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         int ret;
 470         int testend = 1;
 471         unsigned long flags;
 472         int compressed = 0;
 473
 474         WARN_ON(end < start);
 475         if (end == (u64)-1) {
 476                 len = (u64)-1;
 477                 testend = 0;
 478         }
 479         while (1) {
 480                 if (!split)
 481                         split = alloc_extent_map();
 482                 if (!split2)
 483                         split2 = alloc_extent_map();
 484                 BUG_ON(!split || !split2); /* -ENOMEM */
 485
 486                 write_lock(&em_tree->lock);
 487                 em = lookup_extent_mapping(em_tree, start, len);
 488                 if (!em) {
 489                         write_unlock(&em_tree->lock);
 490                         break;
 491                 }
 492                 flags = em->flags;
 493                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 494                         if (testend && em->start + em->len >= start + len) {
 495                                 free_extent_map(em);
 496                                 write_unlock(&em_tree->lock);
 497                                 break;
 498                         }
 499                         start = em->start + em->len;
 500                         if (testend)
 501                                 len = start + len - (em->start + em->len);
 502                         free_extent_map(em);
 503                         write_unlock(&em_tree->lock);
 504                         continue;
 505                 }
 506                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 507                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 508                 remove_extent_mapping(em_tree, em);
 509
 510                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 511                     em->start < start) {
 512                         split->start = em->start;
 513                         split->len = start - em->start;
 514                         split->orig_start = em->orig_start;
 515                         split->block_start = em->block_start;
 516
 517                         if (compressed)
 518                                 split->block_len = em->block_len;
 519                         else
 520                                 split->block_len = split->len;
 521
 522                         split->bdev = em->bdev;
 523                         split->flags = flags;
 524                         split->compress_type = em->compress_type;
 525                         ret = add_extent_mapping(em_tree, split);
 526                         BUG_ON(ret); /* Logic error */
 527                         free_extent_map(split);
 528                         split = split2;
 529                         split2 = NULL;
 530                 }
 531                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 532                     testend && em->start + em->len > start + len) {
 533                         u64 diff = start + len - em->start;
 534
 535                         split->start = start + len;
 536                         split->len = em->start + em->len - (start + len);
 537                         split->bdev = em->bdev;
 538                         split->flags = flags;
 539                         split->compress_type = em->compress_type;
 540
 541                         if (compressed) {
 542                                 split->block_len = em->block_len;
 543                                 split->block_start = em->block_start;
 544                                 split->orig_start = em->orig_start;
 545                         } else {
 546                                 split->block_len = split->len;
 547                                 split->block_start = em->block_start + diff;
 548                                 split->orig_start = split->start;
 549                         }
 550
 551                         ret = add_extent_mapping(em_tree, split);
 552                         BUG_ON(ret); /* Logic error */
 553                         free_extent_map(split);
 554                         split = NULL;
 555                 }
 556                 write_unlock(&em_tree->lock);
 557
 558                 /* once for us */
 559                 free_extent_map(em);
 560                 /* once for the tree*/
 561                 free_extent_map(em);
 562         }
 563         if (split)
 564                 free_extent_map(split);
 565         if (split2)
 566                 free_extent_map(split2);
 567         return 0;
 568 }
 569
 570 /*
 571  * this is very complex, but the basic idea is to drop all extents
 572  * in the range start - end.  hint_block is filled in with a block number
 573  * that would be a good hint to the block allocator for this file.
 574  *
 575  * If an extent intersects the range but is not entirely inside the range
 576  * it is either truncated or split.  Anything entirely inside the range
 577  * is deleted from the tree.
 578  */
 579 int btrfs_drop_extents(struct btrfs_trans_handle *trans, struct inode *inode,
 580                        u64 start, u64 end, u64 *hint_byte, int drop_cache)
 581 {
 582         struct btrfs_root *root = BTRFS_I(inode)->root;
 583         struct extent_buffer *leaf;
 584         struct btrfs_file_extent_item *fi;
 585         struct btrfs_path *path;
 586         struct btrfs_key key;
 587         struct btrfs_key new_key;
 588         u64 ino = btrfs_ino(inode);
 589         u64 search_start = start;
 590         u64 disk_bytenr = 0;
 591         u64 num_bytes = 0;
 592         u64 extent_offset = 0;
 593         u64 extent_end = 0;
 594         int del_nr = 0;
 595         int del_slot = 0;
 596         int extent_type;
 597         int recow;
 598         int ret;
 599         int modify_tree = -1;
 600
 601         if (drop_cache)
 602                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 603
 604         path = btrfs_alloc_path();
 605         if (!path)
 606                 return -ENOMEM;
 607
 608         if (start >= BTRFS_I(inode)->disk_i_size)
 609                 modify_tree = 0;
 610
 611         while (1) {
 612                 recow = 0;
 613                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 614                                                search_start, modify_tree);
 615                 if (ret < 0)
 616                         break;
 617                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 618                         leaf = path->nodes[0];
 619                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 620                         if (key.objectid == ino &&
 621                             key.type == BTRFS_EXTENT_DATA_KEY)
 622                                 path->slots[0]--;
 623                 }
 624                 ret = 0;
 625 next_slot:
 626                 leaf = path->nodes[0];
 627                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 628                         BUG_ON(del_nr > 0);
 629                         ret = btrfs_next_leaf(root, path);
 630                         if (ret < 0)
 631                                 break;
 632                         if (ret > 0) {
 633                                 ret = 0;
 634                                 break;
 635                         }
 636                         leaf = path->nodes[0];
 637                         recow = 1;
 638                 }
 639
 640                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 641                 if (key.objectid > ino ||
 642                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 643                         break;
 644
 645                 fi = btrfs_item_ptr(leaf, path->slots[0],
 646                                     struct btrfs_file_extent_item);
 647                 extent_type = btrfs_file_extent_type(leaf, fi);
 648
 649                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 650                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 651                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 652                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 653                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 654                         extent_end = key.offset +
 655                                 btrfs_file_extent_num_bytes(leaf, fi);
 656                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 657                         extent_end = key.offset +
 658                                 btrfs_file_extent_inline_len(leaf, fi);
 659                 } else {
 660                         WARN_ON(1);
 661                         extent_end = search_start;
 662                 }
 663
 664                 if (extent_end <= search_start) {
 665                         path->slots[0]++;
 666                         goto next_slot;
 667                 }
 668
 669                 search_start = max(key.offset, start);
 670                 if (recow || !modify_tree) {
 671                         modify_tree = -1;
 672                         btrfs_release_path(path);
 673                         continue;
 674                 }
 675
 676                 /*
 677                  *     | - range to drop - |
 678                  *  | -------- extent -------- |
 679                  */
 680                 if (start > key.offset && end < extent_end) {
 681                         BUG_ON(del_nr > 0);
 682                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 683
 684                         memcpy(&new_key, &key, sizeof(new_key));
 685                         new_key.offset = start;
 686                         ret = btrfs_duplicate_item(trans, root, path,
 687                                                    &new_key);
 688                         if (ret == -EAGAIN) {
 689                                 btrfs_release_path(path);
 690                                 continue;
 691                         }
 692                         if (ret < 0)
 693                                 break;
 694
 695                         leaf = path->nodes[0];
 696                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 697                                             struct btrfs_file_extent_item);
 698                         btrfs_set_file_extent_num_bytes(leaf, fi,
 699                                                         start - key.offset);
 700
 701                         fi = btrfs_item_ptr(leaf, path->slots[0],
 702                                             struct btrfs_file_extent_item);
 703
 704                         extent_offset += start - key.offset;
 705                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 706                         btrfs_set_file_extent_num_bytes(leaf, fi,
 707                                                         extent_end - start);
 708                         btrfs_mark_buffer_dirty(leaf);
 709
 710                         if (disk_bytenr > 0) {
 711                                 ret = btrfs_inc_extent_ref(trans, root,
 712                                                 disk_bytenr, num_bytes, 0,
 713                                                 root->root_key.objectid,
 714                                                 new_key.objectid,
 715                                                 start - extent_offset, 0);
 716                                 BUG_ON(ret); /* -ENOMEM */
 717                                 *hint_byte = disk_bytenr;
 718                         }
 719                         key.offset = start;
 720                 }
 721                 /*
 722                  *  | ---- range to drop ----- |
 723                  *      | -------- extent -------- |
 724                  */
 725                 if (start <= key.offset && end < extent_end) {
 726                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 727
 728                         memcpy(&new_key, &key, sizeof(new_key));
 729                         new_key.offset = end;
 730                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 731
 732                         extent_offset += end - key.offset;
 733                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 734                         btrfs_set_file_extent_num_bytes(leaf, fi,
 735                                                         extent_end - end);
 736                         btrfs_mark_buffer_dirty(leaf);
 737                         if (disk_bytenr > 0) {
 738                                 inode_sub_bytes(inode, end - key.offset);
 739                                 *hint_byte = disk_bytenr;
 740                         }
 741                         break;
 742                 }
 743
 744                 search_start = extent_end;
 745                 /*
 746                  *       | ---- range to drop ----- |
 747                  *  | -------- extent -------- |
 748                  */
 749                 if (start > key.offset && end >= extent_end) {
 750                         BUG_ON(del_nr > 0);
 751                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 752
 753                         btrfs_set_file_extent_num_bytes(leaf, fi,
 754                                                         start - key.offset);
 755                         btrfs_mark_buffer_dirty(leaf);
 756                         if (disk_bytenr > 0) {
 757                                 inode_sub_bytes(inode, extent_end - start);
 758                                 *hint_byte = disk_bytenr;
 759                         }
 760                         if (end == extent_end)
 761                                 break;
 762
 763                         path->slots[0]++;
 764                         goto next_slot;
 765                 }
 766
 767                 /*
 768                  *  | ---- range to drop ----- |
 769                  *    | ------ extent ------ |
 770                  */
 771                 if (start <= key.offset && end >= extent_end) {
 772                         if (del_nr == 0) {
 773                                 del_slot = path->slots[0];
 774                                 del_nr = 1;
 775                         } else {
 776                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 777                                 del_nr++;
 778                         }
 779
 780                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 781                                 inode_sub_bytes(inode,
 782                                                 extent_end - key.offset);
 783                                 extent_end = ALIGN(extent_end,
 784                                                    root->sectorsize);
 785                         } else if (disk_bytenr > 0) {
 786                                 ret = btrfs_free_extent(trans, root,
 787                                                 disk_bytenr, num_bytes, 0,
 788                                                 root->root_key.objectid,
 789                                                 key.objectid, key.offset -
 790                                                 extent_offset, 0);
 791                                 BUG_ON(ret); /* -ENOMEM */
 792                                 inode_sub_bytes(inode,
 793                                                 extent_end - key.offset);
 794                                 *hint_byte = disk_bytenr;
 795                         }
 796
 797                         if (end == extent_end)
 798                                 break;
 799
 800                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 801                                 path->slots[0]++;
 802                                 goto next_slot;
 803                         }
 804
 805                         ret = btrfs_del_items(trans, root, path, del_slot,
 806                                               del_nr);
 807                         if (ret) {
 808                                 btrfs_abort_transaction(trans, root, ret);
 809                                 goto out;
 810                         }
 811
 812                         del_nr = 0;
 813                         del_slot = 0;
 814
 815                         btrfs_release_path(path);
 816                         continue;
 817                 }
 818
 819                 BUG_ON(1);
 820         }
 821
 822         if (!ret && del_nr > 0) {
 823                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 824                 if (ret)
 825                         btrfs_abort_transaction(trans, root, ret);
 826         }
 827
 828 out:
 829         btrfs_free_path(path);
 830         return ret;
 831 }
 832
 833 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 834                             u64 objectid, u64 bytenr, u64 orig_offset,
 835                             u64 *start, u64 *end)
 836 {
 837         struct btrfs_file_extent_item *fi;
 838         struct btrfs_key key;
 839         u64 extent_end;
 840
 841         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 842                 return 0;
 843
 844         btrfs_item_key_to_cpu(leaf, &key, slot);
 845         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 846                 return 0;
 847
 848         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 849         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 850             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 851             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 852             btrfs_file_extent_compression(leaf, fi) ||
 853             btrfs_file_extent_encryption(leaf, fi) ||
 854             btrfs_file_extent_other_encoding(leaf, fi))
 855                 return 0;
 856
 857         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 858         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 859                 return 0;
 860
 861         *start = key.offset;
 862         *end = extent_end;
 863         return 1;
 864 }
 865
 866 /*
 867  * Mark extent in the range start - end as written.
 868  *
 869  * This changes extent type from 'pre-allocated' to 'regular'. If only
 870  * part of extent is marked as written, the extent will be split into
 871  * two or three.
 872  */
 873 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 874                               struct inode *inode, u64 start, u64 end)
 875 {
 876         struct btrfs_root *root = BTRFS_I(inode)->root;
 877         struct extent_buffer *leaf;
 878         struct btrfs_path *path;
 879         struct btrfs_file_extent_item *fi;
 880         struct btrfs_key key;
 881         struct btrfs_key new_key;
 882         u64 bytenr;
 883         u64 num_bytes;
 884         u64 extent_end;
 885         u64 orig_offset;
 886         u64 other_start;
 887         u64 other_end;
 888         u64 split;
 889         int del_nr = 0;
 890         int del_slot = 0;
 891         int recow;
 892         int ret;
 893         u64 ino = btrfs_ino(inode);
 894
 895         btrfs_drop_extent_cache(inode, start, end - 1, 0);
 896
 897         path = btrfs_alloc_path();
 898         if (!path)
 899                 return -ENOMEM;
 900 again:
 901         recow = 0;
 902         split = start;
 903         key.objectid = ino;
 904         key.type = BTRFS_EXTENT_DATA_KEY;
 905         key.offset = split;
 906
 907         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 908         if (ret < 0)
 909                 goto out;
 910         if (ret > 0 && path->slots[0] > 0)
 911                 path->slots[0]--;
 912
 913         leaf = path->nodes[0];
 914         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 915         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
 916         fi = btrfs_item_ptr(leaf, path->slots[0],
 917                             struct btrfs_file_extent_item);
 918         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
 919                BTRFS_FILE_EXTENT_PREALLOC);
 920         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 921         BUG_ON(key.offset > start || extent_end < end);
 922
 923         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 924         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 925         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
 926         memcpy(&new_key, &key, sizeof(new_key));
 927
 928         if (start == key.offset && end < extent_end) {
 929                 other_start = 0;
 930                 other_end = start;
 931                 if (extent_mergeable(leaf, path->slots[0] - 1,
 932                                      ino, bytenr, orig_offset,
 933                                      &other_start, &other_end)) {
 934                         new_key.offset = end;
 935                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 936                         fi = btrfs_item_ptr(leaf, path->slots[0],
 937                                             struct btrfs_file_extent_item);
 938                         btrfs_set_file_extent_num_bytes(leaf, fi,
 939                                                         extent_end - end);
 940                         btrfs_set_file_extent_offset(leaf, fi,
 941                                                      end - orig_offset);
 942                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 943                                             struct btrfs_file_extent_item);
 944                         btrfs_set_file_extent_num_bytes(leaf, fi,
 945                                                         end - other_start);
 946                         btrfs_mark_buffer_dirty(leaf);
 947                         goto out;
 948                 }
 949         }
 950
 951         if (start > key.offset && end == extent_end) {
 952                 other_start = end;
 953                 other_end = 0;
 954                 if (extent_mergeable(leaf, path->slots[0] + 1,
 955                                      ino, bytenr, orig_offset,
 956                                      &other_start, &other_end)) {
 957                         fi = btrfs_item_ptr(leaf, path->slots[0],
 958                                             struct btrfs_file_extent_item);
 959                         btrfs_set_file_extent_num_bytes(leaf, fi,
 960                                                         start - key.offset);
 961                         path->slots[0]++;
 962                         new_key.offset = start;
 963                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 964
 965                         fi = btrfs_item_ptr(leaf, path->slots[0],
 966                                             struct btrfs_file_extent_item);
 967                         btrfs_set_file_extent_num_bytes(leaf, fi,
 968                                                         other_end - start);
 969                         btrfs_set_file_extent_offset(leaf, fi,
 970                                                      start - orig_offset);
 971                         btrfs_mark_buffer_dirty(leaf);
 972                         goto out;
 973                 }
 974         }
 975
 976         while (start > key.offset || end < extent_end) {
 977                 if (key.offset == start)
 978                         split = end;
 979
 980                 new_key.offset = split;
 981                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
 982                 if (ret == -EAGAIN) {
 983                         btrfs_release_path(path);
 984                         goto again;
 985                 }
 986                 if (ret < 0) {
 987                         btrfs_abort_transaction(trans, root, ret);
 988                         goto out;
 989                 }
 990
 991                 leaf = path->nodes[0];
 992                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 993                                     struct btrfs_file_extent_item);
 994                 btrfs_set_file_extent_num_bytes(leaf, fi,
 995                                                 split - key.offset);
 996
 997                 fi = btrfs_item_ptr(leaf, path->slots[0],
 998                                     struct btrfs_file_extent_item);
 999
1000                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1001                 btrfs_set_file_extent_num_bytes(leaf, fi,
1002                                                 extent_end - split);
1003                 btrfs_mark_buffer_dirty(leaf);
1004
1005                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1006                                            root->root_key.objectid,
1007                                            ino, orig_offset, 0);
1008                 BUG_ON(ret); /* -ENOMEM */
1009
1010                 if (split == start) {
1011                         key.offset = start;
1012                 } else {
1013                         BUG_ON(start != key.offset);
1014                         path->slots[0]--;
1015                         extent_end = end;
1016                 }
1017                 recow = 1;
1018         }
1019
1020         other_start = end;
1021         other_end = 0;
1022         if (extent_mergeable(leaf, path->slots[0] + 1,
1023                              ino, bytenr, orig_offset,
1024                              &other_start, &other_end)) {
1025                 if (recow) {
1026                         btrfs_release_path(path);
1027                         goto again;
1028                 }
1029                 extent_end = other_end;
1030                 del_slot = path->slots[0] + 1;
1031                 del_nr++;
1032                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1033                                         0, root->root_key.objectid,
1034                                         ino, orig_offset, 0);
1035                 BUG_ON(ret); /* -ENOMEM */
1036         }
1037         other_start = 0;
1038         other_end = start;
1039         if (extent_mergeable(leaf, path->slots[0] - 1,
1040                              ino, bytenr, orig_offset,
1041                              &other_start, &other_end)) {
1042                 if (recow) {
1043                         btrfs_release_path(path);
1044                         goto again;
1045                 }
1046                 key.offset = other_start;
1047                 del_slot = path->slots[0];
1048                 del_nr++;
1049                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1050                                         0, root->root_key.objectid,
1051                                         ino, orig_offset, 0);
1052                 BUG_ON(ret); /* -ENOMEM */
1053         }
1054         if (del_nr == 0) {
1055                 fi = btrfs_item_ptr(leaf, path->slots[0],
1056                            struct btrfs_file_extent_item);
1057                 btrfs_set_file_extent_type(leaf, fi,
1058                                            BTRFS_FILE_EXTENT_REG);
1059                 btrfs_mark_buffer_dirty(leaf);
1060         } else {
1061                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1062                            struct btrfs_file_extent_item);
1063                 btrfs_set_file_extent_type(leaf, fi,
1064                                            BTRFS_FILE_EXTENT_REG);
1065                 btrfs_set_file_extent_num_bytes(leaf, fi,
1066                                                 extent_end - key.offset);
1067                 btrfs_mark_buffer_dirty(leaf);
1068
1069                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1070                 if (ret < 0) {
1071                         btrfs_abort_transaction(trans, root, ret);
1072                         goto out;
1073                 }
1074         }
1075 out:
1076         btrfs_free_path(path);
1077         return 0;
1078 }
1079
1080 /*
1081  * on error we return an unlocked page and the error value
1082  * on success we return a locked page and 0
1083  */
1084 static int prepare_uptodate_page(struct page *page, u64 pos,
1085                                  bool force_uptodate)
1086 {
1087         int ret = 0;
1088
1089         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1090             !PageUptodate(page)) {
1091                 ret = btrfs_readpage(NULL, page);
1092                 if (ret)
1093                         return ret;
1094                 lock_page(page);
1095                 if (!PageUptodate(page)) {
1096                         unlock_page(page);
1097                         return -EIO;
1098                 }
1099         }
1100         return 0;
1101 }
1102
1103 /*
1104  * this gets pages into the page cache and locks them down, it also properly
1105  * waits for data=ordered extents to finish before allowing the pages to be
1106  * modified.
1107  */
1108 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1109                          struct page **pages, size_t num_pages,
1110                          loff_t pos, unsigned long first_index,
1111                          size_t write_bytes, bool force_uptodate)
1112 {
1113         struct extent_state *cached_state = NULL;
1114         int i;
1115         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1116         struct inode *inode = fdentry(file)->d_inode;
1117         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1118         int err = 0;
1119         int faili = 0;
1120         u64 start_pos;
1121         u64 last_pos;
1122
1123         start_pos = pos & ~((u64)root->sectorsize - 1);
1124         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1125
1126 again:
1127         for (i = 0; i < num_pages; i++) {
1128                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1129                                                mask | __GFP_WRITE);
1130                 if (!pages[i]) {
1131                         faili = i - 1;
1132                         err = -ENOMEM;
1133                         goto fail;
1134                 }
1135
1136                 if (i == 0)
1137                         err = prepare_uptodate_page(pages[i], pos,
1138                                                     force_uptodate);
1139                 if (i == num_pages - 1)
1140                         err = prepare_uptodate_page(pages[i],
1141                                                     pos + write_bytes, false);
1142                 if (err) {
1143                         page_cache_release(pages[i]);
1144                         faili = i - 1;
1145                         goto fail;
1146                 }
1147                 wait_on_page_writeback(pages[i]);
1148         }
1149         err = 0;
1150         if (start_pos < inode->i_size) {
1151                 struct btrfs_ordered_extent *ordered;
1152                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1153                                  start_pos, last_pos - 1, 0, &cached_state);
1154                 ordered = btrfs_lookup_first_ordered_extent(inode,
1155                                                             last_pos - 1);
1156                 if (ordered &&
1157                     ordered->file_offset + ordered->len > start_pos &&
1158                     ordered->file_offset < last_pos) {
1159                         btrfs_put_ordered_extent(ordered);
1160                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1161                                              start_pos, last_pos - 1,
1162                                              &cached_state, GFP_NOFS);
1163                         for (i = 0; i < num_pages; i++) {
1164                                 unlock_page(pages[i]);
1165                                 page_cache_release(pages[i]);
1166                         }
1167                         btrfs_wait_ordered_range(inode, start_pos,
1168                                                  last_pos - start_pos);
1169                         goto again;
1170                 }
1171                 if (ordered)
1172                         btrfs_put_ordered_extent(ordered);
1173
1174                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1175                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1176                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
1177                                   GFP_NOFS);
1178                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1179                                      start_pos, last_pos - 1, &cached_state,
1180                                      GFP_NOFS);
1181         }
1182         for (i = 0; i < num_pages; i++) {
1183                 if (clear_page_dirty_for_io(pages[i]))
1184                         account_page_redirty(pages[i]);
1185                 set_page_extent_mapped(pages[i]);
1186                 WARN_ON(!PageLocked(pages[i]));
1187         }
1188         return 0;
1189 fail:
1190         while (faili >= 0) {
1191                 unlock_page(pages[faili]);
1192                 page_cache_release(pages[faili]);
1193                 faili--;
1194         }
1195         return err;
1196
1197 }
1198
1199 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1200                                                struct iov_iter *i,
1201                                                loff_t pos)
1202 {
1203         struct inode *inode = fdentry(file)->d_inode;
1204         struct btrfs_root *root = BTRFS_I(inode)->root;
1205         struct page **pages = NULL;
1206         unsigned long first_index;
1207         size_t num_written = 0;
1208         int nrptrs;
1209         int ret = 0;
1210         bool force_page_uptodate = false;
1211
1212         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1213                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1214                      (sizeof(struct page *)));
1215         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1216         nrptrs = max(nrptrs, 8);
1217         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1218         if (!pages)
1219                 return -ENOMEM;
1220
1221         first_index = pos >> PAGE_CACHE_SHIFT;
1222
1223         while (iov_iter_count(i) > 0) {
1224                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1225                 size_t write_bytes = min(iov_iter_count(i),
1226                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1227                                          offset);
1228                 size_t num_pages = (write_bytes + offset +
1229                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1230                 size_t dirty_pages;
1231                 size_t copied;
1232
1233                 WARN_ON(num_pages > nrptrs);
1234
1235                 /*
1236                  * Fault pages before locking them in prepare_pages
1237                  * to avoid recursive lock
1238                  */
1239                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1240                         ret = -EFAULT;
1241                         break;
1242                 }
1243
1244                 ret = btrfs_delalloc_reserve_space(inode,
1245                                         num_pages << PAGE_CACHE_SHIFT);
1246                 if (ret)
1247                         break;
1248
1249                 /*
1250                  * This is going to setup the pages array with the number of
1251                  * pages we want, so we don't really need to worry about the
1252                  * contents of pages from loop to loop
1253                  */
1254                 ret = prepare_pages(root, file, pages, num_pages,
1255                                     pos, first_index, write_bytes,
1256                                     force_page_uptodate);
1257                 if (ret) {
1258                         btrfs_delalloc_release_space(inode,
1259                                         num_pages << PAGE_CACHE_SHIFT);
1260                         break;
1261                 }
1262
1263                 copied = btrfs_copy_from_user(pos, num_pages,
1264                                            write_bytes, pages, i);
1265
1266                 /*
1267                  * if we have trouble faulting in the pages, fall
1268                  * back to one page at a time
1269                  */
1270                 if (copied < write_bytes)
1271                         nrptrs = 1;
1272
1273                 if (copied == 0) {
1274                         force_page_uptodate = true;
1275                         dirty_pages = 0;
1276                 } else {
1277                         force_page_uptodate = false;
1278                         dirty_pages = (copied + offset +
1279                                        PAGE_CACHE_SIZE - 1) >>
1280                                        PAGE_CACHE_SHIFT;
1281                 }
1282
1283                 /*
1284                  * If we had a short copy we need to release the excess delaloc
1285                  * bytes we reserved.  We need to increment outstanding_extents
1286                  * because btrfs_delalloc_release_space will decrement it, but
1287                  * we still have an outstanding extent for the chunk we actually
1288                  * managed to copy.
1289                  */
1290                 if (num_pages > dirty_pages) {
1291                         if (copied > 0) {
1292                                 spin_lock(&BTRFS_I(inode)->lock);
1293                                 BTRFS_I(inode)->outstanding_extents++;
1294                                 spin_unlock(&BTRFS_I(inode)->lock);
1295                         }
1296                         btrfs_delalloc_release_space(inode,
1297                                         (num_pages - dirty_pages) <<
1298                                         PAGE_CACHE_SHIFT);
1299                 }
1300
1301                 if (copied > 0) {
1302                         ret = btrfs_dirty_pages(root, inode, pages,
1303                                                 dirty_pages, pos, copied,
1304                                                 NULL);
1305                         if (ret) {
1306                                 btrfs_delalloc_release_space(inode,
1307                                         dirty_pages << PAGE_CACHE_SHIFT);
1308                                 btrfs_drop_pages(pages, num_pages);
1309                                 break;
1310                         }
1311                 }
1312
1313                 btrfs_drop_pages(pages, num_pages);
1314
1315                 cond_resched();
1316
1317                 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
1318                                                    dirty_pages);
1319                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1320                         btrfs_btree_balance_dirty(root, 1);
1321
1322                 pos += copied;
1323                 num_written += copied;
1324         }
1325
1326         kfree(pages);
1327
1328         return num_written ? num_written : ret;
1329 }
1330
1331 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1332                                     const struct iovec *iov,
1333                                     unsigned long nr_segs, loff_t pos,
1334                                     loff_t *ppos, size_t count, size_t ocount)
1335 {
1336         struct file *file = iocb->ki_filp;
1337         struct iov_iter i;
1338         ssize_t written;
1339         ssize_t written_buffered;
1340         loff_t endbyte;
1341         int err;
1342
1343         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1344                                             count, ocount);
1345
1346         if (written < 0 || written == count)
1347                 return written;
1348
1349         pos += written;
1350         count -= written;
1351         iov_iter_init(&i, iov, nr_segs, count, written);
1352         written_buffered = __btrfs_buffered_write(file, &i, pos);
1353         if (written_buffered < 0) {
1354                 err = written_buffered;
1355                 goto out;
1356         }
1357         endbyte = pos + written_buffered - 1;
1358         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1359         if (err)
1360                 goto out;
1361         written += written_buffered;
1362         *ppos = pos + written_buffered;
1363         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1364                                  endbyte >> PAGE_CACHE_SHIFT);
1365 out:
1366         return written ? written : err;
1367 }
1368
1369 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1370                                     const struct iovec *iov,
1371                                     unsigned long nr_segs, loff_t pos)
1372 {
1373         struct file *file = iocb->ki_filp;
1374         struct inode *inode = fdentry(file)->d_inode;
1375         struct btrfs_root *root = BTRFS_I(inode)->root;
1376         loff_t *ppos = &iocb->ki_pos;
1377         u64 start_pos;
1378         ssize_t num_written = 0;
1379         ssize_t err = 0;
1380         size_t count, ocount;
1381
1382         sb_start_write(inode->i_sb);
1383
1384         mutex_lock(&inode->i_mutex);
1385
1386         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1387         if (err) {
1388                 mutex_unlock(&inode->i_mutex);
1389                 goto out;
1390         }
1391         count = ocount;
1392
1393         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1394         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1395         if (err) {
1396                 mutex_unlock(&inode->i_mutex);
1397                 goto out;
1398         }
1399
1400         if (count == 0) {
1401                 mutex_unlock(&inode->i_mutex);
1402                 goto out;
1403         }
1404
1405         err = file_remove_suid(file);
1406         if (err) {
1407                 mutex_unlock(&inode->i_mutex);
1408                 goto out;
1409         }
1410
1411         /*
1412          * If BTRFS flips readonly due to some impossible error
1413          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1414          * although we have opened a file as writable, we have
1415          * to stop this write operation to ensure FS consistency.
1416          */
1417         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
1418                 mutex_unlock(&inode->i_mutex);
1419                 err = -EROFS;
1420                 goto out;
1421         }
1422
1423         err = file_update_time(file);
1424         if (err) {
1425                 mutex_unlock(&inode->i_mutex);
1426                 goto out;
1427         }
1428
1429         start_pos = round_down(pos, root->sectorsize);
1430         if (start_pos > i_size_read(inode)) {
1431                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1432                 if (err) {
1433                         mutex_unlock(&inode->i_mutex);
1434                         goto out;
1435                 }
1436         }
1437
1438         if (unlikely(file->f_flags & O_DIRECT)) {
1439                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1440                                                    pos, ppos, count, ocount);
1441         } else {
1442                 struct iov_iter i;
1443
1444                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1445
1446                 num_written = __btrfs_buffered_write(file, &i, pos);
1447                 if (num_written > 0)
1448                         *ppos = pos + num_written;
1449         }
1450
1451         mutex_unlock(&inode->i_mutex);
1452
1453         /*
1454          * we want to make sure fsync finds this change
1455          * but we haven't joined a transaction running right now.
1456          *
1457          * Later on, someone is sure to update the inode and get the
1458          * real transid recorded.
1459          *
1460          * We set last_trans now to the fs_info generation + 1,
1461          * this will either be one more than the running transaction
1462          * or the generation used for the next transaction if there isn't
1463          * one running right now.
1464          */
1465         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1466         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1467                 err = generic_write_sync(file, pos, num_written);
1468                 if (err < 0 && num_written > 0)
1469                         num_written = err;
1470         }
1471 out:
1472         sb_end_write(inode->i_sb);
1473         current->backing_dev_info = NULL;
1474         return num_written ? num_written : err;
1475 }
1476
1477 int btrfs_release_file(struct inode *inode, struct file *filp)
1478 {
1479         /*
1480          * ordered_data_close is set by settattr when we are about to truncate
1481          * a file from a non-zero size to a zero size.  This tries to
1482          * flush down new bytes that may have been written if the
1483          * application were using truncate to replace a file in place.
1484          */
1485         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1486                                &BTRFS_I(inode)->runtime_flags)) {
1487                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1488                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1489                         filemap_flush(inode->i_mapping);
1490         }
1491         if (filp->private_data)
1492                 btrfs_ioctl_trans_end(filp);
1493         return 0;
1494 }
1495
1496 /*
1497  * fsync call for both files and directories.  This logs the inode into
1498  * the tree log instead of forcing full commits whenever possible.
1499  *
1500  * It needs to call filemap_fdatawait so that all ordered extent updates are
1501  * in the metadata btree are up to date for copying to the log.
1502  *
1503  * It drops the inode mutex before doing the tree log commit.  This is an
1504  * important optimization for directories because holding the mutex prevents
1505  * new operations on the dir while we write to disk.
1506  */
1507 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1508 {
1509         struct dentry *dentry = file->f_path.dentry;
1510         struct inode *inode = dentry->d_inode;
1511         struct btrfs_root *root = BTRFS_I(inode)->root;
1512         int ret = 0;
1513         struct btrfs_trans_handle *trans;
1514
1515         trace_btrfs_sync_file(file, datasync);
1516
1517         mutex_lock(&inode->i_mutex);
1518
1519         /*
1520          * we wait first, since the writeback may change the inode, also wait
1521          * ordered range does a filemape_write_and_wait_range which is why we
1522          * don't do it above like other file systems.
1523          */
1524         root->log_batch++;
1525         btrfs_wait_ordered_range(inode, start, end);
1526         root->log_batch++;
1527
1528         /*
1529          * check the transaction that last modified this inode
1530          * and see if its already been committed
1531          */
1532         if (!BTRFS_I(inode)->last_trans) {
1533                 mutex_unlock(&inode->i_mutex);
1534                 goto out;
1535         }
1536
1537         /*
1538          * if the last transaction that changed this file was before
1539          * the current transaction, we can bail out now without any
1540          * syncing
1541          */
1542         smp_mb();
1543         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1544             BTRFS_I(inode)->last_trans <=
1545             root->fs_info->last_trans_committed) {
1546                 BTRFS_I(inode)->last_trans = 0;
1547                 mutex_unlock(&inode->i_mutex);
1548                 goto out;
1549         }
1550
1551         /*
1552          * ok we haven't committed the transaction yet, lets do a commit
1553          */
1554         if (file->private_data)
1555                 btrfs_ioctl_trans_end(file);
1556
1557         trans = btrfs_start_transaction(root, 0);
1558         if (IS_ERR(trans)) {
1559                 ret = PTR_ERR(trans);
1560                 mutex_unlock(&inode->i_mutex);
1561                 goto out;
1562         }
1563
1564         ret = btrfs_log_dentry_safe(trans, root, dentry);
1565         if (ret < 0) {
1566                 mutex_unlock(&inode->i_mutex);
1567                 goto out;
1568         }
1569
1570         /* we've logged all the items and now have a consistent
1571          * version of the file in the log.  It is possible that
1572          * someone will come in and modify the file, but that's
1573          * fine because the log is consistent on disk, and we
1574          * have references to all of the file's extents
1575          *
1576          * It is possible that someone will come in and log the
1577          * file again, but that will end up using the synchronization
1578          * inside btrfs_sync_log to keep things safe.
1579          */
1580         mutex_unlock(&inode->i_mutex);
1581
1582         if (ret != BTRFS_NO_LOG_SYNC) {
1583                 if (ret > 0) {
1584                         ret = btrfs_commit_transaction(trans, root);
1585                 } else {
1586                         ret = btrfs_sync_log(trans, root);
1587                         if (ret == 0)
1588                                 ret = btrfs_end_transaction(trans, root);
1589                         else
1590                                 ret = btrfs_commit_transaction(trans, root);
1591                 }
1592         } else {
1593                 ret = btrfs_end_transaction(trans, root);
1594         }
1595 out:
1596         return ret > 0 ? -EIO : ret;
1597 }
1598
1599 static const struct vm_operations_struct btrfs_file_vm_ops = {
1600         .fault          = filemap_fault,
1601         .page_mkwrite   = btrfs_page_mkwrite,
1602 };
1603
1604 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1605 {
1606         struct address_space *mapping = filp->f_mapping;
1607
1608         if (!mapping->a_ops->readpage)
1609                 return -ENOEXEC;
1610
1611         file_accessed(filp);
1612         vma->vm_ops = &btrfs_file_vm_ops;
1613         vma->vm_flags |= VM_CAN_NONLINEAR;
1614
1615         return 0;
1616 }
1617
1618 static long btrfs_fallocate(struct file *file, int mode,
1619                             loff_t offset, loff_t len)
1620 {
1621         struct inode *inode = file->f_path.dentry->d_inode;
1622         struct extent_state *cached_state = NULL;
1623         u64 cur_offset;
1624         u64 last_byte;
1625         u64 alloc_start;
1626         u64 alloc_end;
1627         u64 alloc_hint = 0;
1628         u64 locked_end;
1629         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1630         struct extent_map *em;
1631         int ret;
1632
1633         alloc_start = offset & ~mask;
1634         alloc_end =  (offset + len + mask) & ~mask;
1635
1636         /* We only support the FALLOC_FL_KEEP_SIZE mode */
1637         if (mode & ~FALLOC_FL_KEEP_SIZE)
1638                 return -EOPNOTSUPP;
1639
1640         /*
1641          * Make sure we have enough space before we do the
1642          * allocation.
1643          */
1644         ret = btrfs_check_data_free_space(inode, len);
1645         if (ret)
1646                 return ret;
1647
1648         /*
1649          * wait for ordered IO before we have any locks.  We'll loop again
1650          * below with the locks held.
1651          */
1652         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
1653
1654         mutex_lock(&inode->i_mutex);
1655         ret = inode_newsize_ok(inode, alloc_end);
1656         if (ret)
1657                 goto out;
1658
1659         if (alloc_start > inode->i_size) {
1660                 ret = btrfs_cont_expand(inode, i_size_read(inode),
1661                                         alloc_start);
1662                 if (ret)
1663                         goto out;
1664         }
1665
1666         locked_end = alloc_end - 1;
1667         while (1) {
1668                 struct btrfs_ordered_extent *ordered;
1669
1670                 /* the extent lock is ordered inside the running
1671                  * transaction
1672                  */
1673                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
1674                                  locked_end, 0, &cached_state);
1675                 ordered = btrfs_lookup_first_ordered_extent(inode,
1676                                                             alloc_end - 1);
1677                 if (ordered &&
1678                     ordered->file_offset + ordered->len > alloc_start &&
1679                     ordered->file_offset < alloc_end) {
1680                         btrfs_put_ordered_extent(ordered);
1681                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1682                                              alloc_start, locked_end,
1683                                              &cached_state, GFP_NOFS);
1684                         /*
1685                          * we can't wait on the range with the transaction
1686                          * running or with the extent lock held
1687                          */
1688                         btrfs_wait_ordered_range(inode, alloc_start,
1689                                                  alloc_end - alloc_start);
1690                 } else {
1691                         if (ordered)
1692                                 btrfs_put_ordered_extent(ordered);
1693                         break;
1694                 }
1695         }
1696
1697         cur_offset = alloc_start;
1698         while (1) {
1699                 u64 actual_end;
1700
1701                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
1702                                       alloc_end - cur_offset, 0);
1703                 if (IS_ERR_OR_NULL(em)) {
1704                         if (!em)
1705                                 ret = -ENOMEM;
1706                         else
1707                                 ret = PTR_ERR(em);
1708                         break;
1709                 }
1710                 last_byte = min(extent_map_end(em), alloc_end);
1711                 actual_end = min_t(u64, extent_map_end(em), offset + len);
1712                 last_byte = (last_byte + mask) & ~mask;
1713
1714                 if (em->block_start == EXTENT_MAP_HOLE ||
1715                     (cur_offset >= inode->i_size &&
1716                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
1717                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
1718                                                         last_byte - cur_offset,
1719                                                         1 << inode->i_blkbits,
1720                                                         offset + len,
1721                                                         &alloc_hint);
1722
1723                         if (ret < 0) {
1724                                 free_extent_map(em);
1725                                 break;
1726                         }
1727                 } else if (actual_end > inode->i_size &&
1728                            !(mode & FALLOC_FL_KEEP_SIZE)) {
1729                         /*
1730                          * We didn't need to allocate any more space, but we
1731                          * still extended the size of the file so we need to
1732                          * update i_size.
1733                          */
1734                         inode->i_ctime = CURRENT_TIME;
1735                         i_size_write(inode, actual_end);
1736                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
1737                 }
1738                 free_extent_map(em);
1739
1740                 cur_offset = last_byte;
1741                 if (cur_offset >= alloc_end) {
1742                         ret = 0;
1743                         break;
1744                 }
1745         }
1746         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
1747                              &cached_state, GFP_NOFS);
1748 out:
1749         mutex_unlock(&inode->i_mutex);
1750         /* Let go of our reservation. */
1751         btrfs_free_reserved_data_space(inode, len);
1752         return ret;
1753 }
1754
1755 static int find_desired_extent(struct inode *inode, loff_t *offset, int origin)
1756 {
1757         struct btrfs_root *root = BTRFS_I(inode)->root;
1758         struct extent_map *em;
1759         struct extent_state *cached_state = NULL;
1760         u64 lockstart = *offset;
1761         u64 lockend = i_size_read(inode);
1762         u64 start = *offset;
1763         u64 orig_start = *offset;
1764         u64 len = i_size_read(inode);
1765         u64 last_end = 0;
1766         int ret = 0;
1767
1768         lockend = max_t(u64, root->sectorsize, lockend);
1769         if (lockend <= lockstart)
1770                 lockend = lockstart + root->sectorsize;
1771
1772         len = lockend - lockstart + 1;
1773
1774         len = max_t(u64, len, root->sectorsize);
1775         if (inode->i_size == 0)
1776                 return -ENXIO;
1777
1778         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
1779                          &cached_state);
1780
1781         /*
1782          * Delalloc is such a pain.  If we have a hole and we have pending
1783          * delalloc for a portion of the hole we will get back a hole that
1784          * exists for the entire range since it hasn't been actually written
1785          * yet.  So to take care of this case we need to look for an extent just
1786          * before the position we want in case there is outstanding delalloc
1787          * going on here.
1788          */
1789         if (origin == SEEK_HOLE && start != 0) {
1790                 if (start <= root->sectorsize)
1791                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
1792                                                      root->sectorsize, 0);
1793                 else
1794                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
1795                                                      start - root->sectorsize,
1796                                                      root->sectorsize, 0);
1797                 if (IS_ERR(em)) {
1798                         ret = PTR_ERR(em);
1799                         goto out;
1800                 }
1801                 last_end = em->start + em->len;
1802                 if (em->block_start == EXTENT_MAP_DELALLOC)
1803                         last_end = min_t(u64, last_end, inode->i_size);
1804                 free_extent_map(em);
1805         }
1806
1807         while (1) {
1808                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
1809                 if (IS_ERR(em)) {
1810                         ret = PTR_ERR(em);
1811                         break;
1812                 }
1813
1814                 if (em->block_start == EXTENT_MAP_HOLE) {
1815                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1816                                 if (last_end <= orig_start) {
1817                                         free_extent_map(em);
1818                                         ret = -ENXIO;
1819                                         break;
1820                                 }
1821                         }
1822
1823                         if (origin == SEEK_HOLE) {
1824                                 *offset = start;
1825                                 free_extent_map(em);
1826                                 break;
1827                         }
1828                 } else {
1829                         if (origin == SEEK_DATA) {
1830                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
1831                                         if (start >= inode->i_size) {
1832                                                 free_extent_map(em);
1833                                                 ret = -ENXIO;
1834                                                 break;
1835                                         }
1836                                 }
1837
1838                                 *offset = start;
1839                                 free_extent_map(em);
1840                                 break;
1841                         }
1842                 }
1843
1844                 start = em->start + em->len;
1845                 last_end = em->start + em->len;
1846
1847                 if (em->block_start == EXTENT_MAP_DELALLOC)
1848                         last_end = min_t(u64, last_end, inode->i_size);
1849
1850                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1851                         free_extent_map(em);
1852                         ret = -ENXIO;
1853                         break;
1854                 }
1855                 free_extent_map(em);
1856                 cond_resched();
1857         }
1858         if (!ret)
1859                 *offset = min(*offset, inode->i_size);
1860 out:
1861         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
1862                              &cached_state, GFP_NOFS);
1863         return ret;
1864 }
1865
1866 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int origin)
1867 {
1868         struct inode *inode = file->f_mapping->host;
1869         int ret;
1870
1871         mutex_lock(&inode->i_mutex);
1872         switch (origin) {
1873         case SEEK_END:
1874         case SEEK_CUR:
1875                 offset = generic_file_llseek(file, offset, origin);
1876                 goto out;
1877         case SEEK_DATA:
1878         case SEEK_HOLE:
1879                 if (offset >= i_size_read(inode)) {
1880                         mutex_unlock(&inode->i_mutex);
1881                         return -ENXIO;
1882                 }
1883
1884                 ret = find_desired_extent(inode, &offset, origin);
1885                 if (ret) {
1886                         mutex_unlock(&inode->i_mutex);
1887                         return ret;
1888                 }
1889         }
1890
1891         if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
1892                 offset = -EINVAL;
1893                 goto out;
1894         }
1895         if (offset > inode->i_sb->s_maxbytes) {
1896                 offset = -EINVAL;
1897                 goto out;
1898         }
1899
1900         /* Special lock needed here? */
1901         if (offset != file->f_pos) {
1902                 file->f_pos = offset;
1903                 file->f_version = 0;
1904         }
1905 out:
1906         mutex_unlock(&inode->i_mutex);
1907         return offset;
1908 }
1909
1910 const struct file_operations btrfs_file_operations = {
1911         .llseek         = btrfs_file_llseek,
1912         .read           = do_sync_read,
1913         .write          = do_sync_write,
1914         .aio_read       = generic_file_aio_read,
1915         .splice_read    = generic_file_splice_read,
1916         .aio_write      = btrfs_file_aio_write,
1917         .mmap           = btrfs_file_mmap,
1918         .open           = generic_file_open,
1919         .release        = btrfs_release_file,
1920         .fsync          = btrfs_sync_file,
1921         .fallocate      = btrfs_fallocate,
1922         .unlocked_ioctl = btrfs_ioctl,
1923 #ifdef CONFIG_COMPAT
1924         .compat_ioctl   = btrfs_ioctl,
1925 #endif
1926 };