree(old);
2650         }
2651         kfree(new);
2652 }
2653
2654 static void relink_file_extents(struct new_sa_defrag_extent *new)
2655 {
2656         struct btrfs_path *path;
2657         struct sa_defrag_extent_backref *backref;
2658         struct sa_defrag_extent_backref *prev = NULL;
2659         struct inode *inode;
2660         struct btrfs_root *root;
2661         struct rb_node *node;
2662         int ret;
2663
2664         inode = new->inode;
2665         root = BTRFS_I(inode)->root;
2666
2667         path = btrfs_alloc_path();
2668         if (!path)
2669                 return;
2670
2671         if (!record_extent_backrefs(path, new)) {
2672                 btrfs_free_path(path);
2673                 goto out;
2674         }
2675         btrfs_release_path(path);
2676
2677         while (1) {
2678                 node = rb_first(&new->root);
2679                 if (!node)
2680                         break;
2681                 rb_erase(node, &new->root);
2682
2683                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2684
2685                 ret = relink_extent_backref(path, prev, backref);
2686                 WARN_ON(ret < 0);
2687
2688                 kfree(prev);
2689
2690                 if (ret == 1)
2691                         prev = backref;
2692                 else
2693                         prev = NULL;
2694                 cond_resched();
2695         }
2696         kfree(prev);
2697
2698         btrfs_free_path(path);
2699 out:
2700         free_sa_defrag_extent(new);
2701
2702         atomic_dec(&root->fs_info->defrag_running);
2703         wake_up(&root->fs_info->transaction_wait);
2704 }
2705
2706 static struct new_sa_defrag_extent *
2707 record_old_file_extents(struct inode *inode,
2708                         struct btrfs_ordered_extent *ordered)
2709 {
2710         struct btrfs_root *root = BTRFS_I(inode)->root;
2711         struct btrfs_path *path;
2712         struct btrfs_key key;
2713         struct old_sa_defrag_extent *old;
2714         struct new_sa_defrag_extent *new;
2715         int ret;
2716
2717         new = kmalloc(sizeof(*new), GFP_NOFS);
2718         if (!new)
2719                 return NULL;
2720
2721         new->inode = inode;
2722         new->file_pos = ordered->file_offset;
2723         new->len = ordered->len;
2724         new->bytenr = ordered->start;
2725         new->disk_len = ordered->disk_len;
2726         new->compress_type = ordered->compress_type;
2727         new->root = RB_ROOT;
2728         INIT_LIST_HEAD(&new->head);
2729
2730         path = btrfs_alloc_path();
2731         if (!path)
2732                 goto out_kfree;
2733
2734         key.objectid = btrfs_ino(inode);
2735         key.type = BTRFS_EXTENT_DATA_KEY;
2736         key.offset = new->file_pos;
2737
2738         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2739         if (ret < 0)
2740                 goto out_free_path;
2741         if (ret > 0 && path->slots[0] > 0)
2742                 path->slots[0]--;
2743
2744         /* find out all the old extents for the file range */
2745         while (1) {
2746                 struct btrfs_file_extent_item *extent;
2747                 struct extent_buffer *l;
2748                 int slot;
2749                 u64 num_bytes;
2750                 u64 offset;
2751                 u64 end;
2752                 u64 disk_bytenr;
2753                 u64 extent_offset;
2754
2755                 l = path->nodes[0];
2756                 slot = path->slots[0];
2757
2758                 if (slot >= btrfs_header_nritems(l)) {
2759                         ret = btrfs_next_leaf(root, path);
2760                         if (ret < 0)
2761                                 goto out_free_path;
2762                         else if (ret > 0)
2763                                 break;
2764                         continue;
2765                 }
2766
2767                 btrfs_item_key_to_cpu(l, &key, slot);
2768
2769                 if (key.objectid != btrfs_ino(inode))
2770                         break;
2771                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2772                         break;
2773                 if (key.offset >= new->file_pos + new->len)
2774                         break;
2775
2776                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2777
2778                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2779                 if (key.offset + num_bytes < new->file_pos)
2780                         goto next;
2781
2782                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2783                 if (!disk_bytenr)
2784                         goto next;
2785
2786                 extent_offset = btrfs_file_extent_offset(l, extent);
2787
2788                 old = kmalloc(sizeof(*old), GFP_NOFS);
2789                 if (!old)
2790                         goto out_free_path;
2791
2792                 offset = max(new->file_pos, key.offset);
2793                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2794
2795                 old->bytenr = disk_bytenr;
2796                 old->extent_offset = extent_offset;
2797                 old->offset = offset - key.offset;
2798                 old->len = end - offset;
2799                 old->new = new;
2800                 old->count = 0;
2801                 list_add_tail(&old->list, &new->head);
2802 next:
2803                 path->slots[0]++;
2804                 cond_resched();
2805         }
2806
2807         btrfs_free_path(path);
2808         atomic_inc(&root->fs_info->defrag_running);
2809
2810         return new;
2811
2812 out_free_path:
2813         btrfs_free_path(path);
2814 out_kfree:
2815         free_sa_defrag_extent(new);
2816         return NULL;
2817 }
2818
2819 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2820                                          u64 start, u64 len)
2821 {
2822         struct btrfs_block_group_cache *cache;
2823
2824         cache = btrfs_lookup_block_group(root->fs_info, start);
2825         ASSERT(cache);
2826
2827         spin_lock(&cache->lock);
2828         cache->delalloc_bytes -= len;
2829         spin_unlock(&cache->lock);
2830
2831         btrfs_put_block_group(cache);
2832 }
2833
2834 /* as ordered data IO finishes, this gets called so we can finish
2835  * an ordered extent if the range of bytes in the file it covers are
2836  * fully written.
2837  */
2838 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2839 {
2840         struct inode *inode = ordered_extent->inode;
2841         struct btrfs_root *root = BTRFS_I(inode)->root;
2842         struct btrfs_trans_handle *trans = NULL;
2843         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2844         struct extent_state *cached_state = NULL;
2845         struct new_sa_defrag_extent *new = NULL;
2846         int compress_type = 0;
2847         int ret = 0;
2848         u64 logical_len = ordered_extent->len;
2849         bool nolock;
2850         bool truncated = false;
2851
2852         nolock = btrfs_is_free_space_inode(inode);
2853
2854         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2855                 ret = -EIO;
2856                 goto out;
2857         }
2858
2859         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2860                                      ordered_extent->file_offset +
2861                                      ordered_extent->len - 1);
2862
2863         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2864                 truncated = true;
2865                 logical_len = ordered_extent->truncated_len;
2866                 /* Truncated the entire extent, don't bother adding */
2867                 if (!logical_len)
2868                         goto out;
2869         }
2870
2871         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2872                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2873
2874                 /*
2875                  * For mwrite(mmap + memset to write) case, we still reserve
2876                  * space for NOCOW range.
2877                  * As NOCOW won't cause a new delayed ref, just free the space
2878                  */
2879                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2880                                        ordered_extent->len);
2881                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2882                 if (nolock)
2883                         trans = btrfs_join_transaction_nolock(root);
2884                 else
2885                         trans = btrfs_join_transaction(root);
2886                 if (IS_ERR(trans)) {
2887                         ret = PTR_ERR(trans);
2888                         trans = NULL;
2889                         goto out;
2890                 }
2891                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2892                 ret = btrfs_update_inode_fallback(trans, root, inode);
2893                 if (ret) /* -ENOMEM or corruption */
2894                         btrfs_abort_transaction(trans, root, ret);
2895                 goto out;
2896         }
2897
2898         lock_extent_bits(io_tree, ordered_extent->file_offset,
2899                          ordered_extent->file_offset + ordered_extent->len - 1,
2900                          &cached_state);
2901
2902         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2903                         ordered_extent->file_offset + ordered_extent->len - 1,
2904                         EXTENT_DEFRAG, 1, cached_state);
2905         if (ret) {
2906                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2907                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2908                         /* the inode is shared */
2909                         new = record_old_file_extents(inode, ordered_extent);
2910
2911                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2912                         ordered_extent->file_offset + ordered_extent->len - 1,
2913                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2914         }
2915
2916         if (nolock)
2917                 trans = btrfs_join_transaction_nolock(root);
2918         else
2919                 trans = btrfs_join_transaction(root);
2920         if (IS_ERR(trans)) {
2921                 ret = PTR_ERR(trans);
2922                 trans = NULL;
2923                 goto out_unlock;
2924         }
2925
2926         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2927
2928         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2929                 compress_type = ordered_extent->compress_type;
2930         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2931                 BUG_ON(compress_type);
2932                 ret = btrfs_mark_extent_written(trans, inode,
2933                                                 ordered_extent->file_offset,
2934                                                 ordered_extent->file_offset +
2935                                                 logical_len);
2936         } else {
2937                 BUG_ON(root == root->fs_info->tree_root);
2938                 ret = insert_reserved_file_extent(trans, inode,
2939                                                 ordered_extent->file_offset,
2940                                                 ordered_extent->start,
2941                                                 ordered_extent->disk_len,
2942                                                 logical_len, logical_len,
2943                                                 compress_type, 0, 0,
2944                                                 BTRFS_FILE_EXTENT_REG);
2945                 if (!ret)
2946                         btrfs_release_delalloc_bytes(root,
2947                                                      ordered_extent->start,
2948                                                      ordered_extent->disk_len);
2949         }
2950         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2951                            ordered_extent->file_offset, ordered_extent->len,
2952                            trans->transid);
2953         if (ret < 0) {
2954                 btrfs_abort_transaction(trans, root, ret);
2955                 goto out_unlock;
2956         }
2957
2958         add_pending_csums(trans, inode, ordered_extent->file_offset,
2959                           &ordered_extent->list);
2960
2961         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2962         ret = btrfs_update_inode_fallback(trans, root, inode);
2963         if (ret) { /* -ENOMEM or corruption */
2964                 btrfs_abort_transaction(trans, root, ret);
2965                 goto out_unlock;
2966         }
2967         ret = 0;
2968 out_unlock:
2969         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2970                              ordered_extent->file_offset +
2971                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2972 out:
2973         if (root != root->fs_info->tree_root)
2974                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2975         if (trans)
2976                 btrfs_end_transaction(trans, root);
2977
2978         if (ret || truncated) {
2979                 u64 start, end;
2980
2981                 if (truncated)
2982                         start = ordered_extent->file_offset + logical_len;
2983                 else
2984                         start = ordered_extent->file_offset;
2985                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2986                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2987
2988                 /* Drop the cache for the part of the extent we didn't write. */
2989                 btrfs_drop_extent_cache(inode, start, end, 0);
2990
2991                 /*
2992                  * If the ordered extent had an IOERR or something else went
2993                  * wrong we need to return the space for this ordered extent
2994                  * back to the allocator.  We only free the extent in the
2995                  * truncated case if we didn't write out the extent at all.
2996                  */
2997                 if ((ret || !logical_len) &&
2998                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2999                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3000                         btrfs_free_reserved_extent(root, ordered_extent->start,
3001                                                    ordered_extent->disk_len, 1);
3002         }
3003
3004
3005         /*
3006          * This needs to be done to make sure anybody waiting knows we are done
3007          * updating everything for this ordered extent.
3008          */
3009         btrfs_remove_ordered_extent(inode, ordered_extent);
3010
3011         /* for snapshot-aware defrag */
3012         if (new) {
3013                 if (ret) {
3014                         free_sa_defrag_extent(new);
3015                         atomic_dec(&root->fs_info->defrag_running);
3016                 } else {
3017                         relink_file_extents(new);
3018                 }
3019         }
3020
3021         /* once for us */
3022         btrfs_put_ordered_extent(ordered_extent);
3023         /* once for the tree */
3024         btrfs_put_ordered_extent(ordered_extent);
3025
3026         return ret;
3027 }
3028
3029 static void finish_ordered_fn(struct btrfs_work *work)
3030 {
3031         struct btrfs_ordered_extent *ordered_extent;
3032         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3033         btrfs_finish_ordered_io(ordered_extent);
3034 }
3035
3036 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3037                                 struct extent_state *state, int uptodate)
3038 {
3039         struct inode *inode = page->mapping->host;
3040         struct btrfs_root *root = BTRFS_I(inode)->root;
3041         struct btrfs_ordered_extent *ordered_extent = NULL;
3042         struct btrfs_workqueue *wq;
3043         btrfs_work_func_t func;
3044
3045         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3046
3047         ClearPagePrivate2(page);
3048         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3049                                             end - start + 1, uptodate))
3050                 return 0;
3051
3052         if (btrfs_is_free_space_inode(inode)) {
3053                 wq = root->fs_info->endio_freespace_worker;
3054                 func = btrfs_freespace_write_helper;
3055         } else {
3056                 wq = root->fs_info->endio_write_workers;
3057                 func = btrfs_endio_write_helper;
3058         }
3059
3060         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3061                         NULL);
3062         btrfs_queue_work(wq, &ordered_extent->work);
3063
3064         return 0;
3065 }
3066
3067 static int __readpage_endio_check(struct inode *inode,
3068                                   struct btrfs_io_bio *io_bio,
3069                                   int icsum, struct page *page,
3070                                   int pgoff, u64 start, size_t len)
3071 {
3072         char *kaddr;
3073         u32 csum_expected;
3074         u32 csum = ~(u32)0;
3075
3076         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3077
3078         kaddr = kmap_atomic(page);
3079         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3080         btrfs_csum_final(csum, (char *)&csum);
3081         if (csum != csum_expected)
3082                 goto zeroit;
3083
3084         kunmap_atomic(kaddr);
3085         return 0;
3086 zeroit:
3087         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3088                 "csum failed ino %llu off %llu csum %u expected csum %u",
3089                            btrfs_ino(inode), start, csum, csum_expected);
3090         memset(kaddr + pgoff, 1, len);
3091         flush_dcache_page(page);
3092         kunmap_atomic(kaddr);
3093         if (csum_expected == 0)
3094                 return 0;
3095         return -EIO;
3096 }
3097
3098 /*
3099  * when reads are done, we need to check csums to verify the data is correct
3100  * if there's a match, we allow the bio to finish.  If not, the code in
3101  * extent_io.c will try to find good copies for us.
3102  */
3103 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3104                                       u64 phy_offset, struct page *page,
3105                                       u64 start, u64 end, int mirror)
3106 {
3107         size_t offset = start - page_offset(page);
3108         struct inode *inode = page->mapping->host;
3109         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3110         struct btrfs_root *root = BTRFS_I(inode)->root;
3111
3112         if (PageChecked(page)) {
3113                 ClearPageChecked(page);
3114                 return 0;
3115         }
3116
3117         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3118                 return 0;
3119
3120         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3121             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3122                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3123                 return 0;
3124         }
3125
3126         phy_offset >>= inode->i_sb->s_blocksize_bits;
3127         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3128                                       start, (size_t)(end - start + 1));
3129 }
3130
3131 void btrfs_add_delayed_iput(struct inode *inode)
3132 {
3133         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3134         struct btrfs_inode *binode = BTRFS_I(inode);
3135
3136         if (atomic_add_unless(&inode->i_count, -1, 1))
3137                 return;
3138
3139         spin_lock(&fs_info->delayed_iput_lock);
3140         if (binode->delayed_iput_count == 0) {
3141                 ASSERT(list_empty(&binode->delayed_iput));
3142                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3143         } else {
3144                 binode->delayed_iput_count++;
3145         }
3146         spin_unlock(&fs_info->delayed_iput_lock);
3147 }
3148
3149 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3150 {
3151         struct btrfs_fs_info *fs_info = root->fs_info;
3152
3153         spin_lock(&fs_info->delayed_iput_lock);
3154         while (!list_empty(&fs_info->delayed_iputs)) {
3155                 struct btrfs_inode *inode;
3156
3157                 inode = list_first_entry(&fs_info->delayed_iputs,
3158                                 struct btrfs_inode, delayed_iput);
3159                 if (inode->delayed_iput_count) {
3160                         inode->delayed_iput_count--;
3161                         list_move_tail(&inode->delayed_iput,
3162                                         &fs_info->delayed_iputs);
3163                 } else {
3164                         list_del_init(&inode->delayed_iput);
3165                 }
3166                 spin_unlock(&fs_info->delayed_iput_lock);
3167                 iput(&inode->vfs_inode);
3168                 spin_lock(&fs_info->delayed_iput_lock);
3169         }
3170         spin_unlock(&fs_info->delayed_iput_lock);
3171 }
3172
3173 /*
3174  * This is called in transaction commit time. If there are no orphan
3175  * files in the subvolume, it removes orphan item and frees block_rsv
3176  * structure.
3177  */
3178 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3179                               struct btrfs_root *root)
3180 {
3181         struct btrfs_block_rsv *block_rsv;
3182         int ret;
3183
3184         if (atomic_read(&root->orphan_inodes) ||
3185             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3186                 return;
3187
3188         spin_lock(&root->orphan_lock);
3189         if (atomic_read(&root->orphan_inodes)) {
3190                 spin_unlock(&root->orphan_lock);
3191                 return;
3192         }
3193
3194         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3195                 spin_unlock(&root->orphan_lock);
3196                 return;
3197         }
3198
3199         block_rsv = root->orphan_block_rsv;
3200         root->orphan_block_rsv = NULL;
3201         spin_unlock(&root->orphan_lock);
3202
3203         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3204             btrfs_root_refs(&root->root_item) > 0) {
3205                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3206                                             root->root_key.objectid);
3207                 if (ret)
3208                         btrfs_abort_transaction(trans, root, ret);
3209                 else
3210                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3211                                   &root->state);
3212         }
3213
3214         if (block_rsv) {
3215                 WARN_ON(block_rsv->size > 0);
3216                 btrfs_free_block_rsv(root, block_rsv);
3217         }
3218 }
3219
3220 /*
3221  * This creates an orphan entry for the given inode in case something goes
3222  * wrong in the middle of an unlink/truncate.
3223  *
3224  * NOTE: caller of this function should reserve 5 units of metadata for
3225  *       this function.
3226  */
3227 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3228 {
3229         struct btrfs_root *root = BTRFS_I(inode)->root;
3230         struct btrfs_block_rsv *block_rsv = NULL;
3231         int reserve = 0;
3232         int insert = 0;
3233         int ret;
3234
3235         if (!root->orphan_block_rsv) {
3236                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3237                 if (!block_rsv)
3238                         return -ENOMEM;
3239         }
3240
3241         spin_lock(&root->orphan_lock);
3242         if (!root->orphan_block_rsv) {
3243                 root->orphan_block_rsv = block_rsv;
3244         } else if (block_rsv) {
3245                 btrfs_free_block_rsv(root, block_rsv);
3246                 block_rsv = NULL;
3247         }
3248
3249         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3250                               &BTRFS_I(inode)->runtime_flags)) {
3251 #if 0
3252                 /*
3253                  * For proper ENOSPC handling, we should do orphan
3254                  * cleanup when mounting. But this introduces backward
3255                  * compatibility issue.
3256                  */
3257                 if (!xchg(&root->orphan_item_inserted, 1))
3258                         insert = 2;
3259                 else
3260                         insert = 1;
3261 #endif
3262                 insert = 1;
3263                 atomic_inc(&root->orphan_inodes);
3264         }
3265
3266         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3267                               &BTRFS_I(inode)->runtime_flags))
3268                 reserve = 1;
3269         spin_unlock(&root->orphan_lock);
3270
3271         /* grab metadata reservation from transaction handle */
3272         if (reserve) {
3273                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3274                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
3275         }
3276
3277         /* insert an orphan item to track this unlinked/truncated file */
3278         if (insert >= 1) {
3279                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3280                 if (ret) {
3281                         atomic_dec(&root->orphan_inodes);
3282                         if (reserve) {
3283                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3284                                           &BTRFS_I(inode)->runtime_flags);
3285                                 btrfs_orphan_release_metadata(inode);
3286                         }
3287                         if (ret != -EEXIST) {
3288                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3289                                           &BTRFS_I(inode)->runtime_flags);
3290                                 btrfs_abort_transaction(trans, root, ret);
3291                                 return ret;
3292                         }
3293                 }
3294                 ret = 0;
3295         }
3296
3297         /* insert an orphan item to track subvolume contains orphan files */
3298         if (insert >= 2) {
3299                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3300                                                root->root_key.objectid);
3301                 if (ret && ret != -EEXIST) {
3302                         btrfs_abort_transaction(trans, root, ret);
3303                         return ret;
3304                 }
3305         }
3306         return 0;
3307 }
3308
3309 /*
3310  * We have done the truncate/delete so we can go ahead and remove the orphan
3311  * item for this particular inode.
3312  */
3313 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3314                             struct inode *inode)
3315 {
3316         struct btrfs_root *root = BTRFS_I(inode)->root;
3317         int delete_item = 0;
3318         int release_rsv = 0;
3319         int ret = 0;
3320
3321         spin_lock(&root->orphan_lock);
3322         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3323                                &BTRFS_I(inode)->runtime_flags))
3324                 delete_item = 1;
3325
3326         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3327                                &BTRFS_I(inode)->runtime_flags))
3328                 release_rsv = 1;
3329         spin_unlock(&root->orphan_lock);
3330
3331         if (delete_item) {
3332                 atomic_dec(&root->orphan_inodes);
3333                 if (trans)
3334                         ret = btrfs_del_orphan_item(trans, root,
3335                                                     btrfs_ino(inode));
3336         }
3337
3338         if (release_rsv)
3339                 btrfs_orphan_release_metadata(inode);
3340
3341         return ret;
3342 }
3343
3344 /*
3345  * this cleans up any orphans that may be left on the list from the last use
3346  * of this root.
3347  */
3348 int btrfs_orphan_cleanup(struct btrfs_root *root)
3349 {
3350         struct btrfs_path *path;
3351         struct extent_buffer *leaf;
3352         struct btrfs_key key, found_key;
3353         struct btrfs_trans_handle *trans;
3354         struct inode *inode;
3355         u64 last_objectid = 0;
3356         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3357
3358         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3359                 return 0;
3360
3361         path = btrfs_alloc_path();
3362         if (!path) {
3363                 ret = -ENOMEM;
3364                 goto out;
3365         }
3366         path->reada = READA_BACK;
3367
3368         key.objectid = BTRFS_ORPHAN_OBJECTID;
3369         key.type = BTRFS_ORPHAN_ITEM_KEY;
3370         key.offset = (u64)-1;
3371
3372         while (1) {
3373                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3374                 if (ret < 0)
3375                         goto out;
3376
3377                 /*
3378                  * if ret == 0 means we found what we were searching for, which
3379                  * is weird, but possible, so only screw with path if we didn't
3380                  * find the key and see if we have stuff that matches
3381                  */
3382                 if (ret > 0) {
3383                         ret = 0;
3384                         if (path->slots[0] == 0)
3385                                 break;
3386                         path->slots[0]--;
3387                 }
3388
3389                 /* pull out the item */
3390                 leaf = path->nodes[0];
3391                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3392
3393                 /* make sure the item matches what we want */
3394                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3395                         break;
3396                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3397                         break;
3398
3399                 /* release the path since we're done with it */
3400                 btrfs_release_path(path);
3401
3402                 /*
3403                  * this is where we are basically btrfs_lookup, without the
3404                  * crossing root thing.  we store the inode number in the
3405                  * offset of the orphan item.
3406                  */
3407
3408                 if (found_key.offset == last_objectid) {
3409                         btrfs_err(root->fs_info,
3410                                 "Error removing orphan entry, stopping orphan cleanup");
3411                         ret = -EINVAL;
3412                         goto out;
3413                 }
3414
3415                 last_objectid = found_key.offset;
3416
3417                 found_key.objectid = found_key.offset;
3418                 found_key.type = BTRFS_INODE_ITEM_KEY;
3419                 found_key.offset = 0;
3420                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3421                 ret = PTR_ERR_OR_ZERO(inode);
3422                 if (ret && ret != -ESTALE)
3423                         goto out;
3424
3425                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3426                         struct btrfs_root *dead_root;
3427                         struct btrfs_fs_info *fs_info = root->fs_info;
3428                         int is_dead_root = 0;
3429
3430                         /*
3431                          * this is an orphan in the tree root. Currently these
3432                          * could come from 2 sources:
3433                          *  a) a snapshot deletion in progress
3434                          *  b) a free space cache inode
3435                          * We need to distinguish those two, as the snapshot
3436                          * orphan must not get deleted.
3437                          * find_dead_roots already ran before us, so if this
3438                          * is a snapshot deletion, we should find the root
3439                          * in the dead_roots list
3440                          */
3441                         spin_lock(&fs_info->trans_lock);
3442                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3443                                             root_list) {
3444                                 if (dead_root->root_key.objectid ==
3445                                     found_key.objectid) {
3446                                         is_dead_root = 1;
3447                                         break;
3448                                 }
3449                         }
3450                         spin_unlock(&fs_info->trans_lock);
3451                         if (is_dead_root) {
3452                                 /* prevent this orphan from being found again */
3453                                 key.offset = found_key.objectid - 1;
3454                                 continue;
3455                         }
3456                 }
3457                 /*
3458                  * Inode is already gone but the orphan item is still there,
3459                  * kill the orphan item.
3460                  */
3461                 if (ret == -ESTALE) {
3462                         trans = btrfs_start_transaction(root, 1);
3463                         if (IS_ERR(trans)) {
3464                                 ret = PTR_ERR(trans);
3465                                 goto out;
3466                         }
3467                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3468                                 found_key.objectid);
3469                         ret = btrfs_del_orphan_item(trans, root,
3470                                                     found_key.objectid);
3471                         btrfs_end_transaction(trans, root);
3472                         if (ret)
3473                                 goto out;
3474                         continue;
3475                 }
3476
3477                 /*
3478                  * add this inode to the orphan list so btrfs_orphan_del does
3479                  * the proper thing when we hit it
3480                  */
3481                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3482                         &BTRFS_I(inode)->runtime_flags);
3483                 atomic_inc(&root->orphan_inodes);
3484
3485                 /* if we have links, this was a truncate, lets do that */
3486                 if (inode->i_nlink) {
3487                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3488                                 iput(inode);
3489                                 continue;
3490                         }
3491                         nr_truncate++;
3492
3493                         /* 1 for the orphan item deletion. */
3494                         trans = btrfs_start_transaction(root, 1);
3495                         if (IS_ERR(trans)) {
3496                                 iput(inode);
3497                                 ret = PTR_ERR(trans);
3498                                 goto out;
3499                         }
3500                         ret = btrfs_orphan_add(trans, inode);
3501                         btrfs_end_transaction(trans, root);
3502                         if (ret) {
3503                                 iput(inode);
3504                                 goto out;
3505                         }
3506
3507                         ret = btrfs_truncate(inode);
3508                         if (ret)
3509                                 btrfs_orphan_del(NULL, inode);
3510                 } else {
3511                         nr_unlink++;
3512                 }
3513
3514                 /* this will do delete_inode and everything for us */
3515                 iput(inode);
3516                 if (ret)
3517                         goto out;
3518         }
3519         /* release the path since we're done with it */
3520         btrfs_release_path(path);
3521
3522         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3523
3524         if (root->orphan_block_rsv)
3525                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3526                                         (u64)-1);
3527
3528         if (root->orphan_block_rsv ||
3529             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3530                 trans = btrfs_join_transaction(root);
3531                 if (!IS_ERR(trans))
3532                         btrfs_end_transaction(trans, root);
3533         }
3534
3535         if (nr_unlink)
3536                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3537         if (nr_truncate)
3538                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3539
3540 out:
3541         if (ret)
3542                 btrfs_err(root->fs_info,
3543                         "could not do orphan cleanup %d", ret);
3544         btrfs_free_path(path);
3545         return ret;
3546 }
3547
3548 /*
3549  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3550  * don't find any xattrs, we know there can't be any acls.
3551  *
3552  * slot is the slot the inode is in, objectid is the objectid of the inode
3553  */
3554 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3555                                           int slot, u64 objectid,
3556                                           int *first_xattr_slot)
3557 {
3558         u32 nritems = btrfs_header_nritems(leaf);
3559         struct btrfs_key found_key;
3560         static u64 xattr_access = 0;
3561         static u64 xattr_default = 0;
3562         int scanned = 0;
3563
3564         if (!xattr_access) {
3565                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3566                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3567                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3568                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3569         }
3570
3571         slot++;
3572         *first_xattr_slot = -1;
3573         while (slot < nritems) {
3574                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3575
3576                 /* we found a different objectid, there must not be acls */
3577                 if (found_key.objectid != objectid)
3578                         return 0;
3579
3580                 /* we found an xattr, assume we've got an acl */
3581                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3582                         if (*first_xattr_slot == -1)
3583                                 *first_xattr_slot = slot;
3584                         if (found_key.offset == xattr_access ||
3585                             found_key.offset == xattr_default)
3586                                 return 1;
3587                 }
3588
3589                 /*
3590                  * we found a key greater than an xattr key, there can't
3591                  * be any acls later on
3592                  */
3593                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3594                         return 0;
3595
3596                 slot++;
3597                 scanned++;
3598
3599                 /*
3600                  * it goes inode, inode backrefs, xattrs, extents,
3601                  * so if there are a ton of hard links to an inode there can
3602                  * be a lot of backrefs.  Don't waste time searching too hard,
3603                  * this is just an optimization
3604                  */
3605                 if (scanned >= 8)
3606                         break;
3607         }
3608         /* we hit the end of the leaf before we found an xattr or
3609          * something larger than an xattr.  We have to assume the inode
3610          * has acls
3611          */
3612         if (*first_xattr_slot == -1)
3613                 *first_xattr_slot = slot;
3614         return 1;
3615 }
3616
3617 /*
3618  * read an inode from the btree into the in-memory inode
3619  */
3620 static void btrfs_read_locked_inode(struct inode *inode)
3621 {
3622         struct btrfs_path *path;
3623         struct extent_buffer *leaf;
3624         struct btrfs_inode_item *inode_item;
3625         struct btrfs_root *root = BTRFS_I(inode)->root;
3626         struct btrfs_key location;
3627         unsigned long ptr;
3628         int maybe_acls;
3629         u32 rdev;
3630         int ret;
3631         bool filled = false;
3632         int first_xattr_slot;
3633
3634         ret = btrfs_fill_inode(inode, &rdev);
3635         if (!ret)
3636                 filled = true;
3637
3638         path = btrfs_alloc_path();
3639         if (!path)
3640                 goto make_bad;
3641
3642         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3643
3644         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3645         if (ret)
3646                 goto make_bad;
3647
3648         leaf = path->nodes[0];
3649
3650         if (filled)
3651                 goto cache_index;
3652
3653         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3654                                     struct btrfs_inode_item);
3655         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3656         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3657         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3658         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3659         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3660
3661         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3662         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3663
3664         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3665         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3666
3667         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3668         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3669
3670         BTRFS_I(inode)->i_otime.tv_sec =
3671                 btrfs_timespec_sec(leaf, &inode_item->otime);
3672         BTRFS_I(inode)->i_otime.tv_nsec =
3673                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3674
3675         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3676         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3677         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3678
3679         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3680         inode->i_generation = BTRFS_I(inode)->generation;
3681         inode->i_rdev = 0;
3682         rdev = btrfs_inode_rdev(leaf, inode_item);
3683
3684         BTRFS_I(inode)->index_cnt = (u64)-1;
3685         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3686
3687 cache_index:
3688         /*
3689          * If we were modified in the current generation and evicted from memory
3690          * and then re-read we need to do a full sync since we don't have any
3691          * idea about which extents were modified before we were evicted from
3692          * cache.
3693          *
3694          * This is required for both inode re-read from disk and delayed inode
3695          * in delayed_nodes_tree.
3696          */
3697         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3698                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3699                         &BTRFS_I(inode)->runtime_flags);
3700
3701         /*
3702          * We don't persist the id of the transaction where an unlink operation
3703          * against the inode was last made. So here we assume the inode might
3704          * have been evicted, and therefore the exact value of last_unlink_trans
3705          * lost, and set it to last_trans to avoid metadata inconsistencies
3706          * between the inode and its parent if the inode is fsync'ed and the log
3707          * replayed. For example, in the scenario:
3708          *
3709          * touch mydir/foo
3710          * ln mydir/foo mydir/bar
3711          * sync
3712          * unlink mydir/bar
3713          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3714          * xfs_io -c fsync mydir/foo
3715          * <power failure>
3716          * mount fs, triggers fsync log replay
3717          *
3718          * We must make sure that when we fsync our inode foo we also log its
3719          * parent inode, otherwise after log replay the parent still has the
3720          * dentry with the "bar" name but our inode foo has a link count of 1
3721          * and doesn't have an inode ref with the name "bar" anymore.
3722          *
3723          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3724          * but it guarantees correctness at the expense of occasional full
3725          * transaction commits on fsync if our inode is a directory, or if our
3726          * inode is not a directory, logging its parent unnecessarily.
3727          */
3728         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3729
3730         path->slots[0]++;
3731         if (inode->i_nlink != 1 ||
3732             path->slots[0] >= btrfs_header_nritems(leaf))
3733                 goto cache_acl;
3734
3735         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3736         if (location.objectid != btrfs_ino(inode))
3737                 goto cache_acl;
3738
3739         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3740         if (location.type == BTRFS_INODE_REF_KEY) {
3741                 struct btrfs_inode_ref *ref;
3742
3743                 ref = (struct btrfs_inode_ref *)ptr;
3744                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3745         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3746                 struct btrfs_inode_extref *extref;
3747
3748                 extref = (struct btrfs_inode_extref *)ptr;
3749                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3750                                                                      extref);
3751         }
3752 cache_acl:
3753         /*
3754          * try to precache a NULL acl entry for files that don't have
3755          * any xattrs or acls
3756          */
3757         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3758                                            btrfs_ino(inode), &first_xattr_slot);
3759         if (first_xattr_slot != -1) {
3760                 path->slots[0] = first_xattr_slot;
3761                 ret = btrfs_load_inode_props(inode, path);
3762                 if (ret)
3763                         btrfs_err(root->fs_info,
3764                                   "error loading props for ino %llu (root %llu): %d",
3765                                   btrfs_ino(inode),
3766                                   root->root_key.objectid, ret);
3767         }
3768         btrfs_free_path(path);
3769
3770         if (!maybe_acls)
3771                 cache_no_acl(inode);
3772
3773         switch (inode->i_mode & S_IFMT) {
3774         case S_IFREG:
3775                 inode->i_mapping->a_ops = &btrfs_aops;
3776                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3777                 inode->i_fop = &btrfs_file_operations;
3778                 inode->i_op = &btrfs_file_inode_operations;
3779                 break;
3780         case S_IFDIR:
3781                 inode->i_fop = &btrfs_dir_file_operations;
3782                 if (root == root->fs_info->tree_root)
3783                         inode->i_op = &btrfs_dir_ro_inode_operations;
3784                 else
3785                         inode->i_op = &btrfs_dir_inode_operations;
3786                 break;
3787         case S_IFLNK:
3788                 inode->i_op = &btrfs_symlink_inode_operations;
3789                 inode_nohighmem(inode);
3790                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3791                 break;
3792         default:
3793                 inode->i_op = &btrfs_special_inode_operations;
3794                 init_special_inode(inode, inode->i_mode, rdev);
3795                 break;
3796         }
3797
3798         btrfs_update_iflags(inode);
3799         return;
3800
3801 make_bad:
3802         btrfs_free_path(path);
3803         make_bad_inode(inode);
3804 }
3805
3806 /*
3807  * given a leaf and an inode, copy the inode fields into the leaf
3808  */
3809 static void fill_inode_item(struct btrfs_trans_handle *trans,
3810                             struct extent_buffer *leaf,
3811                             struct btrfs_inode_item *item,
3812                             struct inode *inode)
3813 {
3814         struct btrfs_map_token token;
3815
3816         btrfs_init_map_token(&token);
3817
3818         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3819         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3820         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3821                                    &token);
3822         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3823         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3824
3825         btrfs_set_token_timespec_sec(leaf, &item->atime,
3826                                      inode->i_atime.tv_sec, &token);
3827         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3828                                       inode->i_atime.tv_nsec, &token);
3829
3830         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3831                                      inode->i_mtime.tv_sec, &token);
3832         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3833                                       inode->i_mtime.tv_nsec, &token);
3834
3835         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3836                                      inode->i_ctime.tv_sec, &token);
3837         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3838                                       inode->i_ctime.tv_nsec, &token);
3839
3840         btrfs_set_token_timespec_sec(leaf, &item->otime,
3841                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3842         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3843                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3844
3845         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3846                                      &token);
3847         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3848                                          &token);
3849         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3850         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3851         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3852         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3853         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3854 }
3855
3856 /*
3857  * copy everything in the in-memory inode into the btree.
3858  */
3859 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3860                                 struct btrfs_root *root, struct inode *inode)
3861 {
3862         struct btrfs_inode_item *inode_item;
3863         struct btrfs_path *path;
3864         struct extent_buffer *leaf;
3865         int ret;
3866
3867         path = btrfs_alloc_path();
3868         if (!path)
3869                 return -ENOMEM;
3870
3871         path->leave_spinning = 1;
3872         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3873                                  1);
3874         if (ret) {
3875                 if (ret > 0)
3876                         ret = -ENOENT;
3877                 goto failed;
3878         }
3879
3880         leaf = path->nodes[0];
3881         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3882                                     struct btrfs_inode_item);
3883
3884         fill_inode_item(trans, leaf, inode_item, inode);
3885         btrfs_mark_buffer_dirty(leaf);
3886         btrfs_set_inode_last_trans(trans, inode);
3887         ret = 0;
3888 failed:
3889         btrfs_free_path(path);
3890         return ret;
3891 }
3892
3893 /*
3894  * copy everything in the in-memory inode into the btree.
3895  */
3896 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3897                                 struct btrfs_root *root, struct inode *inode)
3898 {
3899         int ret;
3900
3901         /*
3902          * If the inode is a free space inode, we can deadlock during commit
3903          * if we put it into the delayed code.
3904          *
3905          * The data relocation inode should also be directly updated
3906          * without delay
3907          */
3908         if (!btrfs_is_free_space_inode(inode)
3909             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3910             && !root->fs_info->log_root_recovering) {
3911                 btrfs_update_root_times(trans, root);
3912
3913                 ret = btrfs_delayed_update_inode(trans, root, inode);
3914                 if (!ret)
3915                         btrfs_set_inode_last_trans(trans, inode);
3916                 return ret;
3917         }
3918
3919         return btrfs_update_inode_item(trans, root, inode);
3920 }
3921
3922 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3923                                          struct btrfs_root *root,
3924                                          struct inode *inode)
3925 {
3926         int ret;
3927
3928         ret = btrfs_update_inode(trans, root, inode);
3929         if (ret == -ENOSPC)
3930                 return btrfs_update_inode_item(trans, root, inode);
3931         return ret;
3932 }
3933
3934 /*
3935  * unlink helper that gets used here in inode.c and in the tree logging
3936  * recovery code.  It remove a link in a directory with a given name, and
3937  * also drops the back refs in the inode to the directory
3938  */
3939 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3940                                 struct btrfs_root *root,
3941                                 struct inode *dir, struct inode *inode,
3942                                 const char *name, int name_len)
3943 {
3944         struct btrfs_path *path;
3945         int ret = 0;
3946         struct extent_buffer *leaf;
3947         struct btrfs_dir_item *di;
3948         struct btrfs_key key;
3949         u64 index;
3950         u64 ino = btrfs_ino(inode);
3951         u64 dir_ino = btrfs_ino(dir);
3952
3953         path = btrfs_alloc_path();
3954         if (!path) {
3955                 ret = -ENOMEM;
3956                 goto out;
3957         }
3958
3959         path->leave_spinning = 1;
3960         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3961                                     name, name_len, -1);
3962         if (IS_ERR(di)) {
3963                 ret = PTR_ERR(di);
3964                 goto err;
3965         }
3966         if (!di) {
3967                 ret = -ENOENT;
3968                 goto err;
3969         }
3970         leaf = path->nodes[0];
3971         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3972         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3973         if (ret)
3974                 goto err;
3975         btrfs_release_path(path);
3976
3977         /*
3978          * If we don't have dir index, we have to get it by looking up
3979          * the inode ref, since we get the inode ref, remove it directly,
3980          * it is unnecessary to do delayed deletion.
3981          *
3982          * But if we have dir index, needn't search inode ref to get it.
3983          * Since the inode ref is close to the inode item, it is better
3984          * that we delay to delete it, and just do this deletion when
3985          * we update the inode item.
3986          */
3987         if (BTRFS_I(inode)->dir_index) {
3988                 ret = btrfs_delayed_delete_inode_ref(inode);
3989                 if (!ret) {
3990                         index = BTRFS_I(inode)->dir_index;
3991                         goto skip_backref;
3992                 }
3993         }
3994
3995         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3996                                   dir_ino, &index);
3997         if (ret) {
3998                 btrfs_info(root->fs_info,
3999                         "failed to delete reference to %.*s, inode %llu parent %llu",
4000                         name_len, name, ino, dir_ino);
4001                 btrfs_abort_transaction(trans, root, ret);
4002                 goto err;
4003         }
4004 skip_backref:
4005         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4006         if (ret) {
4007                 btrfs_abort_transaction(trans, root, ret);
4008                 goto err;
4009         }
4010
4011         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4012                                          inode, dir_ino);
4013         if (ret != 0 && ret != -ENOENT) {
4014                 btrfs_abort_transaction(trans, root, ret);
4015                 goto err;
4016         }
4017
4018         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4019                                            dir, index);
4020         if (ret == -ENOENT)
4021                 ret = 0;
4022         else if (ret)
4023                 btrfs_abort_transaction(trans, root, ret);
4024 err:
4025         btrfs_free_path(path);
4026         if (ret)
4027                 goto out;
4028
4029         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4030         inode_inc_iversion(inode);
4031         inode_inc_iversion(dir);
4032         inode->i_ctime = dir->i_mtime =
4033                 dir->i_ctime = current_fs_time(inode->i_sb);
4034         ret = btrfs_update_inode(trans, root, dir);
4035 out:
4036         return ret;
4037 }
4038
4039 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4040                        struct btrfs_root *root,
4041                        struct inode *dir, struct inode *inode,
4042                        const char *name, int name_len)
4043 {
4044         int ret;
4045         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4046         if (!ret) {
4047                 drop_nlink(inode);
4048                 ret = btrfs_update_inode(trans, root, inode);
4049         }
4050         return ret;
4051 }
4052
4053 /*
4054  * helper to start transaction for unlink and rmdir.
4055  *
4056  * unlink and rmdir are special in btrfs, they do not always free space, so
4057  * if we cannot make our reservations the normal way try and see if there is
4058  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4059  * allow the unlink to occur.
4060  */
4061 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4062 {
4063         struct btrfs_root *root = BTRFS_I(dir)->root;
4064
4065         /*
4066          * 1 for the possible orphan item
4067          * 1 for the dir item
4068          * 1 for the dir index
4069          * 1 for the inode ref
4070          * 1 for the inode
4071          */
4072         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4073 }
4074
4075 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4076 {
4077         struct btrfs_root *root = BTRFS_I(dir)->root;
4078         struct btrfs_trans_handle *trans;
4079         struct inode *inode = d_inode(dentry);
4080         int ret;
4081
4082         trans = __unlink_start_trans(dir);
4083         if (IS_ERR(trans))
4084                 return PTR_ERR(trans);
4085
4086         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4087
4088         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4089                                  dentry->d_name.name, dentry->d_name.len);
4090         if (ret)
4091                 goto out;
4092
4093         if (inode->i_nlink == 0) {
4094                 ret = btrfs_orphan_add(trans, inode);
4095                 if (ret)
4096                         goto out;
4097         }
4098
4099 out:
4100         btrfs_end_transaction(trans, root);
4101         btrfs_btree_balance_dirty(root);
4102         return ret;
4103 }
4104
4105 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4106                         struct btrfs_root *root,
4107                         struct inode *dir, u64 objectid,
4108                         const char *name, int name_len)
4109 {
4110         struct btrfs_path *path;
4111         struct extent_buffer *leaf;
4112         struct btrfs_dir_item *di;
4113         struct btrfs_key key;
4114         u64 index;
4115         int ret;
4116         u64 dir_ino = btrfs_ino(dir);
4117
4118         path = btrfs_alloc_path();
4119         if (!path)
4120                 return -ENOMEM;
4121
4122         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4123                                    name, name_len, -1);
4124         if (IS_ERR_OR_NULL(di)) {
4125                 if (!di)
4126                         ret = -ENOENT;
4127                 else
4128                         ret = PTR_ERR(di);
4129                 goto out;
4130         }
4131
4132         leaf = path->nodes[0];
4133         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4134         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4135         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4136         if (ret) {
4137                 btrfs_abort_transaction(trans, root, ret);
4138                 goto out;
4139         }
4140         btrfs_release_path(path);
4141
4142         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4143                                  objectid, root->root_key.objectid,
4144                                  dir_ino, &index, name, name_len);
4145         if (ret < 0) {
4146                 if (ret != -ENOENT) {
4147                         btrfs_abort_transaction(trans, root, ret);
4148                         goto out;
4149                 }
4150                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4151                                                  name, name_len);
4152                 if (IS_ERR_OR_NULL(di)) {
4153                         if (!di)
4154                                 ret = -ENOENT;
4155                         else
4156                                 ret = PTR_ERR(di);
4157                         btrfs_abort_transaction(trans, root, ret);
4158                         goto out;
4159                 }
4160
4161                 leaf = path->nodes[0];
4162                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4163                 btrfs_release_path(path);
4164                 index = key.offset;
4165         }
4166         btrfs_release_path(path);
4167
4168         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4169         if (ret) {
4170                 btrfs_abort_transaction(trans, root, ret);
4171                 goto out;
4172         }
4173
4174         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4175         inode_inc_iversion(dir);
4176         dir->i_mtime = dir->i_ctime = current_fs_time(dir->i_sb);
4177         ret = btrfs_update_inode_fallback(trans, root, dir);
4178         if (ret)
4179                 btrfs_abort_transaction(trans, root, ret);
4180 out:
4181         btrfs_free_path(path);
4182         return ret;
4183 }
4184
4185 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4186 {
4187         struct inode *inode = d_inode(dentry);
4188         int err = 0;
4189         struct btrfs_root *root = BTRFS_I(dir)->root;
4190         struct btrfs_trans_handle *trans;
4191
4192         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4193                 return -ENOTEMPTY;
4194         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4195                 return -EPERM;
4196
4197         trans = __unlink_start_trans(dir);
4198         if (IS_ERR(trans))
4199                 return PTR_ERR(trans);
4200
4201         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4202                 err = btrfs_unlink_subvol(trans, root, dir,
4203                                           BTRFS_I(inode)->location.objectid,
4204                                           dentry->d_name.name,
4205                                           dentry->d_name.len);
4206                 goto out;
4207         }
4208
4209         err = btrfs_orphan_add(trans, inode);
4210         if (err)
4211                 goto out;
4212
4213         /* now the directory is empty */
4214         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4215                                  dentry->d_name.name, dentry->d_name.len);
4216         if (!err)
4217                 btrfs_i_size_write(inode, 0);
4218 out:
4219         btrfs_end_transaction(trans, root);
4220         btrfs_btree_balance_dirty(root);
4221
4222         return err;
4223 }
4224
4225 static int truncate_space_check(struct btrfs_trans_handle *trans,
4226                                 struct btrfs_root *root,
4227                                 u64 bytes_deleted)
4228 {
4229         int ret;
4230
4231         /*
4232          * This is only used to apply pressure to the enospc system, we don't
4233          * intend to use this reservation at all.
4234          */
4235         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4236         bytes_deleted *= root->nodesize;
4237         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4238                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4239         if (!ret) {
4240                 trace_btrfs_space_reservation(root->fs_info, "transaction",
4241                                               trans->transid,
4242                                               bytes_deleted, 1);
4243                 trans->bytes_reserved += bytes_deleted;
4244         }
4245         return ret;
4246
4247 }
4248
4249 static int truncate_inline_extent(struct inode *inode,
4250                                   struct btrfs_path *path,
4251                                   struct btrfs_key *found_key,
4252                                   const u64 item_end,
4253                                   const u64 new_size)
4254 {
4255         struct extent_buffer *leaf = path->nodes[0];
4256         int slot = path->slots[0];
4257         struct btrfs_file_extent_item *fi;
4258         u32 size = (u32)(new_size - found_key->offset);
4259         struct btrfs_root *root = BTRFS_I(inode)->root;
4260
4261         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4262
4263         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4264                 loff_t offset = new_size;
4265                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4266
4267                 /*
4268                  * Zero out the remaining of the last page of our inline extent,
4269                  * instead of directly truncating our inline extent here - that
4270                  * would be much more complex (decompressing all the data, then
4271                  * compressing the truncated data, which might be bigger than
4272                  * the size of the inline extent, resize the extent, etc).
4273                  * We release the path because to get the page we might need to
4274                  * read the extent item from disk (data not in the page cache).
4275                  */
4276                 btrfs_release_path(path);
4277                 return btrfs_truncate_block(inode, offset, page_end - offset,
4278                                         0);
4279         }
4280
4281         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4282         size = btrfs_file_extent_calc_inline_size(size);
4283         btrfs_truncate_item(root, path, size, 1);
4284
4285         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4286                 inode_sub_bytes(inode, item_end + 1 - new_size);
4287
4288         return 0;
4289 }
4290
4291 /*
4292  * this can truncate away extent items, csum items and directory items.
4293  * It starts at a high offset and removes keys until it can't find
4294  * any higher than new_size
4295  *
4296  * csum items that cross the new i_size are truncated to the new size
4297  * as well.
4298  *
4299  * min_type is the minimum key type to truncate down to.  If set to 0, this
4300  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4301  */
4302 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4303                                struct btrfs_root *root,
4304                                struct inode *inode,
4305                                u64 new_size, u32 min_type)
4306 {
4307         struct btrfs_path *path;
4308         struct extent_buffer *leaf;
4309         struct btrfs_file_extent_item *fi;
4310         struct btrfs_key key;
4311         struct btrfs_key found_key;
4312         u64 extent_start = 0;
4313         u64 extent_num_bytes = 0;
4314         u64 extent_offset = 0;
4315         u64 item_end = 0;
4316         u64 last_size = new_size;
4317         u32 found_type = (u8)-1;
4318         int found_extent;
4319         int del_item;
4320         int pending_del_nr = 0;
4321         int pending_del_slot = 0;
4322         int extent_type = -1;
4323         int ret;
4324         int err = 0;
4325         u64 ino = btrfs_ino(inode);
4326         u64 bytes_deleted = 0;
4327         bool be_nice = 0;
4328         bool should_throttle = 0;
4329         bool should_end = 0;
4330
4331         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4332
4333         /*
4334          * for non-free space inodes and ref cows, we want to back off from
4335          * time to time
4336          */
4337         if (!btrfs_is_free_space_inode(inode) &&
4338             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4339                 be_nice = 1;
4340
4341         path = btrfs_alloc_path();
4342         if (!path)
4343                 return -ENOMEM;
4344         path->reada = READA_BACK;
4345
4346         /*
4347          * We want to drop from the next block forward in case this new size is
4348          * not block aligned since we will be keeping the last block of the
4349          * extent just the way it is.
4350          */
4351         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4352             root == root->fs_info->tree_root)
4353                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4354                                         root->sectorsize), (u64)-1, 0);
4355
4356         /*
4357          * This function is also used to drop the items in the log tree before
4358          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4359          * it is used to drop the loged items. So we shouldn't kill the delayed
4360          * items.
4361          */
4362         if (min_type == 0 && root == BTRFS_I(inode)->root)
4363                 btrfs_kill_delayed_inode_items(inode);
4364
4365         key.objectid = ino;
4366         key.offset = (u64)-1;
4367         key.type = (u8)-1;
4368
4369 search_again:
4370         /*
4371          * with a 16K leaf size and 128MB extents, you can actually queue
4372          * up a huge file in a single leaf.  Most of the time that
4373          * bytes_deleted is > 0, it will be huge by the time we get here
4374          */
4375         if (be_nice && bytes_deleted > SZ_32M) {
4376                 if (btrfs_should_end_transaction(trans, root)) {
4377                         err = -EAGAIN;
4378                         goto error;
4379                 }
4380         }
4381
4382
4383         path->leave_spinning = 1;
4384         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4385         if (ret < 0) {
4386                 err = ret;
4387                 goto out;
4388         }
4389
4390         if (ret > 0) {
4391                 /* there are no items in the tree for us to truncate, we're
4392                  * done
4393                  */
4394                 if (path->slots[0] == 0)
4395                         goto out;
4396                 path->slots[0]--;
4397         }
4398
4399         while (1) {
4400                 fi = NULL;
4401                 leaf = path->nodes[0];
4402                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4403                 found_type = found_key.type;
4404
4405                 if (found_key.objectid != ino)
4406                         break;
4407
4408                 if (found_type < min_type)
4409                         break;
4410
4411                 item_end = found_key.offset;
4412                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4413                         fi = btrfs_item_ptr(leaf, path->slots[0],
4414                                             struct btrfs_file_extent_item);
4415                         extent_type = btrfs_file_extent_type(leaf, fi);
4416                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4417                                 item_end +=
4418                                     btrfs_file_extent_num_bytes(leaf, fi);
4419                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4420                                 item_end += btrfs_file_extent_inline_len(leaf,
4421                                                          path->slots[0], fi);
4422                         }
4423                         item_end--;
4424                 }
4425                 if (found_type > min_type) {
4426                         del_item = 1;
4427                 } else {
4428                         if (item_end < new_size)
4429                                 break;
4430                         if (found_key.offset >= new_size)
4431                                 del_item = 1;
4432                         else
4433                                 del_item = 0;
4434                 }
4435                 found_extent = 0;
4436                 /* FIXME, shrink the extent if the ref count is only 1 */
4437                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4438                         goto delete;
4439
4440                 if (del_item)
4441                         last_size = found_key.offset;
4442                 else
4443                         last_size = new_size;
4444
4445                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4446                         u64 num_dec;
4447                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4448                         if (!del_item) {
4449                                 u64 orig_num_bytes =
4450                                         btrfs_file_extent_num_bytes(leaf, fi);
4451                                 extent_num_bytes = ALIGN(new_size -
4452                                                 found_key.offset,
4453                                                 root->sectorsize);
4454                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4455                                                          extent_num_bytes);
4456                                 num_dec = (orig_num_bytes -
4457                                            extent_num_bytes);
4458                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4459                                              &root->state) &&
4460                                     extent_start != 0)
4461                                         inode_sub_bytes(inode, num_dec);
4462                                 btrfs_mark_buffer_dirty(leaf);
4463                         } else {
4464                                 extent_num_bytes =
4465                                         btrfs_file_extent_disk_num_bytes(leaf,
4466                                                                          fi);
4467                                 extent_offset = found_key.offset -
4468                                         btrfs_file_extent_offset(leaf, fi);
4469
4470                                 /* FIXME blocksize != 4096 */
4471                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4472                                 if (extent_start != 0) {
4473                                         found_extent = 1;
4474                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4475                                                      &root->state))
4476                                                 inode_sub_bytes(inode, num_dec);
4477                                 }
4478                         }
4479                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4480                         /*
4481                          * we can't truncate inline items that have had
4482                          * special encodings
4483                          */
4484                         if (!del_item &&
4485                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4486                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4487
4488                                 /*
4489                                  * Need to release path in order to truncate a
4490                                  * compressed extent. So delete any accumulated
4491                                  * extent items so far.
4492                                  */
4493                                 if (btrfs_file_extent_compression(leaf, fi) !=
4494                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4495                                         err = btrfs_del_items(trans, root, path,
4496                                                               pending_del_slot,
4497                                                               pending_del_nr);
4498                                         if (err) {
4499                                                 btrfs_abort_transaction(trans,
4500                                                                         root,
4501                                                                         err);
4502                                                 goto error;
4503                                         }
4504                                         pending_del_nr = 0;
4505                                 }
4506
4507                                 err = truncate_inline_extent(inode, path,
4508                                                              &found_key,
4509                                                              item_end,
4510                                                              new_size);
4511                                 if (err) {
4512                                         btrfs_abort_transaction(trans,
4513                                                                 root, err);
4514                                         goto error;
4515                                 }
4516                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4517                                             &root->state)) {
4518                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4519                         }
4520                 }
4521 delete:
4522                 if (del_item) {
4523                         if (!pending_del_nr) {
4524                                 /* no pending yet, add ourselves */
4525                                 pending_del_slot = path->slots[0];
4526                                 pending_del_nr = 1;
4527                         } else if (pending_del_nr &&
4528                                    path->slots[0] + 1 == pending_del_slot) {
4529                                 /* hop on the pending chunk */
4530                                 pending_del_nr++;
4531                                 pending_del_slot = path->slots[0];
4532                         } else {
4533                                 BUG();
4534                         }
4535                 } else {
4536                         break;
4537                 }
4538                 should_throttle = 0;
4539
4540                 if (found_extent &&
4541                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4542                      root == root->fs_info->tree_root)) {
4543                         btrfs_set_path_blocking(path);
4544                         bytes_deleted += extent_num_bytes;
4545                         ret = btrfs_free_extent(trans, root, extent_start,
4546                                                 extent_num_bytes, 0,
4547                                                 btrfs_header_owner(leaf),
4548                                                 ino, extent_offset);
4549                         BUG_ON(ret);
4550                         if (btrfs_should_throttle_delayed_refs(trans, root))
4551                                 btrfs_async_run_delayed_refs(root,
4552                                         trans->delayed_ref_updates * 2, 0);
4553                         if (be_nice) {
4554                                 if (truncate_space_check(trans, root,
4555                                                          extent_num_bytes)) {
4556                                         should_end = 1;
4557                                 }
4558                                 if (btrfs_should_throttle_delayed_refs(trans,
4559                                                                        root)) {
4560                                         should_throttle = 1;
4561                                 }
4562                         }
4563                 }
4564
4565                 if (found_type == BTRFS_INODE_ITEM_KEY)
4566                         break;
4567
4568                 if (path->slots[0] == 0 ||
4569                     path->slots[0] != pending_del_slot ||
4570                     should_throttle || should_end) {
4571                         if (pending_del_nr) {
4572                                 ret = btrfs_del_items(trans, root, path,
4573                                                 pending_del_slot,
4574                                                 pending_del_nr);
4575                                 if (ret) {
4576                                         btrfs_abort_transaction(trans,
4577                                                                 root, ret);
4578                                         goto error;
4579                                 }
4580                                 pending_del_nr = 0;
4581                         }
4582                         btrfs_release_path(path);
4583                         if (should_throttle) {
4584                                 unsigned long updates = trans->delayed_ref_updates;
4585                                 if (updates) {
4586                                         trans->delayed_ref_updates = 0;
4587                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4588                                         if (ret && !err)
4589                                                 err = ret;
4590                                 }
4591                         }
4592                         /*
4593                          * if we failed to refill our space rsv, bail out
4594                          * and let the transaction restart
4595                          */
4596                         if (should_end) {
4597                                 err = -EAGAIN;
4598                                 goto error;
4599                         }
4600                         goto search_again;
4601                 } else {
4602                         path->slots[0]--;
4603                 }
4604         }
4605 out:
4606         if (pending_del_nr) {
4607                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4608                                       pending_del_nr);
4609                 if (ret)
4610                         btrfs_abort_transaction(trans, root, ret);
4611         }
4612 error:
4613         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4614                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4615
4616         btrfs_free_path(path);
4617
4618         if (be_nice && bytes_deleted > SZ_32M) {
4619                 unsigned long updates = trans->delayed_ref_updates;
4620                 if (updates) {
4621                         trans->delayed_ref_updates = 0;
4622                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4623                         if (ret && !err)
4624                                 err = ret;
4625                 }
4626         }
4627         return err;
4628 }
4629
4630 /*
4631  * btrfs_truncate_block - read, zero a chunk and write a block
4632  * @inode - inode that we're zeroing
4633  * @from - the offset to start zeroing
4634  * @len - the length to zero, 0 to zero the entire range respective to the
4635  *      offset
4636  * @front - zero up to the offset instead of from the offset on
4637  *
4638  * This will find the block for the "from" offset and cow the block and zero the
4639  * part we want to zero.  This is used with truncate and hole punching.
4640  */
4641 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4642                         int front)
4643 {
4644         struct address_space *mapping = inode->i_mapping;
4645         struct btrfs_root *root = BTRFS_I(inode)->root;
4646         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4647         struct btrfs_ordered_extent *ordered;
4648         struct extent_state *cached_state = NULL;
4649         char *kaddr;
4650         u32 blocksize = root->sectorsize;
4651         pgoff_t index = from >> PAGE_SHIFT;
4652         unsigned offset = from & (blocksize - 1);
4653         struct page *page;
4654         gfp_t mask = btrfs_alloc_write_mask(mapping);
4655         int ret = 0;
4656         u64 block_start;
4657         u64 block_end;
4658
4659         if ((offset & (blocksize - 1)) == 0 &&
4660             (!len || ((len & (blocksize - 1)) == 0)))
4661                 goto out;
4662
4663         ret = btrfs_delalloc_reserve_space(inode,
4664                         round_down(from, blocksize), blocksize);
4665         if (ret)
4666                 goto out;
4667
4668 again:
4669         page = find_or_create_page(mapping, index, mask);
4670         if (!page) {
4671                 btrfs_delalloc_release_space(inode,
4672                                 round_down(from, blocksize),
4673                                 blocksize);
4674                 ret = -ENOMEM;
4675                 goto out;
4676         }
4677
4678         block_start = round_down(from, blocksize);
4679         block_end = block_start + blocksize - 1;
4680
4681         if (!PageUptodate(page)) {
4682                 ret = btrfs_readpage(NULL, page);
4683                 lock_page(page);
4684                 if (page->mapping != mapping) {
4685                         unlock_page(page);
4686                         put_page(page);
4687                         goto again;
4688                 }
4689                 if (!PageUptodate(page)) {
4690                         ret = -EIO;
4691                         goto out_unlock;
4692                 }
4693         }
4694         wait_on_page_writeback(page);
4695
4696         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4697         set_page_extent_mapped(page);
4698
4699         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4700         if (ordered) {
4701                 unlock_extent_cached(io_tree, block_start, block_end,
4702                                      &cached_state, GFP_NOFS);
4703                 unlock_page(page);
4704                 put_page(page);
4705                 btrfs_start_ordered_extent(inode, ordered, 1);
4706                 btrfs_put_ordered_extent(ordered);
4707                 goto again;
4708         }
4709
4710         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4711                           EXTENT_DIRTY | EXTENT_DELALLOC |
4712                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4713                           0, 0, &cached_state, GFP_NOFS);
4714
4715         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4716                                         &cached_state);
4717         if (ret) {
4718                 unlock_extent_cached(io_tree, block_start, block_end,
4719                                      &cached_state, GFP_NOFS);
4720                 goto out_unlock;
4721         }
4722
4723         if (offset != blocksize) {
4724                 if (!len)
4725                         len = blocksize - offset;
4726                 kaddr = kmap(page);
4727                 if (front)
4728                         memset(kaddr + (block_start - page_offset(page)),
4729                                 0, offset);
4730                 else
4731                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4732                                 0, len);
4733                 flush_dcache_page(page);
4734                 kunmap(page);
4735         }
4736         ClearPageChecked(page);
4737         set_page_dirty(page);
4738         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4739                              GFP_NOFS);
4740
4741 out_unlock:
4742         if (ret)
4743                 btrfs_delalloc_release_space(inode, block_start,
4744                                              blocksize);
4745         unlock_page(page);
4746         put_page(page);
4747 out:
4748         return ret;
4749 }
4750
4751 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4752                              u64 offset, u64 len)
4753 {
4754         struct btrfs_trans_handle *trans;
4755         int ret;
4756
4757         /*
4758          * Still need to make sure the inode looks like it's been updated so
4759          * that any holes get logged if we fsync.
4760          */
4761         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4762                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4763                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4764                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4765                 return 0;
4766         }
4767
4768         /*
4769          * 1 - for the one we're dropping
4770          * 1 - for the one we're adding
4771          * 1 - for updating the inode.
4772          */
4773         trans = btrfs_start_transaction(root, 3);
4774         if (IS_ERR(trans))
4775                 return PTR_ERR(trans);
4776
4777         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4778         if (ret) {
4779                 btrfs_abort_transaction(trans, root, ret);
4780                 btrfs_end_transaction(trans, root);
4781                 return ret;
4782         }
4783
4784         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4785                                        0, 0, len, 0, len, 0, 0, 0);
4786         if (ret)
4787                 btrfs_abort_transaction(trans, root, ret);
4788         else
4789                 btrfs_update_inode(trans, root, inode);
4790         btrfs_end_transaction(trans, root);
4791         return ret;
4792 }
4793
4794 /*
4795  * This function puts in dummy file extents for the area we're creating a hole
4796  * for.  So if we are truncating this file to a larger size we need to insert
4797  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4798  * the range between oldsize and size
4799  */
4800 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4801 {
4802         struct btrfs_root *root = BTRFS_I(inode)->root;
4803         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4804         struct extent_map *em = NULL;
4805         struct extent_state *cached_state = NULL;
4806         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4807         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4808         u64 block_end = ALIGN(size, root->sectorsize);
4809         u64 last_byte;
4810         u64 cur_offset;
4811         u64 hole_size;
4812         int err = 0;
4813
4814         /*
4815          * If our size started in the middle of a block we need to zero out the
4816          * rest of the block before we expand the i_size, otherwise we could
4817          * expose stale data.
4818          */
4819         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4820         if (err)
4821                 return err;
4822
4823         if (size <= hole_start)
4824                 return 0;
4825
4826         while (1) {
4827                 struct btrfs_ordered_extent *ordered;
4828
4829                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4830                                  &cached_state);
4831                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4832                                                      block_end - hole_start);
4833                 if (!ordered)
4834                         break;
4835                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4836                                      &cached_state, GFP_NOFS);
4837                 btrfs_start_ordered_extent(inode, ordered, 1);
4838                 btrfs_put_ordered_extent(ordered);
4839         }
4840
4841         cur_offset = hole_start;
4842         while (1) {
4843                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4844                                 block_end - cur_offset, 0);
4845                 if (IS_ERR(em)) {
4846                         err = PTR_ERR(em);
4847                         em = NULL;
4848                         break;
4849                 }
4850                 last_byte = min(extent_map_end(em), block_end);
4851                 last_byte = ALIGN(last_byte , root->sectorsize);
4852                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4853                         struct extent_map *hole_em;
4854                         hole_size = last_byte - cur_offset;
4855
4856                         err = maybe_insert_hole(root, inode, cur_offset,
4857                                                 hole_size);
4858                         if (err)
4859                                 break;
4860                         btrfs_drop_extent_cache(inode, cur_offset,
4861                                                 cur_offset + hole_size - 1, 0);
4862                         hole_em = alloc_extent_map();
4863                         if (!hole_em) {
4864                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4865                                         &BTRFS_I(inode)->runtime_flags);
4866                                 goto next;
4867                         }
4868                         hole_em->start = cur_offset;
4869                         hole_em->len = hole_size;
4870                         hole_em->orig_start = cur_offset;
4871
4872                         hole_em->block_start = EXTENT_MAP_HOLE;
4873                         hole_em->block_len = 0;
4874                         hole_em->orig_block_len = 0;
4875                         hole_em->ram_bytes = hole_size;
4876                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4877                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4878                         hole_em->generation = root->fs_info->generation;
4879
4880                         while (1) {
4881                                 write_lock(&em_tree->lock);
4882                                 err = add_extent_mapping(em_tree, hole_em, 1);
4883                                 write_unlock(&em_tree->lock);
4884                                 if (err != -EEXIST)
4885                                         break;
4886                                 btrfs_drop_extent_cache(inode, cur_offset,
4887                                                         cur_offset +
4888                                                         hole_size - 1, 0);
4889                         }
4890                         free_extent_map(hole_em);
4891                 }
4892 next:
4893                 free_extent_map(em);
4894                 em = NULL;
4895                 cur_offset = last_byte;
4896                 if (cur_offset >= block_end)
4897                         break;
4898         }
4899         free_extent_map(em);
4900         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4901                              GFP_NOFS);
4902         return err;
4903 }
4904
4905 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4906 {
4907         struct btrfs_root *root = BTRFS_I(inode)->root;
4908         struct btrfs_trans_handle *trans;
4909         loff_t oldsize = i_size_read(inode);
4910         loff_t newsize = attr->ia_size;
4911         int mask = attr->ia_valid;
4912         int ret;
4913
4914         /*
4915          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4916          * special case where we need to update the times despite not having
4917          * these flags set.  For all other operations the VFS set these flags
4918          * explicitly if it wants a timestamp update.
4919          */
4920         if (newsize != oldsize) {
4921                 inode_inc_iversion(inode);
4922                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4923                         inode->i_ctime = inode->i_mtime =
4924                                 current_fs_time(inode->i_sb);
4925         }
4926
4927         if (newsize > oldsize) {
4928                 /*
4929                  * Don't do an expanding truncate while snapshoting is ongoing.
4930                  * This is to ensure the snapshot captures a fully consistent
4931                  * state of this file - if the snapshot captures this expanding
4932                  * truncation, it must capture all writes that happened before
4933                  * this truncation.
4934                  */
4935                 btrfs_wait_for_snapshot_creation(root);
4936                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4937                 if (ret) {
4938                         btrfs_end_write_no_snapshoting(root);
4939                         return ret;
4940                 }
4941
4942                 trans = btrfs_start_transaction(root, 1);
4943                 if (IS_ERR(trans)) {
4944                         btrfs_end_write_no_snapshoting(root);
4945                         return PTR_ERR(trans);
4946                 }
4947
4948                 i_size_write(inode, newsize);
4949                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4950                 pagecache_isize_extended(inode, oldsize, newsize);
4951                 ret = btrfs_update_inode(trans, root, inode);
4952                 btrfs_end_write_no_snapshoting(root);
4953                 btrfs_end_transaction(trans, root);
4954         } else {
4955
4956                 /*
4957                  * We're truncating a file that used to have good data down to
4958                  * zero. Make sure it gets into the ordered flush list so that
4959                  * any new writes get down to disk quickly.
4960                  */
4961                 if (newsize == 0)
4962                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4963                                 &BTRFS_I(inode)->runtime_flags);
4964
4965                 /*
4966                  * 1 for the orphan item we're going to add
4967                  * 1 for the orphan item deletion.
4968                  */
4969                 trans = btrfs_start_transaction(root, 2);
4970                 if (IS_ERR(trans))
4971                         return PTR_ERR(trans);
4972
4973                 /*
4974                  * We need to do this in case we fail at _any_ point during the
4975                  * actual truncate.  Once we do the truncate_setsize we could
4976                  * invalidate pages which forces any outstanding ordered io to
4977                  * be instantly completed which will give us extents that need
4978                  * to be truncated.  If we fail to get an orphan inode down we
4979                  * could have left over extents that were never meant to live,
4980                  * so we need to guarantee from this point on that everything
4981                  * will be consistent.
4982                  */
4983                 ret = btrfs_orphan_add(trans, inode);
4984                 btrfs_end_transaction(trans, root);
4985                 if (ret)
4986                         return ret;
4987
4988                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4989                 truncate_setsize(inode, newsize);
4990
4991                 /* Disable nonlocked read DIO to avoid the end less truncate */
4992                 btrfs_inode_block_unlocked_dio(inode);
4993                 inode_dio_wait(inode);
4994                 btrfs_inode_resume_unlocked_dio(inode);
4995
4996                 ret = btrfs_truncate(inode);
4997                 if (ret && inode->i_nlink) {
4998                         int err;
4999
5000                         /*
5001                          * failed to truncate, disk_i_size is only adjusted down
5002                          * as we remove extents, so it should represent the true
5003                          * size of the inode, so reset the in memory size and
5004                          * delete our orphan entry.
5005                          */
5006                         trans = btrfs_join_transaction(root);
5007                         if (IS_ERR(trans)) {
5008                                 btrfs_orphan_del(NULL, inode);
5009                                 return ret;
5010                         }
5011                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5012                         err = btrfs_orphan_del(trans, inode);
5013                         if (err)
5014                                 btrfs_abort_transaction(trans, root, err);
5015                         btrfs_end_transaction(trans, root);
5016                 }
5017         }
5018
5019         return ret;
5020 }
5021
5022 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5023 {
5024         struct inode *inode = d_inode(dentry);
5025         struct btrfs_root *root = BTRFS_I(inode)->root;
5026         int err;
5027
5028         if (btrfs_root_readonly(root))
5029                 return -EROFS;
5030
5031         err = inode_change_ok(inode, attr);
5032         if (err)
5033                 return err;
5034
5035         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5036                 err = btrfs_setsize(inode, attr);
5037                 if (err)
5038                         return err;
5039         }
5040
5041         if (attr->ia_valid) {
5042                 setattr_copy(inode, attr);
5043                 inode_inc_iversion(inode);
5044                 err = btrfs_dirty_inode(inode);
5045
5046                 if (!err && attr->ia_valid & ATTR_MODE)
5047                         err = posix_acl_chmod(inode, inode->i_mode);
5048         }
5049
5050         return err;
5051 }
5052
5053 /*
5054  * While truncating the inode pages during eviction, we get the VFS calling
5055  * btrfs_invalidatepage() against each page of the inode. This is slow because
5056  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5057  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5058  * extent_state structures over and over, wasting lots of time.
5059  *
5060  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5061  * those expensive operations on a per page basis and do only the ordered io
5062  * finishing, while we release here the extent_map and extent_state structures,
5063  * without the excessive merging and splitting.
5064  */
5065 static void evict_inode_truncate_pages(struct inode *inode)
5066 {
5067         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5068         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5069         struct rb_node *node;
5070
5071         ASSERT(inode->i_state & I_FREEING);
5072         truncate_inode_pages_final(&inode->i_data);
5073
5074         write_lock(&map_tree->lock);
5075         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5076                 struct extent_map *em;
5077
5078                 node = rb_first(&map_tree->map);
5079                 em = rb_entry(node, struct extent_map, rb_node);
5080                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5081                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5082                 remove_extent_mapping(map_tree, em);
5083                 free_extent_map(em);
5084                 if (need_resched()) {
5085                         write_unlock(&map_tree->lock);
5086                         cond_resched();
5087                         write_lock(&map_tree->lock);
5088                 }
5089         }
5090         write_unlock(&map_tree->lock);
5091
5092         /*
5093          * Keep looping until we have no more ranges in the io tree.
5094          * We can have ongoing bios started by readpages (called from readahead)
5095          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5096          * still in progress (unlocked the pages in the bio but did not yet
5097          * unlocked the ranges in the io tree). Therefore this means some
5098          * ranges can still be locked and eviction started because before
5099          * submitting those bios, which are executed by a separate task (work
5100          * queue kthread), inode references (inode->i_count) were not taken
5101          * (which would be dropped in the end io callback of each bio).
5102          * Therefore here we effectively end up waiting for those bios and
5103          * anyone else holding locked ranges without having bumped the inode's
5104          * reference count - if we don't do it, when they access the inode's
5105          * io_tree to unlock a range it may be too late, leading to an
5106          * use-after-free issue.
5107          */
5108         spin_lock(&io_tree->lock);
5109         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5110                 struct extent_state *state;
5111                 struct extent_state *cached_state = NULL;
5112                 u64 start;
5113                 u64 end;
5114
5115                 node = rb_first(&io_tree->state);
5116                 state = rb_entry(node, struct extent_state, rb_node);
5117                 start = state->start;
5118                 end = state->end;
5119                 spin_unlock(&io_tree->lock);
5120
5121                 lock_extent_bits(io_tree, start, end, &cached_state);
5122
5123                 /*
5124                  * If still has DELALLOC flag, the extent didn't reach disk,
5125                  * and its reserved space won't be freed by delayed_ref.
5126                  * So we need to free its reserved space here.
5127                  * (Refer to comment in btrfs_invalidatepage, case 2)
5128                  *
5129                  * Note, end is the bytenr of last byte, so we need + 1 here.
5130                  */
5131                 if (state->state & EXTENT_DELALLOC)
5132                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5133
5134                 clear_extent_bit(io_tree, start, end,
5135                                  EXTENT_LOCKED | EXTENT_DIRTY |
5136                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5137                                  EXTENT_DEFRAG, 1, 1,
5138                                  &cached_state, GFP_NOFS);
5139
5140                 cond_resched();
5141                 spin_lock(&io_tree->lock);
5142         }
5143         spin_unlock(&io_tree->lock);
5144 }
5145
5146 void btrfs_evict_inode(struct inode *inode)
5147 {
5148         struct btrfs_trans_handle *trans;
5149         struct btrfs_root *root = BTRFS_I(inode)->root;
5150         struct btrfs_block_rsv *rsv, *global_rsv;
5151         int steal_from_global = 0;
5152         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5153         int ret;
5154
5155         trace_btrfs_inode_evict(inode);
5156
5157         evict_inode_truncate_pages(inode);
5158
5159         if (inode->i_nlink &&
5160             ((btrfs_root_refs(&root->root_item) != 0 &&
5161               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5162              btrfs_is_free_space_inode(inode)))
5163                 goto no_delete;
5164
5165         if (is_bad_inode(inode)) {
5166                 btrfs_orphan_del(NULL, inode);
5167                 goto no_delete;
5168         }
5169         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5170         if (!special_file(inode->i_mode))
5171                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5172
5173         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5174
5175         if (root->fs_info->log_root_recovering) {
5176                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5177                                  &BTRFS_I(inode)->runtime_flags));
5178                 goto no_delete;
5179         }
5180
5181         if (inode->i_nlink > 0) {
5182                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5183                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5184                 goto no_delete;
5185         }
5186
5187         ret = btrfs_commit_inode_delayed_inode(inode);
5188         if (ret) {
5189                 btrfs_orphan_del(NULL, inode);
5190                 goto no_delete;
5191         }
5192
5193         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5194         if (!rsv) {
5195                 btrfs_orphan_del(NULL, inode);
5196                 goto no_delete;
5197         }
5198         rsv->size = min_size;
5199         rsv->failfast = 1;
5200         global_rsv = &root->fs_info->global_block_rsv;
5201
5202         btrfs_i_size_write(inode, 0);
5203
5204         /*
5205          * This is a bit simpler than btrfs_truncate since we've already
5206          * reserved our space for our orphan item in the unlink, so we just
5207          * need to reserve some slack space in case we add bytes and update
5208          * inode item when doing the truncate.
5209          */
5210         while (1) {
5211                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5212                                              BTRFS_RESERVE_FLUSH_LIMIT);
5213
5214                 /*
5215                  * Try and steal from the global reserve since we will
5216                  * likely not use this space anyway, we want to try as
5217                  * hard as possible to get this to work.
5218                  */
5219                 if (ret)
5220                         steal_from_global++;
5221                 else
5222                         steal_from_global = 0;
5223                 ret = 0;
5224
5225                 /*
5226                  * steal_from_global == 0: we reserved stuff, hooray!
5227                  * steal_from_global == 1: we didn't reserve stuff, boo!
5228                  * steal_from_global == 2: we've committed, still not a lot of
5229                  * room but maybe we'll have room in the global reserve this
5230                  * time.
5231                  * steal_from_global == 3: abandon all hope!
5232                  */
5233                 if (steal_from_global > 2) {
5234                         btrfs_warn(root->fs_info,
5235                                 "Could not get space for a delete, will truncate on mount %d",
5236                                 ret);
5237                         btrfs_orphan_del(NULL, inode);
5238                         btrfs_free_block_rsv(root, rsv);
5239                         goto no_delete;
5240                 }
5241
5242                 trans = btrfs_join_transaction(root);
5243                 if (IS_ERR(trans)) {
5244                         btrfs_orphan_del(NULL, inode);
5245                         btrfs_free_block_rsv(root, rsv);
5246                         goto no_delete;
5247                 }
5248
5249                 /*
5250                  * We can't just steal from the global reserve, we need to make
5251                  * sure there is room to do it, if not we need to commit and try
5252                  * again.
5253                  */
5254                 if (steal_from_global) {
5255                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5256                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5257                                                               min_size);
5258                         else
5259                                 ret = -ENOSPC;
5260                 }
5261
5262                 /*
5263                  * Couldn't steal from the global reserve, we have too much
5264                  * pending stuff built up, commit the transaction and try it
5265                  * again.
5266                  */
5267                 if (ret) {
5268                         ret = btrfs_commit_transaction(trans, root);
5269                         if (ret) {
5270                                 btrfs_orphan_del(NULL, inode);
5271                                 btrfs_free_block_rsv(root, rsv);
5272                                 goto no_delete;
5273                         }
5274                         continue;
5275                 } else {
5276                         steal_from_global = 0;
5277                 }
5278
5279                 trans->block_rsv = rsv;
5280
5281                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5282                 if (ret != -ENOSPC && ret != -EAGAIN)
5283                         break;
5284
5285                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5286                 btrfs_end_transaction(trans, root);
5287                 trans = NULL;
5288                 btrfs_btree_balance_dirty(root);
5289         }
5290
5291         btrfs_free_block_rsv(root, rsv);
5292
5293         /*
5294          * Errors here aren't a big deal, it just means we leave orphan items
5295          * in the tree.  They will be cleaned up on the next mount.
5296          */
5297         if (ret == 0) {
5298                 trans->block_rsv = root->orphan_block_rsv;
5299                 btrfs_orphan_del(trans, inode);
5300         } else {
5301                 btrfs_orphan_del(NULL, inode);
5302         }
5303
5304         trans->block_rsv = &root->fs_info->trans_block_rsv;
5305         if (!(root == root->fs_info->tree_root ||
5306               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5307                 btrfs_return_ino(root, btrfs_ino(inode));
5308
5309         btrfs_end_transaction(trans, root);
5310         btrfs_btree_balance_dirty(root);
5311 no_delete:
5312         btrfs_remove_delayed_node(inode);
5313         clear_inode(inode);
5314 }
5315
5316 /*
5317  * this returns the key found in the dir entry in the location pointer.
5318  * If no dir entries were found, location->objectid is 0.
5319  */
5320 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5321                                struct btrfs_key *location)
5322 {
5323         const char *name = dentry->d_name.name;
5324         int namelen = dentry->d_name.len;
5325         struct btrfs_dir_item *di;
5326         struct btrfs_path *path;
5327         struct btrfs_root *root = BTRFS_I(dir)->root;
5328         int ret = 0;
5329
5330         path = btrfs_alloc_path();
5331         if (!path)
5332                 return -ENOMEM;
5333
5334         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5335                                     namelen, 0);
5336         if (IS_ERR(di))
5337                 ret = PTR_ERR(di);
5338
5339         if (IS_ERR_OR_NULL(di))
5340                 goto out_err;
5341
5342         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5343 out:
5344         btrfs_free_path(path);
5345         return ret;
5346 out_err:
5347         location->objectid = 0;
5348         goto out;
5349 }
5350
5351 /*
5352  * when we hit a tree root in a directory, the btrfs part of the inode
5353  * needs to be changed to reflect the root directory of the tree root.  This
5354  * is kind of like crossing a mount point.
5355  */
5356 static int fixup_tree_root_location(struct btrfs_root *root,
5357                                     struct inode *dir,
5358                                     struct dentry *dentry,
5359                                     struct btrfs_key *location,
5360                                     struct btrfs_root **sub_root)
5361 {
5362         struct btrfs_path *path;
5363         struct btrfs_root *new_root;
5364         struct btrfs_root_ref *ref;
5365         struct extent_buffer *leaf;
5366         struct btrfs_key key;
5367         int ret;
5368         int err = 0;
5369
5370         path = btrfs_alloc_path();
5371         if (!path) {
5372                 err = -ENOMEM;
5373                 goto out;
5374         }
5375
5376         err = -ENOENT;
5377         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5378         key.type = BTRFS_ROOT_REF_KEY;
5379         key.offset = location->objectid;
5380
5381         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5382                                 0, 0);
5383         if (ret) {
5384                 if (ret < 0)
5385                         err = ret;
5386                 goto out;
5387         }
5388
5389         leaf = path->nodes[0];
5390         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5391         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5392             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5393                 goto out;
5394
5395         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5396                                    (unsigned long)(ref + 1),
5397                                    dentry->d_name.len);
5398         if (ret)
5399                 goto out;
5400
5401         btrfs_release_path(path);
5402
5403         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5404         if (IS_ERR(new_root)) {
5405                 err = PTR_ERR(new_root);
5406                 goto out;
5407         }
5408
5409         *sub_root = new_root;
5410         location->objectid = btrfs_root_dirid(&new_root->root_item);
5411         location->type = BTRFS_INODE_ITEM_KEY;
5412         location->offset = 0;
5413         err = 0;
5414 out:
5415         btrfs_free_path(path);
5416         return err;
5417 }
5418
5419 static void inode_tree_add(struct inode *inode)
5420 {
5421         struct btrfs_root *root = BTRFS_I(inode)->root;
5422         struct btrfs_inode *entry;
5423         struct rb_node **p;
5424         struct rb_node *parent;
5425         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5426         u64 ino = btrfs_ino(inode);
5427
5428         if (inode_unhashed(inode))
5429                 return;
5430         parent = NULL;
5431         spin_lock(&root->inode_lock);
5432         p = &root->inode_tree.rb_node;
5433         while (*p) {
5434                 parent = *p;
5435                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5436
5437                 if (ino < btrfs_ino(&entry->vfs_inode))
5438                         p = &parent->rb_left;
5439                 else if (ino > btrfs_ino(&entry->vfs_inode))
5440                         p = &parent->rb_right;
5441                 else {
5442                         WARN_ON(!(entry->vfs_inode.i_state &
5443                                   (I_WILL_FREE | I_FREEING)));
5444                         rb_replace_node(parent, new, &root->inode_tree);
5445                         RB_CLEAR_NODE(parent);
5446                         spin_unlock(&root->inode_lock);
5447                         return;
5448                 }
5449         }
5450         rb_link_node(new, parent, p);
5451         rb_insert_color(new, &root->inode_tree);
5452         spin_unlock(&root->inode_lock);
5453 }
5454
5455 static void inode_tree_del(struct inode *inode)
5456 {
5457         struct btrfs_root *root = BTRFS_I(inode)->root;
5458         int empty = 0;
5459
5460         spin_lock(&root->inode_lock);
5461         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5462                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5463                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5464                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5465         }
5466         spin_unlock(&root->inode_lock);
5467
5468         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5469                 synchronize_srcu(&root->fs_info->subvol_srcu);
5470                 spin_lock(&root->inode_lock);
5471                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5472                 spin_unlock(&root->inode_lock);
5473                 if (empty)
5474                         btrfs_add_dead_root(root);
5475         }
5476 }
5477
5478 void btrfs_invalidate_inodes(struct btrfs_root *root)
5479 {
5480         struct rb_node *node;
5481         struct rb_node *prev;
5482         struct btrfs_inode *entry;
5483         struct inode *inode;
5484         u64 objectid = 0;
5485
5486         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5487                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5488
5489         spin_lock(&root->inode_lock);
5490 again:
5491         node = root->inode_tree.rb_node;
5492         prev = NULL;
5493         while (node) {
5494                 prev = node;
5495                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5496
5497                 if (objectid < btrfs_ino(&entry->vfs_inode))
5498                         node = node->rb_left;
5499                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5500                         node = node->rb_right;
5501                 else
5502                         break;
5503         }
5504         if (!node) {
5505                 while (prev) {
5506                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5507                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5508                                 node = prev;
5509                                 break;
5510                         }
5511                         prev = rb_next(prev);
5512                 }
5513         }
5514         while (node) {
5515                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5516                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5517                 inode = igrab(&entry->vfs_inode);
5518                 if (inode) {
5519                         spin_unlock(&root->inode_lock);
5520                         if (atomic_read(&inode->i_count) > 1)
5521                                 d_prune_aliases(inode);
5522                         /*
5523                          * btrfs_drop_inode will have it removed from
5524                          * the inode cache when its usage count
5525                          * hits zero.
5526                          */
5527                         iput(inode);
5528                         cond_resched();
5529                         spin_lock(&root->inode_lock);
5530                         goto again;
5531                 }
5532
5533                 if (cond_resched_lock(&root->inode_lock))
5534                         goto again;
5535
5536                 node = rb_next(node);
5537         }
5538         spin_unlock(&root->inode_lock);
5539 }
5540
5541 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5542 {
5543         struct btrfs_iget_args *args = p;
5544         inode->i_ino = args->location->objectid;
5545         memcpy(&BTRFS_I(inode)->location, args->location,
5546                sizeof(*args->location));
5547         BTRFS_I(inode)->root = args->root;
5548         return 0;
5549 }
5550
5551 static int btrfs_find_actor(struct inode *inode, void *opaque)
5552 {
5553         struct btrfs_iget_args *args = opaque;
5554         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5555                 args->root == BTRFS_I(inode)->root;
5556 }
5557
5558 static struct inode *btrfs_iget_locked(struct super_block *s,
5559                                        struct btrfs_key *location,
5560                                        struct btrfs_root *root)
5561 {
5562         struct inode *inode;
5563         struct btrfs_iget_args args;
5564         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5565
5566         args.location = location;
5567         args.root = root;
5568
5569         inode = iget5_locked(s, hashval, btrfs_find_actor,
5570                              btrfs_init_locked_inode,
5571                              (void *)&args);
5572         return inode;
5573 }
5574
5575 /* Get an inode object given its location and corresponding root.
5576  * Returns in *is_new if the inode was read from disk
5577  */
5578 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5579                          struct btrfs_root *root, int *new)
5580 {
5581         struct inode *inode;
5582
5583         inode = btrfs_iget_locked(s, location, root);
5584         if (!inode)
5585                 return ERR_PTR(-ENOMEM);
5586
5587         if (inode->i_state & I_NEW) {
5588                 btrfs_read_locked_inode(inode);
5589                 if (!is_bad_inode(inode)) {
5590                         inode_tree_add(inode);
5591                         unlock_new_inode(inode);
5592                         if (new)
5593                                 *new = 1;
5594                 } else {
5595                         unlock_new_inode(inode);
5596                         iput(inode);
5597                         inode = ERR_PTR(-ESTALE);
5598                 }
5599         }
5600
5601         return inode;
5602 }
5603
5604 static struct inode *new_simple_dir(struct super_block *s,
5605                                     struct btrfs_key *key,
5606                                     struct btrfs_root *root)
5607 {
5608         struct inode *inode = new_inode(s);
5609
5610         if (!inode)
5611                 return ERR_PTR(-ENOMEM);
5612
5613         BTRFS_I(inode)->root = root;
5614         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5615         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5616
5617         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5618         inode->i_op = &btrfs_dir_ro_inode_operations;
5619         inode->i_fop = &simple_dir_operations;
5620         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5621         inode->i_mtime = current_fs_time(inode->i_sb);
5622         inode->i_atime = inode->i_mtime;
5623         inode->i_ctime = inode->i_mtime;
5624         BTRFS_I(inode)->i_otime = inode->i_mtime;
5625
5626         return inode;
5627 }
5628
5629 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5630 {
5631         struct inode *inode;
5632         struct btrfs_root *root = BTRFS_I(dir)->root;
5633         struct btrfs_root *sub_root = root;
5634         struct btrfs_key location;
5635         int index;
5636         int ret = 0;
5637
5638         if (dentry->d_name.len > BTRFS_NAME_LEN)
5639                 return ERR_PTR(-ENAMETOOLONG);
5640
5641         ret = btrfs_inode_by_name(dir, dentry, &location);
5642         if (ret < 0)
5643                 return ERR_PTR(ret);
5644
5645         if (location.objectid == 0)
5646                 return ERR_PTR(-ENOENT);
5647
5648         if (location.type == BTRFS_INODE_ITEM_KEY) {
5649                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5650                 return inode;
5651         }
5652
5653         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5654
5655         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5656         ret = fixup_tree_root_location(root, dir, dentry,
5657                                        &location, &sub_root);
5658         if (ret < 0) {
5659                 if (ret != -ENOENT)
5660                         inode = ERR_PTR(ret);
5661                 else
5662                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5663         } else {
5664                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5665         }
5666         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5667
5668         if (!IS_ERR(inode) && root != sub_root) {
5669                 down_read(&root->fs_info->cleanup_work_sem);
5670                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5671                         ret = btrfs_orphan_cleanup(sub_root);
5672                 up_read(&root->fs_info->cleanup_work_sem);
5673                 if (ret) {
5674                         iput(inode);
5675                         inode = ERR_PTR(ret);
5676                 }
5677         }
5678
5679         return inode;
5680 }
5681
5682 static int btrfs_dentry_delete(const struct dentry *dentry)
5683 {
5684         struct btrfs_root *root;
5685         struct inode *inode = d_inode(dentry);
5686
5687         if (!inode && !IS_ROOT(dentry))
5688                 inode = d_inode(dentry->d_parent);
5689
5690         if (inode) {
5691                 root = BTRFS_I(inode)->root;
5692                 if (btrfs_root_refs(&root->root_item) == 0)
5693                         return 1;
5694
5695                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5696                         return 1;
5697         }
5698         return 0;
5699 }
5700
5701 static void btrfs_dentry_release(struct dentry *dentry)
5702 {
5703         kfree(dentry->d_fsdata);
5704 }
5705
5706 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5707                                    unsigned int flags)
5708 {
5709         struct inode *inode;
5710
5711         inode = btrfs_lookup_dentry(dir, dentry);
5712         if (IS_ERR(inode)) {
5713                 if (PTR_ERR(inode) == -ENOENT)
5714                         inode = NULL;
5715                 else
5716                         return ERR_CAST(inode);
5717         }
5718
5719         return d_splice_alias(inode, dentry);
5720 }
5721
5722 unsigned char btrfs_filetype_table[] = {
5723         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5724 };
5725
5726 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5727 {
5728         struct inode *inode = file_inode(file);
5729         struct btrfs_root *root = BTRFS_I(inode)->root;
5730         struct btrfs_item *item;
5731         struct btrfs_dir_item *di;
5732         struct btrfs_key key;
5733         struct btrfs_key found_key;
5734         struct btrfs_path *path;
5735         struct list_head ins_list;
5736         struct list_head del_list;
5737         int ret;
5738         struct extent_buffer *leaf;
5739         int slot;
5740         unsigned char d_type;
5741         int over = 0;
5742         u32 di_cur;
5743         u32 di_total;
5744         u32 di_len;
5745         int key_type = BTRFS_DIR_INDEX_KEY;
5746         char tmp_name[32];
5747         char *name_ptr;
5748         int name_len;
5749         int is_curr = 0;        /* ctx->pos points to the current index? */
5750         bool emitted;
5751
5752         /* FIXME, use a real flag for deciding about the key type */
5753         if (root->fs_info->tree_root == root)
5754                 key_type = BTRFS_DIR_ITEM_KEY;
5755
5756         if (!dir_emit_dots(file, ctx))
5757                 return 0;
5758
5759         path = btrfs_alloc_path();
5760         if (!path)
5761                 return -ENOMEM;
5762
5763         path->reada = READA_FORWARD;
5764
5765         if (key_type == BTRFS_DIR_INDEX_KEY) {
5766                 INIT_LIST_HEAD(&ins_list);
5767                 INIT_LIST_HEAD(&del_list);
5768                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5769         }
5770
5771         key.type = key_type;
5772         key.offset = ctx->pos;
5773         key.objectid = btrfs_ino(inode);
5774
5775         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5776         if (ret < 0)
5777                 goto err;
5778
5779         emitted = false;
5780         while (1) {
5781                 leaf = path->nodes[0];
5782                 slot = path->slots[0];
5783                 if (slot >= btrfs_header_nritems(leaf)) {
5784                         ret = btrfs_next_leaf(root, path);
5785                         if (ret < 0)
5786                                 goto err;
5787                         else if (ret > 0)
5788                                 break;
5789                         continue;
5790                 }
5791
5792                 item = btrfs_item_nr(slot);
5793                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5794
5795                 if (found_key.objectid != key.objectid)
5796                         break;
5797                 if (found_key.type != key_type)
5798                         break;
5799                 if (found_key.offset < ctx->pos)
5800                         goto next;
5801                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5802                     btrfs_should_delete_dir_index(&del_list,
5803                                                   found_key.offset))
5804                         goto next;
5805
5806                 ctx->pos = found_key.offset;
5807                 is_curr = 1;
5808
5809                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5810                 di_cur = 0;
5811                 di_total = btrfs_item_size(leaf, item);
5812
5813                 while (di_cur < di_total) {
5814                         struct btrfs_key location;
5815
5816                         if (verify_dir_item(root, leaf, di))
5817                                 break;
5818
5819                         name_len = btrfs_dir_name_len(leaf, di);
5820                         if (name_len <= sizeof(tmp_name)) {
5821                                 name_ptr = tmp_name;
5822                         } else {
5823                                 name_ptr = kmalloc(name_len, GFP_KERNEL);
5824                                 if (!name_ptr) {
5825                                         ret = -ENOMEM;
5826                                         goto err;
5827                                 }
5828                         }
5829                         read_extent_buffer(leaf, name_ptr,
5830                                            (unsigned long)(di + 1), name_len);
5831
5832                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5833                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5834
5835
5836                         /* is this a reference to our own snapshot? If so
5837                          * skip it.
5838                          *
5839                          * In contrast to old kernels, we insert the snapshot's
5840                          * dir item and dir index after it has been created, so
5841                          * we won't find a reference to our own snapshot. We
5842                          * still keep the following code for backward
5843                          * compatibility.
5844                          */
5845                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5846                             location.objectid == root->root_key.objectid) {
5847                                 over = 0;
5848                                 goto skip;
5849                         }
5850                         over = !dir_emit(ctx, name_ptr, name_len,
5851                                        location.objectid, d_type);
5852
5853 skip:
5854                         if (name_ptr != tmp_name)
5855                                 kfree(name_ptr);
5856
5857                         if (over)
5858                                 goto nopos;
5859                         emitted = true;
5860                         di_len = btrfs_dir_name_len(leaf, di) +
5861                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5862                         di_cur += di_len;
5863                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5864                 }
5865 next:
5866                 path->slots[0]++;
5867         }
5868
5869         if (key_type == BTRFS_DIR_INDEX_KEY) {
5870                 if (is_curr)
5871                         ctx->pos++;
5872                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list, &emitted);
5873                 if (ret)
5874                         goto nopos;
5875         }
5876
5877         /*
5878          * If we haven't emitted any dir entry, we must not touch ctx->pos as
5879          * it was was set to the termination value in previous call. We assume
5880          * that "." and ".." were emitted if we reach this point and set the
5881          * termination value as well for an empty directory.
5882          */
5883         if (ctx->pos > 2 && !emitted)
5884                 goto nopos;
5885
5886         /* Reached end of directory/root. Bump pos past the last item. */
5887         ctx->pos++;
5888
5889         /*
5890          * Stop new entries from being returned after we return the last
5891          * entry.
5892          *
5893          * New directory entries are assigned a strictly increasing
5894          * offset.  This means that new entries created during readdir
5895          * are *guaranteed* to be seen in the future by that readdir.
5896          * This has broken buggy programs which operate on names as
5897          * they're returned by readdir.  Until we re-use freed offsets
5898          * we have this hack to stop new entries from being returned
5899          * under the assumption that they'll never reach this huge
5900          * offset.
5901          *
5902          * This is being careful not to overflow 32bit loff_t unless the
5903          * last entry requires it because doing so has broken 32bit apps
5904          * in the past.
5905          */
5906         if (key_type == BTRFS_DIR_INDEX_KEY) {
5907                 if (ctx->pos >= INT_MAX)
5908                         ctx->pos = LLONG_MAX;
5909                 else
5910                         ctx->pos = INT_MAX;
5911         }
5912 nopos:
5913         ret = 0;
5914 err:
5915         if (key_type == BTRFS_DIR_INDEX_KEY)
5916                 btrfs_put_delayed_items(&ins_list, &del_list);
5917         btrfs_free_path(path);
5918         return ret;
5919 }
5920
5921 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5922 {
5923         struct btrfs_root *root = BTRFS_I(inode)->root;
5924         struct btrfs_trans_handle *trans;
5925         int ret = 0;
5926         bool nolock = false;
5927
5928         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5929                 return 0;
5930
5931         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5932                 nolock = true;
5933
5934         if (wbc->sync_mode == WB_SYNC_ALL) {
5935                 if (nolock)
5936                         trans = btrfs_join_transaction_nolock(root);
5937                 else
5938                         trans = btrfs_join_transaction(root);
5939                 if (IS_ERR(trans))
5940                         return PTR_ERR(trans);
5941                 ret = btrfs_commit_transaction(trans, root);
5942         }
5943         return ret;
5944 }
5945
5946 /*
5947  * This is somewhat expensive, updating the tree every time the
5948  * inode changes.  But, it is most likely to find the inode in cache.
5949  * FIXME, needs more benchmarking...there are no reasons other than performance
5950  * to keep or drop this code.
5951  */
5952 static int btrfs_dirty_inode(struct inode *inode)
5953 {
5954         struct btrfs_root *root = BTRFS_I(inode)->root;
5955         struct btrfs_trans_handle *trans;
5956         int ret;
5957
5958         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5959                 return 0;
5960
5961         trans = btrfs_join_transaction(root);
5962         if (IS_ERR(trans))
5963                 return PTR_ERR(trans);
5964
5965         ret = btrfs_update_inode(trans, root, inode);
5966         if (ret && ret == -ENOSPC) {
5967                 /* whoops, lets try again with the full transaction */
5968                 btrfs_end_transaction(trans, root);
5969                 trans = btrfs_start_transaction(root, 1);
5970                 if (IS_ERR(trans))
5971                         return PTR_ERR(trans);
5972
5973                 ret = btrfs_update_inode(trans, root, inode);
5974         }
5975         btrfs_end_transaction(trans, root);
5976         if (BTRFS_I(inode)->delayed_node)
5977                 btrfs_balance_delayed_items(root);
5978
5979         return ret;
5980 }
5981
5982 /*
5983  * This is a copy of file_update_time.  We need this so we can return error on
5984  * ENOSPC for updating the inode in the case of file write and mmap writes.
5985  */
5986 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5987                              int flags)
5988 {
5989         struct btrfs_root *root = BTRFS_I(inode)->root;
5990
5991         if (btrfs_root_readonly(root))
5992                 return -EROFS;
5993
5994         if (flags & S_VERSION)
5995                 inode_inc_iversion(inode);
5996         if (flags & S_CTIME)
5997                 inode->i_ctime = *now;
5998         if (flags & S_MTIME)
5999                 inode->i_mtime = *now;
6000         if (flags & S_ATIME)
6001                 inode->i_atime = *now;
6002         return btrfs_dirty_inode(inode);
6003 }
6004
6005 /*
6006  * find the highest existing sequence number in a directory
6007  * and then set the in-memory index_cnt variable to reflect
6008  * free sequence numbers
6009  */
6010 static int btrfs_set_inode_index_count(struct inode *inode)
6011 {
6012         struct btrfs_root *root = BTRFS_I(inode)->root;
6013         struct btrfs_key key, found_key;
6014         struct btrfs_path *path;
6015         struct extent_buffer *leaf;
6016         int ret;
6017
6018         key.objectid = btrfs_ino(inode);
6019         key.type = BTRFS_DIR_INDEX_KEY;
6020         key.offset = (u64)-1;
6021
6022         path = btrfs_alloc_path();
6023         if (!path)
6024                 return -ENOMEM;
6025
6026         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6027         if (ret < 0)
6028                 goto out;
6029         /* FIXME: we should be able to handle this */
6030         if (ret == 0)
6031                 goto out;
6032         ret = 0;
6033
6034         /*
6035          * MAGIC NUMBER EXPLANATION:
6036          * since we search a directory based on f_pos we have to start at 2
6037          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6038          * else has to start at 2
6039          */
6040         if (path->slots[0] == 0) {
6041                 BTRFS_I(inode)->index_cnt = 2;
6042                 goto out;
6043         }
6044
6045         path->slots[0]--;
6046
6047         leaf = path->nodes[0];
6048         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6049
6050         if (found_key.objectid != btrfs_ino(inode) ||
6051             found_key.type != BTRFS_DIR_INDEX_KEY) {
6052                 BTRFS_I(inode)->index_cnt = 2;
6053                 goto out;
6054         }
6055
6056         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6057 out:
6058         btrfs_free_path(path);
6059         return ret;
6060 }
6061
6062 /*
6063  * helper to find a free sequence number in a given directory.  This current
6064  * code is very simple, later versions will do smarter things in the btree
6065  */
6066 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6067 {
6068         int ret = 0;
6069
6070         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6071                 ret = btrfs_inode_delayed_dir_index_count(dir);
6072                 if (ret) {
6073                         ret = btrfs_set_inode_index_count(dir);
6074                         if (ret)
6075                                 return ret;
6076                 }
6077         }
6078
6079         *index = BTRFS_I(dir)->index_cnt;
6080         BTRFS_I(dir)->index_cnt++;
6081
6082         return ret;
6083 }
6084
6085 static int btrfs_insert_inode_locked(struct inode *inode)
6086 {
6087         struct btrfs_iget_args args;
6088         args.location = &BTRFS_I(inode)->location;
6089         args.root = BTRFS_I(inode)->root;
6090
6091         return insert_inode_locked4(inode,
6092                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6093                    btrfs_find_actor, &args);
6094 }
6095
6096 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6097                                      struct btrfs_root *root,
6098                                      struct inode *dir,
6099                                      const char *name, int name_len,
6100                                      u64 ref_objectid, u64 objectid,
6101                                      umode_t mode, u64 *index)
6102 {
6103         struct inode *inode;
6104         struct btrfs_inode_item *inode_item;
6105         struct btrfs_key *location;
6106         struct btrfs_path *path;
6107         struct btrfs_inode_ref *ref;
6108         struct btrfs_key key[2];
6109         u32 sizes[2];
6110         int nitems = name ? 2 : 1;
6111         unsigned long ptr;
6112         int ret;
6113
6114         path = btrfs_alloc_path();
6115         if (!path)
6116                 return ERR_PTR(-ENOMEM);
6117
6118         inode = new_inode(root->fs_info->sb);
6119         if (!inode) {
6120                 btrfs_free_path(path);
6121                 return ERR_PTR(-ENOMEM);
6122         }
6123
6124         /*
6125          * O_TMPFILE, set link count to 0, so that after this point,
6126          * we fill in an inode item with the correct link count.
6127          */
6128         if (!name)
6129                 set_nlink(inode, 0);
6130
6131         /*
6132          * we have to initialize this early, so we can reclaim the inode
6133          * number if we fail afterwards in this function.
6134          */
6135         inode->i_ino = objectid;
6136
6137         if (dir && name) {
6138                 trace_btrfs_inode_request(dir);
6139
6140                 ret = btrfs_set_inode_index(dir, index);
6141                 if (ret) {
6142                         btrfs_free_path(path);
6143                         iput(inode);
6144                         return ERR_PTR(ret);
6145                 }
6146         } else if (dir) {
6147                 *index = 0;
6148         }
6149         /*
6150          * index_cnt is ignored for everything but a dir,
6151          * btrfs_get_inode_index_count has an explanation for the magic
6152          * number
6153          */
6154         BTRFS_I(inode)->index_cnt = 2;
6155         BTRFS_I(inode)->dir_index = *index;
6156         BTRFS_I(inode)->root = root;
6157         BTRFS_I(inode)->generation = trans->transid;
6158         inode->i_generation = BTRFS_I(inode)->generation;
6159
6160         /*
6161          * We could have gotten an inode number from somebody who was fsynced
6162          * and then removed in this same transaction, so let's just set full
6163          * sync since it will be a full sync anyway and this will blow away the
6164          * old info in the log.
6165          */
6166         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6167
6168         key[0].objectid = objectid;
6169         key[0].type = BTRFS_INODE_ITEM_KEY;
6170         key[0].offset = 0;
6171
6172         sizes[0] = sizeof(struct btrfs_inode_item);
6173
6174         if (name) {
6175                 /*
6176                  * Start new inodes with an inode_ref. This is slightly more
6177                  * efficient for small numbers of hard links since they will
6178                  * be packed into one item. Extended refs will kick in if we
6179                  * add more hard links than can fit in the ref item.
6180                  */
6181                 key[1].objectid = objectid;
6182                 key[1].type = BTRFS_INODE_REF_KEY;
6183                 key[1].offset = ref_objectid;
6184
6185                 sizes[1] = name_len + sizeof(*ref);
6186         }
6187
6188         location = &BTRFS_I(inode)->location;
6189         location->objectid = objectid;
6190         location->offset = 0;
6191         location->type = BTRFS_INODE_ITEM_KEY;
6192
6193         ret = btrfs_insert_inode_locked(inode);
6194         if (ret < 0)
6195                 goto fail;
6196
6197         path->leave_spinning = 1;
6198         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6199         if (ret != 0)
6200                 goto fail_unlock;
6201
6202         inode_init_owner(inode, dir, mode);
6203         inode_set_bytes(inode, 0);
6204
6205         inode->i_mtime = current_fs_time(inode->i_sb);
6206         inode->i_atime = inode->i_mtime;
6207         inode->i_ctime = inode->i_mtime;
6208         BTRFS_I(inode)->i_otime = inode->i_mtime;
6209
6210         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6211                                   struct btrfs_inode_item);
6212         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6213                              sizeof(*inode_item));
6214         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6215
6216         if (name) {
6217                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6218                                      struct btrfs_inode_ref);
6219                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6220                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6221                 ptr = (unsigned long)(ref + 1);
6222                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6223         }
6224
6225         btrfs_mark_buffer_dirty(path->nodes[0]);
6226         btrfs_free_path(path);
6227
6228         btrfs_inherit_iflags(inode, dir);
6229
6230         if (S_ISREG(mode)) {
6231                 if (btrfs_test_opt(root, NODATASUM))
6232                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6233                 if (btrfs_test_opt(root, NODATACOW))
6234                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6235                                 BTRFS_INODE_NODATASUM;
6236         }
6237
6238         inode_tree_add(inode);
6239
6240         trace_btrfs_inode_new(inode);
6241         btrfs_set_inode_last_trans(trans, inode);
6242
6243         btrfs_update_root_times(trans, root);
6244
6245         ret = btrfs_inode_inherit_props(trans, inode, dir);
6246         if (ret)
6247                 btrfs_err(root->fs_info,
6248                           "error inheriting props for ino %llu (root %llu): %d",
6249                           btrfs_ino(inode), root->root_key.objectid, ret);
6250
6251         return inode;
6252
6253 fail_unlock:
6254         unlock_new_inode(inode);
6255 fail:
6256         if (dir && name)
6257                 BTRFS_I(dir)->index_cnt--;
6258         btrfs_free_path(path);
6259         iput(inode);
6260         return ERR_PTR(ret);
6261 }
6262
6263 static inline u8 btrfs_inode_type(struct inode *inode)
6264 {
6265         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6266 }
6267
6268 /*
6269  * utility function to add 'inode' into 'parent_inode' with
6270  * a give name and a given sequence number.
6271  * if 'add_backref' is true, also insert a backref from the
6272  * inode to the parent directory.
6273  */
6274 int btrfs_add_link(struct btrfs_trans_handle *trans,
6275                    struct inode *parent_inode, struct inode *inode,
6276                    const char *name, int name_len, int add_backref, u64 index)
6277 {
6278         int ret = 0;
6279         struct btrfs_key key;
6280         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6281         u64 ino = btrfs_ino(inode);
6282         u64 parent_ino = btrfs_ino(parent_inode);
6283
6284         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6285                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6286         } else {
6287                 key.objectid = ino;
6288                 key.type = BTRFS_INODE_ITEM_KEY;
6289                 key.offset = 0;
6290         }
6291
6292         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6293                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6294                                          key.objectid, root->root_key.objectid,
6295                                          parent_ino, index, name, name_len);
6296         } else if (add_backref) {
6297                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6298                                              parent_ino, index);
6299         }
6300
6301         /* Nothing to clean up yet */
6302         if (ret)
6303                 return ret;
6304
6305         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6306                                     parent_inode, &key,
6307                                     btrfs_inode_type(inode), index);
6308         if (ret == -EEXIST || ret == -EOVERFLOW)
6309                 goto fail_dir_item;
6310         else if (ret) {
6311                 btrfs_abort_transaction(trans, root, ret);
6312                 return ret;
6313         }
6314
6315         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6316                            name_len * 2);
6317         inode_inc_iversion(parent_inode);
6318         parent_inode->i_mtime = parent_inode->i_ctime =
6319                 current_fs_time(parent_inode->i_sb);
6320         ret = btrfs_update_inode(trans, root, parent_inode);
6321         if (ret)
6322                 btrfs_abort_transaction(trans, root, ret);
6323         return ret;
6324
6325 fail_dir_item:
6326         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6327                 u64 local_index;
6328                 int err;
6329                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6330                                  key.objectid, root->root_key.objectid,
6331                                  parent_ino, &local_index, name, name_len);
6332
6333         } else if (add_backref) {
6334                 u64 local_index;
6335                 int err;
6336
6337                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6338                                           ino, parent_ino, &local_index);
6339         }
6340         return ret;
6341 }
6342
6343 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6344                             struct inode *dir, struct dentry *dentry,
6345                             struct inode *inode, int backref, u64 index)
6346 {
6347         int err = btrfs_add_link(trans, dir, inode,
6348                                  dentry->d_name.name, dentry->d_name.len,
6349                                  backref, index);
6350         if (err > 0)
6351                 err = -EEXIST;
6352         return err;
6353 }
6354
6355 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6356                         umode_t mode, dev_t rdev)
6357 {
6358         struct btrfs_trans_handle *trans;
6359         struct btrfs_root *root = BTRFS_I(dir)->root;
6360         struct inode *inode = NULL;
6361         int err;
6362         int drop_inode = 0;
6363         u64 objectid;
6364         u64 index = 0;
6365
6366         /*
6367          * 2 for inode item and ref
6368          * 2 for dir items
6369          * 1 for xattr if selinux is on
6370          */
6371         trans = btrfs_start_transaction(root, 5);
6372         if (IS_ERR(trans))
6373                 return PTR_ERR(trans);
6374
6375         err = btrfs_find_free_ino(root, &objectid);
6376         if (err)
6377                 goto out_unlock;
6378
6379         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6380                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6381                                 mode, &index);
6382         if (IS_ERR(inode)) {
6383                 err = PTR_ERR(inode);
6384                 goto out_unlock;
6385         }
6386
6387         /*
6388         * If the active LSM wants to access the inode during
6389         * d_instantiate it needs these. Smack checks to see
6390         * if the filesystem supports xattrs by looking at the
6391         * ops vector.
6392         */
6393         inode->i_op = &btrfs_special_inode_operations;
6394         init_special_inode(inode, inode->i_mode, rdev);
6395
6396         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6397         if (err)
6398                 goto out_unlock_inode;
6399
6400         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6401         if (err) {
6402                 goto out_unlock_inode;
6403         } else {
6404                 btrfs_update_inode(trans, root, inode);
6405                 unlock_new_inode(inode);
6406                 d_instantiate(dentry, inode);
6407         }
6408
6409 out_unlock:
6410         btrfs_end_transaction(trans, root);
6411         btrfs_balance_delayed_items(root);
6412         btrfs_btree_balance_dirty(root);
6413         if (drop_inode) {
6414                 inode_dec_link_count(inode);
6415                 iput(inode);
6416         }
6417         return err;
6418
6419 out_unlock_inode:
6420         drop_inode = 1;
6421         unlock_new_inode(inode);
6422         goto out_unlock;
6423
6424 }
6425
6426 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6427                         umode_t mode, bool excl)
6428 {
6429         struct btrfs_trans_handle *trans;
6430         struct btrfs_root *root = BTRFS_I(dir)->root;
6431         struct inode *inode = NULL;
6432         int drop_inode_on_err = 0;
6433         int err;
6434         u64 objectid;
6435         u64 index = 0;
6436
6437         /*
6438          * 2 for inode item and ref
6439          * 2 for dir items
6440          * 1 for xattr if selinux is on
6441          */
6442         trans = btrfs_start_transaction(root, 5);
6443         if (IS_ERR(trans))
6444                 return PTR_ERR(trans);
6445
6446         err = btrfs_find_free_ino(root, &objectid);
6447         if (err)
6448                 goto out_unlock;
6449
6450         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6451                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6452                                 mode, &index);
6453         if (IS_ERR(inode)) {
6454                 err = PTR_ERR(inode);
6455                 goto out_unlock;
6456         }
6457         drop_inode_on_err = 1;
6458         /*
6459         * If the active LSM wants to access the inode during
6460         * d_instantiate it needs these. Smack checks to see
6461         * if the filesystem supports xattrs by looking at the
6462         * ops vector.
6463         */
6464         inode->i_fop = &btrfs_file_operations;
6465         inode->i_op = &btrfs_file_inode_operations;
6466         inode->i_mapping->a_ops = &btrfs_aops;
6467
6468         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6469         if (err)
6470                 goto out_unlock_inode;
6471
6472         err = btrfs_update_inode(trans, root, inode);
6473         if (err)
6474                 goto out_unlock_inode;
6475
6476         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6477         if (err)
6478                 goto out_unlock_inode;
6479
6480         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6481         unlock_new_inode(inode);
6482         d_instantiate(dentry, inode);
6483
6484 out_unlock:
6485         btrfs_end_transaction(trans, root);
6486         if (err && drop_inode_on_err) {
6487                 inode_dec_link_count(inode);
6488                 iput(inode);
6489         }
6490         btrfs_balance_delayed_items(root);
6491         btrfs_btree_balance_dirty(root);
6492         return err;
6493
6494 out_unlock_inode:
6495         unlock_new_inode(inode);
6496         goto out_unlock;
6497
6498 }
6499
6500 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6501                       struct dentry *dentry)
6502 {
6503         struct btrfs_trans_handle *trans = NULL;
6504         struct btrfs_root *root = BTRFS_I(dir)->root;
6505         struct inode *inode = d_inode(old_dentry);
6506         u64 index;
6507         int err;
6508         int drop_inode = 0;
6509
6510         /* do not allow sys_link's with other subvols of the same device */
6511         if (root->objectid != BTRFS_I(inode)->root->objectid)
6512                 return -EXDEV;
6513
6514         if (inode->i_nlink >= BTRFS_LINK_MAX)
6515                 return -EMLINK;
6516
6517         err = btrfs_set_inode_index(dir, &index);
6518         if (err)
6519                 goto fail;
6520
6521         /*
6522          * 2 items for inode and inode ref
6523          * 2 items for dir items
6524          * 1 item for parent inode
6525          */
6526         trans = btrfs_start_transaction(root, 5);
6527         if (IS_ERR(trans)) {
6528                 err = PTR_ERR(trans);
6529                 trans = NULL;
6530                 goto fail;
6531         }
6532
6533         /* There are several dir indexes for this inode, clear the cache. */
6534         BTRFS_I(inode)->dir_index = 0ULL;
6535         inc_nlink(inode);
6536         inode_inc_iversion(inode);
6537         inode->i_ctime = current_fs_time(inode->i_sb);
6538         ihold(inode);
6539         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6540
6541         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6542
6543         if (err) {
6544                 drop_inode = 1;
6545         } else {
6546                 struct dentry *parent = dentry->d_parent;
6547                 err = btrfs_update_inode(trans, root, inode);
6548                 if (err)
6549                         goto fail;
6550                 if (inode->i_nlink == 1) {
6551                         /*
6552                          * If new hard link count is 1, it's a file created
6553                          * with open(2) O_TMPFILE flag.
6554                          */
6555                         err = btrfs_orphan_del(trans, inode);
6556                         if (err)
6557                                 goto fail;
6558                 }
6559                 d_instantiate(dentry, inode);
6560                 btrfs_log_new_name(trans, inode, NULL, parent);
6561         }
6562
6563         btrfs_balance_delayed_items(root);
6564 fail:
6565         if (trans)
6566                 btrfs_end_transaction(trans, root);
6567         if (drop_inode) {
6568                 inode_dec_link_count(inode);
6569                 iput(inode);
6570         }
6571         btrfs_btree_balance_dirty(root);
6572         return err;
6573 }
6574
6575 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6576 {
6577         struct inode *inode = NULL;
6578         struct btrfs_trans_handle *trans;
6579         struct btrfs_root *root = BTRFS_I(dir)->root;
6580         int err = 0;
6581         int drop_on_err = 0;
6582         u64 objectid = 0;
6583         u64 index = 0;
6584
6585         /*
6586          * 2 items for inode and ref
6587          * 2 items for dir items
6588          * 1 for xattr if selinux is on
6589          */
6590         trans = btrfs_start_transaction(root, 5);
6591         if (IS_ERR(trans))
6592                 return PTR_ERR(trans);
6593
6594         err = btrfs_find_free_ino(root, &objectid);
6595         if (err)
6596                 goto out_fail;
6597
6598         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6599                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6600                                 S_IFDIR | mode, &index);
6601         if (IS_ERR(inode)) {
6602                 err = PTR_ERR(inode);
6603                 goto out_fail;
6604         }
6605
6606         drop_on_err = 1;
6607         /* these must be set before we unlock the inode */
6608         inode->i_op = &btrfs_dir_inode_operations;
6609         inode->i_fop = &btrfs_dir_file_operations;
6610
6611         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6612         if (err)
6613                 goto out_fail_inode;
6614
6615         btrfs_i_size_write(inode, 0);
6616         err = btrfs_update_inode(trans, root, inode);
6617         if (err)
6618                 goto out_fail_inode;
6619
6620         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6621                              dentry->d_name.len, 0, index);
6622         if (err)
6623                 goto out_fail_inode;
6624
6625         d_instantiate(dentry, inode);
6626         /*
6627          * mkdir is special.  We're unlocking after we call d_instantiate
6628          * to avoid a race with nfsd calling d_instantiate.
6629          */
6630         unlock_new_inode(inode);
6631         drop_on_err = 0;
6632
6633 out_fail:
6634         btrfs_end_transaction(trans, root);
6635         if (drop_on_err) {
6636                 inode_dec_link_count(inode);
6637                 iput(inode);
6638         }
6639         btrfs_balance_delayed_items(root);
6640         btrfs_btree_balance_dirty(root);
6641         return err;
6642
6643 out_fail_inode:
6644         unlock_new_inode(inode);
6645         goto out_fail;
6646 }
6647
6648 /* Find next extent map of a given extent map, caller needs to ensure locks */
6649 static struct extent_map *next_extent_map(struct extent_map *em)
6650 {
6651         struct rb_node *next;
6652
6653         next = rb_next(&em->rb_node);
6654         if (!next)
6655                 return NULL;
6656         return container_of(next, struct extent_map, rb_node);
6657 }
6658
6659 static struct extent_map *prev_extent_map(struct extent_map *em)
6660 {
6661         struct rb_node *prev;
6662
6663         prev = rb_prev(&em->rb_node);
6664         if (!prev)
6665                 return NULL;
6666         return container_of(prev, struct extent_map, rb_node);
6667 }
6668
6669 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6670  * the existing extent is the nearest extent to map_start,
6671  * and an extent that you want to insert, deal with overlap and insert
6672  * the best fitted new extent into the tree.
6673  */
6674 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6675                                 struct extent_map *existing,
6676                                 struct extent_map *em,
6677                                 u64 map_start)
6678 {
6679         struct extent_map *prev;
6680         struct extent_map *next;
6681         u64 start;
6682         u64 end;
6683         u64 start_diff;
6684
6685         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6686
6687         if (existing->start > map_start) {
6688                 next = existing;
6689                 prev = prev_extent_map(next);
6690         } else {
6691                 prev = existing;
6692                 next = next_extent_map(prev);
6693         }
6694
6695         start = prev ? extent_map_end(prev) : em->start;
6696         start = max_t(u64, start, em->start);
6697         end = next ? next->start : extent_map_end(em);
6698         end = min_t(u64, end, extent_map_end(em));
6699         start_diff = start - em->start;
6700         em->start = start;
6701         em->len = end - start;
6702         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6703             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6704                 em->block_start += start_diff;
6705                 em->block_len -= start_diff;
6706         }
6707         return add_extent_mapping(em_tree, em, 0);
6708 }
6709
6710 static noinline int uncompress_inline(struct btrfs_path *path,
6711                                       struct page *page,
6712                                       size_t pg_offset, u64 extent_offset,
6713                                       struct btrfs_file_extent_item *item)
6714 {
6715         int ret;
6716         struct extent_buffer *leaf = path->nodes[0];
6717         char *tmp;
6718         size_t max_size;
6719         unsigned long inline_size;
6720         unsigned long ptr;
6721         int compress_type;
6722
6723         WARN_ON(pg_offset != 0);
6724         compress_type = btrfs_file_extent_compression(leaf, item);
6725         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6726         inline_size = btrfs_file_extent_inline_item_len(leaf,
6727                                         btrfs_item_nr(path->slots[0]));
6728         tmp = kmalloc(inline_size, GFP_NOFS);
6729         if (!tmp)
6730                 return -ENOMEM;
6731         ptr = btrfs_file_extent_inline_start(item);
6732
6733         read_extent_buffer(leaf, tmp, ptr, inline_size);
6734
6735         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6736         ret = btrfs_decompress(compress_type, tmp, page,
6737                                extent_offset, inline_size, max_size);
6738         kfree(tmp);
6739         return ret;
6740 }
6741
6742 /*
6743  * a bit scary, this does extent mapping from logical file offset to the disk.
6744  * the ugly parts come from merging extents from the disk with the in-ram
6745  * representation.  This gets more complex because of the data=ordered code,
6746  * where the in-ram extents might be locked pending data=ordered completion.
6747  *
6748  * This also copies inline extents directly into the page.
6749  */
6750
6751 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6752                                     size_t pg_offset, u64 start, u64 len,
6753                                     int create)
6754 {
6755         int ret;
6756         int err = 0;
6757         u64 extent_start = 0;
6758         u64 extent_end = 0;
6759         u64 objectid = btrfs_ino(inode);
6760         u32 found_type;
6761         struct btrfs_path *path = NULL;
6762         struct btrfs_root *root = BTRFS_I(inode)->root;
6763         struct btrfs_file_extent_item *item;
6764         struct extent_buffer *leaf;
6765         struct btrfs_key found_key;
6766         struct extent_map *em = NULL;
6767         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6768         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6769         struct btrfs_trans_handle *trans = NULL;
6770         const bool new_inline = !page || create;
6771
6772 again:
6773         read_lock(&em_tree->lock);
6774         em = lookup_extent_mapping(em_tree, start, len);
6775         if (em)
6776                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6777         read_unlock(&em_tree->lock);
6778
6779         if (em) {
6780                 if (em->start > start || em->start + em->len <= start)
6781                         free_extent_map(em);
6782                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6783                         free_extent_map(em);
6784                 else
6785                         goto out;
6786         }
6787         em = alloc_extent_map();
6788         if (!em) {
6789                 err = -ENOMEM;
6790                 goto out;
6791         }
6792         em->bdev = root->fs_info->fs_devices->latest_bdev;
6793         em->start = EXTENT_MAP_HOLE;
6794         em->orig_start = EXTENT_MAP_HOLE;
6795         em->len = (u64)-1;
6796         em->block_len = (u64)-1;
6797
6798         if (!path) {
6799                 path = btrfs_alloc_path();
6800                 if (!path) {
6801                         err = -ENOMEM;
6802                         goto out;
6803                 }
6804                 /*
6805                  * Chances are we'll be called again, so go ahead and do
6806                  * readahead
6807                  */
6808                 path->reada = READA_FORWARD;
6809         }
6810
6811         ret = btrfs_lookup_file_extent(trans, root, path,
6812                                        objectid, start, trans != NULL);
6813         if (ret < 0) {
6814                 err = ret;
6815                 goto out;
6816         }
6817
6818         if (ret != 0) {
6819                 if (path->slots[0] == 0)
6820                         goto not_found;
6821                 path->slots[0]--;
6822         }
6823
6824         leaf = path->nodes[0];
6825         item = btrfs_item_ptr(leaf, path->slots[0],
6826                               struct btrfs_file_extent_item);
6827         /* are we inside the extent that was found? */
6828         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6829         found_type = found_key.type;
6830         if (found_key.objectid != objectid ||
6831             found_type != BTRFS_EXTENT_DATA_KEY) {
6832                 /*
6833                  * If we backup past the first extent we want to move forward
6834                  * and see if there is an extent in front of us, otherwise we'll
6835                  * say there is a hole for our whole search range which can
6836                  * cause problems.
6837                  */
6838                 extent_end = start;
6839                 goto next;
6840         }
6841
6842         found_type = btrfs_file_extent_type(leaf, item);
6843         extent_start = found_key.offset;
6844         if (found_type == BTRFS_FILE_EXTENT_REG ||
6845             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6846                 extent_end = extent_start +
6847                        btrfs_file_extent_num_bytes(leaf, item);
6848         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6849                 size_t size;
6850                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6851                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6852         }
6853 next:
6854         if (start >= extent_end) {
6855                 path->slots[0]++;
6856                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6857                         ret = btrfs_next_leaf(root, path);
6858                         if (ret < 0) {
6859                                 err = ret;
6860                                 goto out;
6861                         }
6862                         if (ret > 0)
6863                                 goto not_found;
6864                         leaf = path->nodes[0];
6865                 }
6866                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6867                 if (found_key.objectid != objectid ||
6868                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6869                         goto not_found;
6870                 if (start + len <= found_key.offset)
6871                         goto not_found;
6872                 if (start > found_key.offset)
6873                         goto next;
6874                 em->start = start;
6875                 em->orig_start = start;
6876                 em->len = found_key.offset - start;
6877                 goto not_found_em;
6878         }
6879
6880         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6881
6882         if (found_type == BTRFS_FILE_EXTENT_REG ||
6883             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6884                 goto insert;
6885         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6886                 unsigned long ptr;
6887                 char *map;
6888                 size_t size;
6889                 size_t extent_offset;
6890                 size_t copy_size;
6891
6892                 if (new_inline)
6893                         goto out;
6894
6895                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6896                 extent_offset = page_offset(page) + pg_offset - extent_start;
6897                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6898                                   size - extent_offset);
6899                 em->start = extent_start + extent_offset;
6900                 em->len = ALIGN(copy_size, root->sectorsize);
6901                 em->orig_block_len = em->len;
6902                 em->orig_start = em->start;
6903                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6904                 if (create == 0 && !PageUptodate(page)) {
6905                         if (btrfs_file_extent_compression(leaf, item) !=
6906                             BTRFS_COMPRESS_NONE) {
6907                                 ret = uncompress_inline(path, page, pg_offset,
6908                                                         extent_offset, item);
6909                                 if (ret) {
6910                                         err = ret;
6911                                         goto out;
6912                                 }
6913                         } else {
6914                                 map = kmap(page);
6915                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6916                                                    copy_size);
6917                                 if (pg_offset + copy_size < PAGE_SIZE) {
6918                                         memset(map + pg_offset + copy_size, 0,
6919                                                PAGE_SIZE - pg_offset -
6920                                                copy_size);
6921                                 }
6922                                 kunmap(page);
6923                         }
6924                         flush_dcache_page(page);
6925                 } else if (create && PageUptodate(page)) {
6926                         BUG();
6927                         if (!trans) {
6928                                 kunmap(page);
6929                                 free_extent_map(em);
6930                                 em = NULL;
6931
6932                                 btrfs_release_path(path);
6933                                 trans = btrfs_join_transaction(root);
6934
6935                                 if (IS_ERR(trans))
6936                                         return ERR_CAST(trans);
6937                                 goto again;
6938                         }
6939                         map = kmap(page);
6940                         write_extent_buffer(leaf, map + pg_offset, ptr,
6941                                             copy_size);
6942                         kunmap(page);
6943                         btrfs_mark_buffer_dirty(leaf);
6944                 }
6945                 set_extent_uptodate(io_tree, em->start,
6946                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6947                 goto insert;
6948         }
6949 not_found:
6950         em->start = start;
6951         em->orig_start = start;
6952         em->len = len;
6953 not_found_em:
6954         em->block_start = EXTENT_MAP_HOLE;
6955         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6956 insert:
6957         btrfs_release_path(path);
6958         if (em->start > start || extent_map_end(em) <= start) {
6959                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6960                         em->start, em->len, start, len);
6961                 err = -EIO;
6962                 goto out;
6963         }
6964
6965         err = 0;
6966         write_lock(&em_tree->lock);
6967         ret = add_extent_mapping(em_tree, em, 0);
6968         /* it is possible that someone inserted the extent into the tree
6969          * while we had the lock dropped.  It is also possible that
6970          * an overlapping map exists in the tree
6971          */
6972         if (ret == -EEXIST) {
6973                 struct extent_map *existing;
6974
6975                 ret = 0;
6976
6977                 existing = search_extent_mapping(em_tree, start, len);
6978                 /*
6979                  * existing will always be non-NULL, since there must be
6980                  * extent causing the -EEXIST.
6981                  */
6982                 if (existing->start == em->start &&
6983                     extent_map_end(existing) == extent_map_end(em) &&
6984                     em->block_start == existing->block_start) {
6985                         /*
6986                          * these two extents are the same, it happens
6987                          * with inlines especially
6988                          */
6989                         free_extent_map(em);
6990                         em = existing;
6991                         err = 0;
6992
6993                 } else if (start >= extent_map_end(existing) ||
6994                     start <= existing->start) {
6995                         /*
6996                          * The existing extent map is the one nearest to
6997                          * the [start, start + len) range which overlaps
6998                          */
6999                         err = merge_extent_mapping(em_tree, existing,
7000                                                    em, start);
7001                         free_extent_map(existing);
7002                         if (err) {
7003                                 free_extent_map(em);
7004                                 em = NULL;
7005                         }
7006                 } else {
7007                         free_extent_map(em);
7008                         em = existing;
7009                         err = 0;
7010                 }
7011         }
7012         write_unlock(&em_tree->lock);
7013 out:
7014
7015         trace_btrfs_get_extent(root, em);
7016
7017         btrfs_free_path(path);
7018         if (trans) {
7019                 ret = btrfs_end_transaction(trans, root);
7020                 if (!err)
7021                         err = ret;
7022         }
7023         if (err) {
7024                 free_extent_map(em);
7025                 return ERR_PTR(err);
7026         }
7027         BUG_ON(!em); /* Error is always set */
7028         return em;
7029 }
7030
7031 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7032                                            size_t pg_offset, u64 start, u64 len,
7033                                            int create)
7034 {
7035         struct extent_map *em;
7036         struct extent_map *hole_em = NULL;
7037         u64 range_start = start;
7038         u64 end;
7039         u64 found;
7040         u64 found_end;
7041         int err = 0;
7042
7043         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7044         if (IS_ERR(em))
7045                 return em;
7046         if (em) {
7047                 /*
7048                  * if our em maps to
7049                  * -  a hole or
7050                  * -  a pre-alloc extent,
7051                  * there might actually be delalloc bytes behind it.
7052                  */
7053                 if (em->block_start != EXTENT_MAP_HOLE &&
7054                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7055                         return em;
7056                 else
7057                         hole_em = em;
7058         }
7059
7060         /* check to see if we've wrapped (len == -1 or similar) */
7061         end = start + len;
7062         if (end < start)
7063                 end = (u64)-1;
7064         else
7065                 end -= 1;
7066
7067         em = NULL;
7068
7069         /* ok, we didn't find anything, lets look for delalloc */
7070         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7071                                  end, len, EXTENT_DELALLOC, 1);
7072         found_end = range_start + found;
7073         if (found_end < range_start)
7074                 found_end = (u64)-1;
7075
7076         /*
7077          * we didn't find anything useful, return
7078          * the original results from get_extent()
7079          */
7080         if (range_start > end || found_end <= start) {
7081                 em = hole_em;
7082                 hole_em = NULL;
7083                 goto out;
7084         }
7085
7086         /* adjust the range_start to make sure it doesn't
7087          * go backwards from the start they passed in
7088          */
7089         range_start = max(start, range_start);
7090         found = found_end - range_start;
7091
7092         if (found > 0) {
7093                 u64 hole_start = start;
7094                 u64 hole_len = len;
7095
7096                 em = alloc_extent_map();
7097                 if (!em) {
7098                         err = -ENOMEM;
7099                         goto out;
7100                 }
7101                 /*
7102                  * when btrfs_get_extent can't find anything it
7103                  * returns one huge hole
7104                  *
7105                  * make sure what it found really fits our range, and
7106                  * adjust to make sure it is based on the start from
7107                  * the caller
7108                  */
7109                 if (hole_em) {
7110                         u64 calc_end = extent_map_end(hole_em);
7111
7112                         if (calc_end <= start || (hole_em->start > end)) {
7113                                 free_extent_map(hole_em);
7114                                 hole_em = NULL;
7115                         } else {
7116                                 hole_start = max(hole_em->start, start);
7117                                 hole_len = calc_end - hole_start;
7118                         }
7119                 }
7120                 em->bdev = NULL;
7121                 if (hole_em && range_start > hole_start) {
7122                         /* our hole starts before our delalloc, so we
7123                          * have to return just the parts of the hole
7124                          * that go until  the delalloc starts
7125                          */
7126                         em->len = min(hole_len,
7127                                       range_start - hole_start);
7128                         em->start = hole_start;
7129                         em->orig_start = hole_start;
7130                         /*
7131                          * don't adjust block start at all,
7132                          * it is fixed at EXTENT_MAP_HOLE
7133                          */
7134                         em->block_start = hole_em->block_start;
7135                         em->block_len = hole_len;
7136                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7137                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7138                 } else {
7139                         em->start = range_start;
7140                         em->len = found;
7141                         em->orig_start = range_start;
7142                         em->block_start = EXTENT_MAP_DELALLOC;
7143                         em->block_len = found;
7144                 }
7145         } else if (hole_em) {
7146                 return hole_em;
7147         }
7148 out:
7149
7150         free_extent_map(hole_em);
7151         if (err) {
7152                 free_extent_map(em);
7153                 return ERR_PTR(err);
7154         }
7155         return em;
7156 }
7157
7158 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7159                                                   const u64 start,
7160                                                   const u64 len,
7161                                                   const u64 orig_start,
7162                                                   const u64 block_start,
7163                                                   const u64 block_len,
7164                                                   const u64 orig_block_len,
7165                                                   const u64 ram_bytes,
7166                                                   const int type)
7167 {
7168         struct extent_map *em = NULL;
7169         int ret;
7170
7171         down_read(&BTRFS_I(inode)->dio_sem);
7172         if (type != BTRFS_ORDERED_NOCOW) {
7173                 em = create_pinned_em(inode, start, len, orig_start,
7174                                       block_start, block_len, orig_block_len,
7175                                       ram_bytes, type);
7176                 if (IS_ERR(em))
7177                         goto out;
7178         }
7179         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7180                                            len, block_len, type);
7181         if (ret) {
7182                 if (em) {
7183                         free_extent_map(em);
7184                         btrfs_drop_extent_cache(inode, start,
7185                                                 start + len - 1, 0);
7186                 }
7187                 em = ERR_PTR(ret);
7188         }
7189  out:
7190         up_read(&BTRFS_I(inode)->dio_sem);
7191
7192         return em;
7193 }
7194
7195 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7196                                                   u64 start, u64 len)
7197 {
7198         struct btrfs_root *root = BTRFS_I(inode)->root;
7199         struct extent_map *em;
7200         struct btrfs_key ins;
7201         u64 alloc_hint;
7202         int ret;
7203
7204         alloc_hint = get_extent_allocation_hint(inode, start, len);
7205         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7206                                    alloc_hint, &ins, 1, 1);
7207         if (ret)
7208                 return ERR_PTR(ret);
7209
7210         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7211                                      ins.objectid, ins.offset, ins.offset,
7212                                      ins.offset, 0);
7213         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
7214         if (IS_ERR(em))
7215                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7216
7217         return em;
7218 }
7219
7220 /*
7221  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7222  * block must be cow'd
7223  */
7224 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7225                               u64 *orig_start, u64 *orig_block_len,
7226                               u64 *ram_bytes)
7227 {
7228         struct btrfs_trans_handle *trans;
7229         struct btrfs_path *path;
7230         int ret;
7231         struct extent_buffer *leaf;
7232         struct btrfs_root *root = BTRFS_I(inode)->root;
7233         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7234         struct btrfs_file_extent_item *fi;
7235         struct btrfs_key key;
7236         u64 disk_bytenr;
7237         u64 backref_offset;
7238         u64 extent_end;
7239         u64 num_bytes;
7240         int slot;
7241         int found_type;
7242         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7243
7244         path = btrfs_alloc_path();
7245         if (!path)
7246                 return -ENOMEM;
7247
7248         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7249                                        offset, 0);
7250         if (ret < 0)
7251                 goto out;
7252
7253         slot = path->slots[0];
7254         if (ret == 1) {
7255                 if (slot == 0) {
7256                         /* can't find the item, must cow */
7257                         ret = 0;
7258                         goto out;
7259                 }
7260                 slot--;
7261         }
7262         ret = 0;
7263         leaf = path->nodes[0];
7264         btrfs_item_key_to_cpu(leaf, &key, slot);
7265         if (key.objectid != btrfs_ino(inode) ||
7266             key.type != BTRFS_EXTENT_DATA_KEY) {
7267                 /* not our file or wrong item type, must cow */
7268                 goto out;
7269         }
7270
7271         if (key.offset > offset) {
7272                 /* Wrong offset, must cow */
7273                 goto out;
7274         }
7275
7276         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7277         found_type = btrfs_file_extent_type(leaf, fi);
7278         if (found_type != BTRFS_FILE_EXTENT_REG &&
7279             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7280                 /* not a regular extent, must cow */
7281                 goto out;
7282         }
7283
7284         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7285                 goto out;
7286
7287         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7288         if (extent_end <= offset)
7289                 goto out;
7290
7291         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7292         if (disk_bytenr == 0)
7293                 goto out;
7294
7295         if (btrfs_file_extent_compression(leaf, fi) ||
7296             btrfs_file_extent_encryption(leaf, fi) ||
7297             btrfs_file_extent_other_encoding(leaf, fi))
7298                 goto out;
7299
7300         backref_offset = btrfs_file_extent_offset(leaf, fi);
7301
7302         if (orig_start) {
7303                 *orig_start = key.offset - backref_offset;
7304                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7305                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7306         }
7307
7308         if (btrfs_extent_readonly(root, disk_bytenr))
7309                 goto out;
7310
7311         num_bytes = min(offset + *len, extent_end) - offset;
7312         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7313                 u64 range_end;
7314
7315                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7316                 ret = test_range_bit(io_tree, offset, range_end,
7317                                      EXTENT_DELALLOC, 0, NULL);
7318                 if (ret) {
7319                         ret = -EAGAIN;
7320                         goto out;
7321                 }
7322         }
7323
7324         btrfs_release_path(path);
7325
7326         /*
7327          * look for other files referencing this extent, if we
7328          * find any we must cow
7329          */
7330         trans = btrfs_join_transaction(root);
7331         if (IS_ERR(trans)) {
7332                 ret = 0;
7333                 goto out;
7334         }
7335
7336         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7337                                     key.offset - backref_offset, disk_bytenr);
7338         btrfs_end_transaction(trans, root);
7339         if (ret) {
7340                 ret = 0;
7341                 goto out;
7342         }
7343
7344         /*
7345          * adjust disk_bytenr and num_bytes to cover just the bytes
7346          * in this extent we are about to write.  If there
7347          * are any csums in that range we have to cow in order
7348          * to keep the csums correct
7349          */
7350         disk_bytenr += backref_offset;
7351         disk_bytenr += offset - key.offset;
7352         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7353                                 goto out;
7354         /*
7355          * all of the above have passed, it is safe to overwrite this extent
7356          * without cow
7357          */
7358         *len = num_bytes;
7359         ret = 1;
7360 out:
7361         btrfs_free_path(path);
7362         return ret;
7363 }
7364
7365 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7366 {
7367         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7368         int found = false;
7369         void **pagep = NULL;
7370         struct page *page = NULL;
7371         int start_idx;
7372         int end_idx;
7373
7374         start_idx = start >> PAGE_SHIFT;
7375
7376         /*
7377          * end is the last byte in the last page.  end == start is legal
7378          */
7379         end_idx = end >> PAGE_SHIFT;
7380
7381         rcu_read_lock();
7382
7383         /* Most of the code in this while loop is lifted from
7384          * find_get_page.  It's been modified to begin searching from a
7385          * page and return just the first page found in that range.  If the
7386          * found idx is less than or equal to the end idx then we know that
7387          * a page exists.  If no pages are found or if those pages are
7388          * outside of the range then we're fine (yay!) */
7389         while (page == NULL &&
7390                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7391                 page = radix_tree_deref_slot(pagep);
7392                 if (unlikely(!page))
7393                         break;
7394
7395                 if (radix_tree_exception(page)) {
7396                         if (radix_tree_deref_retry(page)) {
7397                                 page = NULL;
7398                                 continue;
7399                         }
7400                         /*
7401                          * Otherwise, shmem/tmpfs must be storing a swap entry
7402                          * here as an exceptional entry: so return it without
7403                          * attempting to raise page count.
7404                          */
7405                         page = NULL;
7406                         break; /* TODO: Is this relevant for this use case? */
7407                 }
7408
7409                 if (!page_cache_get_speculative(page)) {
7410                         page = NULL;
7411                         continue;
7412                 }
7413
7414                 /*
7415                  * Has the page moved?
7416                  * This is part of the lockless pagecache protocol. See
7417                  * include/linux/pagemap.h for details.
7418                  */
7419                 if (unlikely(page != *pagep)) {
7420                         put_page(page);
7421                         page = NULL;
7422                 }
7423         }
7424
7425         if (page) {
7426                 if (page->index <= end_idx)
7427                         found = true;
7428                 put_page(page);
7429         }
7430
7431         rcu_read_unlock();
7432         return found;
7433 }
7434
7435 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7436                               struct extent_state **cached_state, int writing)
7437 {
7438         struct btrfs_ordered_extent *ordered;
7439         int ret = 0;
7440
7441         while (1) {
7442                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7443                                  cached_state);
7444                 /*
7445                  * We're concerned with the entire range that we're going to be
7446                  * doing DIO to, so we need to make sure there's no ordered
7447                  * extents in this range.
7448                  */
7449                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7450                                                      lockend - lockstart + 1);
7451
7452                 /*
7453                  * We need to make sure there are no buffered pages in this
7454                  * range either, we could have raced between the invalidate in
7455                  * generic_file_direct_write and locking the extent.  The
7456                  * invalidate needs to happen so that reads after a write do not
7457                  * get stale data.
7458                  */
7459                 if (!ordered &&
7460                     (!writing ||
7461                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7462                         break;
7463
7464                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7465                                      cached_state, GFP_NOFS);
7466
7467                 if (ordered) {
7468                         /*
7469                          * If we are doing a DIO read and the ordered extent we
7470                          * found is for a buffered write, we can not wait for it
7471                          * to complete and retry, because if we do so we can
7472                          * deadlock with concurrent buffered writes on page
7473                          * locks. This happens only if our DIO read covers more
7474                          * than one extent map, if at this point has already
7475                          * created an ordered extent for a previous extent map
7476                          * and locked its range in the inode's io tree, and a
7477                          * concurrent write against that previous extent map's
7478                          * range and this range started (we unlock the ranges
7479                          * in the io tree only when the bios complete and
7480                          * buffered writes always lock pages before attempting
7481                          * to lock range in the io tree).
7482                          */
7483                         if (writing ||
7484                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7485                                 btrfs_start_ordered_extent(inode, ordered, 1);
7486                         else
7487                                 ret = -ENOTBLK;
7488                         btrfs_put_ordered_extent(ordered);
7489                 } else {
7490                         /*
7491                          * We could trigger writeback for this range (and wait
7492                          * for it to complete) and then invalidate the pages for
7493                          * this range (through invalidate_inode_pages2_range()),
7494                          * but that can lead us to a deadlock with a concurrent
7495                          * call to readpages() (a buffered read or a defrag call
7496                          * triggered a readahead) on a page lock due to an
7497                          * ordered dio extent we created before but did not have
7498                          * yet a corresponding bio submitted (whence it can not
7499                          * complete), which makes readpages() wait for that
7500                          * ordered extent to complete while holding a lock on
7501                          * that page.
7502                          */
7503                         ret = -ENOTBLK;
7504                 }
7505
7506                 if (ret)
7507                         break;
7508
7509                 cond_resched();
7510         }
7511
7512         return ret;
7513 }
7514
7515 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7516                                            u64 len, u64 orig_start,
7517                                            u64 block_start, u64 block_len,
7518                                            u64 orig_block_len, u64 ram_bytes,
7519                                            int type)
7520 {
7521         struct extent_map_tree *em_tree;
7522         struct extent_map *em;
7523         struct btrfs_root *root = BTRFS_I(inode)->root;
7524         int ret;
7525
7526         em_tree = &BTRFS_I(inode)->extent_tree;
7527         em = alloc_extent_map();
7528         if (!em)
7529                 return ERR_PTR(-ENOMEM);
7530
7531         em->start = start;
7532         em->orig_start = orig_start;
7533         em->mod_start = start;
7534         em->mod_len = len;
7535         em->len = len;
7536         em->block_len = block_len;
7537         em->block_start = block_start;
7538         em->bdev = root->fs_info->fs_devices->latest_bdev;
7539         em->orig_block_len = orig_block_len;
7540         em->ram_bytes = ram_bytes;
7541         em->generation = -1;
7542         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7543         if (type == BTRFS_ORDERED_PREALLOC)
7544                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7545
7546         do {
7547                 btrfs_drop_extent_cache(inode, em->start,
7548                                 em->start + em->len - 1, 0);
7549                 write_lock(&em_tree->lock);
7550                 ret = add_extent_mapping(em_tree, em, 1);
7551                 write_unlock(&em_tree->lock);
7552         } while (ret == -EEXIST);
7553
7554         if (ret) {
7555                 free_extent_map(em);
7556                 return ERR_PTR(ret);
7557         }
7558
7559         return em;
7560 }
7561
7562 static void adjust_dio_outstanding_extents(struct inode *inode,
7563                                            struct btrfs_dio_data *dio_data,
7564                                            const u64 len)
7565 {
7566         unsigned num_extents;
7567
7568         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7569                                            BTRFS_MAX_EXTENT_SIZE);
7570         /*
7571          * If we have an outstanding_extents count still set then we're
7572          * within our reservation, otherwise we need to adjust our inode
7573          * counter appropriately.
7574          */
7575         if (dio_data->outstanding_extents) {
7576                 dio_data->outstanding_extents -= num_extents;
7577         } else {
7578                 spin_lock(&BTRFS_I(inode)->lock);
7579                 BTRFS_I(inode)->outstanding_extents += num_extents;
7580                 spin_unlock(&BTRFS_I(inode)->lock);
7581         }
7582 }
7583
7584 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7585                                    struct buffer_head *bh_result, int create)
7586 {
7587         struct extent_map *em;
7588         struct btrfs_root *root = BTRFS_I(inode)->root;
7589         struct extent_state *cached_state = NULL;
7590         struct btrfs_dio_data *dio_data = NULL;
7591         u64 start = iblock << inode->i_blkbits;
7592         u64 lockstart, lockend;
7593         u64 len = bh_result->b_size;
7594         int unlock_bits = EXTENT_LOCKED;
7595         int ret = 0;
7596
7597         if (create)
7598                 unlock_bits |= EXTENT_DIRTY;
7599         else
7600                 len = min_t(u64, len, root->sectorsize);
7601
7602         lockstart = start;
7603         lockend = start + len - 1;
7604
7605         if (current->journal_info) {
7606                 /*
7607                  * Need to pull our outstanding extents and set journal_info to NULL so
7608                  * that anything that needs to check if there's a transaction doesn't get
7609                  * confused.
7610                  */
7611                 dio_data = current->journal_info;
7612                 current->journal_info = NULL;
7613         }
7614
7615         /*
7616          * If this errors out it's because we couldn't invalidate pagecache for
7617          * this range and we need to fallback to buffered.
7618          */
7619         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7620                                create)) {
7621                 ret = -ENOTBLK;
7622                 goto err;
7623         }
7624
7625         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7626         if (IS_ERR(em)) {
7627                 ret = PTR_ERR(em);
7628                 goto unlock_err;
7629         }
7630
7631         /*
7632          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7633          * io.  INLINE is special, and we could probably kludge it in here, but
7634          * it's still buffered so for safety lets just fall back to the generic
7635          * buffered path.
7636          *
7637          * For COMPRESSED we _have_ to read the entire extent in so we can
7638          * decompress it, so there will be buffering required no matter what we
7639          * do, so go ahead and fallback to buffered.
7640          *
7641          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7642          * to buffered IO.  Don't blame me, this is the price we pay for using
7643          * the generic code.
7644          */
7645         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7646             em->block_start == EXTENT_MAP_INLINE) {
7647                 free_extent_map(em);
7648                 ret = -ENOTBLK;
7649                 goto unlock_err;
7650         }
7651
7652         /* Just a good old fashioned hole, return */
7653         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7654                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7655                 free_extent_map(em);
7656                 goto unlock_err;
7657         }
7658
7659         /*
7660          * We don't allocate a new extent in the following cases
7661          *
7662          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7663          * existing extent.
7664          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7665          * just use the extent.
7666          *
7667          */
7668         if (!create) {
7669                 len = min(len, em->len - (start - em->start));
7670                 lockstart = start + len;
7671                 goto unlock;
7672         }
7673
7674         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7675             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7676              em->block_start != EXTENT_MAP_HOLE)) {
7677                 int type;
7678                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7679
7680                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7681                         type = BTRFS_ORDERED_PREALLOC;
7682                 else
7683                         type = BTRFS_ORDERED_NOCOW;
7684                 len = min(len, em->len - (start - em->start));
7685                 block_start = em->block_start + (start - em->start);
7686
7687                 if (can_nocow_extent(inode, start, &len, &orig_start,
7688                                      &orig_block_len, &ram_bytes) == 1 &&
7689                     btrfs_inc_nocow_writers(root->fs_info, block_start)) {
7690                         struct extent_map *em2;
7691
7692                         em2 = btrfs_create_dio_extent(inode, start, len,
7693                                                       orig_start, block_start,
7694                                                       len, orig_block_len,
7695                                                       ram_bytes, type);
7696                         btrfs_dec_nocow_writers(root->fs_info, block_start);
7697                         if (type == BTRFS_ORDERED_PREALLOC) {
7698                                 free_extent_map(em);
7699                                 em = em2;
7700                         }
7701                         if (em2 && IS_ERR(em2)) {
7702                                 ret = PTR_ERR(em2);
7703                                 goto unlock_err;
7704                         }
7705                         goto unlock;
7706                 }
7707         }
7708
7709         /*
7710          * this will cow the extent, reset the len in case we changed
7711          * it above
7712          */
7713         len = bh_result->b_size;
7714         free_extent_map(em);
7715         em = btrfs_new_extent_direct(inode, start, len);
7716         if (IS_ERR(em)) {
7717                 ret = PTR_ERR(em);
7718                 goto unlock_err;
7719         }
7720         len = min(len, em->len - (start - em->start));
7721 unlock:
7722         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7723                 inode->i_blkbits;
7724         bh_result->b_size = len;
7725         bh_result->b_bdev = em->bdev;
7726         set_buffer_mapped(bh_result);
7727         if (create) {
7728                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7729                         set_buffer_new(bh_result);
7730
7731                 /*
7732                  * Need to update the i_size under the extent lock so buffered
7733                  * readers will get the updated i_size when we unlock.
7734                  */
7735                 if (start + len > i_size_read(inode))
7736                         i_size_write(inode, start + len);
7737
7738                 adjust_dio_outstanding_extents(inode, dio_data, len);
7739                 btrfs_free_reserved_data_space(inode, start, len);
7740                 WARN_ON(dio_data->reserve < len);
7741                 dio_data->reserve -= len;
7742                 dio_data->unsubmitted_oe_range_end = start + len;
7743                 current->journal_info = dio_data;
7744         }
7745
7746         /*
7747          * In the case of write we need to clear and unlock the entire range,
7748          * in the case of read we need to unlock only the end area that we
7749          * aren't using if there is any left over space.
7750          */
7751         if (lockstart < lockend) {
7752                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7753                                  lockend, unlock_bits, 1, 0,
7754                                  &cached_state, GFP_NOFS);
7755         } else {
7756                 free_extent_state(cached_state);
7757         }
7758
7759         free_extent_map(em);
7760
7761         return 0;
7762
7763 unlock_err:
7764         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7765                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7766 err:
7767         if (dio_data)
7768                 current->journal_info = dio_data;
7769         /*
7770          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7771          * write less data then expected, so that we don't underflow our inode's
7772          * outstanding extents counter.
7773          */
7774         if (create && dio_data)
7775                 adjust_dio_outstanding_extents(inode, dio_data, len);
7776
7777         return ret;
7778 }
7779
7780 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7781                                         int rw, int mirror_num)
7782 {
7783         struct btrfs_root *root = BTRFS_I(inode)->root;
7784         int ret;
7785
7786         BUG_ON(rw & REQ_WRITE);
7787
7788         bio_get(bio);
7789
7790         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7791                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7792         if (ret)
7793                 goto err;
7794
7795         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7796 err:
7797         bio_put(bio);
7798         return ret;
7799 }
7800
7801 static int btrfs_check_dio_repairable(struct inode *inode,
7802                                       struct bio *failed_bio,
7803                                       struct io_failure_record *failrec,
7804                                       int failed_mirror)
7805 {
7806         int num_copies;
7807
7808         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7809                                       failrec->logical, failrec->len);
7810         if (num_copies == 1) {
7811                 /*
7812                  * we only have a single copy of the data, so don't bother with
7813                  * all the retry and error correction code that follows. no
7814                  * matter what the error is, it is very likely to persist.
7815                  */
7816                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7817                          num_copies, failrec->this_mirror, failed_mirror);
7818                 return 0;
7819         }
7820
7821         failrec->failed_mirror = failed_mirror;
7822         failrec->this_mirror++;
7823         if (failrec->this_mirror == failed_mirror)
7824                 failrec->this_mirror++;
7825
7826         if (failrec->this_mirror > num_copies) {
7827                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7828                          num_copies, failrec->this_mirror, failed_mirror);
7829                 return 0;
7830         }
7831
7832         return 1;
7833 }
7834
7835 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7836                         struct page *page, unsigned int pgoff,
7837                         u64 start, u64 end, int failed_mirror,
7838                         bio_end_io_t *repair_endio, void *repair_arg)
7839 {
7840         struct io_failure_record *failrec;
7841         struct bio *bio;
7842         int isector;
7843         int read_mode;
7844         int ret;
7845
7846         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7847
7848         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7849         if (ret)
7850                 return ret;
7851
7852         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7853                                          failed_mirror);
7854         if (!ret) {
7855                 free_io_failure(inode, failrec);
7856                 return -EIO;
7857         }
7858
7859         if ((failed_bio->bi_vcnt > 1)
7860                 || (failed_bio->bi_io_vec->bv_len
7861                         > BTRFS_I(inode)->root->sectorsize))
7862                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7863         else
7864                 read_mode = READ_SYNC;
7865
7866         isector = start - btrfs_io_bio(failed_bio)->logical;
7867         isector >>= inode->i_sb->s_blocksize_bits;
7868         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7869                                 pgoff, isector, repair_endio, repair_arg);
7870         if (!bio) {
7871                 free_io_failure(inode, failrec);
7872                 return -EIO;
7873         }
7874
7875         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7876                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7877                     read_mode, failrec->this_mirror, failrec->in_validation);
7878
7879         ret = submit_dio_repair_bio(inode, bio, read_mode,
7880                                     failrec->this_mirror);
7881         if (ret) {
7882                 free_io_failure(inode, failrec);
7883                 bio_put(bio);
7884         }
7885
7886         return ret;
7887 }
7888
7889 struct btrfs_retry_complete {
7890         struct completion done;
7891         struct inode *inode;
7892         u64 start;
7893         int uptodate;
7894 };
7895
7896 static void btrfs_retry_endio_nocsum(struct bio *bio)
7897 {
7898         struct btrfs_retry_complete *done = bio->bi_private;
7899         struct inode *inode;
7900         struct bio_vec *bvec;
7901         int i;
7902
7903         if (bio->bi_error)
7904                 goto end;
7905
7906         ASSERT(bio->bi_vcnt == 1);
7907         inode = bio->bi_io_vec->bv_page->mapping->host;
7908         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7909
7910         done->uptodate = 1;
7911         bio_for_each_segment_all(bvec, bio, i)
7912                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7913 end:
7914         complete(&done->done);
7915         bio_put(bio);
7916 }
7917
7918 static int __btrfs_correct_data_nocsum(struct inode *inode,
7919                                        struct btrfs_io_bio *io_bio)
7920 {
7921         struct btrfs_fs_info *fs_info;
7922         struct bio_vec *bvec;
7923         struct btrfs_retry_complete done;
7924         u64 start;
7925         unsigned int pgoff;
7926         u32 sectorsize;
7927         int nr_sectors;
7928         int i;
7929         int ret;
7930
7931         fs_info = BTRFS_I(inode)->root->fs_info;
7932         sectorsize = BTRFS_I(inode)->root->sectorsize;
7933
7934         start = io_bio->logical;
7935         done.inode = inode;
7936
7937         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7938                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
7939                 pgoff = bvec->bv_offset;
7940
7941 next_block_or_try_again:
7942                 done.uptodate = 0;
7943                 done.start = start;
7944                 init_completion(&done.done);
7945
7946                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
7947                                 pgoff, start, start + sectorsize - 1,
7948                                 io_bio->mirror_num,
7949                                 btrfs_retry_endio_nocsum, &done);
7950                 if (ret)
7951                         return ret;
7952
7953                 wait_for_completion(&done.done);
7954
7955                 if (!done.uptodate) {
7956                         /* We might have another mirror, so try again */
7957                         goto next_block_or_try_again;
7958                 }
7959
7960                 start += sectorsize;
7961
7962                 if (nr_sectors--) {
7963                         pgoff += sectorsize;
7964                         goto next_block_or_try_again;
7965                 }
7966         }
7967
7968         return 0;
7969 }
7970
7971 static void btrfs_retry_endio(struct bio *bio)
7972 {
7973         struct btrfs_retry_complete *done = bio->bi_private;
7974         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7975         struct inode *inode;
7976         struct bio_vec *bvec;
7977         u64 start;
7978         int uptodate;
7979         int ret;
7980         int i;
7981
7982         if (bio->bi_error)
7983                 goto end;
7984
7985         uptodate = 1;
7986
7987         start = done->start;
7988
7989         ASSERT(bio->bi_vcnt == 1);
7990         inode = bio->bi_io_vec->bv_page->mapping->host;
7991         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7992
7993         bio_for_each_segment_all(bvec, bio, i) {
7994                 ret = __readpage_endio_check(done->inode, io_bio, i,
7995                                         bvec->bv_page, bvec->bv_offset,
7996                                         done->start, bvec->bv_len);
7997                 if (!ret)
7998                         clean_io_failure(done->inode, done->start,
7999                                         bvec->bv_page, bvec->bv_offset);
8000                 else
8001                         uptodate = 0;
8002         }
8003
8004         done->uptodate = uptodate;
8005 end:
8006         complete(&done->done);
8007         bio_put(bio);
8008 }
8009
8010 static int __btrfs_subio_endio_read(struct inode *inode,
8011                                     struct btrfs_io_bio *io_bio, int err)
8012 {
8013         struct btrfs_fs_info *fs_info;
8014         struct bio_vec *bvec;
8015         struct btrfs_retry_complete done;
8016         u64 start;
8017         u64 offset = 0;
8018         u32 sectorsize;
8019         int nr_sectors;
8020         unsigned int pgoff;
8021         int csum_pos;
8022         int i;
8023         int ret;
8024
8025         fs_info = BTRFS_I(inode)->root->fs_info;
8026         sectorsize = BTRFS_I(inode)->root->sectorsize;
8027
8028         err = 0;
8029         start = io_bio->logical;
8030         done.inode = inode;
8031
8032         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8033                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8034
8035                 pgoff = bvec->bv_offset;
8036 next_block:
8037                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8038                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8039                                         bvec->bv_page, pgoff, start,
8040                                         sectorsize);
8041                 if (likely(!ret))
8042                         goto next;
8043 try_again:
8044                 done.uptodate = 0;
8045                 done.start = start;
8046                 init_completion(&done.done);
8047
8048                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8049                                 pgoff, start, start + sectorsize - 1,
8050                                 io_bio->mirror_num,
8051                                 btrfs_retry_endio, &done);
8052                 if (ret) {
8053                         err = ret;
8054                         goto next;
8055                 }
8056
8057                 wait_for_completion(&done.done);
8058
8059                 if (!done.uptodate) {
8060                         /* We might have another mirror, so try again */
8061                         goto try_again;
8062                 }
8063 next:
8064                 offset += sectorsize;
8065                 start += sectorsize;
8066
8067                 ASSERT(nr_sectors);
8068
8069                 if (--nr_sectors) {
8070                         pgoff += sectorsize;
8071                         goto next_block;
8072                 }
8073         }
8074
8075         return err;
8076 }
8077
8078 static int btrfs_subio_endio_read(struct inode *inode,
8079                                   struct btrfs_io_bio *io_bio, int err)
8080 {
8081         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8082
8083         if (skip_csum) {
8084                 if (unlikely(err))
8085                         return __btrfs_correct_data_nocsum(inode, io_bio);
8086                 else
8087                         return 0;
8088         } else {
8089                 return __btrfs_subio_endio_read(inode, io_bio, err);
8090         }
8091 }
8092
8093 static void btrfs_endio_direct_read(struct bio *bio)
8094 {
8095         struct btrfs_dio_private *dip = bio->bi_private;
8096         struct inode *inode = dip->inode;
8097         struct bio *dio_bio;
8098         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8099         int err = bio->bi_error;
8100
8101         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8102                 err = btrfs_subio_endio_read(inode, io_bio, err);
8103
8104         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8105                       dip->logical_offset + dip->bytes - 1);
8106         dio_bio = dip->dio_bio;
8107
8108         kfree(dip);
8109
8110         dio_bio->bi_error = bio->bi_error;
8111         dio_end_io(dio_bio, bio->bi_error);
8112
8113         if (io_bio->end_io)
8114                 io_bio->end_io(io_bio, err);
8115         bio_put(bio);
8116 }
8117
8118 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8119                                                     const u64 offset,
8120                                                     const u64 bytes,
8121                                                     const int uptodate)
8122 {
8123         struct btrfs_root *root = BTRFS_I(inode)->root;
8124         struct btrfs_ordered_extent *ordered = NULL;
8125         u64 ordered_offset = offset;
8126         u64 ordered_bytes = bytes;
8127         int ret;
8128
8129 again:
8130         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8131                                                    &ordered_offset,
8132                                                    ordered_bytes,
8133                                                    uptodate);
8134         if (!ret)
8135                 goto out_test;
8136
8137         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8138                         finish_ordered_fn, NULL, NULL);
8139         btrfs_queue_work(root->fs_info->endio_write_workers,
8140                          &ordered->work);
8141 out_test:
8142         /*
8143          * our bio might span multiple ordered extents.  If we haven't
8144          * completed the accounting for the whole dio, go back and try again
8145          */
8146         if (ordered_offset < offset + bytes) {
8147                 ordered_bytes = offset + bytes - ordered_offset;
8148                 ordered = NULL;
8149                 goto again;
8150         }
8151 }
8152
8153 static void btrfs_endio_direct_write(struct bio *bio)
8154 {
8155         struct btrfs_dio_private *dip = bio->bi_private;
8156         struct bio *dio_bio = dip->dio_bio;
8157
8158         btrfs_endio_direct_write_update_ordered(dip->inode,
8159                                                 dip->logical_offset,
8160                                                 dip->bytes,
8161                                                 !bio->bi_error);
8162
8163         kfree(dip);
8164
8165         dio_bio->bi_error = bio->bi_error;
8166         dio_end_io(dio_bio, bio->bi_error);
8167         bio_put(bio);
8168 }
8169
8170 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8171                                     struct bio *bio, int mirror_num,
8172                                     unsigned long bio_flags, u64 offset)
8173 {
8174         int ret;
8175         struct btrfs_root *root = BTRFS_I(inode)->root;
8176         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8177         BUG_ON(ret); /* -ENOMEM */
8178         return 0;
8179 }
8180
8181 static void btrfs_end_dio_bio(struct bio *bio)
8182 {
8183         struct btrfs_dio_private *dip = bio->bi_private;
8184         int err = bio->bi_error;
8185
8186         if (err)
8187                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8188                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8189                            btrfs_ino(dip->inode), bio->bi_rw,
8190                            (unsigned long long)bio->bi_iter.bi_sector,
8191                            bio->bi_iter.bi_size, err);
8192
8193         if (dip->subio_endio)
8194                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8195
8196         if (err) {
8197                 dip->errors = 1;
8198
8199                 /*
8200                  * before atomic variable goto zero, we must make sure
8201                  * dip->errors is perceived to be set.
8202                  */
8203                 smp_mb__before_atomic();
8204         }
8205
8206         /* if there are more bios still pending for this dio, just exit */
8207         if (!atomic_dec_and_test(&dip->pending_bios))
8208                 goto out;
8209
8210         if (dip->errors) {
8211                 bio_io_error(dip->orig_bio);
8212         } else {
8213                 dip->dio_bio->bi_error = 0;
8214                 bio_endio(dip->orig_bio);
8215         }
8216 out:
8217         bio_put(bio);
8218 }
8219
8220 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8221                                        u64 first_sector, gfp_t gfp_flags)
8222 {
8223         struct bio *bio;
8224         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8225         if (bio)
8226                 bio_associate_current(bio);
8227         return bio;
8228 }
8229
8230 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8231                                                  struct inode *inode,
8232                                                  struct btrfs_dio_private *dip,
8233                                                  struct bio *bio,
8234                                                  u64 file_offset)
8235 {
8236         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8237         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8238         int ret;
8239
8240         /*
8241          * We load all the csum data we need when we submit
8242          * the first bio to reduce the csum tree search and
8243          * contention.
8244          */
8245         if (dip->logical_offset == file_offset) {
8246                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8247                                                 file_offset);
8248                 if (ret)
8249                         return ret;
8250         }
8251
8252         if (bio == dip->orig_bio)
8253                 return 0;
8254
8255         file_offset -= dip->logical_offset;
8256         file_offset >>= inode->i_sb->s_blocksize_bits;
8257         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8258
8259         return 0;
8260 }
8261
8262 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8263                                          int rw, u64 file_offset, int skip_sum,
8264                                          int async_submit)
8265 {
8266         struct btrfs_dio_private *dip = bio->bi_private;
8267         int write = rw & REQ_WRITE;
8268         struct btrfs_root *root = BTRFS_I(inode)->root;
8269         int ret;
8270
8271         if (async_submit)
8272                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8273
8274         bio_get(bio);
8275
8276         if (!write) {
8277                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8278                                 BTRFS_WQ_ENDIO_DATA);
8279                 if (ret)
8280                         goto err;
8281         }
8282
8283         if (skip_sum)
8284                 goto map;
8285
8286         if (write && async_submit) {
8287                 ret = btrfs_wq_submit_bio(root->fs_info,
8288                                    inode, rw, bio, 0, 0,
8289                                    file_offset,
8290                                    __btrfs_submit_bio_start_direct_io,
8291                                    __btrfs_submit_bio_done);
8292                 goto err;
8293         } else if (write) {
8294                 /*
8295                  * If we aren't doing async submit, calculate the csum of the
8296                  * bio now.
8297                  */
8298                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8299                 if (ret)
8300                         goto err;
8301         } else {
8302                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8303                                                      file_offset);
8304                 if (ret)
8305                         goto err;
8306         }
8307 map:
8308         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8309 err:
8310         bio_put(bio);
8311         return ret;
8312 }
8313
8314 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8315                                     int skip_sum)
8316 {
8317         struct inode *inode = dip->inode;
8318         struct btrfs_root *root = BTRFS_I(inode)->root;
8319         struct bio *bio;
8320         struct bio *orig_bio = dip->orig_bio;
8321         struct bio_vec *bvec = orig_bio->bi_io_vec;
8322         u64 start_sector = orig_bio->bi_iter.bi_sector;
8323         u64 file_offset = dip->logical_offset;
8324         u64 submit_len = 0;
8325         u64 map_length;
8326         u32 blocksize = root->sectorsize;
8327         int async_submit = 0;
8328         int nr_sectors;
8329         int ret;
8330         int i;
8331
8332         map_length = orig_bio->bi_iter.bi_size;
8333         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8334                               &map_length, NULL, 0);
8335         if (ret)
8336                 return -EIO;
8337
8338         if (map_length >= orig_bio->bi_iter.bi_size) {
8339                 bio = orig_bio;
8340                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8341                 goto submit;
8342         }
8343
8344         /* async crcs make it difficult to collect full stripe writes. */
8345         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8346                 async_submit = 0;
8347         else
8348                 async_submit = 1;
8349
8350         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8351         if (!bio)
8352                 return -ENOMEM;
8353
8354         bio->bi_private = dip;
8355         bio->bi_end_io = btrfs_end_dio_bio;
8356         btrfs_io_bio(bio)->logical = file_offset;
8357         atomic_inc(&dip->pending_bios);
8358
8359         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8360                 nr_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info, bvec->bv_len);
8361                 i = 0;
8362 next_block:
8363                 if (unlikely(map_length < submit_len + blocksize ||
8364                     bio_add_page(bio, bvec->bv_page, blocksize,
8365                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8366                         /*
8367                          * inc the count before we submit the bio so
8368                          * we know the end IO handler won't happen before
8369                          * we inc the count. Otherwise, the dip might get freed
8370                          * before we're done setting it up
8371                          */
8372                         atomic_inc(&dip->pending_bios);
8373                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8374                                                      file_offset, skip_sum,
8375                                                      async_submit);
8376                         if (ret) {
8377                                 bio_put(bio);
8378                                 atomic_dec(&dip->pending_bios);
8379                                 goto out_err;
8380                         }
8381
8382                         start_sector += submit_len >> 9;
8383                         file_offset += submit_len;
8384
8385                         submit_len = 0;
8386
8387                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8388                                                   start_sector, GFP_NOFS);
8389                         if (!bio)
8390                                 goto out_err;
8391                         bio->bi_private = dip;
8392                         bio->bi_end_io = btrfs_end_dio_bio;
8393                         btrfs_io_bio(bio)->logical = file_offset;
8394
8395                         map_length = orig_bio->bi_iter.bi_size;
8396                         ret = btrfs_map_block(root->fs_info, rw,
8397                                               start_sector << 9,
8398                                               &map_length, NULL, 0);
8399                         if (ret) {
8400                                 bio_put(bio);
8401                                 goto out_err;
8402                         }
8403
8404                         goto next_block;
8405                 } else {
8406                         submit_len += blocksize;
8407                         if (--nr_sectors) {
8408                                 i++;
8409                                 goto next_block;
8410                         }
8411                         bvec++;
8412                 }
8413         }
8414
8415 submit:
8416         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8417                                      async_submit);
8418         if (!ret)
8419                 return 0;
8420
8421         bio_put(bio);
8422 out_err:
8423         dip->errors = 1;
8424         /*
8425          * before atomic variable goto zero, we must
8426          * make sure dip->errors is perceived to be set.
8427          */
8428         smp_mb__before_atomic();
8429         if (atomic_dec_and_test(&dip->pending_bios))
8430                 bio_io_error(dip->orig_bio);
8431
8432         /* bio_end_io() will handle error, so we needn't return it */
8433         return 0;
8434 }
8435
8436 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8437                                 struct inode *inode, loff_t file_offset)
8438 {
8439         struct btrfs_dio_private *dip = NULL;
8440         struct bio *io_bio = NULL;
8441         struct btrfs_io_bio *btrfs_bio;
8442         int skip_sum;
8443         int write = rw & REQ_WRITE;
8444         int ret = 0;
8445
8446         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8447
8448         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8449         if (!io_bio) {
8450                 ret = -ENOMEM;
8451                 goto free_ordered;
8452         }
8453
8454         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8455         if (!dip) {
8456                 ret = -ENOMEM;
8457                 goto free_ordered;
8458         }
8459
8460         dip->private = dio_bio->bi_private;
8461         dip->inode = inode;
8462         dip->logical_offset = file_offset;
8463         dip->bytes = dio_bio->bi_iter.bi_size;
8464         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8465         io_bio->bi_private = dip;
8466         dip->orig_bio = io_bio;
8467         dip->dio_bio = dio_bio;
8468         atomic_set(&dip->pending_bios, 0);
8469         btrfs_bio = btrfs_io_bio(io_bio);
8470         btrfs_bio->logical = file_offset;
8471
8472         if (write) {
8473                 io_bio->bi_end_io = btrfs_endio_direct_write;
8474         } else {
8475                 io_bio->bi_end_io = btrfs_endio_direct_read;
8476                 dip->subio_endio = btrfs_subio_endio_read;
8477         }
8478
8479         /*
8480          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8481          * even if we fail to submit a bio, because in such case we do the
8482          * corresponding error handling below and it must not be done a second
8483          * time by btrfs_direct_IO().
8484          */
8485         if (write) {
8486                 struct btrfs_dio_data *dio_data = current->journal_info;
8487
8488                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8489                         dip->bytes;
8490                 dio_data->unsubmitted_oe_range_start =
8491                         dio_data->unsubmitted_oe_range_end;
8492         }
8493
8494         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8495         if (!ret)
8496                 return;
8497
8498         if (btrfs_bio->end_io)
8499                 btrfs_bio->end_io(btrfs_bio, ret);
8500
8501 free_ordered:
8502         /*
8503          * If we arrived here it means either we failed to submit the dip
8504          * or we either failed to clone the dio_bio or failed to allocate the
8505          * dip. If we cloned the dio_bio and allocated the dip, we can just
8506          * call bio_endio against our io_bio so that we get proper resource
8507          * cleanup if we fail to submit the dip, otherwise, we must do the
8508          * same as btrfs_endio_direct_[write|read] because we can't call these
8509          * callbacks - they require an allocated dip and a clone of dio_bio.
8510          */
8511         if (io_bio && dip) {
8512                 io_bio->bi_error = -EIO;
8513                 bio_endio(io_bio);
8514                 /*
8515                  * The end io callbacks free our dip, do the final put on io_bio
8516                  * and all the cleanup and final put for dio_bio (through
8517                  * dio_end_io()).
8518                  */
8519                 dip = NULL;
8520                 io_bio = NULL;
8521         } else {
8522                 if (write)
8523                         btrfs_endio_direct_write_update_ordered(inode,
8524                                                 file_offset,
8525                                                 dio_bio->bi_iter.bi_size,
8526                                                 0);
8527                 else
8528                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8529                               file_offset + dio_bio->bi_iter.bi_size - 1);
8530
8531                 dio_bio->bi_error = -EIO;
8532                 /*
8533                  * Releases and cleans up our dio_bio, no need to bio_put()
8534                  * nor bio_endio()/bio_io_error() against dio_bio.
8535                  */
8536                 dio_end_io(dio_bio, ret);
8537         }
8538         if (io_bio)
8539                 bio_put(io_bio);
8540         kfree(dip);
8541 }
8542
8543 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8544                         const struct iov_iter *iter, loff_t offset)
8545 {
8546         int seg;
8547         int i;
8548         unsigned blocksize_mask = root->sectorsize - 1;
8549         ssize_t retval = -EINVAL;
8550
8551         if (offset & blocksize_mask)
8552                 goto out;
8553
8554         if (iov_iter_alignment(iter) & blocksize_mask)
8555                 goto out;
8556
8557         /* If this is a write we don't need to check anymore */
8558         if (iov_iter_rw(iter) == WRITE)
8559                 return 0;
8560         /*
8561          * Check to make sure we don't have duplicate iov_base's in this
8562          * iovec, if so return EINVAL, otherwise we'll get csum errors
8563          * when reading back.
8564          */
8565         for (seg = 0; seg < iter->nr_segs; seg++) {
8566                 for (i = seg + 1; i < iter->nr_segs; i++) {
8567                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8568                                 goto out;
8569                 }
8570         }
8571         retval = 0;
8572 out:
8573         return retval;
8574 }
8575
8576 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8577 {
8578         struct file *file = iocb->ki_filp;
8579         struct inode *inode = file->f_mapping->host;
8580         struct btrfs_root *root = BTRFS_I(inode)->root;
8581         struct btrfs_dio_data dio_data = { 0 };
8582         loff_t offset = iocb->ki_pos;
8583         size_t count = 0;
8584         int flags = 0;
8585         bool wakeup = true;
8586         bool relock = false;
8587         ssize_t ret;
8588
8589         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8590                 return 0;
8591
8592         inode_dio_begin(inode);
8593         smp_mb__after_atomic();
8594
8595         /*
8596          * The generic stuff only does filemap_write_and_wait_range, which
8597          * isn't enough if we've written compressed pages to this area, so
8598          * we need to flush the dirty pages again to make absolutely sure
8599          * that any outstanding dirty pages are on disk.
8600          */
8601         count = iov_iter_count(iter);
8602         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8603                      &BTRFS_I(inode)->runtime_flags))
8604                 filemap_fdatawrite_range(inode->i_mapping, offset,
8605                                          offset + count - 1);
8606
8607         if (iov_iter_rw(iter) == WRITE) {
8608                 /*
8609                  * If the write DIO is beyond the EOF, we need update
8610                  * the isize, but it is protected by i_mutex. So we can
8611                  * not unlock the i_mutex at this case.
8612                  */
8613                 if (offset + count <= inode->i_size) {
8614                         inode_unlock(inode);
8615                         relock = true;
8616                 }
8617                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8618                 if (ret)
8619                         goto out;
8620                 dio_data.outstanding_extents = div64_u64(count +
8621                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8622                                                 BTRFS_MAX_EXTENT_SIZE);
8623
8624                 /*
8625                  * We need to know how many extents we reserved so that we can
8626                  * do the accounting properly if we go over the number we
8627                  * originally calculated.  Abuse current->journal_info for this.
8628                  */
8629                 dio_data.reserve = round_up(count, root->sectorsize);
8630                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8631                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8632                 current->journal_info = &dio_data;
8633         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8634                                      &BTRFS_I(inode)->runtime_flags)) {
8635                 inode_dio_end(inode);
8636                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8637                 wakeup = false;
8638         }
8639
8640         ret = __blockdev_direct_IO(iocb, inode,
8641                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8642                                    iter, btrfs_get_blocks_direct, NULL,
8643                                    btrfs_submit_direct, flags);
8644         if (iov_iter_rw(iter) == WRITE) {
8645                 current->journal_info = NULL;
8646                 if (ret < 0 && ret != -EIOCBQUEUED) {
8647                         if (dio_data.reserve)
8648                                 btrfs_delalloc_release_space(inode, offset,
8649                                                              dio_data.reserve);
8650                         /*
8651                          * On error we might have left some ordered extents
8652                          * without submitting corresponding bios for them, so
8653                          * cleanup them up to avoid other tasks getting them
8654                          * and waiting for them to complete forever.
8655                          */
8656                         if (dio_data.unsubmitted_oe_range_start <
8657                             dio_data.unsubmitted_oe_range_end)
8658                                 btrfs_endio_direct_write_update_ordered(inode,
8659                                         dio_data.unsubmitted_oe_range_start,
8660                                         dio_data.unsubmitted_oe_range_end -
8661                                         dio_data.unsubmitted_oe_range_start,
8662                                         0);
8663                 } else if (ret >= 0 && (size_t)ret < count)
8664                         btrfs_delalloc_release_space(inode, offset,
8665                                                      count - (size_t)ret);
8666         }
8667 out:
8668         if (wakeup)
8669                 inode_dio_end(inode);
8670         if (relock)
8671                 inode_lock(inode);
8672
8673         return ret;
8674 }
8675
8676 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8677
8678 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8679                 __u64 start, __u64 len)
8680 {
8681         int     ret;
8682
8683         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8684         if (ret)
8685                 return ret;
8686
8687         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8688 }
8689
8690 int btrfs_readpage(struct file *file, struct page *page)
8691 {
8692         struct extent_io_tree *tree;
8693         tree = &BTRFS_I(page->mapping->host)->io_tree;
8694         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8695 }
8696
8697 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8698 {
8699         struct extent_io_tree *tree;
8700         struct inode *inode = page->mapping->host;
8701         int ret;
8702
8703         if (current->flags & PF_MEMALLOC) {
8704                 redirty_page_for_writepage(wbc, page);
8705                 unlock_page(page);
8706                 return 0;
8707         }
8708
8709         /*
8710          * If we are under memory pressure we will call this directly from the
8711          * VM, we need to make sure we have the inode referenced for the ordered
8712          * extent.  If not just return like we didn't do anything.
8713          */
8714         if (!igrab(inode)) {
8715                 redirty_page_for_writepage(wbc, page);
8716                 return AOP_WRITEPAGE_ACTIVATE;
8717         }
8718         tree = &BTRFS_I(page->mapping->host)->io_tree;
8719         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8720         btrfs_add_delayed_iput(inode);
8721         return ret;
8722 }
8723
8724 static int btrfs_writepages(struct address_space *mapping,
8725                             struct writeback_control *wbc)
8726 {
8727         struct extent_io_tree *tree;
8728
8729         tree = &BTRFS_I(mapping->host)->io_tree;
8730         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8731 }
8732
8733 static int
8734 btrfs_readpages(struct file *file, struct address_space *mapping,
8735                 struct list_head *pages, unsigned nr_pages)
8736 {
8737         struct extent_io_tree *tree;
8738         tree = &BTRFS_I(mapping->host)->io_tree;
8739         return extent_readpages(tree, mapping, pages, nr_pages,
8740                                 btrfs_get_extent);
8741 }
8742 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8743 {
8744         struct extent_io_tree *tree;
8745         struct extent_map_tree *map;
8746         int ret;
8747
8748         tree = &BTRFS_I(page->mapping->host)->io_tree;
8749         map = &BTRFS_I(page->mapping->host)->extent_tree;
8750         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8751         if (ret == 1) {
8752                 ClearPagePrivate(page);
8753                 set_page_private(page, 0);
8754                 put_page(page);
8755         }
8756         return ret;
8757 }
8758
8759 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8760 {
8761         if (PageWriteback(page) || PageDirty(page))
8762                 return 0;
8763         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8764 }
8765
8766 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8767                                  unsigned int length)
8768 {
8769         struct inode *inode = page->mapping->host;
8770         struct extent_io_tree *tree;
8771         struct btrfs_ordered_extent *ordered;
8772         struct extent_state *cached_state = NULL;
8773         u64 page_start = page_offset(page);
8774         u64 page_end = page_start + PAGE_SIZE - 1;
8775         u64 start;
8776         u64 end;
8777         int inode_evicting = inode->i_state & I_FREEING;
8778
8779         /*
8780          * we have the page locked, so new writeback can't start,
8781          * and the dirty bit won't be cleared while we are here.
8782          *
8783          * Wait for IO on this page so that we can safely clear
8784          * the PagePrivate2 bit and do ordered accounting
8785          */
8786         wait_on_page_writeback(page);
8787
8788         tree = &BTRFS_I(inode)->io_tree;
8789         if (offset) {
8790                 btrfs_releasepage(page, GFP_NOFS);
8791                 return;
8792         }
8793
8794         if (!inode_evicting)
8795                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8796 again:
8797         start = page_start;
8798         ordered = btrfs_lookup_ordered_range(inode, start,
8799                                         page_end - start + 1);
8800         if (ordered) {
8801                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8802                 /*
8803                  * IO on this page will never be started, so we need
8804                  * to account for any ordered extents now
8805                  */
8806                 if (!inode_evicting)
8807                         clear_extent_bit(tree, start, end,
8808                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8809                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8810                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8811                                          GFP_NOFS);
8812                 /*
8813                  * whoever cleared the private bit is responsible
8814                  * for the finish_ordered_io
8815                  */
8816                 if (TestClearPagePrivate2(page)) {
8817                         struct btrfs_ordered_inode_tree *tree;
8818                         u64 new_len;
8819
8820                         tree = &BTRFS_I(inode)->ordered_tree;
8821
8822                         spin_lock_irq(&tree->lock);
8823                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8824                         new_len = start - ordered->file_offset;
8825                         if (new_len < ordered->truncated_len)
8826                                 ordered->truncated_len = new_len;
8827                         spin_unlock_irq(&tree->lock);
8828
8829                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8830                                                            start,
8831                                                            end - start + 1, 1))
8832                                 btrfs_finish_ordered_io(ordered);
8833                 }
8834                 btrfs_put_ordered_extent(ordered);
8835                 if (!inode_evicting) {
8836                         cached_state = NULL;
8837                         lock_extent_bits(tree, start, end,
8838                                          &cached_state);
8839                 }
8840
8841                 start = end + 1;
8842                 if (start < page_end)
8843                         goto again;
8844         }
8845
8846         /*
8847          * Qgroup reserved space handler
8848          * Page here will be either
8849          * 1) Already written to disk
8850          *    In this case, its reserved space is released from data rsv map
8851          *    and will be freed by delayed_ref handler finally.
8852          *    So even we call qgroup_free_data(), it won't decrease reserved
8853          *    space.
8854          * 2) Not written to disk
8855          *    This means the reserved space should be freed here.
8856          */
8857         btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8858         if (!inode_evicting) {
8859                 clear_extent_bit(tree, page_start, page_end,
8860                                  EXTENT_LOCKED | EXTENT_DIRTY |
8861                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8862                                  EXTENT_DEFRAG, 1, 1,
8863                                  &cached_state, GFP_NOFS);
8864
8865                 __btrfs_releasepage(page, GFP_NOFS);
8866         }
8867
8868         ClearPageChecked(page);
8869         if (PagePrivate(page)) {
8870                 ClearPagePrivate(page);
8871                 set_page_private(page, 0);
8872                 put_page(page);
8873         }
8874 }
8875
8876 /*
8877  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8878  * called from a page fault handler when a page is first dirtied. Hence we must
8879  * be careful to check for EOF conditions here. We set the page up correctly
8880  * for a written page which means we get ENOSPC checking when writing into
8881  * holes and correct delalloc and unwritten extent mapping on filesystems that
8882  * support these features.
8883  *
8884  * We are not allowed to take the i_mutex here so we have to play games to
8885  * protect against truncate races as the page could now be beyond EOF.  Because
8886  * vmtruncate() writes the inode size before removing pages, once we have the
8887  * page lock we can determine safely if the page is beyond EOF. If it is not
8888  * beyond EOF, then the page is guaranteed safe against truncation until we
8889  * unlock the page.
8890  */
8891 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8892 {
8893         struct page *page = vmf->page;
8894         struct inode *inode = file_inode(vma->vm_file);
8895         struct btrfs_root *root = BTRFS_I(inode)->root;
8896         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8897         struct btrfs_ordered_extent *ordered;
8898         struct extent_state *cached_state = NULL;
8899         char *kaddr;
8900         unsigned long zero_start;
8901         loff_t size;
8902         int ret;
8903         int reserved = 0;
8904         u64 reserved_space;
8905         u64 page_start;
8906         u64 page_end;
8907         u64 end;
8908
8909         reserved_space = PAGE_SIZE;
8910
8911         sb_start_pagefault(inode->i_sb);
8912         page_start = page_offset(page);
8913         page_end = page_start + PAGE_SIZE - 1;
8914         end = page_end;
8915
8916         /*
8917          * Reserving delalloc space after obtaining the page lock can lead to
8918          * deadlock. For example, if a dirty page is locked by this function
8919          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8920          * dirty page write out, then the btrfs_writepage() function could
8921          * end up waiting indefinitely to get a lock on the page currently
8922          * being processed by btrfs_page_mkwrite() function.
8923          */
8924         ret = btrfs_delalloc_reserve_space(inode, page_start,
8925                                            reserved_space);
8926         if (!ret) {
8927                 ret = file_update_time(vma->vm_file);
8928                 reserved = 1;
8929         }
8930         if (ret) {
8931                 if (ret == -ENOMEM)
8932                         ret = VM_FAULT_OOM;
8933                 else /* -ENOSPC, -EIO, etc */
8934                         ret = VM_FAULT_SIGBUS;
8935                 if (reserved)
8936                         goto out;
8937                 goto out_noreserve;
8938         }
8939
8940         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8941 again:
8942         lock_page(page);
8943         size = i_size_read(inode);
8944
8945         if ((page->mapping != inode->i_mapping) ||
8946             (page_start >= size)) {
8947                 /* page got truncated out from underneath us */
8948                 goto out_unlock;
8949         }
8950         wait_on_page_writeback(page);
8951
8952         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8953         set_page_extent_mapped(page);
8954
8955         /*
8956          * we can't set the delalloc bits if there are pending ordered
8957          * extents.  Drop our locks and wait for them to finish
8958          */
8959         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
8960         if (ordered) {
8961                 unlock_extent_cached(io_tree, page_start, page_end,
8962                                      &cached_state, GFP_NOFS);
8963                 unlock_page(page);
8964                 btrfs_start_ordered_extent(inode, ordered, 1);
8965                 btrfs_put_ordered_extent(ordered);
8966                 goto again;
8967         }
8968
8969         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8970                 reserved_space = round_up(size - page_start, root->sectorsize);
8971                 if (reserved_space < PAGE_SIZE) {
8972                         end = page_start + reserved_space - 1;
8973                         spin_lock(&BTRFS_I(inode)->lock);
8974                         BTRFS_I(inode)->outstanding_extents++;
8975                         spin_unlock(&BTRFS_I(inode)->lock);
8976                         btrfs_delalloc_release_space(inode, page_start,
8977                                                 PAGE_SIZE - reserved_space);
8978                 }
8979         }
8980
8981         /*
8982          * XXX - page_mkwrite gets called every time the page is dirtied, even
8983          * if it was already dirty, so for space accounting reasons we need to
8984          * clear any delalloc bits for the range we are fixing to save.  There
8985          * is probably a better way to do this, but for now keep consistent with
8986          * prepare_pages in the normal write path.
8987          */
8988         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8989                           EXTENT_DIRTY | EXTENT_DELALLOC |
8990                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8991                           0, 0, &cached_state, GFP_NOFS);
8992
8993         ret = btrfs_set_extent_delalloc(inode, page_start, end,
8994                                         &cached_state);
8995         if (ret) {
8996                 unlock_extent_cached(io_tree, page_start, page_end,
8997                                      &cached_state, GFP_NOFS);
8998                 ret = VM_FAULT_SIGBUS;
8999                 goto out_unlock;
9000         }
9001         ret = 0;
9002
9003         /* page is wholly or partially inside EOF */
9004         if (page_start + PAGE_SIZE > size)
9005                 zero_start = size & ~PAGE_MASK;
9006         else
9007                 zero_start = PAGE_SIZE;
9008
9009         if (zero_start != PAGE_SIZE) {
9010                 kaddr = kmap(page);
9011                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9012                 flush_dcache_page(page);
9013                 kunmap(page);
9014         }
9015         ClearPageChecked(page);
9016         set_page_dirty(page);
9017         SetPageUptodate(page);
9018
9019         BTRFS_I(inode)->last_trans = root->fs_info->generation;
9020         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9021         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9022
9023         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9024
9025 out_unlock:
9026         if (!ret) {
9027                 sb_end_pagefault(inode->i_sb);
9028                 return VM_FAULT_LOCKED;
9029         }
9030         unlock_page(page);
9031 out:
9032         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9033 out_noreserve:
9034         sb_end_pagefault(inode->i_sb);
9035         return ret;
9036 }
9037
9038 static int btrfs_truncate(struct inode *inode)
9039 {
9040         struct btrfs_root *root = BTRFS_I(inode)->root;
9041         struct btrfs_block_rsv *rsv;
9042         int ret = 0;
9043         int err = 0;
9044         struct btrfs_trans_handle *trans;
9045         u64 mask = root->sectorsize - 1;
9046         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
9047
9048         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9049                                        (u64)-1);
9050         if (ret)
9051                 return ret;
9052
9053         /*
9054          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9055          * 3 things going on here
9056          *
9057          * 1) We need to reserve space for our orphan item and the space to
9058          * delete our orphan item.  Lord knows we don't want to have a dangling
9059          * orphan item because we didn't reserve space to remove it.
9060          *
9061          * 2) We need to reserve space to update our inode.
9062          *
9063          * 3) We need to have something to cache all the space that is going to
9064          * be free'd up by the truncate operation, but also have some slack
9065          * space reserved in case it uses space during the truncate (thank you
9066          * very much snapshotting).
9067          *
9068          * And we need these to all be separate.  The fact is we can use a lot of
9069          * space doing the truncate, and we have no earthly idea how much space
9070          * we will use, so we need the truncate reservation to be separate so it
9071          * doesn't end up using space reserved for updating the inode or
9072          * removing the orphan item.  We also need to be able to stop the
9073          * transaction and start a new one, which means we need to be able to
9074          * update the inode several times, and we have no idea of knowing how
9075          * many times that will be, so we can't just reserve 1 item for the
9076          * entirety of the operation, so that has to be done separately as well.
9077          * Then there is the orphan item, which does indeed need to be held on
9078          * to for the whole operation, and we need nobody to touch this reserved
9079          * space except the orphan code.
9080          *
9081          * So that leaves us with
9082          *
9083          * 1) root->orphan_block_rsv - for the orphan deletion.
9084          * 2) rsv - for the truncate reservation, which we will steal from the
9085          * transaction reservation.
9086          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9087          * updating the inode.
9088          */
9089         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
9090         if (!rsv)
9091                 return -ENOMEM;
9092         rsv->size = min_size;
9093         rsv->failfast = 1;
9094
9095         /*
9096          * 1 for the truncate slack space
9097          * 1 for updating the inode.
9098          */
9099         trans = btrfs_start_transaction(root, 2);
9100         if (IS_ERR(trans)) {
9101                 err = PTR_ERR(trans);
9102                 goto out;
9103         }
9104
9105         /* Migrate the slack space for the truncate to our reserve */
9106         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
9107                                       min_size);
9108         BUG_ON(ret);
9109
9110         /*
9111          * So if we truncate and then write and fsync we normally would just
9112          * write the extents that changed, which is a problem if we need to
9113          * first truncate that entire inode.  So set this flag so we write out
9114          * all of the extents in the inode to the sync log so we're completely
9115          * safe.
9116          */
9117         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9118         trans->block_rsv = rsv;
9119
9120         while (1) {
9121                 ret = btrfs_truncate_inode_items(trans, root, inode,
9122                                                  inode->i_size,
9123                                                  BTRFS_EXTENT_DATA_KEY);
9124                 if (ret != -ENOSPC && ret != -EAGAIN) {
9125                         err = ret;
9126                         break;
9127                 }
9128
9129                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9130                 ret = btrfs_update_inode(trans, root, inode);
9131                 if (ret) {
9132                         err = ret;
9133                         break;
9134                 }
9135
9136                 btrfs_end_transaction(trans, root);
9137                 btrfs_btree_balance_dirty(root);
9138
9139                 trans = btrfs_start_transaction(root, 2);
9140                 if (IS_ERR(trans)) {
9141                         ret = err = PTR_ERR(trans);
9142                         trans = NULL;
9143                         break;
9144                 }
9145
9146                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
9147                                               rsv, min_size);
9148                 BUG_ON(ret);    /* shouldn't happen */
9149                 trans->block_rsv = rsv;
9150         }
9151
9152         if (ret == 0 && inode->i_nlink > 0) {
9153                 trans->block_rsv = root->orphan_block_rsv;
9154                 ret = btrfs_orphan_del(trans, inode);
9155                 if (ret)
9156                         err = ret;
9157         }
9158
9159         if (trans) {
9160                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9161                 ret = btrfs_update_inode(trans, root, inode);
9162                 if (ret && !err)
9163                         err = ret;
9164
9165                 ret = btrfs_end_transaction(trans, root);
9166                 btrfs_btree_balance_dirty(root);
9167         }
9168
9169 out:
9170         btrfs_free_block_rsv(root, rsv);
9171
9172         if (ret && !err)
9173                 err = ret;
9174
9175         return err;
9176 }
9177
9178 /*
9179  * create a new subvolume directory/inode (helper for the ioctl).
9180  */
9181 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9182                              struct btrfs_root *new_root,
9183                              struct btrfs_root *parent_root,
9184                              u64 new_dirid)
9185 {
9186         struct inode *inode;
9187         int err;
9188         u64 index = 0;
9189
9190         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9191                                 new_dirid, new_dirid,
9192                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9193                                 &index);
9194         if (IS_ERR(inode))
9195                 return PTR_ERR(inode);
9196         inode->i_op = &btrfs_dir_inode_operations;
9197         inode->i_fop = &btrfs_dir_file_operations;
9198
9199         set_nlink(inode, 1);
9200         btrfs_i_size_write(inode, 0);
9201         unlock_new_inode(inode);
9202
9203         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9204         if (err)
9205                 btrfs_err(new_root->fs_info,
9206                           "error inheriting subvolume %llu properties: %d",
9207                           new_root->root_key.objectid, err);
9208
9209         err = btrfs_update_inode(trans, new_root, inode);
9210
9211         iput(inode);
9212         return err;
9213 }
9214
9215 struct inode *btrfs_alloc_inode(struct super_block *sb)
9216 {
9217         struct btrfs_inode *ei;
9218         struct inode *inode;
9219
9220         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9221         if (!ei)
9222                 return NULL;
9223
9224         ei->root = NULL;
9225         ei->generation = 0;
9226         ei->last_trans = 0;
9227         ei->last_sub_trans = 0;
9228         ei->logged_trans = 0;
9229         ei->delalloc_bytes = 0;
9230         ei->defrag_bytes = 0;
9231         ei->disk_i_size = 0;
9232         ei->flags = 0;
9233         ei->csum_bytes = 0;
9234         ei->index_cnt = (u64)-1;
9235         ei->dir_index = 0;
9236         ei->last_unlink_trans = 0;
9237         ei->last_log_commit = 0;
9238         ei->delayed_iput_count = 0;
9239
9240         spin_lock_init(&ei->lock);
9241         ei->outstanding_extents = 0;
9242         ei->reserved_extents = 0;
9243
9244         ei->runtime_flags = 0;
9245         ei->force_compress = BTRFS_COMPRESS_NONE;
9246
9247         ei->delayed_node = NULL;
9248
9249         ei->i_otime.tv_sec = 0;
9250         ei->i_otime.tv_nsec = 0;
9251
9252         inode = &ei->vfs_inode;
9253         extent_map_tree_init(&ei->extent_tree);
9254         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9255         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9256         ei->io_tree.track_uptodate = 1;
9257         ei->io_failure_tree.track_uptodate = 1;
9258         atomic_set(&ei->sync_writers, 0);
9259         mutex_init(&ei->log_mutex);
9260         mutex_init(&ei->delalloc_mutex);
9261         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9262         INIT_LIST_HEAD(&ei->delalloc_inodes);
9263         INIT_LIST_HEAD(&ei->delayed_iput);
9264         RB_CLEAR_NODE(&ei->rb_node);
9265         init_rwsem(&ei->dio_sem);
9266
9267         return inode;
9268 }
9269
9270 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9271 void btrfs_test_destroy_inode(struct inode *inode)
9272 {
9273         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9274         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9275 }
9276 #endif
9277
9278 static void btrfs_i_callback(struct rcu_head *head)
9279 {
9280         struct inode *inode = container_of(head, struct inode, i_rcu);
9281         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9282 }
9283
9284 void btrfs_destroy_inode(struct inode *inode)
9285 {
9286         struct btrfs_ordered_extent *ordered;
9287         struct btrfs_root *root = BTRFS_I(inode)->root;
9288
9289         WARN_ON(!hlist_empty(&inode->i_dentry));
9290         WARN_ON(inode->i_data.nrpages);
9291         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9292         WARN_ON(BTRFS_I(inode)->reserved_extents);
9293         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9294         WARN_ON(BTRFS_I(inode)->csum_bytes);
9295         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9296
9297         /*
9298          * This can happen where we create an inode, but somebody else also
9299          * created the same inode and we need to destroy the one we already
9300          * created.
9301          */
9302         if (!root)
9303                 goto free;
9304
9305         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9306                      &BTRFS_I(inode)->runtime_flags)) {
9307                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9308                         btrfs_ino(inode));
9309                 atomic_dec(&root->orphan_inodes);
9310         }
9311
9312         while (1) {
9313                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9314                 if (!ordered)
9315                         break;
9316                 else {
9317                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9318                                 ordered->file_offset, ordered->len);
9319                         btrfs_remove_ordered_extent(inode, ordered);
9320                         btrfs_put_ordered_extent(ordered);
9321                         btrfs_put_ordered_extent(ordered);
9322                 }
9323         }
9324         btrfs_qgroup_check_reserved_leak(inode);
9325         inode_tree_del(inode);
9326         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9327 free:
9328         call_rcu(&inode->i_rcu, btrfs_i_callback);
9329 }
9330
9331 int btrfs_drop_inode(struct inode *inode)
9332 {
9333         struct btrfs_root *root = BTRFS_I(inode)->root;
9334
9335         if (root == NULL)
9336                 return 1;
9337
9338         /* the snap/subvol tree is on deleting */
9339         if (btrfs_root_refs(&root->root_item) == 0)
9340                 return 1;
9341         else
9342                 return generic_drop_inode(inode);
9343 }
9344
9345 static void init_once(void *foo)
9346 {
9347         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9348
9349         inode_init_once(&ei->vfs_inode);
9350 }
9351
9352 void btrfs_destroy_cachep(void)
9353 {
9354         /*
9355          * Make sure all delayed rcu free inodes are flushed before we
9356          * destroy cache.
9357          */
9358         rcu_barrier();
9359         kmem_cache_destroy(btrfs_inode_cachep);
9360         kmem_cache_destroy(btrfs_trans_handle_cachep);
9361         kmem_cache_destroy(btrfs_transaction_cachep);
9362         kmem_cache_destroy(btrfs_path_cachep);
9363         kmem_cache_destroy(btrfs_free_space_cachep);
9364 }
9365
9366 int btrfs_init_cachep(void)
9367 {
9368         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9369                         sizeof(struct btrfs_inode), 0,
9370                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9371                         init_once);
9372         if (!btrfs_inode_cachep)
9373                 goto fail;
9374
9375         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9376                         sizeof(struct btrfs_trans_handle), 0,
9377                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9378         if (!btrfs_trans_handle_cachep)
9379                 goto fail;
9380
9381         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9382                         sizeof(struct btrfs_transaction), 0,
9383                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9384         if (!btrfs_transaction_cachep)
9385                 goto fail;
9386
9387         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9388                         sizeof(struct btrfs_path), 0,
9389                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9390         if (!btrfs_path_cachep)
9391                 goto fail;
9392
9393         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9394                         sizeof(struct btrfs_free_space), 0,
9395                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9396         if (!btrfs_free_space_cachep)
9397                 goto fail;
9398
9399         return 0;
9400 fail:
9401         btrfs_destroy_cachep();
9402         return -ENOMEM;
9403 }
9404
9405 static int btrfs_getattr(struct vfsmount *mnt,
9406                          struct dentry *dentry, struct kstat *stat)
9407 {
9408         u64 delalloc_bytes;
9409         struct inode *inode = d_inode(dentry);
9410         u32 blocksize = inode->i_sb->s_blocksize;
9411
9412         generic_fillattr(inode, stat);
9413         stat->dev = BTRFS_I(inode)->root->anon_dev;
9414
9415         spin_lock(&BTRFS_I(inode)->lock);
9416         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9417         spin_unlock(&BTRFS_I(inode)->lock);
9418         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9419                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9420         return 0;
9421 }
9422
9423 static int btrfs_rename_exchange(struct inode *old_dir,
9424                               struct dentry *old_dentry,
9425                               struct inode *new_dir,
9426                               struct dentry *new_dentry)
9427 {
9428         struct btrfs_trans_handle *trans;
9429         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9430         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9431         struct inode *new_inode = new_dentry->d_inode;
9432         struct inode *old_inode = old_dentry->d_inode;
9433         struct timespec ctime = CURRENT_TIME;
9434         struct dentry *parent;
9435         u64 old_ino = btrfs_ino(old_inode);
9436         u64 new_ino = btrfs_ino(new_inode);
9437         u64 old_idx = 0;
9438         u64 new_idx = 0;
9439         u64 root_objectid;
9440         int ret;
9441         bool root_log_pinned = false;
9442         bool dest_log_pinned = false;
9443
9444         /* we only allow rename subvolume link between subvolumes */
9445         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9446                 return -EXDEV;
9447
9448         /* close the race window with snapshot create/destroy ioctl */
9449         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9450                 down_read(&root->fs_info->subvol_sem);
9451         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9452                 down_read(&dest->fs_info->subvol_sem);
9453
9454         /*
9455          * We want to reserve the absolute worst case amount of items.  So if
9456          * both inodes are subvols and we need to unlink them then that would
9457          * require 4 item modifications, but if they are both normal inodes it
9458          * would require 5 item modifications, so we'll assume their normal
9459          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9460          * should cover the worst case number of items we'll modify.
9461          */
9462         trans = btrfs_start_transaction(root, 12);
9463         if (IS_ERR(trans)) {
9464                 ret = PTR_ERR(trans);
9465                 goto out_notrans;
9466         }
9467
9468         /*
9469          * We need to find a free sequence number both in the source and
9470          * in the destination directory for the exchange.
9471          */
9472         ret = btrfs_set_inode_index(new_dir, &old_idx);
9473         if (ret)
9474                 goto out_fail;
9475         ret = btrfs_set_inode_index(old_dir, &new_idx);
9476         if (ret)
9477                 goto out_fail;
9478
9479         BTRFS_I(old_inode)->dir_index = 0ULL;
9480         BTRFS_I(new_inode)->dir_index = 0ULL;
9481
9482         /* Reference for the source. */
9483         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9484                 /* force full log commit if subvolume involved. */
9485                 btrfs_set_log_full_commit(root->fs_info, trans);
9486         } else {
9487                 btrfs_pin_log_trans(root);
9488                 root_log_pinned = true;
9489                 ret = btrfs_insert_inode_ref(trans, dest,
9490                                              new_dentry->d_name.name,
9491                                              new_dentry->d_name.len,
9492                                              old_ino,
9493                                              btrfs_ino(new_dir), old_idx);
9494                 if (ret)
9495                         goto out_fail;
9496         }
9497
9498         /* And now for the dest. */
9499         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9500                 /* force full log commit if subvolume involved. */
9501                 btrfs_set_log_full_commit(dest->fs_info, trans);
9502         } else {
9503                 btrfs_pin_log_trans(dest);
9504                 dest_log_pinned = true;
9505                 ret = btrfs_insert_inode_ref(trans, root,
9506                                              old_dentry->d_name.name,
9507                                              old_dentry->d_name.len,
9508                                              new_ino,
9509                                              btrfs_ino(old_dir), new_idx);
9510                 if (ret)
9511                         goto out_fail;
9512         }
9513
9514         /* Update inode version and ctime/mtime. */
9515         inode_inc_iversion(old_dir);
9516         inode_inc_iversion(new_dir);
9517         inode_inc_iversion(old_inode);
9518         inode_inc_iversion(new_inode);
9519         old_dir->i_ctime = old_dir->i_mtime = ctime;
9520         new_dir->i_ctime = new_dir->i_mtime = ctime;
9521         old_inode->i_ctime = ctime;
9522         new_inode->i_ctime = ctime;
9523
9524         if (old_dentry->d_parent != new_dentry->d_parent) {
9525                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9526                 btrfs_record_unlink_dir(trans, new_dir, new_inode, 1);
9527         }
9528
9529         /* src is a subvolume */
9530         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9531                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9532                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9533                                           root_objectid,
9534                                           old_dentry->d_name.name,
9535                                           old_dentry->d_name.len);
9536         } else { /* src is an inode */
9537                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9538                                            old_dentry->d_inode,
9539                                            old_dentry->d_name.name,
9540                                            old_dentry->d_name.len);
9541                 if (!ret)
9542                         ret = btrfs_update_inode(trans, root, old_inode);
9543         }
9544         if (ret) {
9545                 btrfs_abort_transaction(trans, root, ret);
9546                 goto out_fail;
9547         }
9548
9549         /* dest is a subvolume */
9550         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9551                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9552                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9553                                           root_objectid,
9554                                           new_dentry->d_name.name,
9555                                           new_dentry->d_name.len);
9556         } else { /* dest is an inode */
9557                 ret = __btrfs_unlink_inode(trans, dest, new_dir,
9558                                            new_dentry->d_inode,
9559                                            new_dentry->d_name.name,
9560                                            new_dentry->d_name.len);
9561                 if (!ret)
9562                         ret = btrfs_update_inode(trans, dest, new_inode);
9563         }
9564         if (ret) {
9565                 btrfs_abort_transaction(trans, root, ret);
9566                 goto out_fail;
9567         }
9568
9569         ret = btrfs_add_link(trans, new_dir, old_inode,
9570                              new_dentry->d_name.name,
9571                              new_dentry->d_name.len, 0, old_idx);
9572         if (ret) {
9573                 btrfs_abort_transaction(trans, root, ret);
9574                 goto out_fail;
9575         }
9576
9577         ret = btrfs_add_link(trans, old_dir, new_inode,
9578                              old_dentry->d_name.name,
9579                              old_dentry->d_name.len, 0, new_idx);
9580         if (ret) {
9581                 btrfs_abort_transaction(trans, root, ret);
9582                 goto out_fail;
9583         }
9584
9585         if (old_inode->i_nlink == 1)
9586                 BTRFS_I(old_inode)->dir_index = old_idx;
9587         if (new_inode->i_nlink == 1)
9588                 BTRFS_I(new_inode)->dir_index = new_idx;
9589
9590         if (root_log_pinned) {
9591                 parent = new_dentry->d_parent;
9592                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9593                 btrfs_end_log_trans(root);
9594                 root_log_pinned = false;
9595         }
9596         if (dest_log_pinned) {
9597                 parent = old_dentry->d_parent;
9598                 btrfs_log_new_name(trans, new_inode, new_dir, parent);
9599                 btrfs_end_log_trans(dest);
9600                 dest_log_pinned = false;
9601         }
9602 out_fail:
9603         /*
9604          * If we have pinned a log and an error happened, we unpin tasks
9605          * trying to sync the log and force them to fallback to a transaction
9606          * commit if the log currently contains any of the inodes involved in
9607          * this rename operation (to ensure we do not persist a log with an
9608          * inconsistent state for any of these inodes or leading to any
9609          * inconsistencies when replayed). If the transaction was aborted, the
9610          * abortion reason is propagated to userspace when attempting to commit
9611          * the transaction. If the log does not contain any of these inodes, we
9612          * allow the tasks to sync it.
9613          */
9614         if (ret && (root_log_pinned || dest_log_pinned)) {
9615                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9616                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9617                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9618                     (new_inode &&
9619                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9620                     btrfs_set_log_full_commit(root->fs_info, trans);
9621
9622                 if (root_log_pinned) {
9623                         btrfs_end_log_trans(root);
9624                         root_log_pinned = false;
9625                 }
9626                 if (dest_log_pinned) {
9627                         btrfs_end_log_trans(dest);
9628                         dest_log_pinned = false;
9629                 }
9630         }
9631         ret = btrfs_end_transaction(trans, root);
9632 out_notrans:
9633         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9634                 up_read(&dest->fs_info->subvol_sem);
9635         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9636                 up_read(&root->fs_info->subvol_sem);
9637
9638         return ret;
9639 }
9640
9641 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9642                                      struct btrfs_root *root,
9643                                      struct inode *dir,
9644                                      struct dentry *dentry)
9645 {
9646         int ret;
9647         struct inode *inode;
9648         u64 objectid;
9649         u64 index;
9650
9651         ret = btrfs_find_free_ino(root, &objectid);
9652         if (ret)
9653                 return ret;
9654
9655         inode = btrfs_new_inode(trans, root, dir,
9656                                 dentry->d_name.name,
9657                                 dentry->d_name.len,
9658                                 btrfs_ino(dir),
9659                                 objectid,
9660                                 S_IFCHR | WHITEOUT_MODE,
9661                                 &index);
9662
9663         if (IS_ERR(inode)) {
9664                 ret = PTR_ERR(inode);
9665                 return ret;
9666         }
9667
9668         inode->i_op = &btrfs_special_inode_operations;
9669         init_special_inode(inode, inode->i_mode,
9670                 WHITEOUT_DEV);
9671
9672         ret = btrfs_init_inode_security(trans, inode, dir,
9673                                 &dentry->d_name);
9674         if (ret)
9675                 goto out;
9676
9677         ret = btrfs_add_nondir(trans, dir, dentry,
9678                                 inode, 0, index);
9679         if (ret)
9680                 goto out;
9681
9682         ret = btrfs_update_inode(trans, root, inode);
9683 out:
9684         unlock_new_inode(inode);
9685         if (ret)
9686                 inode_dec_link_count(inode);
9687         iput(inode);
9688
9689         return ret;
9690 }
9691
9692 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9693                            struct inode *new_dir, struct dentry *new_dentry,
9694                            unsigned int flags)
9695 {
9696         struct btrfs_trans_handle *trans;
9697         unsigned int trans_num_items;
9698         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9699         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9700         struct inode *new_inode = d_inode(new_dentry);
9701         struct inode *old_inode = d_inode(old_dentry);
9702         u64 index = 0;
9703         u64 root_objectid;
9704         int ret;
9705         u64 old_ino = btrfs_ino(old_inode);
9706         bool log_pinned = false;
9707
9708         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9709                 return -EPERM;
9710
9711         /* we only allow rename subvolume link between subvolumes */
9712         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9713                 return -EXDEV;
9714
9715         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9716             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9717                 return -ENOTEMPTY;
9718
9719         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9720             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9721                 return -ENOTEMPTY;
9722
9723
9724         /* check for collisions, even if the  name isn't there */
9725         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9726                              new_dentry->d_name.name,
9727                              new_dentry->d_name.len);
9728
9729         if (ret) {
9730                 if (ret == -EEXIST) {
9731                         /* we shouldn't get
9732                          * eexist without a new_inode */
9733                         if (WARN_ON(!new_inode)) {
9734                                 return ret;
9735                         }
9736                 } else {
9737                         /* maybe -EOVERFLOW */
9738                         return ret;
9739                 }
9740         }
9741         ret = 0;
9742
9743         /*
9744          * we're using rename to replace one file with another.  Start IO on it
9745          * now so  we don't add too much work to the end of the transaction
9746          */
9747         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9748                 filemap_flush(old_inode->i_mapping);
9749
9750         /* close the racy window with snapshot create/destroy ioctl */
9751         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9752                 down_read(&root->fs_info->subvol_sem);
9753         /*
9754          * We want to reserve the absolute worst case amount of items.  So if
9755          * both inodes are subvols and we need to unlink them then that would
9756          * require 4 item modifications, but if they are both normal inodes it
9757          * would require 5 item modifications, so we'll assume they are normal
9758          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9759          * should cover the worst case number of items we'll modify.
9760          * If our rename has the whiteout flag, we need more 5 units for the
9761          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9762          * when selinux is enabled).
9763          */
9764         trans_num_items = 11;
9765         if (flags & RENAME_WHITEOUT)
9766                 trans_num_items += 5;
9767         trans = btrfs_start_transaction(root, trans_num_items);
9768         if (IS_ERR(trans)) {
9769                 ret = PTR_ERR(trans);
9770                 goto out_notrans;
9771         }
9772
9773         if (dest != root)
9774                 btrfs_record_root_in_trans(trans, dest);
9775
9776         ret = btrfs_set_inode_index(new_dir, &index);
9777         if (ret)
9778                 goto out_fail;
9779
9780         BTRFS_I(old_inode)->dir_index = 0ULL;
9781         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9782                 /* force full log commit if subvolume involved. */
9783                 btrfs_set_log_full_commit(root->fs_info, trans);
9784         } else {
9785                 btrfs_pin_log_trans(root);
9786                 log_pinned = true;
9787                 ret = btrfs_insert_inode_ref(trans, dest,
9788                                              new_dentry->d_name.name,
9789                                              new_dentry->d_name.len,
9790                                              old_ino,
9791                                              btrfs_ino(new_dir), index);
9792                 if (ret)
9793                         goto out_fail;
9794         }
9795
9796         inode_inc_iversion(old_dir);
9797         inode_inc_iversion(new_dir);
9798         inode_inc_iversion(old_inode);
9799         old_dir->i_ctime = old_dir->i_mtime =
9800         new_dir->i_ctime = new_dir->i_mtime =
9801         old_inode->i_ctime = current_fs_time(old_dir->i_sb);
9802
9803         if (old_dentry->d_parent != new_dentry->d_parent)
9804                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9805
9806         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9807                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9808                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9809                                         old_dentry->d_name.name,
9810                                         old_dentry->d_name.len);
9811         } else {
9812                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9813                                         d_inode(old_dentry),
9814                                         old_dentry->d_name.name,
9815                                         old_dentry->d_name.len);
9816                 if (!ret)
9817                         ret = btrfs_update_inode(trans, root, old_inode);
9818         }
9819         if (ret) {
9820                 btrfs_abort_transaction(trans, root, ret);
9821                 goto out_fail;
9822         }
9823
9824         if (new_inode) {
9825                 inode_inc_iversion(new_inode);
9826                 new_inode->i_ctime = current_fs_time(new_inode->i_sb);
9827                 if (unlikely(btrfs_ino(new_inode) ==
9828                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9829                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9830                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9831                                                 root_objectid,
9832                                                 new_dentry->d_name.name,
9833                                                 new_dentry->d_name.len);
9834                         BUG_ON(new_inode->i_nlink == 0);
9835                 } else {
9836                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9837                                                  d_inode(new_dentry),
9838                                                  new_dentry->d_name.name,
9839                                                  new_dentry->d_name.len);
9840                 }
9841                 if (!ret && new_inode->i_nlink == 0)
9842                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9843                 if (ret) {
9844                         btrfs_abort_transaction(trans, root, ret);
9845                         goto out_fail;
9846                 }
9847         }
9848
9849         ret = btrfs_add_link(trans, new_dir, old_inode,
9850                              new_dentry->d_name.name,
9851                              new_dentry->d_name.len, 0, index);
9852         if (ret) {
9853                 btrfs_abort_transaction(trans, root, ret);
9854                 goto out_fail;
9855         }
9856
9857         if (old_inode->i_nlink == 1)
9858                 BTRFS_I(old_inode)->dir_index = index;
9859
9860         if (log_pinned) {
9861                 struct dentry *parent = new_dentry->d_parent;
9862
9863                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9864                 btrfs_end_log_trans(root);
9865                 log_pinned = false;
9866         }
9867
9868         if (flags & RENAME_WHITEOUT) {
9869                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9870                                                 old_dentry);
9871
9872                 if (ret) {
9873                         btrfs_abort_transaction(trans, root, ret);
9874                         goto out_fail;
9875                 }
9876         }
9877 out_fail:
9878         /*
9879          * If we have pinned the log and an error happened, we unpin tasks
9880          * trying to sync the log and force them to fallback to a transaction
9881          * commit if the log currently contains any of the inodes involved in
9882          * this rename operation (to ensure we do not persist a log with an
9883          * inconsistent state for any of these inodes or leading to any
9884          * inconsistencies when replayed). If the transaction was aborted, the
9885          * abortion reason is propagated to userspace when attempting to commit
9886          * the transaction. If the log does not contain any of these inodes, we
9887          * allow the tasks to sync it.
9888          */
9889         if (ret && log_pinned) {
9890                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9891                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9892                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9893                     (new_inode &&
9894                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9895                     btrfs_set_log_full_commit(root->fs_info, trans);
9896
9897                 btrfs_end_log_trans(root);
9898                 log_pinned = false;
9899         }
9900         btrfs_end_transaction(trans, root);
9901 out_notrans:
9902         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9903                 up_read(&root->fs_info->subvol_sem);
9904
9905         return ret;
9906 }
9907
9908 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9909                          struct inode *new_dir, struct dentry *new_dentry,
9910                          unsigned int flags)
9911 {
9912         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9913                 return -EINVAL;
9914
9915         if (flags & RENAME_EXCHANGE)
9916                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9917                                           new_dentry);
9918
9919         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9920 }
9921
9922 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9923 {
9924         struct btrfs_delalloc_work *delalloc_work;
9925         struct inode *inode;
9926
9927         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9928                                      work);
9929         inode = delalloc_work->inode;
9930         filemap_flush(inode->i_mapping);
9931         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9932                                 &BTRFS_I(inode)->runtime_flags))
9933                 filemap_flush(inode->i_mapping);
9934
9935         if (delalloc_work->delay_iput)
9936                 btrfs_add_delayed_iput(inode);
9937         else
9938                 iput(inode);
9939         complete(&delalloc_work->completion);
9940 }
9941
9942 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9943                                                     int delay_iput)
9944 {
9945         struct btrfs_delalloc_work *work;
9946
9947         work = kmalloc(sizeof(*work), GFP_NOFS);
9948         if (!work)
9949                 return NULL;
9950
9951         init_completion(&work->completion);
9952         INIT_LIST_HEAD(&work->list);
9953         work->inode = inode;
9954         work->delay_iput = delay_iput;
9955         WARN_ON_ONCE(!inode);
9956         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9957                         btrfs_run_delalloc_work, NULL, NULL);
9958
9959         return work;
9960 }
9961
9962 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9963 {
9964         wait_for_completion(&work->completion);
9965         kfree(work);
9966 }
9967
9968 /*
9969  * some fairly slow code that needs optimization. This walks the list
9970  * of all the inodes with pending delalloc and forces them to disk.
9971  */
9972 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9973                                    int nr)
9974 {
9975         struct btrfs_inode *binode;
9976         struct inode *inode;
9977         struct btrfs_delalloc_work *work, *next;
9978         struct list_head works;
9979         struct list_head splice;
9980         int ret = 0;
9981
9982         INIT_LIST_HEAD(&works);
9983         INIT_LIST_HEAD(&splice);
9984
9985         mutex_lock(&root->delalloc_mutex);
9986         spin_lock(&root->delalloc_lock);
9987         list_splice_init(&root->delalloc_inodes, &splice);
9988         while (!list_empty(&splice)) {
9989                 binode = list_entry(splice.next, struct btrfs_inode,
9990                                     delalloc_inodes);
9991
9992                 list_move_tail(&binode->delalloc_inodes,
9993                                &root->delalloc_inodes);
9994                 inode = igrab(&binode->vfs_inode);
9995                 if (!inode) {
9996                         cond_resched_lock(&root->delalloc_lock);
9997                         continue;
9998                 }
9999                 spin_unlock(&root->delalloc_lock);
10000
10001                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10002                 if (!work) {
10003                         if (delay_iput)
10004                                 btrfs_add_delayed_iput(inode);
10005                         else
10006                                 iput(inode);
10007                         ret = -ENOMEM;
10008                         goto out;
10009                 }
10010                 list_add_tail(&work->list, &works);
10011                 btrfs_queue_work(root->fs_info->flush_workers,
10012                                  &work->work);
10013                 ret++;
10014                 if (nr != -1 && ret >= nr)
10015                         goto out;
10016                 cond_resched();
10017                 spin_lock(&root->delalloc_lock);
10018         }
10019         spin_unlock(&root->delalloc_lock);
10020
10021 out:
10022         list_for_each_entry_safe(work, next, &works, list) {
10023                 list_del_init(&work->list);
10024                 btrfs_wait_and_free_delalloc_work(work);
10025         }
10026
10027         if (!list_empty_careful(&splice)) {
10028                 spin_lock(&root->delalloc_lock);
10029                 list_splice_tail(&splice, &root->delalloc_inodes);
10030                 spin_unlock(&root->delalloc_lock);
10031         }
10032         mutex_unlock(&root->delalloc_mutex);
10033         return ret;
10034 }
10035
10036 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10037 {
10038         int ret;
10039
10040         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
10041                 return -EROFS;
10042
10043         ret = __start_delalloc_inodes(root, delay_iput, -1);
10044         if (ret > 0)
10045                 ret = 0;
10046         /*
10047          * the filemap_flush will queue IO into the worker threads, but
10048          * we have to make sure the IO is actually started and that
10049          * ordered extents get created before we return
10050          */
10051         atomic_inc(&root->fs_info->async_submit_draining);
10052         while (atomic_read(&root->fs_info->nr_async_submits) ||
10053               atomic_read(&root->fs_info->async_delalloc_pages)) {
10054                 wait_event(root->fs_info->async_submit_wait,
10055                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
10056                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
10057         }
10058         atomic_dec(&root->fs_info->async_submit_draining);
10059         return ret;
10060 }
10061
10062 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10063                                int nr)
10064 {
10065         struct btrfs_root *root;
10066         struct list_head splice;
10067         int ret;
10068
10069         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10070                 return -EROFS;
10071
10072         INIT_LIST_HEAD(&splice);
10073
10074         mutex_lock(&fs_info->delalloc_root_mutex);
10075         spin_lock(&fs_info->delalloc_root_lock);
10076         list_splice_init(&fs_info->delalloc_roots, &splice);
10077         while (!list_empty(&splice) && nr) {
10078                 root = list_first_entry(&splice, struct btrfs_root,
10079                                         delalloc_root);
10080                 root = btrfs_grab_fs_root(root);
10081                 BUG_ON(!root);
10082                 list_move_tail(&root->delalloc_root,
10083                                &fs_info->delalloc_roots);
10084                 spin_unlock(&fs_info->delalloc_root_lock);
10085
10086                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10087                 btrfs_put_fs_root(root);
10088                 if (ret < 0)
10089                         goto out;
10090
10091                 if (nr != -1) {
10092                         nr -= ret;
10093                         WARN_ON(nr < 0);
10094                 }
10095                 spin_lock(&fs_info->delalloc_root_lock);
10096         }
10097         spin_unlock(&fs_info->delalloc_root_lock);
10098
10099         ret = 0;
10100         atomic_inc(&fs_info->async_submit_draining);
10101         while (atomic_read(&fs_info->nr_async_submits) ||
10102               atomic_read(&fs_info->async_delalloc_pages)) {
10103                 wait_event(fs_info->async_submit_wait,
10104                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10105                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10106         }
10107         atomic_dec(&fs_info->async_submit_draining);
10108 out:
10109         if (!list_empty_careful(&splice)) {
10110                 spin_lock(&fs_info->delalloc_root_lock);
10111                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10112                 spin_unlock(&fs_info->delalloc_root_lock);
10113         }
10114         mutex_unlock(&fs_info->delalloc_root_mutex);
10115         return ret;
10116 }
10117
10118 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10119                          const char *symname)
10120 {
10121         struct btrfs_trans_handle *trans;
10122         struct btrfs_root *root = BTRFS_I(dir)->root;
10123         struct btrfs_path *path;
10124         struct btrfs_key key;
10125         struct inode *inode = NULL;
10126         int err;
10127         int drop_inode = 0;
10128         u64 objectid;
10129         u64 index = 0;
10130         int name_len;
10131         int datasize;
10132         unsigned long ptr;
10133         struct btrfs_file_extent_item *ei;
10134         struct extent_buffer *leaf;
10135
10136         name_len = strlen(symname);
10137         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
10138                 return -ENAMETOOLONG;
10139
10140         /*
10141          * 2 items for inode item and ref
10142          * 2 items for dir items
10143          * 1 item for updating parent inode item
10144          * 1 item for the inline extent item
10145          * 1 item for xattr if selinux is on
10146          */
10147         trans = btrfs_start_transaction(root, 7);
10148         if (IS_ERR(trans))
10149                 return PTR_ERR(trans);
10150
10151         err = btrfs_find_free_ino(root, &objectid);
10152         if (err)
10153                 goto out_unlock;
10154
10155         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10156                                 dentry->d_name.len, btrfs_ino(dir), objectid,
10157                                 S_IFLNK|S_IRWXUGO, &index);
10158         if (IS_ERR(inode)) {
10159                 err = PTR_ERR(inode);
10160                 goto out_unlock;
10161         }
10162
10163         /*
10164         * If the active LSM wants to access the inode during
10165         * d_instantiate it needs these. Smack checks to see
10166         * if the filesystem supports xattrs by looking at the
10167         * ops vector.
10168         */
10169         inode->i_fop = &btrfs_file_operations;
10170         inode->i_op = &btrfs_file_inode_operations;
10171         inode->i_mapping->a_ops = &btrfs_aops;
10172         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10173
10174         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10175         if (err)
10176                 goto out_unlock_inode;
10177
10178         path = btrfs_alloc_path();
10179         if (!path) {
10180                 err = -ENOMEM;
10181                 goto out_unlock_inode;
10182         }
10183         key.objectid = btrfs_ino(inode);
10184         key.offset = 0;
10185         key.type = BTRFS_EXTENT_DATA_KEY;
10186         datasize = btrfs_file_extent_calc_inline_size(name_len);
10187         err = btrfs_insert_empty_item(trans, root, path, &key,
10188                                       datasize);
10189         if (err) {
10190                 btrfs_free_path(path);
10191                 goto out_unlock_inode;
10192         }
10193         leaf = path->nodes[0];
10194         ei = btrfs_item_ptr(leaf, path->slots[0],
10195                             struct btrfs_file_extent_item);
10196         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10197         btrfs_set_file_extent_type(leaf, ei,
10198                                    BTRFS_FILE_EXTENT_INLINE);
10199         btrfs_set_file_extent_encryption(leaf, ei, 0);
10200         btrfs_set_file_extent_compression(leaf, ei, 0);
10201         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10202         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10203
10204         ptr = btrfs_file_extent_inline_start(ei);
10205         write_extent_buffer(leaf, symname, ptr, name_len);
10206         btrfs_mark_buffer_dirty(leaf);
10207         btrfs_free_path(path);
10208
10209         inode->i_op = &btrfs_symlink_inode_operations;
10210         inode_nohighmem(inode);
10211         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10212         inode_set_bytes(inode, name_len);
10213         btrfs_i_size_write(inode, name_len);
10214         err = btrfs_update_inode(trans, root, inode);
10215         /*
10216          * Last step, add directory indexes for our symlink inode. This is the
10217          * last step to avoid extra cleanup of these indexes if an error happens
10218          * elsewhere above.
10219          */
10220         if (!err)
10221                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
10222         if (err) {
10223                 drop_inode = 1;
10224                 goto out_unlock_inode;
10225         }
10226
10227         unlock_new_inode(inode);
10228         d_instantiate(dentry, inode);
10229
10230 out_unlock:
10231         btrfs_end_transaction(trans, root);
10232         if (drop_inode) {
10233                 inode_dec_link_count(inode);
10234                 iput(inode);
10235         }
10236         btrfs_btree_balance_dirty(root);
10237         return err;
10238
10239 out_unlock_inode:
10240         drop_inode = 1;
10241         unlock_new_inode(inode);
10242         goto out_unlock;
10243 }
10244
10245 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10246                                        u64 start, u64 num_bytes, u64 min_size,
10247                                        loff_t actual_len, u64 *alloc_hint,
10248                                        struct btrfs_trans_handle *trans)
10249 {
10250         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10251         struct extent_map *em;
10252         struct btrfs_root *root = BTRFS_I(inode)->root;
10253         struct btrfs_key ins;
10254         u64 cur_offset = start;
10255         u64 i_size;
10256         u64 cur_bytes;
10257         u64 last_alloc = (u64)-1;
10258         int ret = 0;
10259         bool own_trans = true;
10260
10261         if (trans)
10262                 own_trans = false;
10263         while (num_bytes > 0) {
10264                 if (own_trans) {
10265                         trans = btrfs_start_transaction(root, 3);
10266                         if (IS_ERR(trans)) {
10267                                 ret = PTR_ERR(trans);
10268                                 break;
10269                         }
10270                 }
10271
10272                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10273                 cur_bytes = max(cur_bytes, min_size);
10274                 /*
10275                  * If we are severely fragmented we could end up with really
10276                  * small allocations, so if the allocator is returning small
10277                  * chunks lets make its job easier by only searching for those
10278                  * sized chunks.
10279                  */
10280                 cur_bytes = min(cur_bytes, last_alloc);
10281                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
10282                                            *alloc_hint, &ins, 1, 0);
10283                 if (ret) {
10284                         if (own_trans)
10285                                 btrfs_end_transaction(trans, root);
10286                         break;
10287                 }
10288                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
10289
10290                 last_alloc = ins.offset;
10291                 ret = insert_reserved_file_extent(trans, inode,
10292                                                   cur_offset, ins.objectid,
10293                                                   ins.offset, ins.offset,
10294                                                   ins.offset, 0, 0, 0,
10295                                                   BTRFS_FILE_EXTENT_PREALLOC);
10296                 if (ret) {
10297                         btrfs_free_reserved_extent(root, ins.objectid,
10298                                                    ins.offset, 0);
10299                         btrfs_abort_transaction(trans, root, ret);
10300                         if (own_trans)
10301                                 btrfs_end_transaction(trans, root);
10302                         break;
10303                 }
10304
10305                 btrfs_drop_extent_cache(inode, cur_offset,
10306                                         cur_offset + ins.offset -1, 0);
10307
10308                 em = alloc_extent_map();
10309                 if (!em) {
10310                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10311                                 &BTRFS_I(inode)->runtime_flags);
10312                         goto next;
10313                 }
10314
10315                 em->start = cur_offset;
10316                 em->orig_start = cur_offset;
10317                 em->len = ins.offset;
10318                 em->block_start = ins.objectid;
10319                 em->block_len = ins.offset;
10320                 em->orig_block_len = ins.offset;
10321                 em->ram_bytes = ins.offset;
10322                 em->bdev = root->fs_info->fs_devices->latest_bdev;
10323                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10324                 em->generation = trans->transid;
10325
10326                 while (1) {
10327                         write_lock(&em_tree->lock);
10328                         ret = add_extent_mapping(em_tree, em, 1);
10329                         write_unlock(&em_tree->lock);
10330                         if (ret != -EEXIST)
10331                                 break;
10332                         btrfs_drop_extent_cache(inode, cur_offset,
10333                                                 cur_offset + ins.offset - 1,
10334                                                 0);
10335                 }
10336                 free_extent_map(em);
10337 next:
10338                 num_bytes -= ins.offset;
10339                 cur_offset += ins.offset;
10340                 *alloc_hint = ins.objectid + ins.offset;
10341
10342                 inode_inc_iversion(inode);
10343                 inode->i_ctime = current_fs_time(inode->i_sb);
10344                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10345                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10346                     (actual_len > inode->i_size) &&
10347                     (cur_offset > inode->i_size)) {
10348                         if (cur_offset > actual_len)
10349                                 i_size = actual_len;
10350                         else
10351                                 i_size = cur_offset;
10352                         i_size_write(inode, i_size);
10353                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10354                 }
10355
10356                 ret = btrfs_update_inode(trans, root, inode);
10357
10358                 if (ret) {
10359                         btrfs_abort_transaction(trans, root, ret);
10360                         if (own_trans)
10361                                 btrfs_end_transaction(trans, root);
10362                         break;
10363                 }
10364
10365                 if (own_trans)
10366                         btrfs_end_transaction(trans, root);
10367         }
10368         return ret;
10369 }
10370
10371 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10372                               u64 start, u64 num_bytes, u64 min_size,
10373                               loff_t actual_len, u64 *alloc_hint)
10374 {
10375         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10376                                            min_size, actual_len, alloc_hint,
10377                                            NULL);
10378 }
10379
10380 int btrfs_prealloc_file_range_trans(struct inode *inode,
10381                                     struct btrfs_trans_handle *trans, int mode,
10382                                     u64 start, u64 num_bytes, u64 min_size,
10383                                     loff_t actual_len, u64 *alloc_hint)
10384 {
10385         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10386                                            min_size, actual_len, alloc_hint, trans);
10387 }
10388
10389 static int btrfs_set_page_dirty(struct page *page)
10390 {
10391         return __set_page_dirty_nobuffers(page);
10392 }
10393
10394 static int btrfs_permission(struct inode *inode, int mask)
10395 {
10396         struct btrfs_root *root = BTRFS_I(inode)->root;
10397         umode_t mode = inode->i_mode;
10398
10399         if (mask & MAY_WRITE &&
10400             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10401                 if (btrfs_root_readonly(root))
10402                         return -EROFS;
10403                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10404                         return -EACCES;
10405         }
10406         return generic_permission(inode, mask);
10407 }
10408
10409 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10410 {
10411         struct btrfs_trans_handle *trans;
10412         struct btrfs_root *root = BTRFS_I(dir)->root;
10413         struct inode *inode = NULL;
10414         u64 objectid;
10415         u64 index;
10416         int ret = 0;
10417
10418         /*
10419          * 5 units required for adding orphan entry
10420          */
10421         trans = btrfs_start_transaction(root, 5);
10422         if (IS_ERR(trans))
10423                 return PTR_ERR(trans);
10424
10425         ret = btrfs_find_free_ino(root, &objectid);
10426         if (ret)
10427                 goto out;
10428
10429         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10430                                 btrfs_ino(dir), objectid, mode, &index);
10431         if (IS_ERR(inode)) {
10432                 ret = PTR_ERR(inode);
10433                 inode = NULL;
10434                 goto out;
10435         }
10436
10437         inode->i_fop = &btrfs_file_operations;
10438         inode->i_op = &btrfs_file_inode_operations;
10439
10440         inode->i_mapping->a_ops = &btrfs_aops;
10441         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10442
10443         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10444         if (ret)
10445                 goto out_inode;
10446
10447         ret = btrfs_update_inode(trans, root, inode);
10448         if (ret)
10449                 goto out_inode;
10450         ret = btrfs_orphan_add(trans, inode);
10451         if (ret)
10452                 goto out_inode;
10453
10454         /*
10455          * We set number of links to 0 in btrfs_new_inode(), and here we set
10456          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10457          * through:
10458          *
10459          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10460          */
10461         set_nlink(inode, 1);
10462         unlock_new_inode(inode);
10463         d_tmpfile(dentry, inode);
10464         mark_inode_dirty(inode);
10465
10466 out:
10467         btrfs_end_transaction(trans, root);
10468         if (ret)
10469                 iput(inode);
10470         btrfs_balance_delayed_items(root);
10471         btrfs_btree_balance_dirty(root);
10472         return ret;
10473
10474 out_inode:
10475         unlock_new_inode(inode);
10476         goto out;
10477
10478 }
10479
10480 /* Inspired by filemap_check_errors() */
10481 int btrfs_inode_check_errors(struct inode *inode)
10482 {
10483         int ret = 0;
10484
10485         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10486             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10487                 ret = -ENOSPC;
10488         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10489             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10490                 ret = -EIO;
10491
10492         return ret;
10493 }
10494
10495 static const struct inode_operations btrfs_dir_inode_operations = {
10496         .getattr        = btrfs_getattr,
10497         .lookup         = btrfs_lookup,
10498         .create         = btrfs_create,
10499         .unlink         = btrfs_unlink,
10500         .link           = btrfs_link,
10501         .mkdir          = btrfs_mkdir,
10502         .rmdir          = btrfs_rmdir,
10503         .rename2        = btrfs_rename2,
10504         .symlink        = btrfs_symlink,
10505         .setattr        = btrfs_setattr,
10506         .mknod          = btrfs_mknod,
10507         .setxattr       = generic_setxattr,
10508         .getxattr       = generic_getxattr,
10509         .listxattr      = btrfs_listxattr,
10510         .removexattr    = generic_removexattr,
10511         .permission     = btrfs_permission,
10512         .get_acl        = btrfs_get_acl,
10513         .set_acl        = btrfs_set_acl,
10514         .update_time    = btrfs_update_time,
10515         .tmpfile        = btrfs_tmpfile,
10516 };
10517 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10518         .lookup         = btrfs_lookup,
10519         .permission     = btrfs_permission,
10520         .get_acl        = btrfs_get_acl,
10521         .set_acl        = btrfs_set_acl,
10522         .update_time    = btrfs_update_time,
10523 };
10524
10525 static const struct file_operations btrfs_dir_file_operations = {
10526         .llseek         = generic_file_llseek,
10527         .read           = generic_read_dir,
10528         .iterate        = btrfs_real_readdir,
10529         .unlocked_ioctl = btrfs_ioctl,
10530 #ifdef CONFIG_COMPAT
10531         .compat_ioctl   = btrfs_compat_ioctl,
10532 #endif
10533         .release        = btrfs_release_file,
10534         .fsync          = btrfs_sync_file,
10535 };
10536
10537 static const struct extent_io_ops btrfs_extent_io_ops = {
10538         .fill_delalloc = run_delalloc_range,
10539         .submit_bio_hook = btrfs_submit_bio_hook,
10540         .merge_bio_hook = btrfs_merge_bio_hook,
10541         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10542         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10543         .writepage_start_hook = btrfs_writepage_start_hook,
10544         .set_bit_hook = btrfs_set_bit_hook,
10545         .clear_bit_hook = btrfs_clear_bit_hook,
10546         .merge_extent_hook = btrfs_merge_extent_hook,
10547         .split_extent_hook = btrfs_split_extent_hook,
10548 };
10549
10550 /*
10551  * btrfs doesn't support the bmap operation because swapfiles
10552  * use bmap to make a mapping of extents in the file.  They assume
10553  * these extents won't change over the life of the file and they
10554  * use the bmap result to do IO directly to the drive.
10555  *
10556  * the btrfs bmap call would return logical addresses that aren't
10557  * suitable for IO and they also will change frequently as COW
10558  * operations happen.  So, swapfile + btrfs == corruption.
10559  *
10560  * For now we're avoiding this by dropping bmap.
10561  */
10562 static const struct address_space_operations btrfs_aops = {
10563         .readpage       = btrfs_readpage,
10564         .writepage      = btrfs_writepage,
10565         .writepages     = btrfs_writepages,
10566         .readpages      = btrfs_readpages,
10567         .direct_IO      = btrfs_direct_IO,
10568         .invalidatepage = btrfs_invalidatepage,
10569         .releasepage    = btrfs_releasepage,
10570         .set_page_dirty = btrfs_set_page_dirty,
10571         .error_remove_page = generic_error_remove_page,
10572 };
10573
10574 static const struct address_space_operations btrfs_symlink_aops = {
10575         .readpage       = btrfs_readpage,
10576         .writepage      = btrfs_writepage,
10577         .invalidatepage = btrfs_invalidatepage,
10578         .releasepage    = btrfs_releasepage,
10579 };
10580
10581 static const struct inode_operations btrfs_file_inode_operations = {
10582         .getattr        = btrfs_getattr,
10583         .setattr        = btrfs_setattr,
10584         .setxattr       = generic_setxattr,
10585         .getxattr       = generic_getxattr,
10586         .listxattr      = btrfs_listxattr,
10587         .removexattr    = generic_removexattr,
10588         .permission     = btrfs_permission,
10589         .fiemap         = btrfs_fiemap,
10590         .get_acl        = btrfs_get_acl,
10591         .set_acl        = btrfs_set_acl,
10592         .update_time    = btrfs_update_time,
10593 };
10594 static const struct inode_operations btrfs_special_inode_operations = {
10595         .getattr        = btrfs_getattr,
10596         .setattr        = btrfs_setattr,
10597         .permission     = btrfs_permission,
10598         .setxattr       = generic_setxattr,
10599         .getxattr       = generic_getxattr,
10600         .listxattr      = btrfs_listxattr,
10601         .removexattr    = generic_removexattr,
10602         .get_acl        = btrfs_get_acl,
10603         .set_acl        = btrfs_set_acl,
10604         .update_time    = btrfs_update_time,
10605 };
10606 static const struct inode_operations btrfs_symlink_inode_operations = {
10607         .readlink       = generic_readlink,
10608         .get_link       = page_get_link,
10609         .getattr        = btrfs_getattr,
10610         .setattr        = btrfs_setattr,
10611         .permission     = btrfs_permission,
10612         .setxattr       = generic_setxattr,
10613         .getxattr       = generic_getxattr,
10614         .listxattr      = btrfs_listxattr,
10615         .removexattr    = generic_removexattr,
10616         .update_time    = btrfs_update_time,
10617 };
10618
10619 const struct dentry_operations btrfs_dentry_operations = {
10620         .d_delete       = btrfs_dentry_delete,
10621         .d_release      = btrfs_dentry_release,
10622 };