fs/btrfs/file.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/fs.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/time.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/mpage.h>
  27 #include <linux/falloc.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/statfs.h>
  31 #include <linux/compat.h>
  32 #include <linux/slab.h>
  33 #include "ctree.h"
  34 #include "disk-io.h"
  35 #include "transaction.h"
  36 #include "btrfs_inode.h"
  37 #include "ioctl.h"
  38 #include "print-tree.h"
  39 #include "tree-log.h"
  40 #include "locking.h"
  41 #include "compat.h"
  42
  43 /*
  44  * when auto defrag is enabled we
  45  * queue up these defrag structs to remember which
  46  * inodes need defragging passes
  47  */
  48 struct inode_defrag {
  49         struct rb_node rb_node;
  50         /* objectid */
  51         u64 ino;
  52         /*
  53          * transid where the defrag was added, we search for
  54          * extents newer than this
  55          */
  56         u64 transid;
  57
  58         /* root objectid */
  59         u64 root;
  60
  61         /* last offset we were able to defrag */
  62         u64 last_offset;
  63
  64         /* if we've wrapped around back to zero once already */
  65         int cycled;
  66 };
  67
  68 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  69                                   struct inode_defrag *defrag2)
  70 {
  71         if (defrag1->root > defrag2->root)
  72                 return 1;
  73         else if (defrag1->root < defrag2->root)
  74                 return -1;
  75         else if (defrag1->ino > defrag2->ino)
  76                 return 1;
  77         else if (defrag1->ino < defrag2->ino)
  78                 return -1;
  79         else
  80                 return 0;
  81 }
  82
  83 /* pop a record for an inode into the defrag tree.  The lock
  84  * must be held already
  85  *
  86  * If you're inserting a record for an older transid than an
  87  * existing record, the transid already in the tree is lowered
  88  *
  89  * If an existing record is found the defrag item you
  90  * pass in is freed
  91  */
  92 static void __btrfs_add_inode_defrag(struct inode *inode,
  93                                     struct inode_defrag *defrag)
  94 {
  95         struct btrfs_root *root = BTRFS_I(inode)->root;
  96         struct inode_defrag *entry;
  97         struct rb_node **p;
  98         struct rb_node *parent = NULL;
  99         int ret;
 100
 101         p = &root->fs_info->defrag_inodes.rb_node;
 102         while (*p) {
 103                 parent = *p;
 104                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 105
 106                 ret = __compare_inode_defrag(defrag, entry);
 107                 if (ret < 0)
 108                         p = &parent->rb_left;
 109                 else if (ret > 0)
 110                         p = &parent->rb_right;
 111                 else {
 112                         /* if we're reinserting an entry for
 113                          * an old defrag run, make sure to
 114                          * lower the transid of our existing record
 115                          */
 116                         if (defrag->transid < entry->transid)
 117                                 entry->transid = defrag->transid;
 118                         if (defrag->last_offset > entry->last_offset)
 119                                 entry->last_offset = defrag->last_offset;
 120                         goto exists;
 121                 }
 122         }
 123         set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 124         rb_link_node(&defrag->rb_node, parent, p);
 125         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 126         return;
 127
 128 exists:
 129         kfree(defrag);
 130         return;
 131
 132 }
 133
 134 /*
 135  * insert a defrag record for this inode if auto defrag is
 136  * enabled
 137  */
 138 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 139                            struct inode *inode)
 140 {
 141         struct btrfs_root *root = BTRFS_I(inode)->root;
 142         struct inode_defrag *defrag;
 143         u64 transid;
 144
 145         if (!btrfs_test_opt(root, AUTO_DEFRAG))
 146                 return 0;
 147
 148         if (btrfs_fs_closing(root->fs_info))
 149                 return 0;
 150
 151         if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 152                 return 0;
 153
 154         if (trans)
 155                 transid = trans->transid;
 156         else
 157                 transid = BTRFS_I(inode)->root->last_trans;
 158
 159         defrag = kzalloc(sizeof(*defrag), GFP_NOFS);
 160         if (!defrag)
 161                 return -ENOMEM;
 162
 163         defrag->ino = btrfs_ino(inode);
 164         defrag->transid = transid;
 165         defrag->root = root->root_key.objectid;
 166
 167         spin_lock(&root->fs_info->defrag_inodes_lock);
 168         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 169                 __btrfs_add_inode_defrag(inode, defrag);
 170         else
 171                 kfree(defrag);
 172         spin_unlock(&root->fs_info->defrag_inodes_lock);
 173         return 0;
 174 }
 175
 176 /*
 177  * must be called with the defrag_inodes lock held
 178  */
 179 struct inode_defrag *btrfs_find_defrag_inode(struct btrfs_fs_info *info,
 180                                              u64 root, u64 ino,
 181                                              struct rb_node **next)
 182 {
 183         struct inode_defrag *entry = NULL;
 184         struct inode_defrag tmp;
 185         struct rb_node *p;
 186         struct rb_node *parent = NULL;
 187         int ret;
 188
 189         tmp.ino = ino;
 190         tmp.root = root;
 191
 192         p = info->defrag_inodes.rb_node;
 193         while (p) {
 194                 parent = p;
 195                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 196
 197                 ret = __compare_inode_defrag(&tmp, entry);
 198                 if (ret < 0)
 199                         p = parent->rb_left;
 200                 else if (ret > 0)
 201                         p = parent->rb_right;
 202                 else
 203                         return entry;
 204         }
 205
 206         if (next) {
 207                 while (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 208                         parent = rb_next(parent);
 209                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 210                 }
 211                 *next = parent;
 212         }
 213         return NULL;
 214 }
 215
 216 /*
 217  * run through the list of inodes in the FS that need
 218  * defragging
 219  */
 220 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 221 {
 222         struct inode_defrag *defrag;
 223         struct btrfs_root *inode_root;
 224         struct inode *inode;
 225         struct rb_node *n;
 226         struct btrfs_key key;
 227         struct btrfs_ioctl_defrag_range_args range;
 228         u64 first_ino = 0;
 229         u64 root_objectid = 0;
 230         int num_defrag;
 231         int defrag_batch = 1024;
 232
 233         memset(&range, 0, sizeof(range));
 234         range.len = (u64)-1;
 235
 236         atomic_inc(&fs_info->defrag_running);
 237         spin_lock(&fs_info->defrag_inodes_lock);
 238         while(1) {
 239                 n = NULL;
 240
 241                 /* find an inode to defrag */
 242                 defrag = btrfs_find_defrag_inode(fs_info, root_objectid,
 243                                                  first_ino, &n);
 244                 if (!defrag) {
 245                         if (n) {
 246                                 defrag = rb_entry(n, struct inode_defrag,
 247                                                   rb_node);
 248                         } else if (root_objectid || first_ino) {
 249                                 root_objectid = 0;
 250                                 first_ino = 0;
 251                                 continue;
 252                         } else {
 253                                 break;
 254                         }
 255                 }
 256
 257                 /* remove it from the rbtree */
 258                 first_ino = defrag->ino + 1;
 259                 root_objectid = defrag->root;
 260                 rb_erase(&defrag->rb_node, &fs_info->defrag_inodes);
 261
 262                 if (btrfs_fs_closing(fs_info))
 263                         goto next_free;
 264
 265                 spin_unlock(&fs_info->defrag_inodes_lock);
 266
 267                 /* get the inode */
 268                 key.objectid = defrag->root;
 269                 btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 270                 key.offset = (u64)-1;
 271                 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 272                 if (IS_ERR(inode_root))
 273                         goto next;
 274
 275                 key.objectid = defrag->ino;
 276                 btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 277                 key.offset = 0;
 278
 279                 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 280                 if (IS_ERR(inode))
 281                         goto next;
 282
 283                 /* do a chunk of defrag */
 284                 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 285                 range.start = defrag->last_offset;
 286                 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 287                                                defrag_batch);
 288                 /*
 289                  * if we filled the whole defrag batch, there
 290                  * must be more work to do.  Queue this defrag
 291                  * again
 292                  */
 293                 if (num_defrag == defrag_batch) {
 294                         defrag->last_offset = range.start;
 295                         __btrfs_add_inode_defrag(inode, defrag);
 296                         /*
 297                          * we don't want to kfree defrag, we added it back to
 298                          * the rbtree
 299                          */
 300                         defrag = NULL;
 301                 } else if (defrag->last_offset && !defrag->cycled) {
 302                         /*
 303                          * we didn't fill our defrag batch, but
 304                          * we didn't start at zero.  Make sure we loop
 305                          * around to the start of the file.
 306                          */
 307                         defrag->last_offset = 0;
 308                         defrag->cycled = 1;
 309                         __btrfs_add_inode_defrag(inode, defrag);
 310                         defrag = NULL;
 311                 }
 312
 313                 iput(inode);
 314 next:
 315                 spin_lock(&fs_info->defrag_inodes_lock);
 316 next_free:
 317                 kfree(defrag);
 318         }
 319         spin_unlock(&fs_info->defrag_inodes_lock);
 320
 321         atomic_dec(&fs_info->defrag_running);
 322
 323         /*
 324          * during unmount, we use the transaction_wait queue to
 325          * wait for the defragger to stop
 326          */
 327         wake_up(&fs_info->transaction_wait);
 328         return 0;
 329 }
 330
 331 /* simple helper to fault in pages and copy.  This should go away
 332  * and be replaced with calls into generic code.
 333  */
 334 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 335                                          size_t write_bytes,
 336                                          struct page **prepared_pages,
 337                                          struct iov_iter *i)
 338 {
 339         size_t copied = 0;
 340         size_t total_copied = 0;
 341         int pg = 0;
 342         int offset = pos & (PAGE_CACHE_SIZE - 1);
 343
 344         while (write_bytes > 0) {
 345                 size_t count = min_t(size_t,
 346                                      PAGE_CACHE_SIZE - offset, write_bytes);
 347                 struct page *page = prepared_pages[pg];
 348                 /*
 349                  * Copy data from userspace to the current page
 350                  *
 351                  * Disable pagefault to avoid recursive lock since
 352                  * the pages are already locked
 353                  */
 354                 pagefault_disable();
 355                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 356                 pagefault_enable();
 357
 358                 /* Flush processor's dcache for this page */
 359                 flush_dcache_page(page);
 360
 361                 /*
 362                  * if we get a partial write, we can end up with
 363                  * partially up to date pages.  These add
 364                  * a lot of complexity, so make sure they don't
 365                  * happen by forcing this copy to be retried.
 366                  *
 367                  * The rest of the btrfs_file_write code will fall
 368                  * back to page at a time copies after we return 0.
 369                  */
 370                 if (!PageUptodate(page) && copied < count)
 371                         copied = 0;
 372
 373                 iov_iter_advance(i, copied);
 374                 write_bytes -= copied;
 375                 total_copied += copied;
 376
 377                 /* Return to btrfs_file_aio_write to fault page */
 378                 if (unlikely(copied == 0))
 379                         break;
 380
 381                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 382                         offset += copied;
 383                 } else {
 384                         pg++;
 385                         offset = 0;
 386                 }
 387         }
 388         return total_copied;
 389 }
 390
 391 /*
 392  * unlocks pages after btrfs_file_write is done with them
 393  */
 394 void btrfs_drop_pages(struct page **pages, size_t num_pages)
 395 {
 396         size_t i;
 397         for (i = 0; i < num_pages; i++) {
 398                 /* page checked is some magic around finding pages that
 399                  * have been modified without going through btrfs_set_page_dirty
 400                  * clear it here
 401                  */
 402                 ClearPageChecked(pages[i]);
 403                 unlock_page(pages[i]);
 404                 mark_page_accessed(pages[i]);
 405                 page_cache_release(pages[i]);
 406         }
 407 }
 408
 409 /*
 410  * after copy_from_user, pages need to be dirtied and we need to make
 411  * sure holes are created between the current EOF and the start of
 412  * any next extents (if required).
 413  *
 414  * this also makes the decision about creating an inline extent vs
 415  * doing real data extents, marking pages dirty and delalloc as required.
 416  */
 417 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 418                       struct page **pages, size_t num_pages,
 419                       loff_t pos, size_t write_bytes,
 420                       struct extent_state **cached)
 421 {
 422         int err = 0;
 423         int i;
 424         u64 num_bytes;
 425         u64 start_pos;
 426         u64 end_of_last_block;
 427         u64 end_pos = pos + write_bytes;
 428         loff_t isize = i_size_read(inode);
 429
 430         start_pos = pos & ~((u64)root->sectorsize - 1);
 431         num_bytes = (write_bytes + pos - start_pos +
 432                     root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
 433
 434         end_of_last_block = start_pos + num_bytes - 1;
 435         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 436                                         cached);
 437         if (err)
 438                 return err;
 439
 440         for (i = 0; i < num_pages; i++) {
 441                 struct page *p = pages[i];
 442                 SetPageUptodate(p);
 443                 ClearPageChecked(p);
 444                 set_page_dirty(p);
 445         }
 446
 447         /*
 448          * we've only changed i_size in ram, and we haven't updated
 449          * the disk i_size.  There is no need to log the inode
 450          * at this time.
 451          */
 452         if (end_pos > isize)
 453                 i_size_write(inode, end_pos);
 454         return 0;
 455 }
 456
 457 /*
 458  * this drops all the extents in the cache that intersect the range
 459  * [start, end].  Existing extents are split as required.
 460  */
 461 int btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 462                                int skip_pinned)
 463 {
 464         struct extent_map *em;
 465         struct extent_map *split = NULL;
 466         struct extent_map *split2 = NULL;
 467         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 468         u64 len = end - start + 1;
 469         u64 gen;
 470         int ret;
 471         int testend = 1;
 472         unsigned long flags;
 473         int compressed = 0;
 474
 475         WARN_ON(end < start);
 476         if (end == (u64)-1) {
 477                 len = (u64)-1;
 478                 testend = 0;
 479         }
 480         while (1) {
 481                 if (!split)
 482                         split = alloc_extent_map();
 483                 if (!split2)
 484                         split2 = alloc_extent_map();
 485                 BUG_ON(!split || !split2); /* -ENOMEM */
 486
 487                 write_lock(&em_tree->lock);
 488                 em = lookup_extent_mapping(em_tree, start, len);
 489                 if (!em) {
 490                         write_unlock(&em_tree->lock);
 491                         break;
 492                 }
 493                 flags = em->flags;
 494                 gen = em->generation;
 495                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 496                         if (testend && em->start + em->len >= start + len) {
 497                                 free_extent_map(em);
 498                                 write_unlock(&em_tree->lock);
 499                                 break;
 500                         }
 501                         start = em->start + em->len;
 502                         if (testend)
 503                                 len = start + len - (em->start + em->len);
 504                         free_extent_map(em);
 505                         write_unlock(&em_tree->lock);
 506                         continue;
 507                 }
 508                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 509                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 510                 remove_extent_mapping(em_tree, em);
 511
 512                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 513                     em->start < start) {
 514                         split->start = em->start;
 515                         split->len = start - em->start;
 516                         split->orig_start = em->orig_start;
 517                         split->block_start = em->block_start;
 518
 519                         if (compressed)
 520                                 split->block_len = em->block_len;
 521                         else
 522                                 split->block_len = split->len;
 523                         split->generation = gen;
 524                         split->bdev = em->bdev;
 525                         split->flags = flags;
 526                         split->compress_type = em->compress_type;
 527                         ret = add_extent_mapping(em_tree, split);
 528                         BUG_ON(ret); /* Logic error */
 529                         list_move(&split->list, &em_tree->modified_extents);
 530                         free_extent_map(split);
 531                         split = split2;
 532                         split2 = NULL;
 533                 }
 534                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 535                     testend && em->start + em->len > start + len) {
 536                         u64 diff = start + len - em->start;
 537
 538                         split->start = start + len;
 539                         split->len = em->start + em->len - (start + len);
 540                         split->bdev = em->bdev;
 541                         split->flags = flags;
 542                         split->compress_type = em->compress_type;
 543                         split->generation = gen;
 544
 545                         if (compressed) {
 546                                 split->block_len = em->block_len;
 547                                 split->block_start = em->block_start;
 548                                 split->orig_start = em->orig_start;
 549                         } else {
 550                                 split->block_len = split->len;
 551                                 split->block_start = em->block_start + diff;
 552                                 split->orig_start = split->start;
 553                         }
 554
 555                         ret = add_extent_mapping(em_tree, split);
 556                         BUG_ON(ret); /* Logic error */
 557                         list_move(&split->list, &em_tree->modified_extents);
 558                         free_extent_map(split);
 559                         split = NULL;
 560                 }
 561                 write_unlock(&em_tree->lock);
 562
 563                 /* once for us */
 564                 free_extent_map(em);
 565                 /* once for the tree*/
 566                 free_extent_map(em);
 567         }
 568         if (split)
 569                 free_extent_map(split);
 570         if (split2)
 571                 free_extent_map(split2);
 572         return 0;
 573 }
 574
 575 /*
 576  * this is very complex, but the basic idea is to drop all extents
 577  * in the range start - end.  hint_block is filled in with a block number
 578  * that would be a good hint to the block allocator for this file.
 579  *
 580  * If an extent intersects the range but is not entirely inside the range
 581  * it is either truncated or split.  Anything entirely inside the range
 582  * is deleted from the tree.
 583  */
 584 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 585                          struct btrfs_root *root, struct inode *inode,
 586                          struct btrfs_path *path, u64 start, u64 end,
 587                          u64 *hint_byte, int drop_cache)
 588 {
 589         struct extent_buffer *leaf;
 590         struct btrfs_file_extent_item *fi;
 591         struct btrfs_key key;
 592         struct btrfs_key new_key;
 593         u64 ino = btrfs_ino(inode);
 594         u64 search_start = start;
 595         u64 disk_bytenr = 0;
 596         u64 num_bytes = 0;
 597         u64 extent_offset = 0;
 598         u64 extent_end = 0;
 599         int del_nr = 0;
 600         int del_slot = 0;
 601         int extent_type;
 602         int recow;
 603         int ret;
 604         int modify_tree = -1;
 605         int update_refs = (root->ref_cows || root == root->fs_info->tree_root);
 606
 607         if (drop_cache)
 608                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 609
 610         if (start >= BTRFS_I(inode)->disk_i_size)
 611                 modify_tree = 0;
 612
 613         while (1) {
 614                 recow = 0;
 615                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 616                                                search_start, modify_tree);
 617                 if (ret < 0)
 618                         break;
 619                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 620                         leaf = path->nodes[0];
 621                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 622                         if (key.objectid == ino &&
 623                             key.type == BTRFS_EXTENT_DATA_KEY)
 624                                 path->slots[0]--;
 625                 }
 626                 ret = 0;
 627 next_slot:
 628                 leaf = path->nodes[0];
 629                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 630                         BUG_ON(del_nr > 0);
 631                         ret = btrfs_next_leaf(root, path);
 632                         if (ret < 0)
 633                                 break;
 634                         if (ret > 0) {
 635                                 ret = 0;
 636                                 break;
 637                         }
 638                         leaf = path->nodes[0];
 639                         recow = 1;
 640                 }
 641
 642                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 643                 if (key.objectid > ino ||
 644                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 645                         break;
 646
 647                 fi = btrfs_item_ptr(leaf, path->slots[0],
 648                                     struct btrfs_file_extent_item);
 649                 extent_type = btrfs_file_extent_type(leaf, fi);
 650
 651                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 652                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 653                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 654                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 655                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 656                         extent_end = key.offset +
 657                                 btrfs_file_extent_num_bytes(leaf, fi);
 658                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 659                         extent_end = key.offset +
 660                                 btrfs_file_extent_inline_len(leaf, fi);
 661                 } else {
 662                         WARN_ON(1);
 663                         extent_end = search_start;
 664                 }
 665
 666                 if (extent_end <= search_start) {
 667                         path->slots[0]++;
 668                         goto next_slot;
 669                 }
 670
 671                 search_start = max(key.offset, start);
 672                 if (recow || !modify_tree) {
 673                         modify_tree = -1;
 674                         btrfs_release_path(path);
 675                         continue;
 676                 }
 677
 678                 /*
 679                  *     | - range to drop - |
 680                  *  | -------- extent -------- |
 681                  */
 682                 if (start > key.offset && end < extent_end) {
 683                         BUG_ON(del_nr > 0);
 684                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 685
 686                         memcpy(&new_key, &key, sizeof(new_key));
 687                         new_key.offset = start;
 688                         ret = btrfs_duplicate_item(trans, root, path,
 689                                                    &new_key);
 690                         if (ret == -EAGAIN) {
 691                                 btrfs_release_path(path);
 692                                 continue;
 693                         }
 694                         if (ret < 0)
 695                                 break;
 696
 697                         leaf = path->nodes[0];
 698                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 699                                             struct btrfs_file_extent_item);
 700                         btrfs_set_file_extent_num_bytes(leaf, fi,
 701                                                         start - key.offset);
 702
 703                         fi = btrfs_item_ptr(leaf, path->slots[0],
 704                                             struct btrfs_file_extent_item);
 705
 706                         extent_offset += start - key.offset;
 707                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 708                         btrfs_set_file_extent_num_bytes(leaf, fi,
 709                                                         extent_end - start);
 710                         btrfs_mark_buffer_dirty(leaf);
 711
 712                         if (update_refs && disk_bytenr > 0) {
 713                                 ret = btrfs_inc_extent_ref(trans, root,
 714                                                 disk_bytenr, num_bytes, 0,
 715                                                 root->root_key.objectid,
 716                                                 new_key.objectid,
 717                                                 start - extent_offset, 0);
 718                                 BUG_ON(ret); /* -ENOMEM */
 719                                 *hint_byte = disk_bytenr;
 720                         }
 721                         key.offset = start;
 722                 }
 723                 /*
 724                  *  | ---- range to drop ----- |
 725                  *      | -------- extent -------- |
 726                  */
 727                 if (start <= key.offset && end < extent_end) {
 728                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 729
 730                         memcpy(&new_key, &key, sizeof(new_key));
 731                         new_key.offset = end;
 732                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 733
 734                         extent_offset += end - key.offset;
 735                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 736                         btrfs_set_file_extent_num_bytes(leaf, fi,
 737                                                         extent_end - end);
 738                         btrfs_mark_buffer_dirty(leaf);
 739                         if (update_refs && disk_bytenr > 0) {
 740                                 inode_sub_bytes(inode, end - key.offset);
 741                                 *hint_byte = disk_bytenr;
 742                         }
 743                         break;
 744                 }
 745
 746                 search_start = extent_end;
 747                 /*
 748                  *       | ---- range to drop ----- |
 749                  *  | -------- extent -------- |
 750                  */
 751                 if (start > key.offset && end >= extent_end) {
 752                         BUG_ON(del_nr > 0);
 753                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 754
 755                         btrfs_set_file_extent_num_bytes(leaf, fi,
 756                                                         start - key.offset);
 757                         btrfs_mark_buffer_dirty(leaf);
 758                         if (update_refs && disk_bytenr > 0) {
 759                                 inode_sub_bytes(inode, extent_end - start);
 760                                 *hint_byte = disk_bytenr;
 761                         }
 762                         if (end == extent_end)
 763                                 break;
 764
 765                         path->slots[0]++;
 766                         goto next_slot;
 767                 }
 768
 769                 /*
 770                  *  | ---- range to drop ----- |
 771                  *    | ------ extent ------ |
 772                  */
 773                 if (start <= key.offset && end >= extent_end) {
 774                         if (del_nr == 0) {
 775                                 del_slot = path->slots[0];
 776                                 del_nr = 1;
 777                         } else {
 778                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 779                                 del_nr++;
 780                         }
 781
 782                         if (update_refs &&
 783                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 784                                 inode_sub_bytes(inode,
 785                                                 extent_end - key.offset);
 786                                 extent_end = ALIGN(extent_end,
 787                                                    root->sectorsize);
 788                         } else if (update_refs && disk_bytenr > 0) {
 789                                 ret = btrfs_free_extent(trans, root,
 790                                                 disk_bytenr, num_bytes, 0,
 791                                                 root->root_key.objectid,
 792                                                 key.objectid, key.offset -
 793                                                 extent_offset, 0);
 794                                 BUG_ON(ret); /* -ENOMEM */
 795                                 inode_sub_bytes(inode,
 796                                                 extent_end - key.offset);
 797                                 *hint_byte = disk_bytenr;
 798                         }
 799
 800                         if (end == extent_end)
 801                                 break;
 802
 803                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 804                                 path->slots[0]++;
 805                                 goto next_slot;
 806                         }
 807
 808                         ret = btrfs_del_items(trans, root, path, del_slot,
 809                                               del_nr);
 810                         if (ret) {
 811                                 btrfs_abort_transaction(trans, root, ret);
 812                                 break;
 813                         }
 814
 815                         del_nr = 0;
 816                         del_slot = 0;
 817
 818                         btrfs_release_path(path);
 819                         continue;
 820                 }
 821
 822                 BUG_ON(1);
 823         }
 824
 825         if (!ret && del_nr > 0) {
 826                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 827                 if (ret)
 828                         btrfs_abort_transaction(trans, root, ret);
 829         }
 830
 831         btrfs_release_path(path);
 832         return ret;
 833 }
 834
 835 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 836                        struct btrfs_root *root, struct inode *inode, u64 start,
 837                        u64 end, u64 *hint_byte, int drop_cache)
 838 {
 839         struct btrfs_path *path;
 840         int ret;
 841
 842         path = btrfs_alloc_path();
 843         if (!path)
 844                 return -ENOMEM;
 845         ret = __btrfs_drop_extents(trans, root, inode, path, start, end,
 846                                    hint_byte, drop_cache);
 847         btrfs_free_path(path);
 848         return ret;
 849 }
 850
 851 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 852                             u64 objectid, u64 bytenr, u64 orig_offset,
 853                             u64 *start, u64 *end)
 854 {
 855         struct btrfs_file_extent_item *fi;
 856         struct btrfs_key key;
 857         u64 extent_end;
 858
 859         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 860                 return 0;
 861
 862         btrfs_item_key_to_cpu(leaf, &key, slot);
 863         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 864                 return 0;
 865
 866         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 867         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 868             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 869             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 870             btrfs_file_extent_compression(leaf, fi) ||
 871             btrfs_file_extent_encryption(leaf, fi) ||
 872             btrfs_file_extent_other_encoding(leaf, fi))
 873                 return 0;
 874
 875         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 876         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 877                 return 0;
 878
 879         *start = key.offset;
 880         *end = extent_end;
 881         return 1;
 882 }
 883
 884 /*
 885  * Mark extent in the range start - end as written.
 886  *
 887  * This changes extent type from 'pre-allocated' to 'regular'. If only
 888  * part of extent is marked as written, the extent will be split into
 889  * two or three.
 890  */
 891 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 892                               struct inode *inode, u64 start, u64 end)
 893 {
 894         struct btrfs_root *root = BTRFS_I(inode)->root;
 895         struct extent_buffer *leaf;
 896         struct btrfs_path *path;
 897         struct btrfs_file_extent_item *fi;
 898         struct btrfs_key key;
 899         struct btrfs_key new_key;
 900         u64 bytenr;
 901         u64 num_bytes;
 902         u64 extent_end;
 903         u64 orig_offset;
 904         u64 other_start;
 905         u64 other_end;
 906         u64 split;
 907         int del_nr = 0;
 908         int del_slot = 0;
 909         int recow;
 910         int ret;
 911         u64 ino = btrfs_ino(inode);
 912
 913         path = btrfs_alloc_path();
 914         if (!path)
 915                 return -ENOMEM;
 916 again:
 917         recow = 0;
 918         split = start;
 919         key.objectid = ino;
 920         key.type = BTRFS_EXTENT_DATA_KEY;
 921         key.offset = split;
 922
 923         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 924         if (ret < 0)
 925                 goto out;
 926         if (ret > 0 && path->slots[0] > 0)
 927                 path->slots[0]--;
 928
 929         leaf = path->nodes[0];
 930         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 931         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
 932         fi = btrfs_item_ptr(leaf, path->slots[0],
 933                             struct btrfs_file_extent_item);
 934         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
 935                BTRFS_FILE_EXTENT_PREALLOC);
 936         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 937         BUG_ON(key.offset > start || extent_end < end);
 938
 939         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 940         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 941         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
 942         memcpy(&new_key, &key, sizeof(new_key));
 943
 944         if (start == key.offset && end < extent_end) {
 945                 other_start = 0;
 946                 other_end = start;
 947                 if (extent_mergeable(leaf, path->slots[0] - 1,
 948                                      ino, bytenr, orig_offset,
 949                                      &other_start, &other_end)) {
 950                         new_key.offset = end;
 951                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 952                         fi = btrfs_item_ptr(leaf, path->slots[0],
 953                                             struct btrfs_file_extent_item);
 954                         btrfs_set_file_extent_generation(leaf, fi,
 955                                                          trans->transid);
 956                         btrfs_set_file_extent_num_bytes(leaf, fi,
 957                                                         extent_end - end);
 958                         btrfs_set_file_extent_offset(leaf, fi,
 959                                                      end - orig_offset);
 960                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 961                                             struct btrfs_file_extent_item);
 962                         btrfs_set_file_extent_generation(leaf, fi,
 963                                                          trans->transid);
 964                         btrfs_set_file_extent_num_bytes(leaf, fi,
 965                                                         end - other_start);
 966                         btrfs_mark_buffer_dirty(leaf);
 967                         goto out;
 968                 }
 969         }
 970
 971         if (start > key.offset && end == extent_end) {
 972                 other_start = end;
 973                 other_end = 0;
 974                 if (extent_mergeable(leaf, path->slots[0] + 1,
 975                                      ino, bytenr, orig_offset,
 976                                      &other_start, &other_end)) {
 977                         fi = btrfs_item_ptr(leaf, path->slots[0],
 978                                             struct btrfs_file_extent_item);
 979                         btrfs_set_file_extent_num_bytes(leaf, fi,
 980                                                         start - key.offset);
 981                         btrfs_set_file_extent_generation(leaf, fi,
 982                                                          trans->transid);
 983                         path->slots[0]++;
 984                         new_key.offset = start;
 985                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 986
 987                         fi = btrfs_item_ptr(leaf, path->slots[0],
 988                                             struct btrfs_file_extent_item);
 989                         btrfs_set_file_extent_generation(leaf, fi,
 990                                                          trans->transid);
 991                         btrfs_set_file_extent_num_bytes(leaf, fi,
 992                                                         other_end - start);
 993                         btrfs_set_file_extent_offset(leaf, fi,
 994                                                      start - orig_offset);
 995                         btrfs_mark_buffer_dirty(leaf);
 996                         goto out;
 997                 }
 998         }
 999
1000         while (start > key.offset || end < extent_end) {
1001                 if (key.offset == start)
1002                         split = end;
1003
1004                 new_key.offset = split;
1005                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1006                 if (ret == -EAGAIN) {
1007                         btrfs_release_path(path);
1008                         goto again;
1009                 }
1010                 if (ret < 0) {
1011                         btrfs_abort_transaction(trans, root, ret);
1012                         goto out;
1013                 }
1014
1015                 leaf = path->nodes[0];
1016                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1017                                     struct btrfs_file_extent_item);
1018                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1019                 btrfs_set_file_extent_num_bytes(leaf, fi,
1020                                                 split - key.offset);
1021
1022                 fi = btrfs_item_ptr(leaf, path->slots[0],
1023                                     struct btrfs_file_extent_item);
1024
1025                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1026                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1027                 btrfs_set_file_extent_num_bytes(leaf, fi,
1028                                                 extent_end - split);
1029                 btrfs_mark_buffer_dirty(leaf);
1030
1031                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1032                                            root->root_key.objectid,
1033                                            ino, orig_offset, 0);
1034                 BUG_ON(ret); /* -ENOMEM */
1035
1036                 if (split == start) {
1037                         key.offset = start;
1038                 } else {
1039                         BUG_ON(start != key.offset);
1040                         path->slots[0]--;
1041                         extent_end = end;
1042                 }
1043                 recow = 1;
1044         }
1045
1046         other_start = end;
1047         other_end = 0;
1048         if (extent_mergeable(leaf, path->slots[0] + 1,
1049                              ino, bytenr, orig_offset,
1050                              &other_start, &other_end)) {
1051                 if (recow) {
1052                         btrfs_release_path(path);
1053                         goto again;
1054                 }
1055                 extent_end = other_end;
1056                 del_slot = path->slots[0] + 1;
1057                 del_nr++;
1058                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1059                                         0, root->root_key.objectid,
1060                                         ino, orig_offset, 0);
1061                 BUG_ON(ret); /* -ENOMEM */
1062         }
1063         other_start = 0;
1064         other_end = start;
1065         if (extent_mergeable(leaf, path->slots[0] - 1,
1066                              ino, bytenr, orig_offset,
1067                              &other_start, &other_end)) {
1068                 if (recow) {
1069                         btrfs_release_path(path);
1070                         goto again;
1071                 }
1072                 key.offset = other_start;
1073                 del_slot = path->slots[0];
1074                 del_nr++;
1075                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1076                                         0, root->root_key.objectid,
1077                                         ino, orig_offset, 0);
1078                 BUG_ON(ret); /* -ENOMEM */
1079         }
1080         if (del_nr == 0) {
1081                 fi = btrfs_item_ptr(leaf, path->slots[0],
1082                            struct btrfs_file_extent_item);
1083                 btrfs_set_file_extent_type(leaf, fi,
1084                                            BTRFS_FILE_EXTENT_REG);
1085                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1086                 btrfs_mark_buffer_dirty(leaf);
1087         } else {
1088                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1089                            struct btrfs_file_extent_item);
1090                 btrfs_set_file_extent_type(leaf, fi,
1091                                            BTRFS_FILE_EXTENT_REG);
1092                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1093                 btrfs_set_file_extent_num_bytes(leaf, fi,
1094                                                 extent_end - key.offset);
1095                 btrfs_mark_buffer_dirty(leaf);
1096
1097                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1098                 if (ret < 0) {
1099                         btrfs_abort_transaction(trans, root, ret);
1100                         goto out;
1101                 }
1102         }
1103 out:
1104         btrfs_free_path(path);
1105         return 0;
1106 }
1107
1108 /*
1109  * on error we return an unlocked page and the error value
1110  * on success we return a locked page and 0
1111  */
1112 static int prepare_uptodate_page(struct page *page, u64 pos,
1113                                  bool force_uptodate)
1114 {
1115         int ret = 0;
1116
1117         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1118             !PageUptodate(page)) {
1119                 ret = btrfs_readpage(NULL, page);
1120                 if (ret)
1121                         return ret;
1122                 lock_page(page);
1123                 if (!PageUptodate(page)) {
1124                         unlock_page(page);
1125                         return -EIO;
1126                 }
1127         }
1128         return 0;
1129 }
1130
1131 /*
1132  * this gets pages into the page cache and locks them down, it also properly
1133  * waits for data=ordered extents to finish before allowing the pages to be
1134  * modified.
1135  */
1136 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1137                          struct page **pages, size_t num_pages,
1138                          loff_t pos, unsigned long first_index,
1139                          size_t write_bytes, bool force_uptodate)
1140 {
1141         struct extent_state *cached_state = NULL;
1142         int i;
1143         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1144         struct inode *inode = fdentry(file)->d_inode;
1145         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1146         int err = 0;
1147         int faili = 0;
1148         u64 start_pos;
1149         u64 last_pos;
1150
1151         start_pos = pos & ~((u64)root->sectorsize - 1);
1152         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1153
1154 again:
1155         for (i = 0; i < num_pages; i++) {
1156                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1157                                                mask | __GFP_WRITE);
1158                 if (!pages[i]) {
1159                         faili = i - 1;
1160                         err = -ENOMEM;
1161                         goto fail;
1162                 }
1163
1164                 if (i == 0)
1165                         err = prepare_uptodate_page(pages[i], pos,
1166                                                     force_uptodate);
1167                 if (i == num_pages - 1)
1168                         err = prepare_uptodate_page(pages[i],
1169                                                     pos + write_bytes, false);
1170                 if (err) {
1171                         page_cache_release(pages[i]);
1172                         faili = i - 1;
1173                         goto fail;
1174                 }
1175                 wait_on_page_writeback(pages[i]);
1176         }
1177         err = 0;
1178         if (start_pos < inode->i_size) {
1179                 struct btrfs_ordered_extent *ordered;
1180                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1181                                  start_pos, last_pos - 1, 0, &cached_state);
1182                 ordered = btrfs_lookup_first_ordered_extent(inode,
1183                                                             last_pos - 1);
1184                 if (ordered &&
1185                     ordered->file_offset + ordered->len > start_pos &&
1186                     ordered->file_offset < last_pos) {
1187                         btrfs_put_ordered_extent(ordered);
1188                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1189                                              start_pos, last_pos - 1,
1190                                              &cached_state, GFP_NOFS);
1191                         for (i = 0; i < num_pages; i++) {
1192                                 unlock_page(pages[i]);
1193                                 page_cache_release(pages[i]);
1194                         }
1195                         btrfs_wait_ordered_range(inode, start_pos,
1196                                                  last_pos - start_pos);
1197                         goto again;
1198                 }
1199                 if (ordered)
1200                         btrfs_put_ordered_extent(ordered);
1201
1202                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1203                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1204                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
1205                                   GFP_NOFS);
1206                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1207                                      start_pos, last_pos - 1, &cached_state,
1208                                      GFP_NOFS);
1209         }
1210         for (i = 0; i < num_pages; i++) {
1211                 if (clear_page_dirty_for_io(pages[i]))
1212                         account_page_redirty(pages[i]);
1213                 set_page_extent_mapped(pages[i]);
1214                 WARN_ON(!PageLocked(pages[i]));
1215         }
1216         return 0;
1217 fail:
1218         while (faili >= 0) {
1219                 unlock_page(pages[faili]);
1220                 page_cache_release(pages[faili]);
1221                 faili--;
1222         }
1223         return err;
1224
1225 }
1226
1227 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1228                                                struct iov_iter *i,
1229                                                loff_t pos)
1230 {
1231         struct inode *inode = fdentry(file)->d_inode;
1232         struct btrfs_root *root = BTRFS_I(inode)->root;
1233         struct page **pages = NULL;
1234         unsigned long first_index;
1235         size_t num_written = 0;
1236         int nrptrs;
1237         int ret = 0;
1238         bool force_page_uptodate = false;
1239
1240         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1241                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1242                      (sizeof(struct page *)));
1243         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1244         nrptrs = max(nrptrs, 8);
1245         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1246         if (!pages)
1247                 return -ENOMEM;
1248
1249         first_index = pos >> PAGE_CACHE_SHIFT;
1250
1251         while (iov_iter_count(i) > 0) {
1252                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1253                 size_t write_bytes = min(iov_iter_count(i),
1254                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1255                                          offset);
1256                 size_t num_pages = (write_bytes + offset +
1257                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1258                 size_t dirty_pages;
1259                 size_t copied;
1260
1261                 WARN_ON(num_pages > nrptrs);
1262
1263                 /*
1264                  * Fault pages before locking them in prepare_pages
1265                  * to avoid recursive lock
1266                  */
1267                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1268                         ret = -EFAULT;
1269                         break;
1270                 }
1271
1272                 ret = btrfs_delalloc_reserve_space(inode,
1273                                         num_pages << PAGE_CACHE_SHIFT);
1274                 if (ret)
1275                         break;
1276
1277                 /*
1278                  * This is going to setup the pages array with the number of
1279                  * pages we want, so we don't really need to worry about the
1280                  * contents of pages from loop to loop
1281                  */
1282                 ret = prepare_pages(root, file, pages, num_pages,
1283                                     pos, first_index, write_bytes,
1284                                     force_page_uptodate);
1285                 if (ret) {
1286                         btrfs_delalloc_release_space(inode,
1287                                         num_pages << PAGE_CACHE_SHIFT);
1288                         break;
1289                 }
1290
1291                 copied = btrfs_copy_from_user(pos, num_pages,
1292                                            write_bytes, pages, i);
1293
1294                 /*
1295                  * if we have trouble faulting in the pages, fall
1296                  * back to one page at a time
1297                  */
1298                 if (copied < write_bytes)
1299                         nrptrs = 1;
1300
1301                 if (copied == 0) {
1302                         force_page_uptodate = true;
1303                         dirty_pages = 0;
1304                 } else {
1305                         force_page_uptodate = false;
1306                         dirty_pages = (copied + offset +
1307                                        PAGE_CACHE_SIZE - 1) >>
1308                                        PAGE_CACHE_SHIFT;
1309                 }
1310
1311                 /*
1312                  * If we had a short copy we need to release the excess delaloc
1313                  * bytes we reserved.  We need to increment outstanding_extents
1314                  * because btrfs_delalloc_release_space will decrement it, but
1315                  * we still have an outstanding extent for the chunk we actually
1316                  * managed to copy.
1317                  */
1318                 if (num_pages > dirty_pages) {
1319                         if (copied > 0) {
1320                                 spin_lock(&BTRFS_I(inode)->lock);
1321                                 BTRFS_I(inode)->outstanding_extents++;
1322                                 spin_unlock(&BTRFS_I(inode)->lock);
1323                         }
1324                         btrfs_delalloc_release_space(inode,
1325                                         (num_pages - dirty_pages) <<
1326                                         PAGE_CACHE_SHIFT);
1327                 }
1328
1329                 if (copied > 0) {
1330                         ret = btrfs_dirty_pages(root, inode, pages,
1331                                                 dirty_pages, pos, copied,
1332                                                 NULL);
1333                         if (ret) {
1334                                 btrfs_delalloc_release_space(inode,
1335                                         dirty_pages << PAGE_CACHE_SHIFT);
1336                                 btrfs_drop_pages(pages, num_pages);
1337                                 break;
1338                         }
1339                 }
1340
1341                 btrfs_drop_pages(pages, num_pages);
1342
1343                 cond_resched();
1344
1345                 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
1346                                                    dirty_pages);
1347                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1348                         btrfs_btree_balance_dirty(root, 1);
1349
1350                 pos += copied;
1351                 num_written += copied;
1352         }
1353
1354         kfree(pages);
1355
1356         return num_written ? num_written : ret;
1357 }
1358
1359 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1360                                     const struct iovec *iov,
1361                                     unsigned long nr_segs, loff_t pos,
1362                                     loff_t *ppos, size_t count, size_t ocount)
1363 {
1364         struct file *file = iocb->ki_filp;
1365         struct iov_iter i;
1366         ssize_t written;
1367         ssize_t written_buffered;
1368         loff_t endbyte;
1369         int err;
1370
1371         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1372                                             count, ocount);
1373
1374         if (written < 0 || written == count)
1375                 return written;
1376
1377         pos += written;
1378         count -= written;
1379         iov_iter_init(&i, iov, nr_segs, count, written);
1380         written_buffered = __btrfs_buffered_write(file, &i, pos);
1381         if (written_buffered < 0) {
1382                 err = written_buffered;
1383                 goto out;
1384         }
1385         endbyte = pos + written_buffered - 1;
1386         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1387         if (err)
1388                 goto out;
1389         written += written_buffered;
1390         *ppos = pos + written_buffered;
1391         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1392                                  endbyte >> PAGE_CACHE_SHIFT);
1393 out:
1394         return written ? written : err;
1395 }
1396
1397 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1398                                     const struct iovec *iov,
1399                                     unsigned long nr_segs, loff_t pos)
1400 {
1401         struct file *file = iocb->ki_filp;
1402         struct inode *inode = fdentry(file)->d_inode;
1403         struct btrfs_root *root = BTRFS_I(inode)->root;
1404         loff_t *ppos = &iocb->ki_pos;
1405         u64 start_pos;
1406         ssize_t num_written = 0;
1407         ssize_t err = 0;
1408         size_t count, ocount;
1409
1410         sb_start_write(inode->i_sb);
1411
1412         mutex_lock(&inode->i_mutex);
1413
1414         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1415         if (err) {
1416                 mutex_unlock(&inode->i_mutex);
1417                 goto out;
1418         }
1419         count = ocount;
1420
1421         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1422         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1423         if (err) {
1424                 mutex_unlock(&inode->i_mutex);
1425                 goto out;
1426         }
1427
1428         if (count == 0) {
1429                 mutex_unlock(&inode->i_mutex);
1430                 goto out;
1431         }
1432
1433         err = file_remove_suid(file);
1434         if (err) {
1435                 mutex_unlock(&inode->i_mutex);
1436                 goto out;
1437         }
1438
1439         /*
1440          * If BTRFS flips readonly due to some impossible error
1441          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1442          * although we have opened a file as writable, we have
1443          * to stop this write operation to ensure FS consistency.
1444          */
1445         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
1446                 mutex_unlock(&inode->i_mutex);
1447                 err = -EROFS;
1448                 goto out;
1449         }
1450
1451         err = file_update_time(file);
1452         if (err) {
1453                 mutex_unlock(&inode->i_mutex);
1454                 goto out;
1455         }
1456
1457         start_pos = round_down(pos, root->sectorsize);
1458         if (start_pos > i_size_read(inode)) {
1459                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1460                 if (err) {
1461                         mutex_unlock(&inode->i_mutex);
1462                         goto out;
1463                 }
1464         }
1465
1466         if (unlikely(file->f_flags & O_DIRECT)) {
1467                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1468                                                    pos, ppos, count, ocount);
1469         } else {
1470                 struct iov_iter i;
1471
1472                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1473
1474                 num_written = __btrfs_buffered_write(file, &i, pos);
1475                 if (num_written > 0)
1476                         *ppos = pos + num_written;
1477         }
1478
1479         mutex_unlock(&inode->i_mutex);
1480
1481         /*
1482          * we want to make sure fsync finds this change
1483          * but we haven't joined a transaction running right now.
1484          *
1485          * Later on, someone is sure to update the inode and get the
1486          * real transid recorded.
1487          *
1488          * We set last_trans now to the fs_info generation + 1,
1489          * this will either be one more than the running transaction
1490          * or the generation used for the next transaction if there isn't
1491          * one running right now.
1492          */
1493         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1494         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1495                 err = generic_write_sync(file, pos, num_written);
1496                 if (err < 0 && num_written > 0)
1497                         num_written = err;
1498         }
1499 out:
1500         sb_end_write(inode->i_sb);
1501         current->backing_dev_info = NULL;
1502         return num_written ? num_written : err;
1503 }
1504
1505 int btrfs_release_file(struct inode *inode, struct file *filp)
1506 {
1507         /*
1508          * ordered_data_close is set by settattr when we are about to truncate
1509          * a file from a non-zero size to a zero size.  This tries to
1510          * flush down new bytes that may have been written if the
1511          * application were using truncate to replace a file in place.
1512          */
1513         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1514                                &BTRFS_I(inode)->runtime_flags)) {
1515                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1516                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1517                         filemap_flush(inode->i_mapping);
1518         }
1519         if (filp->private_data)
1520                 btrfs_ioctl_trans_end(filp);
1521         return 0;
1522 }
1523
1524 /*
1525  * fsync call for both files and directories.  This logs the inode into
1526  * the tree log instead of forcing full commits whenever possible.
1527  *
1528  * It needs to call filemap_fdatawait so that all ordered extent updates are
1529  * in the metadata btree are up to date for copying to the log.
1530  *
1531  * It drops the inode mutex before doing the tree log commit.  This is an
1532  * important optimization for directories because holding the mutex prevents
1533  * new operations on the dir while we write to disk.
1534  */
1535 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1536 {
1537         struct dentry *dentry = file->f_path.dentry;
1538         struct inode *inode = dentry->d_inode;
1539         struct btrfs_root *root = BTRFS_I(inode)->root;
1540         int ret = 0;
1541         struct btrfs_trans_handle *trans;
1542
1543         trace_btrfs_sync_file(file, datasync);
1544
1545         mutex_lock(&inode->i_mutex);
1546
1547         /*
1548          * we wait first, since the writeback may change the inode, also wait
1549          * ordered range does a filemape_write_and_wait_range which is why we
1550          * don't do it above like other file systems.
1551          */
1552         root->log_batch++;
1553         btrfs_wait_ordered_range(inode, start, end);
1554         root->log_batch++;
1555
1556         /*
1557          * check the transaction that last modified this inode
1558          * and see if its already been committed
1559          */
1560         if (!BTRFS_I(inode)->last_trans) {
1561                 mutex_unlock(&inode->i_mutex);
1562                 goto out;
1563         }
1564
1565         /*
1566          * if the last transaction that changed this file was before
1567          * the current transaction, we can bail out now without any
1568          * syncing
1569          */
1570         smp_mb();
1571         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1572             BTRFS_I(inode)->last_trans <=
1573             root->fs_info->last_trans_committed) {
1574                 BTRFS_I(inode)->last_trans = 0;
1575
1576                 /*
1577                  * We'v had everything committed since the last time we were
1578                  * modified so clear this flag in case it was set for whatever
1579                  * reason, it's no longer relevant.
1580                  */
1581                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1582                           &BTRFS_I(inode)->runtime_flags);
1583                 mutex_unlock(&inode->i_mutex);
1584                 goto out;
1585         }
1586
1587         /*
1588          * ok we haven't committed the transaction yet, lets do a commit
1589          */
1590         if (file->private_data)
1591                 btrfs_ioctl_trans_end(file);
1592
1593         trans = btrfs_start_transaction(root, 0);
1594         if (IS_ERR(trans)) {
1595                 ret = PTR_ERR(trans);
1596                 mutex_unlock(&inode->i_mutex);
1597                 goto out;
1598         }
1599
1600         ret = btrfs_log_dentry_safe(trans, root, dentry);
1601         if (ret < 0) {
1602                 mutex_unlock(&inode->i_mutex);
1603                 goto out;
1604         }
1605
1606         /* we've logged all the items and now have a consistent
1607          * version of the file in the log.  It is possible that
1608          * someone will come in and modify the file, but that's
1609          * fine because the log is consistent on disk, and we
1610          * have references to all of the file's extents
1611          *
1612          * It is possible that someone will come in and log the
1613          * file again, but that will end up using the synchronization
1614          * inside btrfs_sync_log to keep things safe.
1615          */
1616         mutex_unlock(&inode->i_mutex);
1617
1618         if (ret != BTRFS_NO_LOG_SYNC) {
1619                 if (ret > 0) {
1620                         ret = btrfs_commit_transaction(trans, root);
1621                 } else {
1622                         ret = btrfs_sync_log(trans, root);
1623                         if (ret == 0)
1624                                 ret = btrfs_end_transaction(trans, root);
1625                         else
1626                                 ret = btrfs_commit_transaction(trans, root);
1627                 }
1628         } else {
1629                 ret = btrfs_end_transaction(trans, root);
1630         }
1631 out:
1632         return ret > 0 ? -EIO : ret;
1633 }
1634
1635 static const struct vm_operations_struct btrfs_file_vm_ops = {
1636         .fault          = filemap_fault,
1637         .page_mkwrite   = btrfs_page_mkwrite,
1638 };
1639
1640 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1641 {
1642         struct address_space *mapping = filp->f_mapping;
1643
1644         if (!mapping->a_ops->readpage)
1645                 return -ENOEXEC;
1646
1647         file_accessed(filp);
1648         vma->vm_ops = &btrfs_file_vm_ops;
1649         vma->vm_flags |= VM_CAN_NONLINEAR;
1650
1651         return 0;
1652 }
1653
1654 static long btrfs_fallocate(struct file *file, int mode,
1655                             loff_t offset, loff_t len)
1656 {
1657         struct inode *inode = file->f_path.dentry->d_inode;
1658         struct extent_state *cached_state = NULL;
1659         u64 cur_offset;
1660         u64 last_byte;
1661         u64 alloc_start;
1662         u64 alloc_end;
1663         u64 alloc_hint = 0;
1664         u64 locked_end;
1665         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1666         struct extent_map *em;
1667         int ret;
1668
1669         alloc_start = offset & ~mask;
1670         alloc_end =  (offset + len + mask) & ~mask;
1671
1672         /* We only support the FALLOC_FL_KEEP_SIZE mode */
1673         if (mode & ~FALLOC_FL_KEEP_SIZE)
1674                 return -EOPNOTSUPP;
1675
1676         /*
1677          * Make sure we have enough space before we do the
1678          * allocation.
1679          */
1680         ret = btrfs_check_data_free_space(inode, len);
1681         if (ret)
1682                 return ret;
1683
1684         /*
1685          * wait for ordered IO before we have any locks.  We'll loop again
1686          * below with the locks held.
1687          */
1688         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
1689
1690         mutex_lock(&inode->i_mutex);
1691         ret = inode_newsize_ok(inode, alloc_end);
1692         if (ret)
1693                 goto out;
1694
1695         if (alloc_start > inode->i_size) {
1696                 ret = btrfs_cont_expand(inode, i_size_read(inode),
1697                                         alloc_start);
1698                 if (ret)
1699                         goto out;
1700         }
1701
1702         locked_end = alloc_end - 1;
1703         while (1) {
1704                 struct btrfs_ordered_extent *ordered;
1705
1706                 /* the extent lock is ordered inside the running
1707                  * transaction
1708                  */
1709                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
1710                                  locked_end, 0, &cached_state);
1711                 ordered = btrfs_lookup_first_ordered_extent(inode,
1712                                                             alloc_end - 1);
1713                 if (ordered &&
1714                     ordered->file_offset + ordered->len > alloc_start &&
1715                     ordered->file_offset < alloc_end) {
1716                         btrfs_put_ordered_extent(ordered);
1717                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1718                                              alloc_start, locked_end,
1719                                              &cached_state, GFP_NOFS);
1720                         /*
1721                          * we can't wait on the range with the transaction
1722                          * running or with the extent lock held
1723                          */
1724                         btrfs_wait_ordered_range(inode, alloc_start,
1725                                                  alloc_end - alloc_start);
1726                 } else {
1727                         if (ordered)
1728                                 btrfs_put_ordered_extent(ordered);
1729                         break;
1730                 }
1731         }
1732
1733         cur_offset = alloc_start;
1734         while (1) {
1735                 u64 actual_end;
1736
1737                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
1738                                       alloc_end - cur_offset, 0);
1739                 if (IS_ERR_OR_NULL(em)) {
1740                         if (!em)
1741                                 ret = -ENOMEM;
1742                         else
1743                                 ret = PTR_ERR(em);
1744                         break;
1745                 }
1746                 last_byte = min(extent_map_end(em), alloc_end);
1747                 actual_end = min_t(u64, extent_map_end(em), offset + len);
1748                 last_byte = (last_byte + mask) & ~mask;
1749
1750                 if (em->block_start == EXTENT_MAP_HOLE ||
1751                     (cur_offset >= inode->i_size &&
1752                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
1753                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
1754                                                         last_byte - cur_offset,
1755                                                         1 << inode->i_blkbits,
1756                                                         offset + len,
1757                                                         &alloc_hint);
1758
1759                         if (ret < 0) {
1760                                 free_extent_map(em);
1761                                 break;
1762                         }
1763                 } else if (actual_end > inode->i_size &&
1764                            !(mode & FALLOC_FL_KEEP_SIZE)) {
1765                         /*
1766                          * We didn't need to allocate any more space, but we
1767                          * still extended the size of the file so we need to
1768                          * update i_size.
1769                          */
1770                         inode->i_ctime = CURRENT_TIME;
1771                         i_size_write(inode, actual_end);
1772                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
1773                 }
1774                 free_extent_map(em);
1775
1776                 cur_offset = last_byte;
1777                 if (cur_offset >= alloc_end) {
1778                         ret = 0;
1779                         break;
1780                 }
1781         }
1782         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
1783                              &cached_state, GFP_NOFS);
1784 out:
1785         mutex_unlock(&inode->i_mutex);
1786         /* Let go of our reservation. */
1787         btrfs_free_reserved_data_space(inode, len);
1788         return ret;
1789 }
1790
1791 static int find_desired_extent(struct inode *inode, loff_t *offset, int origin)
1792 {
1793         struct btrfs_root *root = BTRFS_I(inode)->root;
1794         struct extent_map *em;
1795         struct extent_state *cached_state = NULL;
1796         u64 lockstart = *offset;
1797         u64 lockend = i_size_read(inode);
1798         u64 start = *offset;
1799         u64 orig_start = *offset;
1800         u64 len = i_size_read(inode);
1801         u64 last_end = 0;
1802         int ret = 0;
1803
1804         lockend = max_t(u64, root->sectorsize, lockend);
1805         if (lockend <= lockstart)
1806                 lockend = lockstart + root->sectorsize;
1807
1808         len = lockend - lockstart + 1;
1809
1810         len = max_t(u64, len, root->sectorsize);
1811         if (inode->i_size == 0)
1812                 return -ENXIO;
1813
1814         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
1815                          &cached_state);
1816
1817         /*
1818          * Delalloc is such a pain.  If we have a hole and we have pending
1819          * delalloc for a portion of the hole we will get back a hole that
1820          * exists for the entire range since it hasn't been actually written
1821          * yet.  So to take care of this case we need to look for an extent just
1822          * before the position we want in case there is outstanding delalloc
1823          * going on here.
1824          */
1825         if (origin == SEEK_HOLE && start != 0) {
1826                 if (start <= root->sectorsize)
1827                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
1828                                                      root->sectorsize, 0);
1829                 else
1830                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
1831                                                      start - root->sectorsize,
1832                                                      root->sectorsize, 0);
1833                 if (IS_ERR(em)) {
1834                         ret = PTR_ERR(em);
1835                         goto out;
1836                 }
1837                 last_end = em->start + em->len;
1838                 if (em->block_start == EXTENT_MAP_DELALLOC)
1839                         last_end = min_t(u64, last_end, inode->i_size);
1840                 free_extent_map(em);
1841         }
1842
1843         while (1) {
1844                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
1845                 if (IS_ERR(em)) {
1846                         ret = PTR_ERR(em);
1847                         break;
1848                 }
1849
1850                 if (em->block_start == EXTENT_MAP_HOLE) {
1851                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1852                                 if (last_end <= orig_start) {
1853                                         free_extent_map(em);
1854                                         ret = -ENXIO;
1855                                         break;
1856                                 }
1857                         }
1858
1859                         if (origin == SEEK_HOLE) {
1860                                 *offset = start;
1861                                 free_extent_map(em);
1862                                 break;
1863                         }
1864                 } else {
1865                         if (origin == SEEK_DATA) {
1866                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
1867                                         if (start >= inode->i_size) {
1868                                                 free_extent_map(em);
1869                                                 ret = -ENXIO;
1870                                                 break;
1871                                         }
1872                                 }
1873
1874                                 *offset = start;
1875                                 free_extent_map(em);
1876                                 break;
1877                         }
1878                 }
1879
1880                 start = em->start + em->len;
1881                 last_end = em->start + em->len;
1882
1883                 if (em->block_start == EXTENT_MAP_DELALLOC)
1884                         last_end = min_t(u64, last_end, inode->i_size);
1885
1886                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1887                         free_extent_map(em);
1888                         ret = -ENXIO;
1889                         break;
1890                 }
1891                 free_extent_map(em);
1892                 cond_resched();
1893         }
1894         if (!ret)
1895                 *offset = min(*offset, inode->i_size);
1896 out:
1897         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
1898                              &cached_state, GFP_NOFS);
1899         return ret;
1900 }
1901
1902 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int origin)
1903 {
1904         struct inode *inode = file->f_mapping->host;
1905         int ret;
1906
1907         mutex_lock(&inode->i_mutex);
1908         switch (origin) {
1909         case SEEK_END:
1910         case SEEK_CUR:
1911                 offset = generic_file_llseek(file, offset, origin);
1912                 goto out;
1913         case SEEK_DATA:
1914         case SEEK_HOLE:
1915                 if (offset >= i_size_read(inode)) {
1916                         mutex_unlock(&inode->i_mutex);
1917                         return -ENXIO;
1918                 }
1919
1920                 ret = find_desired_extent(inode, &offset, origin);
1921                 if (ret) {
1922                         mutex_unlock(&inode->i_mutex);
1923                         return ret;
1924                 }
1925         }
1926
1927         if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
1928                 offset = -EINVAL;
1929                 goto out;
1930         }
1931         if (offset > inode->i_sb->s_maxbytes) {
1932                 offset = -EINVAL;
1933                 goto out;
1934         }
1935
1936         /* Special lock needed here? */
1937         if (offset != file->f_pos) {
1938                 file->f_pos = offset;
1939                 file->f_version = 0;
1940         }
1941 out:
1942         mutex_unlock(&inode->i_mutex);
1943         return offset;
1944 }
1945
1946 const struct file_operations btrfs_file_operations = {
1947         .llseek         = btrfs_file_llseek,
1948         .read           = do_sync_read,
1949         .write          = do_sync_write,
1950         .aio_read       = generic_file_aio_read,
1951         .splice_read    = generic_file_splice_read,
1952         .aio_write      = btrfs_file_aio_write,
1953         .mmap           = btrfs_file_mmap,
1954         .open           = generic_file_open,
1955         .release        = btrfs_release_file,
1956         .fsync          = btrfs_sync_file,
1957         .fallocate      = btrfs_fallocate,
1958         .unlocked_ioctl = btrfs_ioctl,
1959 #ifdef CONFIG_COMPAT
1960         .compat_ioctl   = btrfs_ioctl,
1961 #endif
1962 };