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 <linux/btrfs.h>
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.h"
  38 #include "print-tree.h"
  39 #include "tree-log.h"
  40 #include "locking.h"
  41 #include "compat.h"
  42 #include "volumes.h"
  43
  44 static struct kmem_cache *btrfs_inode_defrag_cachep;
  45 /*
  46  * when auto defrag is enabled we
  47  * queue up these defrag structs to remember which
  48  * inodes need defragging passes
  49  */
  50 struct inode_defrag {
  51         struct rb_node rb_node;
  52         /* objectid */
  53         u64 ino;
  54         /*
  55          * transid where the defrag was added, we search for
  56          * extents newer than this
  57          */
  58         u64 transid;
  59
  60         /* root objectid */
  61         u64 root;
  62
  63         /* last offset we were able to defrag */
  64         u64 last_offset;
  65
  66         /* if we've wrapped around back to zero once already */
  67         int cycled;
  68 };
  69
  70 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  71                                   struct inode_defrag *defrag2)
  72 {
  73         if (defrag1->root > defrag2->root)
  74                 return 1;
  75         else if (defrag1->root < defrag2->root)
  76                 return -1;
  77         else if (defrag1->ino > defrag2->ino)
  78                 return 1;
  79         else if (defrag1->ino < defrag2->ino)
  80                 return -1;
  81         else
  82                 return 0;
  83 }
  84
  85 /* pop a record for an inode into the defrag tree.  The lock
  86  * must be held already
  87  *
  88  * If you're inserting a record for an older transid than an
  89  * existing record, the transid already in the tree is lowered
  90  *
  91  * If an existing record is found the defrag item you
  92  * pass in is freed
  93  */
  94 static int __btrfs_add_inode_defrag(struct inode *inode,
  95                                     struct inode_defrag *defrag)
  96 {
  97         struct btrfs_root *root = BTRFS_I(inode)->root;
  98         struct inode_defrag *entry;
  99         struct rb_node **p;
 100         struct rb_node *parent = NULL;
 101         int ret;
 102
 103         p = &root->fs_info->defrag_inodes.rb_node;
 104         while (*p) {
 105                 parent = *p;
 106                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 107
 108                 ret = __compare_inode_defrag(defrag, entry);
 109                 if (ret < 0)
 110                         p = &parent->rb_left;
 111                 else if (ret > 0)
 112                         p = &parent->rb_right;
 113                 else {
 114                         /* if we're reinserting an entry for
 115                          * an old defrag run, make sure to
 116                          * lower the transid of our existing record
 117                          */
 118                         if (defrag->transid < entry->transid)
 119                                 entry->transid = defrag->transid;
 120                         if (defrag->last_offset > entry->last_offset)
 121                                 entry->last_offset = defrag->last_offset;
 122                         return -EEXIST;
 123                 }
 124         }
 125         set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 126         rb_link_node(&defrag->rb_node, parent, p);
 127         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 128         return 0;
 129 }
 130
 131 static inline int __need_auto_defrag(struct btrfs_root *root)
 132 {
 133         if (!btrfs_test_opt(root, AUTO_DEFRAG))
 134                 return 0;
 135
 136         if (btrfs_fs_closing(root->fs_info))
 137                 return 0;
 138
 139         return 1;
 140 }
 141
 142 /*
 143  * insert a defrag record for this inode if auto defrag is
 144  * enabled
 145  */
 146 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 147                            struct inode *inode)
 148 {
 149         struct btrfs_root *root = BTRFS_I(inode)->root;
 150         struct inode_defrag *defrag;
 151         u64 transid;
 152         int ret;
 153
 154         if (!__need_auto_defrag(root))
 155                 return 0;
 156
 157         if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 158                 return 0;
 159
 160         if (trans)
 161                 transid = trans->transid;
 162         else
 163                 transid = BTRFS_I(inode)->root->last_trans;
 164
 165         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 166         if (!defrag)
 167                 return -ENOMEM;
 168
 169         defrag->ino = btrfs_ino(inode);
 170         defrag->transid = transid;
 171         defrag->root = root->root_key.objectid;
 172
 173         spin_lock(&root->fs_info->defrag_inodes_lock);
 174         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
 175                 /*
 176                  * If we set IN_DEFRAG flag and evict the inode from memory,
 177                  * and then re-read this inode, this new inode doesn't have
 178                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 179                  */
 180                 ret = __btrfs_add_inode_defrag(inode, defrag);
 181                 if (ret)
 182                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 183         } else {
 184                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 185         }
 186         spin_unlock(&root->fs_info->defrag_inodes_lock);
 187         return 0;
 188 }
 189
 190 /*
 191  * Requeue the defrag object. If there is a defrag object that points to
 192  * the same inode in the tree, we will merge them together (by
 193  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 194  */
 195 void btrfs_requeue_inode_defrag(struct inode *inode,
 196                                 struct inode_defrag *defrag)
 197 {
 198         struct btrfs_root *root = BTRFS_I(inode)->root;
 199         int ret;
 200
 201         if (!__need_auto_defrag(root))
 202                 goto out;
 203
 204         /*
 205          * Here we don't check the IN_DEFRAG flag, because we need merge
 206          * them together.
 207          */
 208         spin_lock(&root->fs_info->defrag_inodes_lock);
 209         ret = __btrfs_add_inode_defrag(inode, defrag);
 210         spin_unlock(&root->fs_info->defrag_inodes_lock);
 211         if (ret)
 212                 goto out;
 213         return;
 214 out:
 215         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 216 }
 217
 218 /*
 219  * pick the defragable inode that we want, if it doesn't exist, we will get
 220  * the next one.
 221  */
 222 static struct inode_defrag *
 223 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 224 {
 225         struct inode_defrag *entry = NULL;
 226         struct inode_defrag tmp;
 227         struct rb_node *p;
 228         struct rb_node *parent = NULL;
 229         int ret;
 230
 231         tmp.ino = ino;
 232         tmp.root = root;
 233
 234         spin_lock(&fs_info->defrag_inodes_lock);
 235         p = fs_info->defrag_inodes.rb_node;
 236         while (p) {
 237                 parent = p;
 238                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 239
 240                 ret = __compare_inode_defrag(&tmp, entry);
 241                 if (ret < 0)
 242                         p = parent->rb_left;
 243                 else if (ret > 0)
 244                         p = parent->rb_right;
 245                 else
 246                         goto out;
 247         }
 248
 249         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 250                 parent = rb_next(parent);
 251                 if (parent)
 252                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 253                 else
 254                         entry = NULL;
 255         }
 256 out:
 257         if (entry)
 258                 rb_erase(parent, &fs_info->defrag_inodes);
 259         spin_unlock(&fs_info->defrag_inodes_lock);
 260         return entry;
 261 }
 262
 263 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 264 {
 265         struct inode_defrag *defrag;
 266         struct rb_node *node;
 267
 268         spin_lock(&fs_info->defrag_inodes_lock);
 269         node = rb_first(&fs_info->defrag_inodes);
 270         while (node) {
 271                 rb_erase(node, &fs_info->defrag_inodes);
 272                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 273                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 274
 275                 if (need_resched()) {
 276                         spin_unlock(&fs_info->defrag_inodes_lock);
 277                         cond_resched();
 278                         spin_lock(&fs_info->defrag_inodes_lock);
 279                 }
 280
 281                 node = rb_first(&fs_info->defrag_inodes);
 282         }
 283         spin_unlock(&fs_info->defrag_inodes_lock);
 284 }
 285
 286 #define BTRFS_DEFRAG_BATCH      1024
 287
 288 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 289                                     struct inode_defrag *defrag)
 290 {
 291         struct btrfs_root *inode_root;
 292         struct inode *inode;
 293         struct btrfs_key key;
 294         struct btrfs_ioctl_defrag_range_args range;
 295         int num_defrag;
 296         int index;
 297         int ret;
 298
 299         /* get the inode */
 300         key.objectid = defrag->root;
 301         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 302         key.offset = (u64)-1;
 303
 304         index = srcu_read_lock(&fs_info->subvol_srcu);
 305
 306         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 307         if (IS_ERR(inode_root)) {
 308                 ret = PTR_ERR(inode_root);
 309                 goto cleanup;
 310         }
 311         if (btrfs_root_refs(&inode_root->root_item) == 0) {
 312                 ret = -ENOENT;
 313                 goto cleanup;
 314         }
 315
 316         key.objectid = defrag->ino;
 317         btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 318         key.offset = 0;
 319         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 320         if (IS_ERR(inode)) {
 321                 ret = PTR_ERR(inode);
 322                 goto cleanup;
 323         }
 324         srcu_read_unlock(&fs_info->subvol_srcu, index);
 325
 326         /* do a chunk of defrag */
 327         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 328         memset(&range, 0, sizeof(range));
 329         range.len = (u64)-1;
 330         range.start = defrag->last_offset;
 331
 332         sb_start_write(fs_info->sb);
 333         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 334                                        BTRFS_DEFRAG_BATCH);
 335         sb_end_write(fs_info->sb);
 336         /*
 337          * if we filled the whole defrag batch, there
 338          * must be more work to do.  Queue this defrag
 339          * again
 340          */
 341         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 342                 defrag->last_offset = range.start;
 343                 btrfs_requeue_inode_defrag(inode, defrag);
 344         } else if (defrag->last_offset && !defrag->cycled) {
 345                 /*
 346                  * we didn't fill our defrag batch, but
 347                  * we didn't start at zero.  Make sure we loop
 348                  * around to the start of the file.
 349                  */
 350                 defrag->last_offset = 0;
 351                 defrag->cycled = 1;
 352                 btrfs_requeue_inode_defrag(inode, defrag);
 353         } else {
 354                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 355         }
 356
 357         iput(inode);
 358         return 0;
 359 cleanup:
 360         srcu_read_unlock(&fs_info->subvol_srcu, index);
 361         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 362         return ret;
 363 }
 364
 365 /*
 366  * run through the list of inodes in the FS that need
 367  * defragging
 368  */
 369 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 370 {
 371         struct inode_defrag *defrag;
 372         u64 first_ino = 0;
 373         u64 root_objectid = 0;
 374
 375         atomic_inc(&fs_info->defrag_running);
 376         while(1) {
 377                 /* Pause the auto defragger. */
 378                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 379                              &fs_info->fs_state))
 380                         break;
 381
 382                 if (!__need_auto_defrag(fs_info->tree_root))
 383                         break;
 384
 385                 /* find an inode to defrag */
 386                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 387                                                  first_ino);
 388                 if (!defrag) {
 389                         if (root_objectid || first_ino) {
 390                                 root_objectid = 0;
 391                                 first_ino = 0;
 392                                 continue;
 393                         } else {
 394                                 break;
 395                         }
 396                 }
 397
 398                 first_ino = defrag->ino + 1;
 399                 root_objectid = defrag->root;
 400
 401                 __btrfs_run_defrag_inode(fs_info, defrag);
 402         }
 403         atomic_dec(&fs_info->defrag_running);
 404
 405         /*
 406          * during unmount, we use the transaction_wait queue to
 407          * wait for the defragger to stop
 408          */
 409         wake_up(&fs_info->transaction_wait);
 410         return 0;
 411 }
 412
 413 /* simple helper to fault in pages and copy.  This should go away
 414  * and be replaced with calls into generic code.
 415  */
 416 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 417                                          size_t write_bytes,
 418                                          struct page **prepared_pages,
 419                                          struct iov_iter *i)
 420 {
 421         size_t copied = 0;
 422         size_t total_copied = 0;
 423         int pg = 0;
 424         int offset = pos & (PAGE_CACHE_SIZE - 1);
 425
 426         while (write_bytes > 0) {
 427                 size_t count = min_t(size_t,
 428                                      PAGE_CACHE_SIZE - offset, write_bytes);
 429                 struct page *page = prepared_pages[pg];
 430                 /*
 431                  * Copy data from userspace to the current page
 432                  *
 433                  * Disable pagefault to avoid recursive lock since
 434                  * the pages are already locked
 435                  */
 436                 pagefault_disable();
 437                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 438                 pagefault_enable();
 439
 440                 /* Flush processor's dcache for this page */
 441                 flush_dcache_page(page);
 442
 443                 /*
 444                  * if we get a partial write, we can end up with
 445                  * partially up to date pages.  These add
 446                  * a lot of complexity, so make sure they don't
 447                  * happen by forcing this copy to be retried.
 448                  *
 449                  * The rest of the btrfs_file_write code will fall
 450                  * back to page at a time copies after we return 0.
 451                  */
 452                 if (!PageUptodate(page) && copied < count)
 453                         copied = 0;
 454
 455                 iov_iter_advance(i, copied);
 456                 write_bytes -= copied;
 457                 total_copied += copied;
 458
 459                 /* Return to btrfs_file_aio_write to fault page */
 460                 if (unlikely(copied == 0))
 461                         break;
 462
 463                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 464                         offset += copied;
 465                 } else {
 466                         pg++;
 467                         offset = 0;
 468                 }
 469         }
 470         return total_copied;
 471 }
 472
 473 /*
 474  * unlocks pages after btrfs_file_write is done with them
 475  */
 476 void btrfs_drop_pages(struct page **pages, size_t num_pages)
 477 {
 478         size_t i;
 479         for (i = 0; i < num_pages; i++) {
 480                 /* page checked is some magic around finding pages that
 481                  * have been modified without going through btrfs_set_page_dirty
 482                  * clear it here
 483                  */
 484                 ClearPageChecked(pages[i]);
 485                 unlock_page(pages[i]);
 486                 mark_page_accessed(pages[i]);
 487                 page_cache_release(pages[i]);
 488         }
 489 }
 490
 491 /*
 492  * after copy_from_user, pages need to be dirtied and we need to make
 493  * sure holes are created between the current EOF and the start of
 494  * any next extents (if required).
 495  *
 496  * this also makes the decision about creating an inline extent vs
 497  * doing real data extents, marking pages dirty and delalloc as required.
 498  */
 499 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 500                       struct page **pages, size_t num_pages,
 501                       loff_t pos, size_t write_bytes,
 502                       struct extent_state **cached)
 503 {
 504         int err = 0;
 505         int i;
 506         u64 num_bytes;
 507         u64 start_pos;
 508         u64 end_of_last_block;
 509         u64 end_pos = pos + write_bytes;
 510         loff_t isize = i_size_read(inode);
 511
 512         start_pos = pos & ~((u64)root->sectorsize - 1);
 513         num_bytes = ALIGN(write_bytes + pos - start_pos, root->sectorsize);
 514
 515         end_of_last_block = start_pos + num_bytes - 1;
 516         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 517                                         cached);
 518         if (err)
 519                 return err;
 520
 521         for (i = 0; i < num_pages; i++) {
 522                 struct page *p = pages[i];
 523                 SetPageUptodate(p);
 524                 ClearPageChecked(p);
 525                 set_page_dirty(p);
 526         }
 527
 528         /*
 529          * we've only changed i_size in ram, and we haven't updated
 530          * the disk i_size.  There is no need to log the inode
 531          * at this time.
 532          */
 533         if (end_pos > isize)
 534                 i_size_write(inode, end_pos);
 535         return 0;
 536 }
 537
 538 /*
 539  * this drops all the extents in the cache that intersect the range
 540  * [start, end].  Existing extents are split as required.
 541  */
 542 void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 543                              int skip_pinned)
 544 {
 545         struct extent_map *em;
 546         struct extent_map *split = NULL;
 547         struct extent_map *split2 = NULL;
 548         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 549         u64 len = end - start + 1;
 550         u64 gen;
 551         int ret;
 552         int testend = 1;
 553         unsigned long flags;
 554         int compressed = 0;
 555
 556         WARN_ON(end < start);
 557         if (end == (u64)-1) {
 558                 len = (u64)-1;
 559                 testend = 0;
 560         }
 561         while (1) {
 562                 int no_splits = 0;
 563
 564                 if (!split)
 565                         split = alloc_extent_map();
 566                 if (!split2)
 567                         split2 = alloc_extent_map();
 568                 if (!split || !split2)
 569                         no_splits = 1;
 570
 571                 write_lock(&em_tree->lock);
 572                 em = lookup_extent_mapping(em_tree, start, len);
 573                 if (!em) {
 574                         write_unlock(&em_tree->lock);
 575                         break;
 576                 }
 577                 flags = em->flags;
 578                 gen = em->generation;
 579                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 580                         if (testend && em->start + em->len >= start + len) {
 581                                 free_extent_map(em);
 582                                 write_unlock(&em_tree->lock);
 583                                 break;
 584                         }
 585                         start = em->start + em->len;
 586                         if (testend)
 587                                 len = start + len - (em->start + em->len);
 588                         free_extent_map(em);
 589                         write_unlock(&em_tree->lock);
 590                         continue;
 591                 }
 592                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 593                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 594                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 595                 remove_extent_mapping(em_tree, em);
 596                 if (no_splits)
 597                         goto next;
 598
 599                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 600                     em->start < start) {
 601                         split->start = em->start;
 602                         split->len = start - em->start;
 603                         split->orig_start = em->orig_start;
 604                         split->block_start = em->block_start;
 605
 606                         if (compressed)
 607                                 split->block_len = em->block_len;
 608                         else
 609                                 split->block_len = split->len;
 610                         split->orig_block_len = max(split->block_len,
 611                                                     em->orig_block_len);
 612                         split->generation = gen;
 613                         split->bdev = em->bdev;
 614                         split->flags = flags;
 615                         split->compress_type = em->compress_type;
 616                         ret = add_extent_mapping(em_tree, split);
 617                         BUG_ON(ret); /* Logic error */
 618                         list_move(&split->list, &em_tree->modified_extents);
 619                         free_extent_map(split);
 620                         split = split2;
 621                         split2 = NULL;
 622                 }
 623                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 624                     testend && em->start + em->len > start + len) {
 625                         u64 diff = start + len - em->start;
 626
 627                         split->start = start + len;
 628                         split->len = em->start + em->len - (start + len);
 629                         split->bdev = em->bdev;
 630                         split->flags = flags;
 631                         split->compress_type = em->compress_type;
 632                         split->generation = gen;
 633                         split->orig_block_len = max(em->block_len,
 634                                                     em->orig_block_len);
 635
 636                         if (compressed) {
 637                                 split->block_len = em->block_len;
 638                                 split->block_start = em->block_start;
 639                                 split->orig_start = em->orig_start;
 640                         } else {
 641                                 split->block_len = split->len;
 642                                 split->block_start = em->block_start + diff;
 643                                 split->orig_start = em->orig_start;
 644                         }
 645
 646                         ret = add_extent_mapping(em_tree, split);
 647                         BUG_ON(ret); /* Logic error */
 648                         list_move(&split->list, &em_tree->modified_extents);
 649                         free_extent_map(split);
 650                         split = NULL;
 651                 }
 652 next:
 653                 write_unlock(&em_tree->lock);
 654
 655                 /* once for us */
 656                 free_extent_map(em);
 657                 /* once for the tree*/
 658                 free_extent_map(em);
 659         }
 660         if (split)
 661                 free_extent_map(split);
 662         if (split2)
 663                 free_extent_map(split2);
 664 }
 665
 666 /*
 667  * this is very complex, but the basic idea is to drop all extents
 668  * in the range start - end.  hint_block is filled in with a block number
 669  * that would be a good hint to the block allocator for this file.
 670  *
 671  * If an extent intersects the range but is not entirely inside the range
 672  * it is either truncated or split.  Anything entirely inside the range
 673  * is deleted from the tree.
 674  */
 675 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 676                          struct btrfs_root *root, struct inode *inode,
 677                          struct btrfs_path *path, u64 start, u64 end,
 678                          u64 *drop_end, int drop_cache)
 679 {
 680         struct extent_buffer *leaf;
 681         struct btrfs_file_extent_item *fi;
 682         struct btrfs_key key;
 683         struct btrfs_key new_key;
 684         u64 ino = btrfs_ino(inode);
 685         u64 search_start = start;
 686         u64 disk_bytenr = 0;
 687         u64 num_bytes = 0;
 688         u64 extent_offset = 0;
 689         u64 extent_end = 0;
 690         int del_nr = 0;
 691         int del_slot = 0;
 692         int extent_type;
 693         int recow;
 694         int ret;
 695         int modify_tree = -1;
 696         int update_refs = (root->ref_cows || root == root->fs_info->tree_root);
 697         int found = 0;
 698
 699         if (drop_cache)
 700                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 701
 702         if (start >= BTRFS_I(inode)->disk_i_size)
 703                 modify_tree = 0;
 704
 705         while (1) {
 706                 recow = 0;
 707                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 708                                                search_start, modify_tree);
 709                 if (ret < 0)
 710                         break;
 711                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 712                         leaf = path->nodes[0];
 713                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 714                         if (key.objectid == ino &&
 715                             key.type == BTRFS_EXTENT_DATA_KEY)
 716                                 path->slots[0]--;
 717                 }
 718                 ret = 0;
 719 next_slot:
 720                 leaf = path->nodes[0];
 721                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 722                         BUG_ON(del_nr > 0);
 723                         ret = btrfs_next_leaf(root, path);
 724                         if (ret < 0)
 725                                 break;
 726                         if (ret > 0) {
 727                                 ret = 0;
 728                                 break;
 729                         }
 730                         leaf = path->nodes[0];
 731                         recow = 1;
 732                 }
 733
 734                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 735                 if (key.objectid > ino ||
 736                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 737                         break;
 738
 739                 fi = btrfs_item_ptr(leaf, path->slots[0],
 740                                     struct btrfs_file_extent_item);
 741                 extent_type = btrfs_file_extent_type(leaf, fi);
 742
 743                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 744                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 745                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 746                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 747                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 748                         extent_end = key.offset +
 749                                 btrfs_file_extent_num_bytes(leaf, fi);
 750                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 751                         extent_end = key.offset +
 752                                 btrfs_file_extent_inline_len(leaf, fi);
 753                 } else {
 754                         WARN_ON(1);
 755                         extent_end = search_start;
 756                 }
 757
 758                 if (extent_end <= search_start) {
 759                         path->slots[0]++;
 760                         goto next_slot;
 761                 }
 762
 763                 found = 1;
 764                 search_start = max(key.offset, start);
 765                 if (recow || !modify_tree) {
 766                         modify_tree = -1;
 767                         btrfs_release_path(path);
 768                         continue;
 769                 }
 770
 771                 /*
 772                  *     | - range to drop - |
 773                  *  | -------- extent -------- |
 774                  */
 775                 if (start > key.offset && end < extent_end) {
 776                         BUG_ON(del_nr > 0);
 777                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 778
 779                         memcpy(&new_key, &key, sizeof(new_key));
 780                         new_key.offset = start;
 781                         ret = btrfs_duplicate_item(trans, root, path,
 782                                                    &new_key);
 783                         if (ret == -EAGAIN) {
 784                                 btrfs_release_path(path);
 785                                 continue;
 786                         }
 787                         if (ret < 0)
 788                                 break;
 789
 790                         leaf = path->nodes[0];
 791                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 792                                             struct btrfs_file_extent_item);
 793                         btrfs_set_file_extent_num_bytes(leaf, fi,
 794                                                         start - key.offset);
 795
 796                         fi = btrfs_item_ptr(leaf, path->slots[0],
 797                                             struct btrfs_file_extent_item);
 798
 799                         extent_offset += start - key.offset;
 800                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 801                         btrfs_set_file_extent_num_bytes(leaf, fi,
 802                                                         extent_end - start);
 803                         btrfs_mark_buffer_dirty(leaf);
 804
 805                         if (update_refs && disk_bytenr > 0) {
 806                                 ret = btrfs_inc_extent_ref(trans, root,
 807                                                 disk_bytenr, num_bytes, 0,
 808                                                 root->root_key.objectid,
 809                                                 new_key.objectid,
 810                                                 start - extent_offset, 0);
 811                                 BUG_ON(ret); /* -ENOMEM */
 812                         }
 813                         key.offset = start;
 814                 }
 815                 /*
 816                  *  | ---- range to drop ----- |
 817                  *      | -------- extent -------- |
 818                  */
 819                 if (start <= key.offset && end < extent_end) {
 820                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 821
 822                         memcpy(&new_key, &key, sizeof(new_key));
 823                         new_key.offset = end;
 824                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 825
 826                         extent_offset += end - key.offset;
 827                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 828                         btrfs_set_file_extent_num_bytes(leaf, fi,
 829                                                         extent_end - end);
 830                         btrfs_mark_buffer_dirty(leaf);
 831                         if (update_refs && disk_bytenr > 0)
 832                                 inode_sub_bytes(inode, end - key.offset);
 833                         break;
 834                 }
 835
 836                 search_start = extent_end;
 837                 /*
 838                  *       | ---- range to drop ----- |
 839                  *  | -------- extent -------- |
 840                  */
 841                 if (start > key.offset && end >= extent_end) {
 842                         BUG_ON(del_nr > 0);
 843                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 844
 845                         btrfs_set_file_extent_num_bytes(leaf, fi,
 846                                                         start - key.offset);
 847                         btrfs_mark_buffer_dirty(leaf);
 848                         if (update_refs && disk_bytenr > 0)
 849                                 inode_sub_bytes(inode, extent_end - start);
 850                         if (end == extent_end)
 851                                 break;
 852
 853                         path->slots[0]++;
 854                         goto next_slot;
 855                 }
 856
 857                 /*
 858                  *  | ---- range to drop ----- |
 859                  *    | ------ extent ------ |
 860                  */
 861                 if (start <= key.offset && end >= extent_end) {
 862                         if (del_nr == 0) {
 863                                 del_slot = path->slots[0];
 864                                 del_nr = 1;
 865                         } else {
 866                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 867                                 del_nr++;
 868                         }
 869
 870                         if (update_refs &&
 871                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 872                                 inode_sub_bytes(inode,
 873                                                 extent_end - key.offset);
 874                                 extent_end = ALIGN(extent_end,
 875                                                    root->sectorsize);
 876                         } else if (update_refs && disk_bytenr > 0) {
 877                                 ret = btrfs_free_extent(trans, root,
 878                                                 disk_bytenr, num_bytes, 0,
 879                                                 root->root_key.objectid,
 880                                                 key.objectid, key.offset -
 881                                                 extent_offset, 0);
 882                                 BUG_ON(ret); /* -ENOMEM */
 883                                 inode_sub_bytes(inode,
 884                                                 extent_end - key.offset);
 885                         }
 886
 887                         if (end == extent_end)
 888                                 break;
 889
 890                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 891                                 path->slots[0]++;
 892                                 goto next_slot;
 893                         }
 894
 895                         ret = btrfs_del_items(trans, root, path, del_slot,
 896                                               del_nr);
 897                         if (ret) {
 898                                 btrfs_abort_transaction(trans, root, ret);
 899                                 break;
 900                         }
 901
 902                         del_nr = 0;
 903                         del_slot = 0;
 904
 905                         btrfs_release_path(path);
 906                         continue;
 907                 }
 908
 909                 BUG_ON(1);
 910         }
 911
 912         if (!ret && del_nr > 0) {
 913                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 914                 if (ret)
 915                         btrfs_abort_transaction(trans, root, ret);
 916         }
 917
 918         if (drop_end)
 919                 *drop_end = found ? min(end, extent_end) : end;
 920         btrfs_release_path(path);
 921         return ret;
 922 }
 923
 924 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 925                        struct btrfs_root *root, struct inode *inode, u64 start,
 926                        u64 end, int drop_cache)
 927 {
 928         struct btrfs_path *path;
 929         int ret;
 930
 931         path = btrfs_alloc_path();
 932         if (!path)
 933                 return -ENOMEM;
 934         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
 935                                    drop_cache);
 936         btrfs_free_path(path);
 937         return ret;
 938 }
 939
 940 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 941                             u64 objectid, u64 bytenr, u64 orig_offset,
 942                             u64 *start, u64 *end)
 943 {
 944         struct btrfs_file_extent_item *fi;
 945         struct btrfs_key key;
 946         u64 extent_end;
 947
 948         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 949                 return 0;
 950
 951         btrfs_item_key_to_cpu(leaf, &key, slot);
 952         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 953                 return 0;
 954
 955         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 956         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 957             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 958             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 959             btrfs_file_extent_compression(leaf, fi) ||
 960             btrfs_file_extent_encryption(leaf, fi) ||
 961             btrfs_file_extent_other_encoding(leaf, fi))
 962                 return 0;
 963
 964         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 965         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 966                 return 0;
 967
 968         *start = key.offset;
 969         *end = extent_end;
 970         return 1;
 971 }
 972
 973 /*
 974  * Mark extent in the range start - end as written.
 975  *
 976  * This changes extent type from 'pre-allocated' to 'regular'. If only
 977  * part of extent is marked as written, the extent will be split into
 978  * two or three.
 979  */
 980 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 981                               struct inode *inode, u64 start, u64 end)
 982 {
 983         struct btrfs_root *root = BTRFS_I(inode)->root;
 984         struct extent_buffer *leaf;
 985         struct btrfs_path *path;
 986         struct btrfs_file_extent_item *fi;
 987         struct btrfs_key key;
 988         struct btrfs_key new_key;
 989         u64 bytenr;
 990         u64 num_bytes;
 991         u64 extent_end;
 992         u64 orig_offset;
 993         u64 other_start;
 994         u64 other_end;
 995         u64 split;
 996         int del_nr = 0;
 997         int del_slot = 0;
 998         int recow;
 999         int ret;
1000         u64 ino = btrfs_ino(inode);
1001
1002         path = btrfs_alloc_path();
1003         if (!path)
1004                 return -ENOMEM;
1005 again:
1006         recow = 0;
1007         split = start;
1008         key.objectid = ino;
1009         key.type = BTRFS_EXTENT_DATA_KEY;
1010         key.offset = split;
1011
1012         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1013         if (ret < 0)
1014                 goto out;
1015         if (ret > 0 && path->slots[0] > 0)
1016                 path->slots[0]--;
1017
1018         leaf = path->nodes[0];
1019         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1020         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1021         fi = btrfs_item_ptr(leaf, path->slots[0],
1022                             struct btrfs_file_extent_item);
1023         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1024                BTRFS_FILE_EXTENT_PREALLOC);
1025         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1026         BUG_ON(key.offset > start || extent_end < end);
1027
1028         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1029         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1030         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1031         memcpy(&new_key, &key, sizeof(new_key));
1032
1033         if (start == key.offset && end < extent_end) {
1034                 other_start = 0;
1035                 other_end = start;
1036                 if (extent_mergeable(leaf, path->slots[0] - 1,
1037                                      ino, bytenr, orig_offset,
1038                                      &other_start, &other_end)) {
1039                         new_key.offset = end;
1040                         btrfs_set_item_key_safe(trans, root, path, &new_key);
1041                         fi = btrfs_item_ptr(leaf, path->slots[0],
1042                                             struct btrfs_file_extent_item);
1043                         btrfs_set_file_extent_generation(leaf, fi,
1044                                                          trans->transid);
1045                         btrfs_set_file_extent_num_bytes(leaf, fi,
1046                                                         extent_end - end);
1047                         btrfs_set_file_extent_offset(leaf, fi,
1048                                                      end - orig_offset);
1049                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1050                                             struct btrfs_file_extent_item);
1051                         btrfs_set_file_extent_generation(leaf, fi,
1052                                                          trans->transid);
1053                         btrfs_set_file_extent_num_bytes(leaf, fi,
1054                                                         end - other_start);
1055                         btrfs_mark_buffer_dirty(leaf);
1056                         goto out;
1057                 }
1058         }
1059
1060         if (start > key.offset && end == extent_end) {
1061                 other_start = end;
1062                 other_end = 0;
1063                 if (extent_mergeable(leaf, path->slots[0] + 1,
1064                                      ino, bytenr, orig_offset,
1065                                      &other_start, &other_end)) {
1066                         fi = btrfs_item_ptr(leaf, path->slots[0],
1067                                             struct btrfs_file_extent_item);
1068                         btrfs_set_file_extent_num_bytes(leaf, fi,
1069                                                         start - key.offset);
1070                         btrfs_set_file_extent_generation(leaf, fi,
1071                                                          trans->transid);
1072                         path->slots[0]++;
1073                         new_key.offset = start;
1074                         btrfs_set_item_key_safe(trans, root, path, &new_key);
1075
1076                         fi = btrfs_item_ptr(leaf, path->slots[0],
1077                                             struct btrfs_file_extent_item);
1078                         btrfs_set_file_extent_generation(leaf, fi,
1079                                                          trans->transid);
1080                         btrfs_set_file_extent_num_bytes(leaf, fi,
1081                                                         other_end - start);
1082                         btrfs_set_file_extent_offset(leaf, fi,
1083                                                      start - orig_offset);
1084                         btrfs_mark_buffer_dirty(leaf);
1085                         goto out;
1086                 }
1087         }
1088
1089         while (start > key.offset || end < extent_end) {
1090                 if (key.offset == start)
1091                         split = end;
1092
1093                 new_key.offset = split;
1094                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1095                 if (ret == -EAGAIN) {
1096                         btrfs_release_path(path);
1097                         goto again;
1098                 }
1099                 if (ret < 0) {
1100                         btrfs_abort_transaction(trans, root, ret);
1101                         goto out;
1102                 }
1103
1104                 leaf = path->nodes[0];
1105                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1106                                     struct btrfs_file_extent_item);
1107                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1108                 btrfs_set_file_extent_num_bytes(leaf, fi,
1109                                                 split - key.offset);
1110
1111                 fi = btrfs_item_ptr(leaf, path->slots[0],
1112                                     struct btrfs_file_extent_item);
1113
1114                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1115                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1116                 btrfs_set_file_extent_num_bytes(leaf, fi,
1117                                                 extent_end - split);
1118                 btrfs_mark_buffer_dirty(leaf);
1119
1120                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1121                                            root->root_key.objectid,
1122                                            ino, orig_offset, 0);
1123                 BUG_ON(ret); /* -ENOMEM */
1124
1125                 if (split == start) {
1126                         key.offset = start;
1127                 } else {
1128                         BUG_ON(start != key.offset);
1129                         path->slots[0]--;
1130                         extent_end = end;
1131                 }
1132                 recow = 1;
1133         }
1134
1135         other_start = end;
1136         other_end = 0;
1137         if (extent_mergeable(leaf, path->slots[0] + 1,
1138                              ino, bytenr, orig_offset,
1139                              &other_start, &other_end)) {
1140                 if (recow) {
1141                         btrfs_release_path(path);
1142                         goto again;
1143                 }
1144                 extent_end = other_end;
1145                 del_slot = path->slots[0] + 1;
1146                 del_nr++;
1147                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1148                                         0, root->root_key.objectid,
1149                                         ino, orig_offset, 0);
1150                 BUG_ON(ret); /* -ENOMEM */
1151         }
1152         other_start = 0;
1153         other_end = start;
1154         if (extent_mergeable(leaf, path->slots[0] - 1,
1155                              ino, bytenr, orig_offset,
1156                              &other_start, &other_end)) {
1157                 if (recow) {
1158                         btrfs_release_path(path);
1159                         goto again;
1160                 }
1161                 key.offset = other_start;
1162                 del_slot = path->slots[0];
1163                 del_nr++;
1164                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1165                                         0, root->root_key.objectid,
1166                                         ino, orig_offset, 0);
1167                 BUG_ON(ret); /* -ENOMEM */
1168         }
1169         if (del_nr == 0) {
1170                 fi = btrfs_item_ptr(leaf, path->slots[0],
1171                            struct btrfs_file_extent_item);
1172                 btrfs_set_file_extent_type(leaf, fi,
1173                                            BTRFS_FILE_EXTENT_REG);
1174                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1175                 btrfs_mark_buffer_dirty(leaf);
1176         } else {
1177                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1178                            struct btrfs_file_extent_item);
1179                 btrfs_set_file_extent_type(leaf, fi,
1180                                            BTRFS_FILE_EXTENT_REG);
1181                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1182                 btrfs_set_file_extent_num_bytes(leaf, fi,
1183                                                 extent_end - key.offset);
1184                 btrfs_mark_buffer_dirty(leaf);
1185
1186                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1187                 if (ret < 0) {
1188                         btrfs_abort_transaction(trans, root, ret);
1189                         goto out;
1190                 }
1191         }
1192 out:
1193         btrfs_free_path(path);
1194         return 0;
1195 }
1196
1197 /*
1198  * on error we return an unlocked page and the error value
1199  * on success we return a locked page and 0
1200  */
1201 static int prepare_uptodate_page(struct page *page, u64 pos,
1202                                  bool force_uptodate)
1203 {
1204         int ret = 0;
1205
1206         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1207             !PageUptodate(page)) {
1208                 ret = btrfs_readpage(NULL, page);
1209                 if (ret)
1210                         return ret;
1211                 lock_page(page);
1212                 if (!PageUptodate(page)) {
1213                         unlock_page(page);
1214                         return -EIO;
1215                 }
1216         }
1217         return 0;
1218 }
1219
1220 /*
1221  * this gets pages into the page cache and locks them down, it also properly
1222  * waits for data=ordered extents to finish before allowing the pages to be
1223  * modified.
1224  */
1225 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1226                          struct page **pages, size_t num_pages,
1227                          loff_t pos, unsigned long first_index,
1228                          size_t write_bytes, bool force_uptodate)
1229 {
1230         struct extent_state *cached_state = NULL;
1231         int i;
1232         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1233         struct inode *inode = file_inode(file);
1234         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1235         int err = 0;
1236         int faili = 0;
1237         u64 start_pos;
1238         u64 last_pos;
1239
1240         start_pos = pos & ~((u64)root->sectorsize - 1);
1241         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1242
1243 again:
1244         for (i = 0; i < num_pages; i++) {
1245                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1246                                                mask | __GFP_WRITE);
1247                 if (!pages[i]) {
1248                         faili = i - 1;
1249                         err = -ENOMEM;
1250                         goto fail;
1251                 }
1252
1253                 if (i == 0)
1254                         err = prepare_uptodate_page(pages[i], pos,
1255                                                     force_uptodate);
1256                 if (i == num_pages - 1)
1257                         err = prepare_uptodate_page(pages[i],
1258                                                     pos + write_bytes, false);
1259                 if (err) {
1260                         page_cache_release(pages[i]);
1261                         faili = i - 1;
1262                         goto fail;
1263                 }
1264                 wait_on_page_writeback(pages[i]);
1265         }
1266         err = 0;
1267         if (start_pos < inode->i_size) {
1268                 struct btrfs_ordered_extent *ordered;
1269                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1270                                  start_pos, last_pos - 1, 0, &cached_state);
1271                 ordered = btrfs_lookup_first_ordered_extent(inode,
1272                                                             last_pos - 1);
1273                 if (ordered &&
1274                     ordered->file_offset + ordered->len > start_pos &&
1275                     ordered->file_offset < last_pos) {
1276                         btrfs_put_ordered_extent(ordered);
1277                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1278                                              start_pos, last_pos - 1,
1279                                              &cached_state, GFP_NOFS);
1280                         for (i = 0; i < num_pages; i++) {
1281                                 unlock_page(pages[i]);
1282                                 page_cache_release(pages[i]);
1283                         }
1284                         btrfs_wait_ordered_range(inode, start_pos,
1285                                                  last_pos - start_pos);
1286                         goto again;
1287                 }
1288                 if (ordered)
1289                         btrfs_put_ordered_extent(ordered);
1290
1291                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1292                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1293                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1294                                   0, 0, &cached_state, GFP_NOFS);
1295                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1296                                      start_pos, last_pos - 1, &cached_state,
1297                                      GFP_NOFS);
1298         }
1299         for (i = 0; i < num_pages; i++) {
1300                 if (clear_page_dirty_for_io(pages[i]))
1301                         account_page_redirty(pages[i]);
1302                 set_page_extent_mapped(pages[i]);
1303                 WARN_ON(!PageLocked(pages[i]));
1304         }
1305         return 0;
1306 fail:
1307         while (faili >= 0) {
1308                 unlock_page(pages[faili]);
1309                 page_cache_release(pages[faili]);
1310                 faili--;
1311         }
1312         return err;
1313
1314 }
1315
1316 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1317                                                struct iov_iter *i,
1318                                                loff_t pos)
1319 {
1320         struct inode *inode = file_inode(file);
1321         struct btrfs_root *root = BTRFS_I(inode)->root;
1322         struct page **pages = NULL;
1323         unsigned long first_index;
1324         size_t num_written = 0;
1325         int nrptrs;
1326         int ret = 0;
1327         bool force_page_uptodate = false;
1328
1329         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1330                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1331                      (sizeof(struct page *)));
1332         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1333         nrptrs = max(nrptrs, 8);
1334         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1335         if (!pages)
1336                 return -ENOMEM;
1337
1338         first_index = pos >> PAGE_CACHE_SHIFT;
1339
1340         while (iov_iter_count(i) > 0) {
1341                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1342                 size_t write_bytes = min(iov_iter_count(i),
1343                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1344                                          offset);
1345                 size_t num_pages = (write_bytes + offset +
1346                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1347                 size_t dirty_pages;
1348                 size_t copied;
1349
1350                 WARN_ON(num_pages > nrptrs);
1351
1352                 /*
1353                  * Fault pages before locking them in prepare_pages
1354                  * to avoid recursive lock
1355                  */
1356                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1357                         ret = -EFAULT;
1358                         break;
1359                 }
1360
1361                 ret = btrfs_delalloc_reserve_space(inode,
1362                                         num_pages << PAGE_CACHE_SHIFT);
1363                 if (ret)
1364                         break;
1365
1366                 /*
1367                  * This is going to setup the pages array with the number of
1368                  * pages we want, so we don't really need to worry about the
1369                  * contents of pages from loop to loop
1370                  */
1371                 ret = prepare_pages(root, file, pages, num_pages,
1372                                     pos, first_index, write_bytes,
1373                                     force_page_uptodate);
1374                 if (ret) {
1375                         btrfs_delalloc_release_space(inode,
1376                                         num_pages << PAGE_CACHE_SHIFT);
1377                         break;
1378                 }
1379
1380                 copied = btrfs_copy_from_user(pos, num_pages,
1381                                            write_bytes, pages, i);
1382
1383                 /*
1384                  * if we have trouble faulting in the pages, fall
1385                  * back to one page at a time
1386                  */
1387                 if (copied < write_bytes)
1388                         nrptrs = 1;
1389
1390                 if (copied == 0) {
1391                         force_page_uptodate = true;
1392                         dirty_pages = 0;
1393                 } else {
1394                         force_page_uptodate = false;
1395                         dirty_pages = (copied + offset +
1396                                        PAGE_CACHE_SIZE - 1) >>
1397                                        PAGE_CACHE_SHIFT;
1398                 }
1399
1400                 /*
1401                  * If we had a short copy we need to release the excess delaloc
1402                  * bytes we reserved.  We need to increment outstanding_extents
1403                  * because btrfs_delalloc_release_space will decrement it, but
1404                  * we still have an outstanding extent for the chunk we actually
1405                  * managed to copy.
1406                  */
1407                 if (num_pages > dirty_pages) {
1408                         if (copied > 0) {
1409                                 spin_lock(&BTRFS_I(inode)->lock);
1410                                 BTRFS_I(inode)->outstanding_extents++;
1411                                 spin_unlock(&BTRFS_I(inode)->lock);
1412                         }
1413                         btrfs_delalloc_release_space(inode,
1414                                         (num_pages - dirty_pages) <<
1415                                         PAGE_CACHE_SHIFT);
1416                 }
1417
1418                 if (copied > 0) {
1419                         ret = btrfs_dirty_pages(root, inode, pages,
1420                                                 dirty_pages, pos, copied,
1421                                                 NULL);
1422                         if (ret) {
1423                                 btrfs_delalloc_release_space(inode,
1424                                         dirty_pages << PAGE_CACHE_SHIFT);
1425                                 btrfs_drop_pages(pages, num_pages);
1426                                 break;
1427                         }
1428                 }
1429
1430                 btrfs_drop_pages(pages, num_pages);
1431
1432                 cond_resched();
1433
1434                 balance_dirty_pages_ratelimited(inode->i_mapping);
1435                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1436                         btrfs_btree_balance_dirty(root);
1437
1438                 pos += copied;
1439                 num_written += copied;
1440         }
1441
1442         kfree(pages);
1443
1444         return num_written ? num_written : ret;
1445 }
1446
1447 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1448                                     const struct iovec *iov,
1449                                     unsigned long nr_segs, loff_t pos,
1450                                     loff_t *ppos, size_t count, size_t ocount)
1451 {
1452         struct file *file = iocb->ki_filp;
1453         struct iov_iter i;
1454         ssize_t written;
1455         ssize_t written_buffered;
1456         loff_t endbyte;
1457         int err;
1458
1459         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1460                                             count, ocount);
1461
1462         if (written < 0 || written == count)
1463                 return written;
1464
1465         pos += written;
1466         count -= written;
1467         iov_iter_init(&i, iov, nr_segs, count, written);
1468         written_buffered = __btrfs_buffered_write(file, &i, pos);
1469         if (written_buffered < 0) {
1470                 err = written_buffered;
1471                 goto out;
1472         }
1473         endbyte = pos + written_buffered - 1;
1474         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1475         if (err)
1476                 goto out;
1477         written += written_buffered;
1478         *ppos = pos + written_buffered;
1479         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1480                                  endbyte >> PAGE_CACHE_SHIFT);
1481 out:
1482         return written ? written : err;
1483 }
1484
1485 static void update_time_for_write(struct inode *inode)
1486 {
1487         struct timespec now;
1488
1489         if (IS_NOCMTIME(inode))
1490                 return;
1491
1492         now = current_fs_time(inode->i_sb);
1493         if (!timespec_equal(&inode->i_mtime, &now))
1494                 inode->i_mtime = now;
1495
1496         if (!timespec_equal(&inode->i_ctime, &now))
1497                 inode->i_ctime = now;
1498
1499         if (IS_I_VERSION(inode))
1500                 inode_inc_iversion(inode);
1501 }
1502
1503 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1504                                     const struct iovec *iov,
1505                                     unsigned long nr_segs, loff_t pos)
1506 {
1507         struct file *file = iocb->ki_filp;
1508         struct inode *inode = file_inode(file);
1509         struct btrfs_root *root = BTRFS_I(inode)->root;
1510         loff_t *ppos = &iocb->ki_pos;
1511         u64 start_pos;
1512         ssize_t num_written = 0;
1513         ssize_t err = 0;
1514         size_t count, ocount;
1515         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1516
1517         sb_start_write(inode->i_sb);
1518
1519         mutex_lock(&inode->i_mutex);
1520
1521         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1522         if (err) {
1523                 mutex_unlock(&inode->i_mutex);
1524                 goto out;
1525         }
1526         count = ocount;
1527
1528         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1529         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1530         if (err) {
1531                 mutex_unlock(&inode->i_mutex);
1532                 goto out;
1533         }
1534
1535         if (count == 0) {
1536                 mutex_unlock(&inode->i_mutex);
1537                 goto out;
1538         }
1539
1540         err = file_remove_suid(file);
1541         if (err) {
1542                 mutex_unlock(&inode->i_mutex);
1543                 goto out;
1544         }
1545
1546         /*
1547          * If BTRFS flips readonly due to some impossible error
1548          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1549          * although we have opened a file as writable, we have
1550          * to stop this write operation to ensure FS consistency.
1551          */
1552         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1553                 mutex_unlock(&inode->i_mutex);
1554                 err = -EROFS;
1555                 goto out;
1556         }
1557
1558         /*
1559          * We reserve space for updating the inode when we reserve space for the
1560          * extent we are going to write, so we will enospc out there.  We don't
1561          * need to start yet another transaction to update the inode as we will
1562          * update the inode when we finish writing whatever data we write.
1563          */
1564         update_time_for_write(inode);
1565
1566         start_pos = round_down(pos, root->sectorsize);
1567         if (start_pos > i_size_read(inode)) {
1568                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1569                 if (err) {
1570                         mutex_unlock(&inode->i_mutex);
1571                         goto out;
1572                 }
1573         }
1574
1575         if (sync)
1576                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1577
1578         if (unlikely(file->f_flags & O_DIRECT)) {
1579                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1580                                                    pos, ppos, count, ocount);
1581         } else {
1582                 struct iov_iter i;
1583
1584                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1585
1586                 num_written = __btrfs_buffered_write(file, &i, pos);
1587                 if (num_written > 0)
1588                         *ppos = pos + num_written;
1589         }
1590
1591         mutex_unlock(&inode->i_mutex);
1592
1593         /*
1594          * we want to make sure fsync finds this change
1595          * but we haven't joined a transaction running right now.
1596          *
1597          * Later on, someone is sure to update the inode and get the
1598          * real transid recorded.
1599          *
1600          * We set last_trans now to the fs_info generation + 1,
1601          * this will either be one more than the running transaction
1602          * or the generation used for the next transaction if there isn't
1603          * one running right now.
1604          *
1605          * We also have to set last_sub_trans to the current log transid,
1606          * otherwise subsequent syncs to a file that's been synced in this
1607          * transaction will appear to have already occured.
1608          */
1609         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1610         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1611         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1612                 err = generic_write_sync(file, pos, num_written);
1613                 if (err < 0 && num_written > 0)
1614                         num_written = err;
1615         }
1616
1617         if (sync)
1618                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1619 out:
1620         sb_end_write(inode->i_sb);
1621         current->backing_dev_info = NULL;
1622         return num_written ? num_written : err;
1623 }
1624
1625 int btrfs_release_file(struct inode *inode, struct file *filp)
1626 {
1627         /*
1628          * ordered_data_close is set by settattr when we are about to truncate
1629          * a file from a non-zero size to a zero size.  This tries to
1630          * flush down new bytes that may have been written if the
1631          * application were using truncate to replace a file in place.
1632          */
1633         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1634                                &BTRFS_I(inode)->runtime_flags)) {
1635                 struct btrfs_trans_handle *trans;
1636                 struct btrfs_root *root = BTRFS_I(inode)->root;
1637
1638                 /*
1639                  * We need to block on a committing transaction to keep us from
1640                  * throwing a ordered operation on to the list and causing
1641                  * something like sync to deadlock trying to flush out this
1642                  * inode.
1643                  */
1644                 trans = btrfs_start_transaction(root, 0);
1645                 if (IS_ERR(trans))
1646                         return PTR_ERR(trans);
1647                 btrfs_add_ordered_operation(trans, BTRFS_I(inode)->root, inode);
1648                 btrfs_end_transaction(trans, root);
1649                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1650                         filemap_flush(inode->i_mapping);
1651         }
1652         if (filp->private_data)
1653                 btrfs_ioctl_trans_end(filp);
1654         return 0;
1655 }
1656
1657 /*
1658  * fsync call for both files and directories.  This logs the inode into
1659  * the tree log instead of forcing full commits whenever possible.
1660  *
1661  * It needs to call filemap_fdatawait so that all ordered extent updates are
1662  * in the metadata btree are up to date for copying to the log.
1663  *
1664  * It drops the inode mutex before doing the tree log commit.  This is an
1665  * important optimization for directories because holding the mutex prevents
1666  * new operations on the dir while we write to disk.
1667  */
1668 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1669 {
1670         struct dentry *dentry = file->f_path.dentry;
1671         struct inode *inode = dentry->d_inode;
1672         struct btrfs_root *root = BTRFS_I(inode)->root;
1673         int ret = 0;
1674         struct btrfs_trans_handle *trans;
1675         bool full_sync = 0;
1676
1677         trace_btrfs_sync_file(file, datasync);
1678
1679         /*
1680          * We write the dirty pages in the range and wait until they complete
1681          * out of the ->i_mutex. If so, we can flush the dirty pages by
1682          * multi-task, and make the performance up.  See
1683          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1684          */
1685         atomic_inc(&BTRFS_I(inode)->sync_writers);
1686         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1687         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1688                              &BTRFS_I(inode)->runtime_flags))
1689                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1690         atomic_dec(&BTRFS_I(inode)->sync_writers);
1691         if (ret)
1692                 return ret;
1693
1694         mutex_lock(&inode->i_mutex);
1695
1696         /*
1697          * We flush the dirty pages again to avoid some dirty pages in the
1698          * range being left.
1699          */
1700         atomic_inc(&root->log_batch);
1701         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1702                              &BTRFS_I(inode)->runtime_flags);
1703         if (full_sync)
1704                 btrfs_wait_ordered_range(inode, start, end - start + 1);
1705         atomic_inc(&root->log_batch);
1706
1707         /*
1708          * check the transaction that last modified this inode
1709          * and see if its already been committed
1710          */
1711         if (!BTRFS_I(inode)->last_trans) {
1712                 mutex_unlock(&inode->i_mutex);
1713                 goto out;
1714         }
1715
1716         /*
1717          * if the last transaction that changed this file was before
1718          * the current transaction, we can bail out now without any
1719          * syncing
1720          */
1721         smp_mb();
1722         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1723             BTRFS_I(inode)->last_trans <=
1724             root->fs_info->last_trans_committed) {
1725                 BTRFS_I(inode)->last_trans = 0;
1726
1727                 /*
1728                  * We'v had everything committed since the last time we were
1729                  * modified so clear this flag in case it was set for whatever
1730                  * reason, it's no longer relevant.
1731                  */
1732                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1733                           &BTRFS_I(inode)->runtime_flags);
1734                 mutex_unlock(&inode->i_mutex);
1735                 goto out;
1736         }
1737
1738         /*
1739          * ok we haven't committed the transaction yet, lets do a commit
1740          */
1741         if (file->private_data)
1742                 btrfs_ioctl_trans_end(file);
1743
1744         trans = btrfs_start_transaction(root, 0);
1745         if (IS_ERR(trans)) {
1746                 ret = PTR_ERR(trans);
1747                 mutex_unlock(&inode->i_mutex);
1748                 goto out;
1749         }
1750
1751         ret = btrfs_log_dentry_safe(trans, root, dentry);
1752         if (ret < 0) {
1753                 mutex_unlock(&inode->i_mutex);
1754                 goto out;
1755         }
1756
1757         /* we've logged all the items and now have a consistent
1758          * version of the file in the log.  It is possible that
1759          * someone will come in and modify the file, but that's
1760          * fine because the log is consistent on disk, and we
1761          * have references to all of the file's extents
1762          *
1763          * It is possible that someone will come in and log the
1764          * file again, but that will end up using the synchronization
1765          * inside btrfs_sync_log to keep things safe.
1766          */
1767         mutex_unlock(&inode->i_mutex);
1768
1769         if (ret != BTRFS_NO_LOG_SYNC) {
1770                 if (ret > 0) {
1771                         /*
1772                          * If we didn't already wait for ordered extents we need
1773                          * to do that now.
1774                          */
1775                         if (!full_sync)
1776                                 btrfs_wait_ordered_range(inode, start,
1777                                                          end - start + 1);
1778                         ret = btrfs_commit_transaction(trans, root);
1779                 } else {
1780                         ret = btrfs_sync_log(trans, root);
1781                         if (ret == 0) {
1782                                 ret = btrfs_end_transaction(trans, root);
1783                         } else {
1784                                 if (!full_sync)
1785                                         btrfs_wait_ordered_range(inode, start,
1786                                                                  end -
1787                                                                  start + 1);
1788                                 ret = btrfs_commit_transaction(trans, root);
1789                         }
1790                 }
1791         } else {
1792                 ret = btrfs_end_transaction(trans, root);
1793         }
1794 out:
1795         return ret > 0 ? -EIO : ret;
1796 }
1797
1798 static const struct vm_operations_struct btrfs_file_vm_ops = {
1799         .fault          = filemap_fault,
1800         .page_mkwrite   = btrfs_page_mkwrite,
1801         .remap_pages    = generic_file_remap_pages,
1802 };
1803
1804 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1805 {
1806         struct address_space *mapping = filp->f_mapping;
1807
1808         if (!mapping->a_ops->readpage)
1809                 return -ENOEXEC;
1810
1811         file_accessed(filp);
1812         vma->vm_ops = &btrfs_file_vm_ops;
1813
1814         return 0;
1815 }
1816
1817 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
1818                           int slot, u64 start, u64 end)
1819 {
1820         struct btrfs_file_extent_item *fi;
1821         struct btrfs_key key;
1822
1823         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1824                 return 0;
1825
1826         btrfs_item_key_to_cpu(leaf, &key, slot);
1827         if (key.objectid != btrfs_ino(inode) ||
1828             key.type != BTRFS_EXTENT_DATA_KEY)
1829                 return 0;
1830
1831         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1832
1833         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
1834                 return 0;
1835
1836         if (btrfs_file_extent_disk_bytenr(leaf, fi))
1837                 return 0;
1838
1839         if (key.offset == end)
1840                 return 1;
1841         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
1842                 return 1;
1843         return 0;
1844 }
1845
1846 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
1847                       struct btrfs_path *path, u64 offset, u64 end)
1848 {
1849         struct btrfs_root *root = BTRFS_I(inode)->root;
1850         struct extent_buffer *leaf;
1851         struct btrfs_file_extent_item *fi;
1852         struct extent_map *hole_em;
1853         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1854         struct btrfs_key key;
1855         int ret;
1856
1857         key.objectid = btrfs_ino(inode);
1858         key.type = BTRFS_EXTENT_DATA_KEY;
1859         key.offset = offset;
1860
1861
1862         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1863         if (ret < 0)
1864                 return ret;
1865         BUG_ON(!ret);
1866
1867         leaf = path->nodes[0];
1868         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
1869                 u64 num_bytes;
1870
1871                 path->slots[0]--;
1872                 fi = btrfs_item_ptr(leaf, path->slots[0],
1873                                     struct btrfs_file_extent_item);
1874                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
1875                         end - offset;
1876                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1877                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
1878                 btrfs_set_file_extent_offset(leaf, fi, 0);
1879                 btrfs_mark_buffer_dirty(leaf);
1880                 goto out;
1881         }
1882
1883         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
1884                 u64 num_bytes;
1885
1886                 path->slots[0]++;
1887                 key.offset = offset;
1888                 btrfs_set_item_key_safe(trans, root, path, &key);
1889                 fi = btrfs_item_ptr(leaf, path->slots[0],
1890                                     struct btrfs_file_extent_item);
1891                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
1892                         offset;
1893                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1894                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
1895                 btrfs_set_file_extent_offset(leaf, fi, 0);
1896                 btrfs_mark_buffer_dirty(leaf);
1897                 goto out;
1898         }
1899         btrfs_release_path(path);
1900
1901         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
1902                                        0, 0, end - offset, 0, end - offset,
1903                                        0, 0, 0);
1904         if (ret)
1905                 return ret;
1906
1907 out:
1908         btrfs_release_path(path);
1909
1910         hole_em = alloc_extent_map();
1911         if (!hole_em) {
1912                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
1913                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1914                         &BTRFS_I(inode)->runtime_flags);
1915         } else {
1916                 hole_em->start = offset;
1917                 hole_em->len = end - offset;
1918                 hole_em->orig_start = offset;
1919
1920                 hole_em->block_start = EXTENT_MAP_HOLE;
1921                 hole_em->block_len = 0;
1922                 hole_em->orig_block_len = 0;
1923                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
1924                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
1925                 hole_em->generation = trans->transid;
1926
1927                 do {
1928                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
1929                         write_lock(&em_tree->lock);
1930                         ret = add_extent_mapping(em_tree, hole_em);
1931                         if (!ret)
1932                                 list_move(&hole_em->list,
1933                                           &em_tree->modified_extents);
1934                         write_unlock(&em_tree->lock);
1935                 } while (ret == -EEXIST);
1936                 free_extent_map(hole_em);
1937                 if (ret)
1938                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1939                                 &BTRFS_I(inode)->runtime_flags);
1940         }
1941
1942         return 0;
1943 }
1944
1945 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
1946 {
1947         struct btrfs_root *root = BTRFS_I(inode)->root;
1948         struct extent_state *cached_state = NULL;
1949         struct btrfs_path *path;
1950         struct btrfs_block_rsv *rsv;
1951         struct btrfs_trans_handle *trans;
1952         u64 lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
1953         u64 lockend = round_down(offset + len,
1954                                  BTRFS_I(inode)->root->sectorsize) - 1;
1955         u64 cur_offset = lockstart;
1956         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
1957         u64 drop_end;
1958         int ret = 0;
1959         int err = 0;
1960         bool same_page = ((offset >> PAGE_CACHE_SHIFT) ==
1961                           ((offset + len - 1) >> PAGE_CACHE_SHIFT));
1962
1963         btrfs_wait_ordered_range(inode, offset, len);
1964
1965         mutex_lock(&inode->i_mutex);
1966         /*
1967          * We needn't truncate any page which is beyond the end of the file
1968          * because we are sure there is no data there.
1969          */
1970         /*
1971          * Only do this if we are in the same page and we aren't doing the
1972          * entire page.
1973          */
1974         if (same_page && len < PAGE_CACHE_SIZE) {
1975                 if (offset < round_up(inode->i_size, PAGE_CACHE_SIZE))
1976                         ret = btrfs_truncate_page(inode, offset, len, 0);
1977                 mutex_unlock(&inode->i_mutex);
1978                 return ret;
1979         }
1980
1981         /* zero back part of the first page */
1982         if (offset < round_up(inode->i_size, PAGE_CACHE_SIZE)) {
1983                 ret = btrfs_truncate_page(inode, offset, 0, 0);
1984                 if (ret) {
1985                         mutex_unlock(&inode->i_mutex);
1986                         return ret;
1987                 }
1988         }
1989
1990         /* zero the front end of the last page */
1991         if (offset + len < round_up(inode->i_size, PAGE_CACHE_SIZE)) {
1992                 ret = btrfs_truncate_page(inode, offset + len, 0, 1);
1993                 if (ret) {
1994                         mutex_unlock(&inode->i_mutex);
1995                         return ret;
1996                 }
1997         }
1998
1999         if (lockend < lockstart) {
2000                 mutex_unlock(&inode->i_mutex);
2001                 return 0;
2002         }
2003
2004         while (1) {
2005                 struct btrfs_ordered_extent *ordered;
2006
2007                 truncate_pagecache_range(inode, lockstart, lockend);
2008
2009                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2010                                  0, &cached_state);
2011                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2012
2013                 /*
2014                  * We need to make sure we have no ordered extents in this range
2015                  * and nobody raced in and read a page in this range, if we did
2016                  * we need to try again.
2017                  */
2018                 if ((!ordered ||
2019                     (ordered->file_offset + ordered->len < lockstart ||
2020                      ordered->file_offset > lockend)) &&
2021                      !test_range_bit(&BTRFS_I(inode)->io_tree, lockstart,
2022                                      lockend, EXTENT_UPTODATE, 0,
2023                                      cached_state)) {
2024                         if (ordered)
2025                                 btrfs_put_ordered_extent(ordered);
2026                         break;
2027                 }
2028                 if (ordered)
2029                         btrfs_put_ordered_extent(ordered);
2030                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2031                                      lockend, &cached_state, GFP_NOFS);
2032                 btrfs_wait_ordered_range(inode, lockstart,
2033                                          lockend - lockstart + 1);
2034         }
2035
2036         path = btrfs_alloc_path();
2037         if (!path) {
2038                 ret = -ENOMEM;
2039                 goto out;
2040         }
2041
2042         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2043         if (!rsv) {
2044                 ret = -ENOMEM;
2045                 goto out_free;
2046         }
2047         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2048         rsv->failfast = 1;
2049
2050         /*
2051          * 1 - update the inode
2052          * 1 - removing the extents in the range
2053          * 1 - adding the hole extent
2054          */
2055         trans = btrfs_start_transaction(root, 3);
2056         if (IS_ERR(trans)) {
2057                 err = PTR_ERR(trans);
2058                 goto out_free;
2059         }
2060
2061         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2062                                       min_size);
2063         BUG_ON(ret);
2064         trans->block_rsv = rsv;
2065
2066         while (cur_offset < lockend) {
2067                 ret = __btrfs_drop_extents(trans, root, inode, path,
2068                                            cur_offset, lockend + 1,
2069                                            &drop_end, 1);
2070                 if (ret != -ENOSPC)
2071                         break;
2072
2073                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2074
2075                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2076                 if (ret) {
2077                         err = ret;
2078                         break;
2079                 }
2080
2081                 cur_offset = drop_end;
2082
2083                 ret = btrfs_update_inode(trans, root, inode);
2084                 if (ret) {
2085                         err = ret;
2086                         break;
2087                 }
2088
2089                 btrfs_end_transaction(trans, root);
2090                 btrfs_btree_balance_dirty(root);
2091
2092                 trans = btrfs_start_transaction(root, 3);
2093                 if (IS_ERR(trans)) {
2094                         ret = PTR_ERR(trans);
2095                         trans = NULL;
2096                         break;
2097                 }
2098
2099                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2100                                               rsv, min_size);
2101                 BUG_ON(ret);    /* shouldn't happen */
2102                 trans->block_rsv = rsv;
2103         }
2104
2105         if (ret) {
2106                 err = ret;
2107                 goto out_trans;
2108         }
2109
2110         trans->block_rsv = &root->fs_info->trans_block_rsv;
2111         ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2112         if (ret) {
2113                 err = ret;
2114                 goto out_trans;
2115         }
2116
2117 out_trans:
2118         if (!trans)
2119                 goto out_free;
2120
2121         inode_inc_iversion(inode);
2122         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2123
2124         trans->block_rsv = &root->fs_info->trans_block_rsv;
2125         ret = btrfs_update_inode(trans, root, inode);
2126         btrfs_end_transaction(trans, root);
2127         btrfs_btree_balance_dirty(root);
2128 out_free:
2129         btrfs_free_path(path);
2130         btrfs_free_block_rsv(root, rsv);
2131 out:
2132         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2133                              &cached_state, GFP_NOFS);
2134         mutex_unlock(&inode->i_mutex);
2135         if (ret && !err)
2136                 err = ret;
2137         return err;
2138 }
2139
2140 static long btrfs_fallocate(struct file *file, int mode,
2141                             loff_t offset, loff_t len)
2142 {
2143         struct inode *inode = file_inode(file);
2144         struct extent_state *cached_state = NULL;
2145         u64 cur_offset;
2146         u64 last_byte;
2147         u64 alloc_start;
2148         u64 alloc_end;
2149         u64 alloc_hint = 0;
2150         u64 locked_end;
2151         struct extent_map *em;
2152         int blocksize = BTRFS_I(inode)->root->sectorsize;
2153         int ret;
2154
2155         alloc_start = round_down(offset, blocksize);
2156         alloc_end = round_up(offset + len, blocksize);
2157
2158         /* Make sure we aren't being give some crap mode */
2159         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2160                 return -EOPNOTSUPP;
2161
2162         if (mode & FALLOC_FL_PUNCH_HOLE)
2163                 return btrfs_punch_hole(inode, offset, len);
2164
2165         /*
2166          * Make sure we have enough space before we do the
2167          * allocation.
2168          */
2169         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2170         if (ret)
2171                 return ret;
2172
2173         /*
2174          * wait for ordered IO before we have any locks.  We'll loop again
2175          * below with the locks held.
2176          */
2177         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
2178
2179         mutex_lock(&inode->i_mutex);
2180         ret = inode_newsize_ok(inode, alloc_end);
2181         if (ret)
2182                 goto out;
2183
2184         if (alloc_start > inode->i_size) {
2185                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2186                                         alloc_start);
2187                 if (ret)
2188                         goto out;
2189         }
2190
2191         locked_end = alloc_end - 1;
2192         while (1) {
2193                 struct btrfs_ordered_extent *ordered;
2194
2195                 /* the extent lock is ordered inside the running
2196                  * transaction
2197                  */
2198                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2199                                  locked_end, 0, &cached_state);
2200                 ordered = btrfs_lookup_first_ordered_extent(inode,
2201                                                             alloc_end - 1);
2202                 if (ordered &&
2203                     ordered->file_offset + ordered->len > alloc_start &&
2204                     ordered->file_offset < alloc_end) {
2205                         btrfs_put_ordered_extent(ordered);
2206                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2207                                              alloc_start, locked_end,
2208                                              &cached_state, GFP_NOFS);
2209                         /*
2210                          * we can't wait on the range with the transaction
2211                          * running or with the extent lock held
2212                          */
2213                         btrfs_wait_ordered_range(inode, alloc_start,
2214                                                  alloc_end - alloc_start);
2215                 } else {
2216                         if (ordered)
2217                                 btrfs_put_ordered_extent(ordered);
2218                         break;
2219                 }
2220         }
2221
2222         cur_offset = alloc_start;
2223         while (1) {
2224                 u64 actual_end;
2225
2226                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2227                                       alloc_end - cur_offset, 0);
2228                 if (IS_ERR_OR_NULL(em)) {
2229                         if (!em)
2230                                 ret = -ENOMEM;
2231                         else
2232                                 ret = PTR_ERR(em);
2233                         break;
2234                 }
2235                 last_byte = min(extent_map_end(em), alloc_end);
2236                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2237                 last_byte = ALIGN(last_byte, blocksize);
2238
2239                 if (em->block_start == EXTENT_MAP_HOLE ||
2240                     (cur_offset >= inode->i_size &&
2241                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2242                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2243                                                         last_byte - cur_offset,
2244                                                         1 << inode->i_blkbits,
2245                                                         offset + len,
2246                                                         &alloc_hint);
2247
2248                         if (ret < 0) {
2249                                 free_extent_map(em);
2250                                 break;
2251                         }
2252                 } else if (actual_end > inode->i_size &&
2253                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2254                         /*
2255                          * We didn't need to allocate any more space, but we
2256                          * still extended the size of the file so we need to
2257                          * update i_size.
2258                          */
2259                         inode->i_ctime = CURRENT_TIME;
2260                         i_size_write(inode, actual_end);
2261                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2262                 }
2263                 free_extent_map(em);
2264
2265                 cur_offset = last_byte;
2266                 if (cur_offset >= alloc_end) {
2267                         ret = 0;
2268                         break;
2269                 }
2270         }
2271         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2272                              &cached_state, GFP_NOFS);
2273 out:
2274         mutex_unlock(&inode->i_mutex);
2275         /* Let go of our reservation. */
2276         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2277         return ret;
2278 }
2279
2280 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2281 {
2282         struct btrfs_root *root = BTRFS_I(inode)->root;
2283         struct extent_map *em;
2284         struct extent_state *cached_state = NULL;
2285         u64 lockstart = *offset;
2286         u64 lockend = i_size_read(inode);
2287         u64 start = *offset;
2288         u64 orig_start = *offset;
2289         u64 len = i_size_read(inode);
2290         u64 last_end = 0;
2291         int ret = 0;
2292
2293         lockend = max_t(u64, root->sectorsize, lockend);
2294         if (lockend <= lockstart)
2295                 lockend = lockstart + root->sectorsize;
2296
2297         lockend--;
2298         len = lockend - lockstart + 1;
2299
2300         len = max_t(u64, len, root->sectorsize);
2301         if (inode->i_size == 0)
2302                 return -ENXIO;
2303
2304         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2305                          &cached_state);
2306
2307         /*
2308          * Delalloc is such a pain.  If we have a hole and we have pending
2309          * delalloc for a portion of the hole we will get back a hole that
2310          * exists for the entire range since it hasn't been actually written
2311          * yet.  So to take care of this case we need to look for an extent just
2312          * before the position we want in case there is outstanding delalloc
2313          * going on here.
2314          */
2315         if (whence == SEEK_HOLE && start != 0) {
2316                 if (start <= root->sectorsize)
2317                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
2318                                                      root->sectorsize, 0);
2319                 else
2320                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
2321                                                      start - root->sectorsize,
2322                                                      root->sectorsize, 0);
2323                 if (IS_ERR(em)) {
2324                         ret = PTR_ERR(em);
2325                         goto out;
2326                 }
2327                 last_end = em->start + em->len;
2328                 if (em->block_start == EXTENT_MAP_DELALLOC)
2329                         last_end = min_t(u64, last_end, inode->i_size);
2330                 free_extent_map(em);
2331         }
2332
2333         while (1) {
2334                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2335                 if (IS_ERR(em)) {
2336                         ret = PTR_ERR(em);
2337                         break;
2338                 }
2339
2340                 if (em->block_start == EXTENT_MAP_HOLE) {
2341                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2342                                 if (last_end <= orig_start) {
2343                                         free_extent_map(em);
2344                                         ret = -ENXIO;
2345                                         break;
2346                                 }
2347                         }
2348
2349                         if (whence == SEEK_HOLE) {
2350                                 *offset = start;
2351                                 free_extent_map(em);
2352                                 break;
2353                         }
2354                 } else {
2355                         if (whence == SEEK_DATA) {
2356                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
2357                                         if (start >= inode->i_size) {
2358                                                 free_extent_map(em);
2359                                                 ret = -ENXIO;
2360                                                 break;
2361                                         }
2362                                 }
2363
2364                                 if (!test_bit(EXTENT_FLAG_PREALLOC,
2365                                               &em->flags)) {
2366                                         *offset = start;
2367                                         free_extent_map(em);
2368                                         break;
2369                                 }
2370                         }
2371                 }
2372
2373                 start = em->start + em->len;
2374                 last_end = em->start + em->len;
2375
2376                 if (em->block_start == EXTENT_MAP_DELALLOC)
2377                         last_end = min_t(u64, last_end, inode->i_size);
2378
2379                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2380                         free_extent_map(em);
2381                         ret = -ENXIO;
2382                         break;
2383                 }
2384                 free_extent_map(em);
2385                 cond_resched();
2386         }
2387         if (!ret)
2388                 *offset = min(*offset, inode->i_size);
2389 out:
2390         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2391                              &cached_state, GFP_NOFS);
2392         return ret;
2393 }
2394
2395 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2396 {
2397         struct inode *inode = file->f_mapping->host;
2398         int ret;
2399
2400         mutex_lock(&inode->i_mutex);
2401         switch (whence) {
2402         case SEEK_END:
2403         case SEEK_CUR:
2404                 offset = generic_file_llseek(file, offset, whence);
2405                 goto out;
2406         case SEEK_DATA:
2407         case SEEK_HOLE:
2408                 if (offset >= i_size_read(inode)) {
2409                         mutex_unlock(&inode->i_mutex);
2410                         return -ENXIO;
2411                 }
2412
2413                 ret = find_desired_extent(inode, &offset, whence);
2414                 if (ret) {
2415                         mutex_unlock(&inode->i_mutex);
2416                         return ret;
2417                 }
2418         }
2419
2420         if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
2421                 offset = -EINVAL;
2422                 goto out;
2423         }
2424         if (offset > inode->i_sb->s_maxbytes) {
2425                 offset = -EINVAL;
2426                 goto out;
2427         }
2428
2429         /* Special lock needed here? */
2430         if (offset != file->f_pos) {
2431                 file->f_pos = offset;
2432                 file->f_version = 0;
2433         }
2434 out:
2435         mutex_unlock(&inode->i_mutex);
2436         return offset;
2437 }
2438
2439 const struct file_operations btrfs_file_operations = {
2440         .llseek         = btrfs_file_llseek,
2441         .read           = do_sync_read,
2442         .write          = do_sync_write,
2443         .aio_read       = generic_file_aio_read,
2444         .splice_read    = generic_file_splice_read,
2445         .aio_write      = btrfs_file_aio_write,
2446         .mmap           = btrfs_file_mmap,
2447         .open           = generic_file_open,
2448         .release        = btrfs_release_file,
2449         .fsync          = btrfs_sync_file,
2450         .fallocate      = btrfs_fallocate,
2451         .unlocked_ioctl = btrfs_ioctl,
2452 #ifdef CONFIG_COMPAT
2453         .compat_ioctl   = btrfs_ioctl,
2454 #endif
2455 };
2456
2457 void btrfs_auto_defrag_exit(void)
2458 {
2459         if (btrfs_inode_defrag_cachep)
2460                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2461 }
2462
2463 int btrfs_auto_defrag_init(void)
2464 {
2465         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2466                                         sizeof(struct inode_defrag), 0,
2467                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2468                                         NULL);
2469         if (!btrfs_inode_defrag_cachep)
2470                 return -ENOMEM;
2471
2472         return 0;
2473 }