fs/btrfs/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/fs.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/time.h>
   9 #include <linux/init.h>
  10 #include <linux/string.h>
  11 #include <linux/backing-dev.h>
  12 #include <linux/falloc.h>
  13 #include <linux/writeback.h>
  14 #include <linux/compat.h>
  15 #include <linux/slab.h>
  16 #include <linux/btrfs.h>
  17 #include <linux/uio.h>
  18 #include <linux/iversion.h>
  19 #include "ctree.h"
  20 #include "disk-io.h"
  21 #include "transaction.h"
  22 #include "btrfs_inode.h"
  23 #include "print-tree.h"
  24 #include "tree-log.h"
  25 #include "locking.h"
  26 #include "volumes.h"
  27 #include "qgroup.h"
  28 #include "compression.h"
  29 #include "delalloc-space.h"
  30 #include "reflink.h"
  31
  32 static struct kmem_cache *btrfs_inode_defrag_cachep;
  33 /*
  34  * when auto defrag is enabled we
  35  * queue up these defrag structs to remember which
  36  * inodes need defragging passes
  37  */
  38 struct inode_defrag {
  39         struct rb_node rb_node;
  40         /* objectid */
  41         u64 ino;
  42         /*
  43          * transid where the defrag was added, we search for
  44          * extents newer than this
  45          */
  46         u64 transid;
  47
  48         /* root objectid */
  49         u64 root;
  50
  51         /* last offset we were able to defrag */
  52         u64 last_offset;
  53
  54         /* if we've wrapped around back to zero once already */
  55         int cycled;
  56 };
  57
  58 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  59                                   struct inode_defrag *defrag2)
  60 {
  61         if (defrag1->root > defrag2->root)
  62                 return 1;
  63         else if (defrag1->root < defrag2->root)
  64                 return -1;
  65         else if (defrag1->ino > defrag2->ino)
  66                 return 1;
  67         else if (defrag1->ino < defrag2->ino)
  68                 return -1;
  69         else
  70                 return 0;
  71 }
  72
  73 /* pop a record for an inode into the defrag tree.  The lock
  74  * must be held already
  75  *
  76  * If you're inserting a record for an older transid than an
  77  * existing record, the transid already in the tree is lowered
  78  *
  79  * If an existing record is found the defrag item you
  80  * pass in is freed
  81  */
  82 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  83                                     struct inode_defrag *defrag)
  84 {
  85         struct btrfs_fs_info *fs_info = inode->root->fs_info;
  86         struct inode_defrag *entry;
  87         struct rb_node **p;
  88         struct rb_node *parent = NULL;
  89         int ret;
  90
  91         p = &fs_info->defrag_inodes.rb_node;
  92         while (*p) {
  93                 parent = *p;
  94                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  95
  96                 ret = __compare_inode_defrag(defrag, entry);
  97                 if (ret < 0)
  98                         p = &parent->rb_left;
  99                 else if (ret > 0)
 100                         p = &parent->rb_right;
 101                 else {
 102                         /* if we're reinserting an entry for
 103                          * an old defrag run, make sure to
 104                          * lower the transid of our existing record
 105                          */
 106                         if (defrag->transid < entry->transid)
 107                                 entry->transid = defrag->transid;
 108                         if (defrag->last_offset > entry->last_offset)
 109                                 entry->last_offset = defrag->last_offset;
 110                         return -EEXIST;
 111                 }
 112         }
 113         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 114         rb_link_node(&defrag->rb_node, parent, p);
 115         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 116         return 0;
 117 }
 118
 119 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 120 {
 121         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 122                 return 0;
 123
 124         if (btrfs_fs_closing(fs_info))
 125                 return 0;
 126
 127         return 1;
 128 }
 129
 130 /*
 131  * insert a defrag record for this inode if auto defrag is
 132  * enabled
 133  */
 134 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 135                            struct btrfs_inode *inode)
 136 {
 137         struct btrfs_root *root = inode->root;
 138         struct btrfs_fs_info *fs_info = root->fs_info;
 139         struct inode_defrag *defrag;
 140         u64 transid;
 141         int ret;
 142
 143         if (!__need_auto_defrag(fs_info))
 144                 return 0;
 145
 146         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 147                 return 0;
 148
 149         if (trans)
 150                 transid = trans->transid;
 151         else
 152                 transid = inode->root->last_trans;
 153
 154         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 155         if (!defrag)
 156                 return -ENOMEM;
 157
 158         defrag->ino = btrfs_ino(inode);
 159         defrag->transid = transid;
 160         defrag->root = root->root_key.objectid;
 161
 162         spin_lock(&fs_info->defrag_inodes_lock);
 163         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 164                 /*
 165                  * If we set IN_DEFRAG flag and evict the inode from memory,
 166                  * and then re-read this inode, this new inode doesn't have
 167                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 168                  */
 169                 ret = __btrfs_add_inode_defrag(inode, defrag);
 170                 if (ret)
 171                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 172         } else {
 173                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 174         }
 175         spin_unlock(&fs_info->defrag_inodes_lock);
 176         return 0;
 177 }
 178
 179 /*
 180  * Requeue the defrag object. If there is a defrag object that points to
 181  * the same inode in the tree, we will merge them together (by
 182  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 183  */
 184 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 185                                        struct inode_defrag *defrag)
 186 {
 187         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 188         int ret;
 189
 190         if (!__need_auto_defrag(fs_info))
 191                 goto out;
 192
 193         /*
 194          * Here we don't check the IN_DEFRAG flag, because we need merge
 195          * them together.
 196          */
 197         spin_lock(&fs_info->defrag_inodes_lock);
 198         ret = __btrfs_add_inode_defrag(inode, defrag);
 199         spin_unlock(&fs_info->defrag_inodes_lock);
 200         if (ret)
 201                 goto out;
 202         return;
 203 out:
 204         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 205 }
 206
 207 /*
 208  * pick the defragable inode that we want, if it doesn't exist, we will get
 209  * the next one.
 210  */
 211 static struct inode_defrag *
 212 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 213 {
 214         struct inode_defrag *entry = NULL;
 215         struct inode_defrag tmp;
 216         struct rb_node *p;
 217         struct rb_node *parent = NULL;
 218         int ret;
 219
 220         tmp.ino = ino;
 221         tmp.root = root;
 222
 223         spin_lock(&fs_info->defrag_inodes_lock);
 224         p = fs_info->defrag_inodes.rb_node;
 225         while (p) {
 226                 parent = p;
 227                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 228
 229                 ret = __compare_inode_defrag(&tmp, entry);
 230                 if (ret < 0)
 231                         p = parent->rb_left;
 232                 else if (ret > 0)
 233                         p = parent->rb_right;
 234                 else
 235                         goto out;
 236         }
 237
 238         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 239                 parent = rb_next(parent);
 240                 if (parent)
 241                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 242                 else
 243                         entry = NULL;
 244         }
 245 out:
 246         if (entry)
 247                 rb_erase(parent, &fs_info->defrag_inodes);
 248         spin_unlock(&fs_info->defrag_inodes_lock);
 249         return entry;
 250 }
 251
 252 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 253 {
 254         struct inode_defrag *defrag;
 255         struct rb_node *node;
 256
 257         spin_lock(&fs_info->defrag_inodes_lock);
 258         node = rb_first(&fs_info->defrag_inodes);
 259         while (node) {
 260                 rb_erase(node, &fs_info->defrag_inodes);
 261                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 262                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 263
 264                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 265
 266                 node = rb_first(&fs_info->defrag_inodes);
 267         }
 268         spin_unlock(&fs_info->defrag_inodes_lock);
 269 }
 270
 271 #define BTRFS_DEFRAG_BATCH      1024
 272
 273 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 274                                     struct inode_defrag *defrag)
 275 {
 276         struct btrfs_root *inode_root;
 277         struct inode *inode;
 278         struct btrfs_ioctl_defrag_range_args range;
 279         int num_defrag;
 280         int ret;
 281
 282         /* get the inode */
 283         inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
 284         if (IS_ERR(inode_root)) {
 285                 ret = PTR_ERR(inode_root);
 286                 goto cleanup;
 287         }
 288
 289         inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
 290         btrfs_put_root(inode_root);
 291         if (IS_ERR(inode)) {
 292                 ret = PTR_ERR(inode);
 293                 goto cleanup;
 294         }
 295
 296         /* do a chunk of defrag */
 297         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 298         memset(&range, 0, sizeof(range));
 299         range.len = (u64)-1;
 300         range.start = defrag->last_offset;
 301
 302         sb_start_write(fs_info->sb);
 303         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 304                                        BTRFS_DEFRAG_BATCH);
 305         sb_end_write(fs_info->sb);
 306         /*
 307          * if we filled the whole defrag batch, there
 308          * must be more work to do.  Queue this defrag
 309          * again
 310          */
 311         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 312                 defrag->last_offset = range.start;
 313                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 314         } else if (defrag->last_offset && !defrag->cycled) {
 315                 /*
 316                  * we didn't fill our defrag batch, but
 317                  * we didn't start at zero.  Make sure we loop
 318                  * around to the start of the file.
 319                  */
 320                 defrag->last_offset = 0;
 321                 defrag->cycled = 1;
 322                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 323         } else {
 324                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 325         }
 326
 327         iput(inode);
 328         return 0;
 329 cleanup:
 330         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 331         return ret;
 332 }
 333
 334 /*
 335  * run through the list of inodes in the FS that need
 336  * defragging
 337  */
 338 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 339 {
 340         struct inode_defrag *defrag;
 341         u64 first_ino = 0;
 342         u64 root_objectid = 0;
 343
 344         atomic_inc(&fs_info->defrag_running);
 345         while (1) {
 346                 /* Pause the auto defragger. */
 347                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 348                              &fs_info->fs_state))
 349                         break;
 350
 351                 if (!__need_auto_defrag(fs_info))
 352                         break;
 353
 354                 /* find an inode to defrag */
 355                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 356                                                  first_ino);
 357                 if (!defrag) {
 358                         if (root_objectid || first_ino) {
 359                                 root_objectid = 0;
 360                                 first_ino = 0;
 361                                 continue;
 362                         } else {
 363                                 break;
 364                         }
 365                 }
 366
 367                 first_ino = defrag->ino + 1;
 368                 root_objectid = defrag->root;
 369
 370                 __btrfs_run_defrag_inode(fs_info, defrag);
 371         }
 372         atomic_dec(&fs_info->defrag_running);
 373
 374         /*
 375          * during unmount, we use the transaction_wait queue to
 376          * wait for the defragger to stop
 377          */
 378         wake_up(&fs_info->transaction_wait);
 379         return 0;
 380 }
 381
 382 /* simple helper to fault in pages and copy.  This should go away
 383  * and be replaced with calls into generic code.
 384  */
 385 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 386                                          struct page **prepared_pages,
 387                                          struct iov_iter *i)
 388 {
 389         size_t copied = 0;
 390         size_t total_copied = 0;
 391         int pg = 0;
 392         int offset = offset_in_page(pos);
 393
 394         while (write_bytes > 0) {
 395                 size_t count = min_t(size_t,
 396                                      PAGE_SIZE - offset, write_bytes);
 397                 struct page *page = prepared_pages[pg];
 398                 /*
 399                  * Copy data from userspace to the current page
 400                  */
 401                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 402
 403                 /* Flush processor's dcache for this page */
 404                 flush_dcache_page(page);
 405
 406                 /*
 407                  * if we get a partial write, we can end up with
 408                  * partially up to date pages.  These add
 409                  * a lot of complexity, so make sure they don't
 410                  * happen by forcing this copy to be retried.
 411                  *
 412                  * The rest of the btrfs_file_write code will fall
 413                  * back to page at a time copies after we return 0.
 414                  */
 415                 if (!PageUptodate(page) && copied < count)
 416                         copied = 0;
 417
 418                 iov_iter_advance(i, copied);
 419                 write_bytes -= copied;
 420                 total_copied += copied;
 421
 422                 /* Return to btrfs_file_write_iter to fault page */
 423                 if (unlikely(copied == 0))
 424                         break;
 425
 426                 if (copied < PAGE_SIZE - offset) {
 427                         offset += copied;
 428                 } else {
 429                         pg++;
 430                         offset = 0;
 431                 }
 432         }
 433         return total_copied;
 434 }
 435
 436 /*
 437  * unlocks pages after btrfs_file_write is done with them
 438  */
 439 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 440 {
 441         size_t i;
 442         for (i = 0; i < num_pages; i++) {
 443                 /* page checked is some magic around finding pages that
 444                  * have been modified without going through btrfs_set_page_dirty
 445                  * clear it here. There should be no need to mark the pages
 446                  * accessed as prepare_pages should have marked them accessed
 447                  * in prepare_pages via find_or_create_page()
 448                  */
 449                 ClearPageChecked(pages[i]);
 450                 unlock_page(pages[i]);
 451                 put_page(pages[i]);
 452         }
 453 }
 454
 455 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
 456                                          const u64 start,
 457                                          const u64 len,
 458                                          struct extent_state **cached_state)
 459 {
 460         u64 search_start = start;
 461         const u64 end = start + len - 1;
 462
 463         while (search_start < end) {
 464                 const u64 search_len = end - search_start + 1;
 465                 struct extent_map *em;
 466                 u64 em_len;
 467                 int ret = 0;
 468
 469                 em = btrfs_get_extent(inode, NULL, 0, search_start, search_len);
 470                 if (IS_ERR(em))
 471                         return PTR_ERR(em);
 472
 473                 if (em->block_start != EXTENT_MAP_HOLE)
 474                         goto next;
 475
 476                 em_len = em->len;
 477                 if (em->start < search_start)
 478                         em_len -= search_start - em->start;
 479                 if (em_len > search_len)
 480                         em_len = search_len;
 481
 482                 ret = set_extent_bit(&inode->io_tree, search_start,
 483                                      search_start + em_len - 1,
 484                                      EXTENT_DELALLOC_NEW,
 485                                      NULL, cached_state, GFP_NOFS);
 486 next:
 487                 search_start = extent_map_end(em);
 488                 free_extent_map(em);
 489                 if (ret)
 490                         return ret;
 491         }
 492         return 0;
 493 }
 494
 495 /*
 496  * after copy_from_user, pages need to be dirtied and we need to make
 497  * sure holes are created between the current EOF and the start of
 498  * any next extents (if required).
 499  *
 500  * this also makes the decision about creating an inline extent vs
 501  * doing real data extents, marking pages dirty and delalloc as required.
 502  */
 503 int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
 504                       size_t num_pages, loff_t pos, size_t write_bytes,
 505                       struct extent_state **cached)
 506 {
 507         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 508         int err = 0;
 509         int i;
 510         u64 num_bytes;
 511         u64 start_pos;
 512         u64 end_of_last_block;
 513         u64 end_pos = pos + write_bytes;
 514         loff_t isize = i_size_read(&inode->vfs_inode);
 515         unsigned int extra_bits = 0;
 516
 517         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 518         num_bytes = round_up(write_bytes + pos - start_pos,
 519                              fs_info->sectorsize);
 520
 521         end_of_last_block = start_pos + num_bytes - 1;
 522
 523         /*
 524          * The pages may have already been dirty, clear out old accounting so
 525          * we can set things up properly
 526          */
 527         clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
 528                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
 529                          0, 0, cached);
 530
 531         if (!btrfs_is_free_space_inode(inode)) {
 532                 if (start_pos >= isize &&
 533                     !(inode->flags & BTRFS_INODE_PREALLOC)) {
 534                         /*
 535                          * There can't be any extents following eof in this case
 536                          * so just set the delalloc new bit for the range
 537                          * directly.
 538                          */
 539                         extra_bits |= EXTENT_DELALLOC_NEW;
 540                 } else {
 541                         err = btrfs_find_new_delalloc_bytes(inode, start_pos,
 542                                                             num_bytes, cached);
 543                         if (err)
 544                                 return err;
 545                 }
 546         }
 547
 548         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 549                                         extra_bits, cached);
 550         if (err)
 551                 return err;
 552
 553         for (i = 0; i < num_pages; i++) {
 554                 struct page *p = pages[i];
 555                 SetPageUptodate(p);
 556                 ClearPageChecked(p);
 557                 set_page_dirty(p);
 558         }
 559
 560         /*
 561          * we've only changed i_size in ram, and we haven't updated
 562          * the disk i_size.  There is no need to log the inode
 563          * at this time.
 564          */
 565         if (end_pos > isize)
 566                 i_size_write(&inode->vfs_inode, end_pos);
 567         return 0;
 568 }
 569
 570 /*
 571  * this drops all the extents in the cache that intersect the range
 572  * [start, end].  Existing extents are split as required.
 573  */
 574 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 575                              int skip_pinned)
 576 {
 577         struct extent_map *em;
 578         struct extent_map *split = NULL;
 579         struct extent_map *split2 = NULL;
 580         struct extent_map_tree *em_tree = &inode->extent_tree;
 581         u64 len = end - start + 1;
 582         u64 gen;
 583         int ret;
 584         int testend = 1;
 585         unsigned long flags;
 586         int compressed = 0;
 587         bool modified;
 588
 589         WARN_ON(end < start);
 590         if (end == (u64)-1) {
 591                 len = (u64)-1;
 592                 testend = 0;
 593         }
 594         while (1) {
 595                 int no_splits = 0;
 596
 597                 modified = false;
 598                 if (!split)
 599                         split = alloc_extent_map();
 600                 if (!split2)
 601                         split2 = alloc_extent_map();
 602                 if (!split || !split2)
 603                         no_splits = 1;
 604
 605                 write_lock(&em_tree->lock);
 606                 em = lookup_extent_mapping(em_tree, start, len);
 607                 if (!em) {
 608                         write_unlock(&em_tree->lock);
 609                         break;
 610                 }
 611                 flags = em->flags;
 612                 gen = em->generation;
 613                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 614                         if (testend && em->start + em->len >= start + len) {
 615                                 free_extent_map(em);
 616                                 write_unlock(&em_tree->lock);
 617                                 break;
 618                         }
 619                         start = em->start + em->len;
 620                         if (testend)
 621                                 len = start + len - (em->start + em->len);
 622                         free_extent_map(em);
 623                         write_unlock(&em_tree->lock);
 624                         continue;
 625                 }
 626                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 627                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 628                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 629                 modified = !list_empty(&em->list);
 630                 if (no_splits)
 631                         goto next;
 632
 633                 if (em->start < start) {
 634                         split->start = em->start;
 635                         split->len = start - em->start;
 636
 637                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 638                                 split->orig_start = em->orig_start;
 639                                 split->block_start = em->block_start;
 640
 641                                 if (compressed)
 642                                         split->block_len = em->block_len;
 643                                 else
 644                                         split->block_len = split->len;
 645                                 split->orig_block_len = max(split->block_len,
 646                                                 em->orig_block_len);
 647                                 split->ram_bytes = em->ram_bytes;
 648                         } else {
 649                                 split->orig_start = split->start;
 650                                 split->block_len = 0;
 651                                 split->block_start = em->block_start;
 652                                 split->orig_block_len = 0;
 653                                 split->ram_bytes = split->len;
 654                         }
 655
 656                         split->generation = gen;
 657                         split->flags = flags;
 658                         split->compress_type = em->compress_type;
 659                         replace_extent_mapping(em_tree, em, split, modified);
 660                         free_extent_map(split);
 661                         split = split2;
 662                         split2 = NULL;
 663                 }
 664                 if (testend && em->start + em->len > start + len) {
 665                         u64 diff = start + len - em->start;
 666
 667                         split->start = start + len;
 668                         split->len = em->start + em->len - (start + len);
 669                         split->flags = flags;
 670                         split->compress_type = em->compress_type;
 671                         split->generation = gen;
 672
 673                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 674                                 split->orig_block_len = max(em->block_len,
 675                                                     em->orig_block_len);
 676
 677                                 split->ram_bytes = em->ram_bytes;
 678                                 if (compressed) {
 679                                         split->block_len = em->block_len;
 680                                         split->block_start = em->block_start;
 681                                         split->orig_start = em->orig_start;
 682                                 } else {
 683                                         split->block_len = split->len;
 684                                         split->block_start = em->block_start
 685                                                 + diff;
 686                                         split->orig_start = em->orig_start;
 687                                 }
 688                         } else {
 689                                 split->ram_bytes = split->len;
 690                                 split->orig_start = split->start;
 691                                 split->block_len = 0;
 692                                 split->block_start = em->block_start;
 693                                 split->orig_block_len = 0;
 694                         }
 695
 696                         if (extent_map_in_tree(em)) {
 697                                 replace_extent_mapping(em_tree, em, split,
 698                                                        modified);
 699                         } else {
 700                                 ret = add_extent_mapping(em_tree, split,
 701                                                          modified);
 702                                 ASSERT(ret == 0); /* Logic error */
 703                         }
 704                         free_extent_map(split);
 705                         split = NULL;
 706                 }
 707 next:
 708                 if (extent_map_in_tree(em))
 709                         remove_extent_mapping(em_tree, em);
 710                 write_unlock(&em_tree->lock);
 711
 712                 /* once for us */
 713                 free_extent_map(em);
 714                 /* once for the tree*/
 715                 free_extent_map(em);
 716         }
 717         if (split)
 718                 free_extent_map(split);
 719         if (split2)
 720                 free_extent_map(split2);
 721 }
 722
 723 /*
 724  * this is very complex, but the basic idea is to drop all extents
 725  * in the range start - end.  hint_block is filled in with a block number
 726  * that would be a good hint to the block allocator for this file.
 727  *
 728  * If an extent intersects the range but is not entirely inside the range
 729  * it is either truncated or split.  Anything entirely inside the range
 730  * is deleted from the tree.
 731  */
 732 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 733                          struct btrfs_root *root, struct btrfs_inode *inode,
 734                          struct btrfs_path *path, u64 start, u64 end,
 735                          u64 *drop_end, int drop_cache,
 736                          int replace_extent,
 737                          u32 extent_item_size,
 738                          int *key_inserted)
 739 {
 740         struct btrfs_fs_info *fs_info = root->fs_info;
 741         struct extent_buffer *leaf;
 742         struct btrfs_file_extent_item *fi;
 743         struct btrfs_ref ref = { 0 };
 744         struct btrfs_key key;
 745         struct btrfs_key new_key;
 746         struct inode *vfs_inode = &inode->vfs_inode;
 747         u64 ino = btrfs_ino(inode);
 748         u64 search_start = start;
 749         u64 disk_bytenr = 0;
 750         u64 num_bytes = 0;
 751         u64 extent_offset = 0;
 752         u64 extent_end = 0;
 753         u64 last_end = start;
 754         int del_nr = 0;
 755         int del_slot = 0;
 756         int extent_type;
 757         int recow;
 758         int ret;
 759         int modify_tree = -1;
 760         int update_refs;
 761         int found = 0;
 762         int leafs_visited = 0;
 763
 764         if (drop_cache)
 765                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 766
 767         if (start >= inode->disk_i_size && !replace_extent)
 768                 modify_tree = 0;
 769
 770         update_refs = (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 771                        root == fs_info->tree_root);
 772         while (1) {
 773                 recow = 0;
 774                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 775                                                search_start, modify_tree);
 776                 if (ret < 0)
 777                         break;
 778                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 779                         leaf = path->nodes[0];
 780                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 781                         if (key.objectid == ino &&
 782                             key.type == BTRFS_EXTENT_DATA_KEY)
 783                                 path->slots[0]--;
 784                 }
 785                 ret = 0;
 786                 leafs_visited++;
 787 next_slot:
 788                 leaf = path->nodes[0];
 789                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 790                         BUG_ON(del_nr > 0);
 791                         ret = btrfs_next_leaf(root, path);
 792                         if (ret < 0)
 793                                 break;
 794                         if (ret > 0) {
 795                                 ret = 0;
 796                                 break;
 797                         }
 798                         leafs_visited++;
 799                         leaf = path->nodes[0];
 800                         recow = 1;
 801                 }
 802
 803                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 804
 805                 if (key.objectid > ino)
 806                         break;
 807                 if (WARN_ON_ONCE(key.objectid < ino) ||
 808                     key.type < BTRFS_EXTENT_DATA_KEY) {
 809                         ASSERT(del_nr == 0);
 810                         path->slots[0]++;
 811                         goto next_slot;
 812                 }
 813                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 814                         break;
 815
 816                 fi = btrfs_item_ptr(leaf, path->slots[0],
 817                                     struct btrfs_file_extent_item);
 818                 extent_type = btrfs_file_extent_type(leaf, fi);
 819
 820                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 821                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 822                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 823                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 824                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 825                         extent_end = key.offset +
 826                                 btrfs_file_extent_num_bytes(leaf, fi);
 827                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 828                         extent_end = key.offset +
 829                                 btrfs_file_extent_ram_bytes(leaf, fi);
 830                 } else {
 831                         /* can't happen */
 832                         BUG();
 833                 }
 834
 835                 /*
 836                  * Don't skip extent items representing 0 byte lengths. They
 837                  * used to be created (bug) if while punching holes we hit
 838                  * -ENOSPC condition. So if we find one here, just ensure we
 839                  * delete it, otherwise we would insert a new file extent item
 840                  * with the same key (offset) as that 0 bytes length file
 841                  * extent item in the call to setup_items_for_insert() later
 842                  * in this function.
 843                  */
 844                 if (extent_end == key.offset && extent_end >= search_start) {
 845                         last_end = extent_end;
 846                         goto delete_extent_item;
 847                 }
 848
 849                 if (extent_end <= search_start) {
 850                         path->slots[0]++;
 851                         goto next_slot;
 852                 }
 853
 854                 found = 1;
 855                 search_start = max(key.offset, start);
 856                 if (recow || !modify_tree) {
 857                         modify_tree = -1;
 858                         btrfs_release_path(path);
 859                         continue;
 860                 }
 861
 862                 /*
 863                  *     | - range to drop - |
 864                  *  | -------- extent -------- |
 865                  */
 866                 if (start > key.offset && end < extent_end) {
 867                         BUG_ON(del_nr > 0);
 868                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 869                                 ret = -EOPNOTSUPP;
 870                                 break;
 871                         }
 872
 873                         memcpy(&new_key, &key, sizeof(new_key));
 874                         new_key.offset = start;
 875                         ret = btrfs_duplicate_item(trans, root, path,
 876                                                    &new_key);
 877                         if (ret == -EAGAIN) {
 878                                 btrfs_release_path(path);
 879                                 continue;
 880                         }
 881                         if (ret < 0)
 882                                 break;
 883
 884                         leaf = path->nodes[0];
 885                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 886                                             struct btrfs_file_extent_item);
 887                         btrfs_set_file_extent_num_bytes(leaf, fi,
 888                                                         start - key.offset);
 889
 890                         fi = btrfs_item_ptr(leaf, path->slots[0],
 891                                             struct btrfs_file_extent_item);
 892
 893                         extent_offset += start - key.offset;
 894                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 895                         btrfs_set_file_extent_num_bytes(leaf, fi,
 896                                                         extent_end - start);
 897                         btrfs_mark_buffer_dirty(leaf);
 898
 899                         if (update_refs && disk_bytenr > 0) {
 900                                 btrfs_init_generic_ref(&ref,
 901                                                 BTRFS_ADD_DELAYED_REF,
 902                                                 disk_bytenr, num_bytes, 0);
 903                                 btrfs_init_data_ref(&ref,
 904                                                 root->root_key.objectid,
 905                                                 new_key.objectid,
 906                                                 start - extent_offset);
 907                                 ret = btrfs_inc_extent_ref(trans, &ref);
 908                                 BUG_ON(ret); /* -ENOMEM */
 909                         }
 910                         key.offset = start;
 911                 }
 912                 /*
 913                  * From here on out we will have actually dropped something, so
 914                  * last_end can be updated.
 915                  */
 916                 last_end = extent_end;
 917
 918                 /*
 919                  *  | ---- range to drop ----- |
 920                  *      | -------- extent -------- |
 921                  */
 922                 if (start <= key.offset && end < extent_end) {
 923                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 924                                 ret = -EOPNOTSUPP;
 925                                 break;
 926                         }
 927
 928                         memcpy(&new_key, &key, sizeof(new_key));
 929                         new_key.offset = end;
 930                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 931
 932                         extent_offset += end - key.offset;
 933                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 934                         btrfs_set_file_extent_num_bytes(leaf, fi,
 935                                                         extent_end - end);
 936                         btrfs_mark_buffer_dirty(leaf);
 937                         if (update_refs && disk_bytenr > 0)
 938                                 inode_sub_bytes(vfs_inode, end - key.offset);
 939                         break;
 940                 }
 941
 942                 search_start = extent_end;
 943                 /*
 944                  *       | ---- range to drop ----- |
 945                  *  | -------- extent -------- |
 946                  */
 947                 if (start > key.offset && end >= extent_end) {
 948                         BUG_ON(del_nr > 0);
 949                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 950                                 ret = -EOPNOTSUPP;
 951                                 break;
 952                         }
 953
 954                         btrfs_set_file_extent_num_bytes(leaf, fi,
 955                                                         start - key.offset);
 956                         btrfs_mark_buffer_dirty(leaf);
 957                         if (update_refs && disk_bytenr > 0)
 958                                 inode_sub_bytes(vfs_inode, extent_end - start);
 959                         if (end == extent_end)
 960                                 break;
 961
 962                         path->slots[0]++;
 963                         goto next_slot;
 964                 }
 965
 966                 /*
 967                  *  | ---- range to drop ----- |
 968                  *    | ------ extent ------ |
 969                  */
 970                 if (start <= key.offset && end >= extent_end) {
 971 delete_extent_item:
 972                         if (del_nr == 0) {
 973                                 del_slot = path->slots[0];
 974                                 del_nr = 1;
 975                         } else {
 976                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 977                                 del_nr++;
 978                         }
 979
 980                         if (update_refs &&
 981                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 982                                 inode_sub_bytes(vfs_inode,
 983                                                 extent_end - key.offset);
 984                                 extent_end = ALIGN(extent_end,
 985                                                    fs_info->sectorsize);
 986                         } else if (update_refs && disk_bytenr > 0) {
 987                                 btrfs_init_generic_ref(&ref,
 988                                                 BTRFS_DROP_DELAYED_REF,
 989                                                 disk_bytenr, num_bytes, 0);
 990                                 btrfs_init_data_ref(&ref,
 991                                                 root->root_key.objectid,
 992                                                 key.objectid,
 993                                                 key.offset - extent_offset);
 994                                 ret = btrfs_free_extent(trans, &ref);
 995                                 BUG_ON(ret); /* -ENOMEM */
 996                                 inode_sub_bytes(vfs_inode,
 997                                                 extent_end - key.offset);
 998                         }
 999
1000                         if (end == extent_end)
1001                                 break;
1002
1003                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1004                                 path->slots[0]++;
1005                                 goto next_slot;
1006                         }
1007
1008                         ret = btrfs_del_items(trans, root, path, del_slot,
1009                                               del_nr);
1010                         if (ret) {
1011                                 btrfs_abort_transaction(trans, ret);
1012                                 break;
1013                         }
1014
1015                         del_nr = 0;
1016                         del_slot = 0;
1017
1018                         btrfs_release_path(path);
1019                         continue;
1020                 }
1021
1022                 BUG();
1023         }
1024
1025         if (!ret && del_nr > 0) {
1026                 /*
1027                  * Set path->slots[0] to first slot, so that after the delete
1028                  * if items are move off from our leaf to its immediate left or
1029                  * right neighbor leafs, we end up with a correct and adjusted
1030                  * path->slots[0] for our insertion (if replace_extent != 0).
1031                  */
1032                 path->slots[0] = del_slot;
1033                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1034                 if (ret)
1035                         btrfs_abort_transaction(trans, ret);
1036         }
1037
1038         leaf = path->nodes[0];
1039         /*
1040          * If btrfs_del_items() was called, it might have deleted a leaf, in
1041          * which case it unlocked our path, so check path->locks[0] matches a
1042          * write lock.
1043          */
1044         if (!ret && replace_extent && leafs_visited == 1 &&
1045             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1046              path->locks[0] == BTRFS_WRITE_LOCK) &&
1047             btrfs_leaf_free_space(leaf) >=
1048             sizeof(struct btrfs_item) + extent_item_size) {
1049
1050                 key.objectid = ino;
1051                 key.type = BTRFS_EXTENT_DATA_KEY;
1052                 key.offset = start;
1053                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1054                         struct btrfs_key slot_key;
1055
1056                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1057                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1058                                 path->slots[0]++;
1059                 }
1060                 setup_items_for_insert(root, path, &key,
1061                                        &extent_item_size,
1062                                        extent_item_size,
1063                                        sizeof(struct btrfs_item) +
1064                                        extent_item_size, 1);
1065                 *key_inserted = 1;
1066         }
1067
1068         if (!replace_extent || !(*key_inserted))
1069                 btrfs_release_path(path);
1070         if (drop_end)
1071                 *drop_end = found ? min(end, last_end) : end;
1072         return ret;
1073 }
1074
1075 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1076                        struct btrfs_root *root, struct inode *inode, u64 start,
1077                        u64 end, int drop_cache)
1078 {
1079         struct btrfs_path *path;
1080         int ret;
1081
1082         path = btrfs_alloc_path();
1083         if (!path)
1084                 return -ENOMEM;
1085         ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path, start,
1086                                    end, NULL, drop_cache, 0, 0, NULL);
1087         btrfs_free_path(path);
1088         return ret;
1089 }
1090
1091 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1092                             u64 objectid, u64 bytenr, u64 orig_offset,
1093                             u64 *start, u64 *end)
1094 {
1095         struct btrfs_file_extent_item *fi;
1096         struct btrfs_key key;
1097         u64 extent_end;
1098
1099         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1100                 return 0;
1101
1102         btrfs_item_key_to_cpu(leaf, &key, slot);
1103         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1104                 return 0;
1105
1106         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1107         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1108             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1109             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1110             btrfs_file_extent_compression(leaf, fi) ||
1111             btrfs_file_extent_encryption(leaf, fi) ||
1112             btrfs_file_extent_other_encoding(leaf, fi))
1113                 return 0;
1114
1115         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1116         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1117                 return 0;
1118
1119         *start = key.offset;
1120         *end = extent_end;
1121         return 1;
1122 }
1123
1124 /*
1125  * Mark extent in the range start - end as written.
1126  *
1127  * This changes extent type from 'pre-allocated' to 'regular'. If only
1128  * part of extent is marked as written, the extent will be split into
1129  * two or three.
1130  */
1131 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1132                               struct btrfs_inode *inode, u64 start, u64 end)
1133 {
1134         struct btrfs_fs_info *fs_info = trans->fs_info;
1135         struct btrfs_root *root = inode->root;
1136         struct extent_buffer *leaf;
1137         struct btrfs_path *path;
1138         struct btrfs_file_extent_item *fi;
1139         struct btrfs_ref ref = { 0 };
1140         struct btrfs_key key;
1141         struct btrfs_key new_key;
1142         u64 bytenr;
1143         u64 num_bytes;
1144         u64 extent_end;
1145         u64 orig_offset;
1146         u64 other_start;
1147         u64 other_end;
1148         u64 split;
1149         int del_nr = 0;
1150         int del_slot = 0;
1151         int recow;
1152         int ret;
1153         u64 ino = btrfs_ino(inode);
1154
1155         path = btrfs_alloc_path();
1156         if (!path)
1157                 return -ENOMEM;
1158 again:
1159         recow = 0;
1160         split = start;
1161         key.objectid = ino;
1162         key.type = BTRFS_EXTENT_DATA_KEY;
1163         key.offset = split;
1164
1165         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1166         if (ret < 0)
1167                 goto out;
1168         if (ret > 0 && path->slots[0] > 0)
1169                 path->slots[0]--;
1170
1171         leaf = path->nodes[0];
1172         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1173         if (key.objectid != ino ||
1174             key.type != BTRFS_EXTENT_DATA_KEY) {
1175                 ret = -EINVAL;
1176                 btrfs_abort_transaction(trans, ret);
1177                 goto out;
1178         }
1179         fi = btrfs_item_ptr(leaf, path->slots[0],
1180                             struct btrfs_file_extent_item);
1181         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1182                 ret = -EINVAL;
1183                 btrfs_abort_transaction(trans, ret);
1184                 goto out;
1185         }
1186         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1187         if (key.offset > start || extent_end < end) {
1188                 ret = -EINVAL;
1189                 btrfs_abort_transaction(trans, ret);
1190                 goto out;
1191         }
1192
1193         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1194         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1195         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1196         memcpy(&new_key, &key, sizeof(new_key));
1197
1198         if (start == key.offset && end < extent_end) {
1199                 other_start = 0;
1200                 other_end = start;
1201                 if (extent_mergeable(leaf, path->slots[0] - 1,
1202                                      ino, bytenr, orig_offset,
1203                                      &other_start, &other_end)) {
1204                         new_key.offset = end;
1205                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1206                         fi = btrfs_item_ptr(leaf, path->slots[0],
1207                                             struct btrfs_file_extent_item);
1208                         btrfs_set_file_extent_generation(leaf, fi,
1209                                                          trans->transid);
1210                         btrfs_set_file_extent_num_bytes(leaf, fi,
1211                                                         extent_end - end);
1212                         btrfs_set_file_extent_offset(leaf, fi,
1213                                                      end - orig_offset);
1214                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1215                                             struct btrfs_file_extent_item);
1216                         btrfs_set_file_extent_generation(leaf, fi,
1217                                                          trans->transid);
1218                         btrfs_set_file_extent_num_bytes(leaf, fi,
1219                                                         end - other_start);
1220                         btrfs_mark_buffer_dirty(leaf);
1221                         goto out;
1222                 }
1223         }
1224
1225         if (start > key.offset && end == extent_end) {
1226                 other_start = end;
1227                 other_end = 0;
1228                 if (extent_mergeable(leaf, path->slots[0] + 1,
1229                                      ino, bytenr, orig_offset,
1230                                      &other_start, &other_end)) {
1231                         fi = btrfs_item_ptr(leaf, path->slots[0],
1232                                             struct btrfs_file_extent_item);
1233                         btrfs_set_file_extent_num_bytes(leaf, fi,
1234                                                         start - key.offset);
1235                         btrfs_set_file_extent_generation(leaf, fi,
1236                                                          trans->transid);
1237                         path->slots[0]++;
1238                         new_key.offset = start;
1239                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1240
1241                         fi = btrfs_item_ptr(leaf, path->slots[0],
1242                                             struct btrfs_file_extent_item);
1243                         btrfs_set_file_extent_generation(leaf, fi,
1244                                                          trans->transid);
1245                         btrfs_set_file_extent_num_bytes(leaf, fi,
1246                                                         other_end - start);
1247                         btrfs_set_file_extent_offset(leaf, fi,
1248                                                      start - orig_offset);
1249                         btrfs_mark_buffer_dirty(leaf);
1250                         goto out;
1251                 }
1252         }
1253
1254         while (start > key.offset || end < extent_end) {
1255                 if (key.offset == start)
1256                         split = end;
1257
1258                 new_key.offset = split;
1259                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1260                 if (ret == -EAGAIN) {
1261                         btrfs_release_path(path);
1262                         goto again;
1263                 }
1264                 if (ret < 0) {
1265                         btrfs_abort_transaction(trans, ret);
1266                         goto out;
1267                 }
1268
1269                 leaf = path->nodes[0];
1270                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1271                                     struct btrfs_file_extent_item);
1272                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1273                 btrfs_set_file_extent_num_bytes(leaf, fi,
1274                                                 split - key.offset);
1275
1276                 fi = btrfs_item_ptr(leaf, path->slots[0],
1277                                     struct btrfs_file_extent_item);
1278
1279                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1280                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1281                 btrfs_set_file_extent_num_bytes(leaf, fi,
1282                                                 extent_end - split);
1283                 btrfs_mark_buffer_dirty(leaf);
1284
1285                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1286                                        num_bytes, 0);
1287                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1288                                     orig_offset);
1289                 ret = btrfs_inc_extent_ref(trans, &ref);
1290                 if (ret) {
1291                         btrfs_abort_transaction(trans, ret);
1292                         goto out;
1293                 }
1294
1295                 if (split == start) {
1296                         key.offset = start;
1297                 } else {
1298                         if (start != key.offset) {
1299                                 ret = -EINVAL;
1300                                 btrfs_abort_transaction(trans, ret);
1301                                 goto out;
1302                         }
1303                         path->slots[0]--;
1304                         extent_end = end;
1305                 }
1306                 recow = 1;
1307         }
1308
1309         other_start = end;
1310         other_end = 0;
1311         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1312                                num_bytes, 0);
1313         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1314         if (extent_mergeable(leaf, path->slots[0] + 1,
1315                              ino, bytenr, orig_offset,
1316                              &other_start, &other_end)) {
1317                 if (recow) {
1318                         btrfs_release_path(path);
1319                         goto again;
1320                 }
1321                 extent_end = other_end;
1322                 del_slot = path->slots[0] + 1;
1323                 del_nr++;
1324                 ret = btrfs_free_extent(trans, &ref);
1325                 if (ret) {
1326                         btrfs_abort_transaction(trans, ret);
1327                         goto out;
1328                 }
1329         }
1330         other_start = 0;
1331         other_end = start;
1332         if (extent_mergeable(leaf, path->slots[0] - 1,
1333                              ino, bytenr, orig_offset,
1334                              &other_start, &other_end)) {
1335                 if (recow) {
1336                         btrfs_release_path(path);
1337                         goto again;
1338                 }
1339                 key.offset = other_start;
1340                 del_slot = path->slots[0];
1341                 del_nr++;
1342                 ret = btrfs_free_extent(trans, &ref);
1343                 if (ret) {
1344                         btrfs_abort_transaction(trans, ret);
1345                         goto out;
1346                 }
1347         }
1348         if (del_nr == 0) {
1349                 fi = btrfs_item_ptr(leaf, path->slots[0],
1350                            struct btrfs_file_extent_item);
1351                 btrfs_set_file_extent_type(leaf, fi,
1352                                            BTRFS_FILE_EXTENT_REG);
1353                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1354                 btrfs_mark_buffer_dirty(leaf);
1355         } else {
1356                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1357                            struct btrfs_file_extent_item);
1358                 btrfs_set_file_extent_type(leaf, fi,
1359                                            BTRFS_FILE_EXTENT_REG);
1360                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1361                 btrfs_set_file_extent_num_bytes(leaf, fi,
1362                                                 extent_end - key.offset);
1363                 btrfs_mark_buffer_dirty(leaf);
1364
1365                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1366                 if (ret < 0) {
1367                         btrfs_abort_transaction(trans, ret);
1368                         goto out;
1369                 }
1370         }
1371 out:
1372         btrfs_free_path(path);
1373         return 0;
1374 }
1375
1376 /*
1377  * on error we return an unlocked page and the error value
1378  * on success we return a locked page and 0
1379  */
1380 static int prepare_uptodate_page(struct inode *inode,
1381                                  struct page *page, u64 pos,
1382                                  bool force_uptodate)
1383 {
1384         int ret = 0;
1385
1386         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1387             !PageUptodate(page)) {
1388                 ret = btrfs_readpage(NULL, page);
1389                 if (ret)
1390                         return ret;
1391                 lock_page(page);
1392                 if (!PageUptodate(page)) {
1393                         unlock_page(page);
1394                         return -EIO;
1395                 }
1396                 if (page->mapping != inode->i_mapping) {
1397                         unlock_page(page);
1398                         return -EAGAIN;
1399                 }
1400         }
1401         return 0;
1402 }
1403
1404 /*
1405  * this just gets pages into the page cache and locks them down.
1406  */
1407 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1408                                   size_t num_pages, loff_t pos,
1409                                   size_t write_bytes, bool force_uptodate)
1410 {
1411         int i;
1412         unsigned long index = pos >> PAGE_SHIFT;
1413         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1414         int err = 0;
1415         int faili;
1416
1417         for (i = 0; i < num_pages; i++) {
1418 again:
1419                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1420                                                mask | __GFP_WRITE);
1421                 if (!pages[i]) {
1422                         faili = i - 1;
1423                         err = -ENOMEM;
1424                         goto fail;
1425                 }
1426
1427                 if (i == 0)
1428                         err = prepare_uptodate_page(inode, pages[i], pos,
1429                                                     force_uptodate);
1430                 if (!err && i == num_pages - 1)
1431                         err = prepare_uptodate_page(inode, pages[i],
1432                                                     pos + write_bytes, false);
1433                 if (err) {
1434                         put_page(pages[i]);
1435                         if (err == -EAGAIN) {
1436                                 err = 0;
1437                                 goto again;
1438                         }
1439                         faili = i - 1;
1440                         goto fail;
1441                 }
1442                 wait_on_page_writeback(pages[i]);
1443         }
1444
1445         return 0;
1446 fail:
1447         while (faili >= 0) {
1448                 unlock_page(pages[faili]);
1449                 put_page(pages[faili]);
1450                 faili--;
1451         }
1452         return err;
1453
1454 }
1455
1456 /*
1457  * This function locks the extent and properly waits for data=ordered extents
1458  * to finish before allowing the pages to be modified if need.
1459  *
1460  * The return value:
1461  * 1 - the extent is locked
1462  * 0 - the extent is not locked, and everything is OK
1463  * -EAGAIN - need re-prepare the pages
1464  * the other < 0 number - Something wrong happens
1465  */
1466 static noinline int
1467 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1468                                 size_t num_pages, loff_t pos,
1469                                 size_t write_bytes,
1470                                 u64 *lockstart, u64 *lockend,
1471                                 struct extent_state **cached_state)
1472 {
1473         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1474         u64 start_pos;
1475         u64 last_pos;
1476         int i;
1477         int ret = 0;
1478
1479         start_pos = round_down(pos, fs_info->sectorsize);
1480         last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1481
1482         if (start_pos < inode->vfs_inode.i_size) {
1483                 struct btrfs_ordered_extent *ordered;
1484
1485                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1486                                 cached_state);
1487                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1488                                                      last_pos - start_pos + 1);
1489                 if (ordered &&
1490                     ordered->file_offset + ordered->num_bytes > start_pos &&
1491                     ordered->file_offset <= last_pos) {
1492                         unlock_extent_cached(&inode->io_tree, start_pos,
1493                                         last_pos, cached_state);
1494                         for (i = 0; i < num_pages; i++) {
1495                                 unlock_page(pages[i]);
1496                                 put_page(pages[i]);
1497                         }
1498                         btrfs_start_ordered_extent(&inode->vfs_inode,
1499                                         ordered, 1);
1500                         btrfs_put_ordered_extent(ordered);
1501                         return -EAGAIN;
1502                 }
1503                 if (ordered)
1504                         btrfs_put_ordered_extent(ordered);
1505
1506                 *lockstart = start_pos;
1507                 *lockend = last_pos;
1508                 ret = 1;
1509         }
1510
1511         /*
1512          * It's possible the pages are dirty right now, but we don't want
1513          * to clean them yet because copy_from_user may catch a page fault
1514          * and we might have to fall back to one page at a time.  If that
1515          * happens, we'll unlock these pages and we'd have a window where
1516          * reclaim could sneak in and drop the once-dirty page on the floor
1517          * without writing it.
1518          *
1519          * We have the pages locked and the extent range locked, so there's
1520          * no way someone can start IO on any dirty pages in this range.
1521          *
1522          * We'll call btrfs_dirty_pages() later on, and that will flip around
1523          * delalloc bits and dirty the pages as required.
1524          */
1525         for (i = 0; i < num_pages; i++) {
1526                 set_page_extent_mapped(pages[i]);
1527                 WARN_ON(!PageLocked(pages[i]));
1528         }
1529
1530         return ret;
1531 }
1532
1533 static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1534                            size_t *write_bytes, bool nowait)
1535 {
1536         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1537         struct btrfs_root *root = inode->root;
1538         u64 lockstart, lockend;
1539         u64 num_bytes;
1540         int ret;
1541
1542         if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1543                 return 0;
1544
1545         if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1546                 return -EAGAIN;
1547
1548         lockstart = round_down(pos, fs_info->sectorsize);
1549         lockend = round_up(pos + *write_bytes,
1550                            fs_info->sectorsize) - 1;
1551         num_bytes = lockend - lockstart + 1;
1552
1553         if (nowait) {
1554                 struct btrfs_ordered_extent *ordered;
1555
1556                 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1557                         return -EAGAIN;
1558
1559                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1560                                                      num_bytes);
1561                 if (ordered) {
1562                         btrfs_put_ordered_extent(ordered);
1563                         ret = -EAGAIN;
1564                         goto out_unlock;
1565                 }
1566         } else {
1567                 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1568                                                    lockend, NULL);
1569         }
1570
1571         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1572                         NULL, NULL, NULL, false);
1573         if (ret <= 0) {
1574                 ret = 0;
1575                 if (!nowait)
1576                         btrfs_drew_write_unlock(&root->snapshot_lock);
1577         } else {
1578                 *write_bytes = min_t(size_t, *write_bytes ,
1579                                      num_bytes - pos + lockstart);
1580         }
1581 out_unlock:
1582         unlock_extent(&inode->io_tree, lockstart, lockend);
1583
1584         return ret;
1585 }
1586
1587 static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1588                               size_t *write_bytes)
1589 {
1590         return check_can_nocow(inode, pos, write_bytes, true);
1591 }
1592
1593 /*
1594  * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1595  *
1596  * @pos:         File offset
1597  * @write_bytes: The length to write, will be updated to the nocow writeable
1598  *               range
1599  *
1600  * This function will flush ordered extents in the range to ensure proper
1601  * nocow checks.
1602  *
1603  * Return:
1604  * >0           and update @write_bytes if we can do nocow write
1605  *  0           if we can't do nocow write
1606  * -EAGAIN      if we can't get the needed lock or there are ordered extents
1607  *              for * (nowait == true) case
1608  * <0           if other error happened
1609  *
1610  * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1611  */
1612 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1613                            size_t *write_bytes)
1614 {
1615         return check_can_nocow(inode, pos, write_bytes, false);
1616 }
1617
1618 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1619 {
1620         btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1621 }
1622
1623 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1624                                                struct iov_iter *i)
1625 {
1626         struct file *file = iocb->ki_filp;
1627         loff_t pos = iocb->ki_pos;
1628         struct inode *inode = file_inode(file);
1629         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1630         struct page **pages = NULL;
1631         struct extent_changeset *data_reserved = NULL;
1632         u64 release_bytes = 0;
1633         u64 lockstart;
1634         u64 lockend;
1635         size_t num_written = 0;
1636         int nrptrs;
1637         int ret = 0;
1638         bool only_release_metadata = false;
1639         bool force_page_uptodate = false;
1640
1641         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1642                         PAGE_SIZE / (sizeof(struct page *)));
1643         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1644         nrptrs = max(nrptrs, 8);
1645         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1646         if (!pages)
1647                 return -ENOMEM;
1648
1649         while (iov_iter_count(i) > 0) {
1650                 struct extent_state *cached_state = NULL;
1651                 size_t offset = offset_in_page(pos);
1652                 size_t sector_offset;
1653                 size_t write_bytes = min(iov_iter_count(i),
1654                                          nrptrs * (size_t)PAGE_SIZE -
1655                                          offset);
1656                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1657                                                 PAGE_SIZE);
1658                 size_t reserve_bytes;
1659                 size_t dirty_pages;
1660                 size_t copied;
1661                 size_t dirty_sectors;
1662                 size_t num_sectors;
1663                 int extents_locked;
1664
1665                 WARN_ON(num_pages > nrptrs);
1666
1667                 /*
1668                  * Fault pages before locking them in prepare_pages
1669                  * to avoid recursive lock
1670                  */
1671                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1672                         ret = -EFAULT;
1673                         break;
1674                 }
1675
1676                 only_release_metadata = false;
1677                 sector_offset = pos & (fs_info->sectorsize - 1);
1678                 reserve_bytes = round_up(write_bytes + sector_offset,
1679                                 fs_info->sectorsize);
1680
1681                 extent_changeset_release(data_reserved);
1682                 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1683                                                   &data_reserved, pos,
1684                                                   write_bytes);
1685                 if (ret < 0) {
1686                         if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1687                                                    &write_bytes) > 0) {
1688                                 /*
1689                                  * For nodata cow case, no need to reserve
1690                                  * data space.
1691                                  */
1692                                 only_release_metadata = true;
1693                                 /*
1694                                  * our prealloc extent may be smaller than
1695                                  * write_bytes, so scale down.
1696                                  */
1697                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1698                                                          PAGE_SIZE);
1699                                 reserve_bytes = round_up(write_bytes +
1700                                                          sector_offset,
1701                                                          fs_info->sectorsize);
1702                         } else {
1703                                 break;
1704                         }
1705                 }
1706
1707                 WARN_ON(reserve_bytes == 0);
1708                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1709                                 reserve_bytes);
1710                 if (ret) {
1711                         if (!only_release_metadata)
1712                                 btrfs_free_reserved_data_space(BTRFS_I(inode),
1713                                                 data_reserved, pos,
1714                                                 write_bytes);
1715                         else
1716                                 btrfs_check_nocow_unlock(BTRFS_I(inode));
1717                         break;
1718                 }
1719
1720                 release_bytes = reserve_bytes;
1721 again:
1722                 /*
1723                  * This is going to setup the pages array with the number of
1724                  * pages we want, so we don't really need to worry about the
1725                  * contents of pages from loop to loop
1726                  */
1727                 ret = prepare_pages(inode, pages, num_pages,
1728                                     pos, write_bytes,
1729                                     force_page_uptodate);
1730                 if (ret) {
1731                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1732                                                        reserve_bytes);
1733                         break;
1734                 }
1735
1736                 extents_locked = lock_and_cleanup_extent_if_need(
1737                                 BTRFS_I(inode), pages,
1738                                 num_pages, pos, write_bytes, &lockstart,
1739                                 &lockend, &cached_state);
1740                 if (extents_locked < 0) {
1741                         if (extents_locked == -EAGAIN)
1742                                 goto again;
1743                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1744                                                        reserve_bytes);
1745                         ret = extents_locked;
1746                         break;
1747                 }
1748
1749                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1750
1751                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1752                 dirty_sectors = round_up(copied + sector_offset,
1753                                         fs_info->sectorsize);
1754                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1755
1756                 /*
1757                  * if we have trouble faulting in the pages, fall
1758                  * back to one page at a time
1759                  */
1760                 if (copied < write_bytes)
1761                         nrptrs = 1;
1762
1763                 if (copied == 0) {
1764                         force_page_uptodate = true;
1765                         dirty_sectors = 0;
1766                         dirty_pages = 0;
1767                 } else {
1768                         force_page_uptodate = false;
1769                         dirty_pages = DIV_ROUND_UP(copied + offset,
1770                                                    PAGE_SIZE);
1771                 }
1772
1773                 if (num_sectors > dirty_sectors) {
1774                         /* release everything except the sectors we dirtied */
1775                         release_bytes -= dirty_sectors <<
1776                                                 fs_info->sb->s_blocksize_bits;
1777                         if (only_release_metadata) {
1778                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1779                                                         release_bytes, true);
1780                         } else {
1781                                 u64 __pos;
1782
1783                                 __pos = round_down(pos,
1784                                                    fs_info->sectorsize) +
1785                                         (dirty_pages << PAGE_SHIFT);
1786                                 btrfs_delalloc_release_space(BTRFS_I(inode),
1787                                                 data_reserved, __pos,
1788                                                 release_bytes, true);
1789                         }
1790                 }
1791
1792                 release_bytes = round_up(copied + sector_offset,
1793                                         fs_info->sectorsize);
1794
1795                 if (copied > 0)
1796                         ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1797                                                 dirty_pages, pos, copied,
1798                                                 &cached_state);
1799
1800                 /*
1801                  * If we have not locked the extent range, because the range's
1802                  * start offset is >= i_size, we might still have a non-NULL
1803                  * cached extent state, acquired while marking the extent range
1804                  * as delalloc through btrfs_dirty_pages(). Therefore free any
1805                  * possible cached extent state to avoid a memory leak.
1806                  */
1807                 if (extents_locked)
1808                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1809                                              lockstart, lockend, &cached_state);
1810                 else
1811                         free_extent_state(cached_state);
1812
1813                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1814                 if (ret) {
1815                         btrfs_drop_pages(pages, num_pages);
1816                         break;
1817                 }
1818
1819                 release_bytes = 0;
1820                 if (only_release_metadata)
1821                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1822
1823                 if (only_release_metadata && copied > 0) {
1824                         lockstart = round_down(pos,
1825                                                fs_info->sectorsize);
1826                         lockend = round_up(pos + copied,
1827                                            fs_info->sectorsize) - 1;
1828
1829                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1830                                        lockend, EXTENT_NORESERVE, NULL,
1831                                        NULL, GFP_NOFS);
1832                 }
1833
1834                 btrfs_drop_pages(pages, num_pages);
1835
1836                 cond_resched();
1837
1838                 balance_dirty_pages_ratelimited(inode->i_mapping);
1839
1840                 pos += copied;
1841                 num_written += copied;
1842         }
1843
1844         kfree(pages);
1845
1846         if (release_bytes) {
1847                 if (only_release_metadata) {
1848                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1849                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1850                                         release_bytes, true);
1851                 } else {
1852                         btrfs_delalloc_release_space(BTRFS_I(inode),
1853                                         data_reserved,
1854                                         round_down(pos, fs_info->sectorsize),
1855                                         release_bytes, true);
1856                 }
1857         }
1858
1859         extent_changeset_free(data_reserved);
1860         return num_written ? num_written : ret;
1861 }
1862
1863 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1864 {
1865         struct file *file = iocb->ki_filp;
1866         struct inode *inode = file_inode(file);
1867         loff_t pos;
1868         ssize_t written;
1869         ssize_t written_buffered;
1870         loff_t endbyte;
1871         int err;
1872
1873         written = btrfs_direct_IO(iocb, from);
1874
1875         if (written < 0 || !iov_iter_count(from))
1876                 return written;
1877
1878         pos = iocb->ki_pos;
1879         written_buffered = btrfs_buffered_write(iocb, from);
1880         if (written_buffered < 0) {
1881                 err = written_buffered;
1882                 goto out;
1883         }
1884         /*
1885          * Ensure all data is persisted. We want the next direct IO read to be
1886          * able to read what was just written.
1887          */
1888         endbyte = pos + written_buffered - 1;
1889         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1890         if (err)
1891                 goto out;
1892         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1893         if (err)
1894                 goto out;
1895         written += written_buffered;
1896         iocb->ki_pos = pos + written_buffered;
1897         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1898                                  endbyte >> PAGE_SHIFT);
1899 out:
1900         return written ? written : err;
1901 }
1902
1903 static void update_time_for_write(struct inode *inode)
1904 {
1905         struct timespec64 now;
1906
1907         if (IS_NOCMTIME(inode))
1908                 return;
1909
1910         now = current_time(inode);
1911         if (!timespec64_equal(&inode->i_mtime, &now))
1912                 inode->i_mtime = now;
1913
1914         if (!timespec64_equal(&inode->i_ctime, &now))
1915                 inode->i_ctime = now;
1916
1917         if (IS_I_VERSION(inode))
1918                 inode_inc_iversion(inode);
1919 }
1920
1921 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1922                                     struct iov_iter *from)
1923 {
1924         struct file *file = iocb->ki_filp;
1925         struct inode *inode = file_inode(file);
1926         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1927         struct btrfs_root *root = BTRFS_I(inode)->root;
1928         u64 start_pos;
1929         u64 end_pos;
1930         ssize_t num_written = 0;
1931         const bool sync = iocb->ki_flags & IOCB_DSYNC;
1932         ssize_t err;
1933         loff_t pos;
1934         size_t count;
1935         loff_t oldsize;
1936         int clean_page = 0;
1937
1938         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1939             (iocb->ki_flags & IOCB_NOWAIT))
1940                 return -EOPNOTSUPP;
1941
1942         if (iocb->ki_flags & IOCB_NOWAIT) {
1943                 if (!inode_trylock(inode))
1944                         return -EAGAIN;
1945         } else {
1946                 inode_lock(inode);
1947         }
1948
1949         err = generic_write_checks(iocb, from);
1950         if (err <= 0) {
1951                 inode_unlock(inode);
1952                 return err;
1953         }
1954
1955         pos = iocb->ki_pos;
1956         count = iov_iter_count(from);
1957         if (iocb->ki_flags & IOCB_NOWAIT) {
1958                 size_t nocow_bytes = count;
1959
1960                 /*
1961                  * We will allocate space in case nodatacow is not set,
1962                  * so bail
1963                  */
1964                 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes)
1965                     <= 0) {
1966                         inode_unlock(inode);
1967                         return -EAGAIN;
1968                 }
1969                 /*
1970                  * There are holes in the range or parts of the range that must
1971                  * be COWed (shared extents, RO block groups, etc), so just bail
1972                  * out.
1973                  */
1974                 if (nocow_bytes < count) {
1975                         inode_unlock(inode);
1976                         return -EAGAIN;
1977                 }
1978         }
1979
1980         current->backing_dev_info = inode_to_bdi(inode);
1981         err = file_remove_privs(file);
1982         if (err) {
1983                 inode_unlock(inode);
1984                 goto out;
1985         }
1986
1987         /*
1988          * If BTRFS flips readonly due to some impossible error
1989          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1990          * although we have opened a file as writable, we have
1991          * to stop this write operation to ensure FS consistency.
1992          */
1993         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1994                 inode_unlock(inode);
1995                 err = -EROFS;
1996                 goto out;
1997         }
1998
1999         /*
2000          * We reserve space for updating the inode when we reserve space for the
2001          * extent we are going to write, so we will enospc out there.  We don't
2002          * need to start yet another transaction to update the inode as we will
2003          * update the inode when we finish writing whatever data we write.
2004          */
2005         update_time_for_write(inode);
2006
2007         start_pos = round_down(pos, fs_info->sectorsize);
2008         oldsize = i_size_read(inode);
2009         if (start_pos > oldsize) {
2010                 /* Expand hole size to cover write data, preventing empty gap */
2011                 end_pos = round_up(pos + count,
2012                                    fs_info->sectorsize);
2013                 err = btrfs_cont_expand(inode, oldsize, end_pos);
2014                 if (err) {
2015                         inode_unlock(inode);
2016                         goto out;
2017                 }
2018                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
2019                         clean_page = 1;
2020         }
2021
2022         if (sync)
2023                 atomic_inc(&BTRFS_I(inode)->sync_writers);
2024
2025         if (iocb->ki_flags & IOCB_DIRECT) {
2026                 /*
2027                  * 1. We must always clear IOCB_DSYNC in order to not deadlock
2028                  *    in iomap, as it calls generic_write_sync() in this case.
2029                  * 2. If we are async, we can call iomap_dio_complete() either
2030                  *    in
2031                  *
2032                  *    2.1. A worker thread from the last bio completed.  In this
2033                  *         case we need to mark the btrfs_dio_data that it is
2034                  *         async in order to call generic_write_sync() properly.
2035                  *         This is handled by setting BTRFS_DIO_SYNC_STUB in the
2036                  *         current->journal_info.
2037                  *    2.2  The submitter context, because all IO completed
2038                  *         before we exited iomap_dio_rw().  In this case we can
2039                  *         just re-set the IOCB_DSYNC on the iocb and we'll do
2040                  *         the sync below.  If our ->end_io() gets called and
2041                  *         current->journal_info is set, then we know we're in
2042                  *         our current context and we will clear
2043                  *         current->journal_info to indicate that we need to
2044                  *         sync below.
2045                  */
2046                 if (sync) {
2047                         ASSERT(current->journal_info == NULL);
2048                         iocb->ki_flags &= ~IOCB_DSYNC;
2049                         current->journal_info = BTRFS_DIO_SYNC_STUB;
2050                 }
2051                 num_written = __btrfs_direct_write(iocb, from);
2052
2053                 /*
2054                  * As stated above, we cleared journal_info, so we need to do
2055                  * the sync ourselves.
2056                  */
2057                 if (sync && current->journal_info == NULL)
2058                         iocb->ki_flags |= IOCB_DSYNC;
2059                 current->journal_info = NULL;
2060         } else {
2061                 num_written = btrfs_buffered_write(iocb, from);
2062                 if (num_written > 0)
2063                         iocb->ki_pos = pos + num_written;
2064                 if (clean_page)
2065                         pagecache_isize_extended(inode, oldsize,
2066                                                 i_size_read(inode));
2067         }
2068
2069         inode_unlock(inode);
2070
2071         /*
2072          * We also have to set last_sub_trans to the current log transid,
2073          * otherwise subsequent syncs to a file that's been synced in this
2074          * transaction will appear to have already occurred.
2075          */
2076         spin_lock(&BTRFS_I(inode)->lock);
2077         BTRFS_I(inode)->last_sub_trans = root->log_transid;
2078         spin_unlock(&BTRFS_I(inode)->lock);
2079         if (num_written > 0)
2080                 num_written = generic_write_sync(iocb, num_written);
2081
2082         if (sync)
2083                 atomic_dec(&BTRFS_I(inode)->sync_writers);
2084 out:
2085         current->backing_dev_info = NULL;
2086         return num_written ? num_written : err;
2087 }
2088
2089 int btrfs_release_file(struct inode *inode, struct file *filp)
2090 {
2091         struct btrfs_file_private *private = filp->private_data;
2092
2093         if (private && private->filldir_buf)
2094                 kfree(private->filldir_buf);
2095         kfree(private);
2096         filp->private_data = NULL;
2097
2098         /*
2099          * ordered_data_close is set by setattr when we are about to truncate
2100          * a file from a non-zero size to a zero size.  This tries to
2101          * flush down new bytes that may have been written if the
2102          * application were using truncate to replace a file in place.
2103          */
2104         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2105                                &BTRFS_I(inode)->runtime_flags))
2106                         filemap_flush(inode->i_mapping);
2107         return 0;
2108 }
2109
2110 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2111 {
2112         int ret;
2113         struct blk_plug plug;
2114
2115         /*
2116          * This is only called in fsync, which would do synchronous writes, so
2117          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2118          * multiple disks using raid profile, a large IO can be split to
2119          * several segments of stripe length (currently 64K).
2120          */
2121         blk_start_plug(&plug);
2122         atomic_inc(&BTRFS_I(inode)->sync_writers);
2123         ret = btrfs_fdatawrite_range(inode, start, end);
2124         atomic_dec(&BTRFS_I(inode)->sync_writers);
2125         blk_finish_plug(&plug);
2126
2127         return ret;
2128 }
2129
2130 /*
2131  * fsync call for both files and directories.  This logs the inode into
2132  * the tree log instead of forcing full commits whenever possible.
2133  *
2134  * It needs to call filemap_fdatawait so that all ordered extent updates are
2135  * in the metadata btree are up to date for copying to the log.
2136  *
2137  * It drops the inode mutex before doing the tree log commit.  This is an
2138  * important optimization for directories because holding the mutex prevents
2139  * new operations on the dir while we write to disk.
2140  */
2141 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2142 {
2143         struct dentry *dentry = file_dentry(file);
2144         struct inode *inode = d_inode(dentry);
2145         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2146         struct btrfs_root *root = BTRFS_I(inode)->root;
2147         struct btrfs_trans_handle *trans;
2148         struct btrfs_log_ctx ctx;
2149         int ret = 0, err;
2150         u64 len;
2151         bool full_sync;
2152
2153         trace_btrfs_sync_file(file, datasync);
2154
2155         btrfs_init_log_ctx(&ctx, inode);
2156
2157         /*
2158          * Always set the range to a full range, otherwise we can get into
2159          * several problems, from missing file extent items to represent holes
2160          * when not using the NO_HOLES feature, to log tree corruption due to
2161          * races between hole detection during logging and completion of ordered
2162          * extents outside the range, to missing checksums due to ordered extents
2163          * for which we flushed only a subset of their pages.
2164          */
2165         start = 0;
2166         end = LLONG_MAX;
2167         len = (u64)LLONG_MAX + 1;
2168
2169         /*
2170          * We write the dirty pages in the range and wait until they complete
2171          * out of the ->i_mutex. If so, we can flush the dirty pages by
2172          * multi-task, and make the performance up.  See
2173          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2174          */
2175         ret = start_ordered_ops(inode, start, end);
2176         if (ret)
2177                 goto out;
2178
2179         inode_lock(inode);
2180
2181         /*
2182          * We take the dio_sem here because the tree log stuff can race with
2183          * lockless dio writes and get an extent map logged for an extent we
2184          * never waited on.  We need it this high up for lockdep reasons.
2185          */
2186         down_write(&BTRFS_I(inode)->dio_sem);
2187
2188         atomic_inc(&root->log_batch);
2189
2190         /*
2191          * Always check for the full sync flag while holding the inode's lock,
2192          * to avoid races with other tasks. The flag must be either set all the
2193          * time during logging or always off all the time while logging.
2194          */
2195         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2196                              &BTRFS_I(inode)->runtime_flags);
2197
2198         /*
2199          * Before we acquired the inode's lock, someone may have dirtied more
2200          * pages in the target range. We need to make sure that writeback for
2201          * any such pages does not start while we are logging the inode, because
2202          * if it does, any of the following might happen when we are not doing a
2203          * full inode sync:
2204          *
2205          * 1) We log an extent after its writeback finishes but before its
2206          *    checksums are added to the csum tree, leading to -EIO errors
2207          *    when attempting to read the extent after a log replay.
2208          *
2209          * 2) We can end up logging an extent before its writeback finishes.
2210          *    Therefore after the log replay we will have a file extent item
2211          *    pointing to an unwritten extent (and no data checksums as well).
2212          *
2213          * So trigger writeback for any eventual new dirty pages and then we
2214          * wait for all ordered extents to complete below.
2215          */
2216         ret = start_ordered_ops(inode, start, end);
2217         if (ret) {
2218                 up_write(&BTRFS_I(inode)->dio_sem);
2219                 inode_unlock(inode);
2220                 goto out;
2221         }
2222
2223         /*
2224          * We have to do this here to avoid the priority inversion of waiting on
2225          * IO of a lower priority task while holding a transaction open.
2226          *
2227          * For a full fsync we wait for the ordered extents to complete while
2228          * for a fast fsync we wait just for writeback to complete, and then
2229          * attach the ordered extents to the transaction so that a transaction
2230          * commit waits for their completion, to avoid data loss if we fsync,
2231          * the current transaction commits before the ordered extents complete
2232          * and a power failure happens right after that.
2233          */
2234         if (full_sync) {
2235                 ret = btrfs_wait_ordered_range(inode, start, len);
2236         } else {
2237                 /*
2238                  * Get our ordered extents as soon as possible to avoid doing
2239                  * checksum lookups in the csum tree, and use instead the
2240                  * checksums attached to the ordered extents.
2241                  */
2242                 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2243                                                       &ctx.ordered_extents);
2244                 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2245         }
2246
2247         if (ret)
2248                 goto out_release_extents;
2249
2250         atomic_inc(&root->log_batch);
2251
2252         /*
2253          * If we are doing a fast fsync we can not bail out if the inode's
2254          * last_trans is <= then the last committed transaction, because we only
2255          * update the last_trans of the inode during ordered extent completion,
2256          * and for a fast fsync we don't wait for that, we only wait for the
2257          * writeback to complete.
2258          */
2259         smp_mb();
2260         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2261             (BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed &&
2262              (full_sync || list_empty(&ctx.ordered_extents)))) {
2263                 /*
2264                  * We've had everything committed since the last time we were
2265                  * modified so clear this flag in case it was set for whatever
2266                  * reason, it's no longer relevant.
2267                  */
2268                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2269                           &BTRFS_I(inode)->runtime_flags);
2270                 /*
2271                  * An ordered extent might have started before and completed
2272                  * already with io errors, in which case the inode was not
2273                  * updated and we end up here. So check the inode's mapping
2274                  * for any errors that might have happened since we last
2275                  * checked called fsync.
2276                  */
2277                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2278                 goto out_release_extents;
2279         }
2280
2281         /*
2282          * We use start here because we will need to wait on the IO to complete
2283          * in btrfs_sync_log, which could require joining a transaction (for
2284          * example checking cross references in the nocow path).  If we use join
2285          * here we could get into a situation where we're waiting on IO to
2286          * happen that is blocked on a transaction trying to commit.  With start
2287          * we inc the extwriter counter, so we wait for all extwriters to exit
2288          * before we start blocking joiners.  This comment is to keep somebody
2289          * from thinking they are super smart and changing this to
2290          * btrfs_join_transaction *cough*Josef*cough*.
2291          */
2292         trans = btrfs_start_transaction(root, 0);
2293         if (IS_ERR(trans)) {
2294                 ret = PTR_ERR(trans);
2295                 goto out_release_extents;
2296         }
2297
2298         ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2299         btrfs_release_log_ctx_extents(&ctx);
2300         if (ret < 0) {
2301                 /* Fallthrough and commit/free transaction. */
2302                 ret = 1;
2303         }
2304
2305         /* we've logged all the items and now have a consistent
2306          * version of the file in the log.  It is possible that
2307          * someone will come in and modify the file, but that's
2308          * fine because the log is consistent on disk, and we
2309          * have references to all of the file's extents
2310          *
2311          * It is possible that someone will come in and log the
2312          * file again, but that will end up using the synchronization
2313          * inside btrfs_sync_log to keep things safe.
2314          */
2315         up_write(&BTRFS_I(inode)->dio_sem);
2316         inode_unlock(inode);
2317
2318         if (ret != BTRFS_NO_LOG_SYNC) {
2319                 if (!ret) {
2320                         ret = btrfs_sync_log(trans, root, &ctx);
2321                         if (!ret) {
2322                                 ret = btrfs_end_transaction(trans);
2323                                 goto out;
2324                         }
2325                 }
2326                 if (!full_sync) {
2327                         ret = btrfs_wait_ordered_range(inode, start, len);
2328                         if (ret) {
2329                                 btrfs_end_transaction(trans);
2330                                 goto out;
2331                         }
2332                 }
2333                 ret = btrfs_commit_transaction(trans);
2334         } else {
2335                 ret = btrfs_end_transaction(trans);
2336         }
2337 out:
2338         ASSERT(list_empty(&ctx.list));
2339         err = file_check_and_advance_wb_err(file);
2340         if (!ret)
2341                 ret = err;
2342         return ret > 0 ? -EIO : ret;
2343
2344 out_release_extents:
2345         btrfs_release_log_ctx_extents(&ctx);
2346         up_write(&BTRFS_I(inode)->dio_sem);
2347         inode_unlock(inode);
2348         goto out;
2349 }
2350
2351 static const struct vm_operations_struct btrfs_file_vm_ops = {
2352         .fault          = filemap_fault,
2353         .map_pages      = filemap_map_pages,
2354         .page_mkwrite   = btrfs_page_mkwrite,
2355 };
2356
2357 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2358 {
2359         struct address_space *mapping = filp->f_mapping;
2360
2361         if (!mapping->a_ops->readpage)
2362                 return -ENOEXEC;
2363
2364         file_accessed(filp);
2365         vma->vm_ops = &btrfs_file_vm_ops;
2366
2367         return 0;
2368 }
2369
2370 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2371                           int slot, u64 start, u64 end)
2372 {
2373         struct btrfs_file_extent_item *fi;
2374         struct btrfs_key key;
2375
2376         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2377                 return 0;
2378
2379         btrfs_item_key_to_cpu(leaf, &key, slot);
2380         if (key.objectid != btrfs_ino(inode) ||
2381             key.type != BTRFS_EXTENT_DATA_KEY)
2382                 return 0;
2383
2384         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2385
2386         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2387                 return 0;
2388
2389         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2390                 return 0;
2391
2392         if (key.offset == end)
2393                 return 1;
2394         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2395                 return 1;
2396         return 0;
2397 }
2398
2399 static int fill_holes(struct btrfs_trans_handle *trans,
2400                 struct btrfs_inode *inode,
2401                 struct btrfs_path *path, u64 offset, u64 end)
2402 {
2403         struct btrfs_fs_info *fs_info = trans->fs_info;
2404         struct btrfs_root *root = inode->root;
2405         struct extent_buffer *leaf;
2406         struct btrfs_file_extent_item *fi;
2407         struct extent_map *hole_em;
2408         struct extent_map_tree *em_tree = &inode->extent_tree;
2409         struct btrfs_key key;
2410         int ret;
2411
2412         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2413                 goto out;
2414
2415         key.objectid = btrfs_ino(inode);
2416         key.type = BTRFS_EXTENT_DATA_KEY;
2417         key.offset = offset;
2418
2419         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2420         if (ret <= 0) {
2421                 /*
2422                  * We should have dropped this offset, so if we find it then
2423                  * something has gone horribly wrong.
2424                  */
2425                 if (ret == 0)
2426                         ret = -EINVAL;
2427                 return ret;
2428         }
2429
2430         leaf = path->nodes[0];
2431         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2432                 u64 num_bytes;
2433
2434                 path->slots[0]--;
2435                 fi = btrfs_item_ptr(leaf, path->slots[0],
2436                                     struct btrfs_file_extent_item);
2437                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2438                         end - offset;
2439                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2440                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2441                 btrfs_set_file_extent_offset(leaf, fi, 0);
2442                 btrfs_mark_buffer_dirty(leaf);
2443                 goto out;
2444         }
2445
2446         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2447                 u64 num_bytes;
2448
2449                 key.offset = offset;
2450                 btrfs_set_item_key_safe(fs_info, path, &key);
2451                 fi = btrfs_item_ptr(leaf, path->slots[0],
2452                                     struct btrfs_file_extent_item);
2453                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2454                         offset;
2455                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2456                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2457                 btrfs_set_file_extent_offset(leaf, fi, 0);
2458                 btrfs_mark_buffer_dirty(leaf);
2459                 goto out;
2460         }
2461         btrfs_release_path(path);
2462
2463         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2464                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2465         if (ret)
2466                 return ret;
2467
2468 out:
2469         btrfs_release_path(path);
2470
2471         hole_em = alloc_extent_map();
2472         if (!hole_em) {
2473                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2474                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2475         } else {
2476                 hole_em->start = offset;
2477                 hole_em->len = end - offset;
2478                 hole_em->ram_bytes = hole_em->len;
2479                 hole_em->orig_start = offset;
2480
2481                 hole_em->block_start = EXTENT_MAP_HOLE;
2482                 hole_em->block_len = 0;
2483                 hole_em->orig_block_len = 0;
2484                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2485                 hole_em->generation = trans->transid;
2486
2487                 do {
2488                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2489                         write_lock(&em_tree->lock);
2490                         ret = add_extent_mapping(em_tree, hole_em, 1);
2491                         write_unlock(&em_tree->lock);
2492                 } while (ret == -EEXIST);
2493                 free_extent_map(hole_em);
2494                 if (ret)
2495                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2496                                         &inode->runtime_flags);
2497         }
2498
2499         return 0;
2500 }
2501
2502 /*
2503  * Find a hole extent on given inode and change start/len to the end of hole
2504  * extent.(hole/vacuum extent whose em->start <= start &&
2505  *         em->start + em->len > start)
2506  * When a hole extent is found, return 1 and modify start/len.
2507  */
2508 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2509 {
2510         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2511         struct extent_map *em;
2512         int ret = 0;
2513
2514         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2515                               round_down(*start, fs_info->sectorsize),
2516                               round_up(*len, fs_info->sectorsize));
2517         if (IS_ERR(em))
2518                 return PTR_ERR(em);
2519
2520         /* Hole or vacuum extent(only exists in no-hole mode) */
2521         if (em->block_start == EXTENT_MAP_HOLE) {
2522                 ret = 1;
2523                 *len = em->start + em->len > *start + *len ?
2524                        0 : *start + *len - em->start - em->len;
2525                 *start = em->start + em->len;
2526         }
2527         free_extent_map(em);
2528         return ret;
2529 }
2530
2531 static int btrfs_punch_hole_lock_range(struct inode *inode,
2532                                        const u64 lockstart,
2533                                        const u64 lockend,
2534                                        struct extent_state **cached_state)
2535 {
2536         while (1) {
2537                 struct btrfs_ordered_extent *ordered;
2538                 int ret;
2539
2540                 truncate_pagecache_range(inode, lockstart, lockend);
2541
2542                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2543                                  cached_state);
2544                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2545                                                             lockend);
2546
2547                 /*
2548                  * We need to make sure we have no ordered extents in this range
2549                  * and nobody raced in and read a page in this range, if we did
2550                  * we need to try again.
2551                  */
2552                 if ((!ordered ||
2553                     (ordered->file_offset + ordered->num_bytes <= lockstart ||
2554                      ordered->file_offset > lockend)) &&
2555                      !filemap_range_has_page(inode->i_mapping,
2556                                              lockstart, lockend)) {
2557                         if (ordered)
2558                                 btrfs_put_ordered_extent(ordered);
2559                         break;
2560                 }
2561                 if (ordered)
2562                         btrfs_put_ordered_extent(ordered);
2563                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2564                                      lockend, cached_state);
2565                 ret = btrfs_wait_ordered_range(inode, lockstart,
2566                                                lockend - lockstart + 1);
2567                 if (ret)
2568                         return ret;
2569         }
2570         return 0;
2571 }
2572
2573 static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2574                                      struct inode *inode,
2575                                      struct btrfs_path *path,
2576                                      struct btrfs_clone_extent_info *clone_info,
2577                                      const u64 clone_len)
2578 {
2579         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2580         struct btrfs_root *root = BTRFS_I(inode)->root;
2581         struct btrfs_file_extent_item *extent;
2582         struct extent_buffer *leaf;
2583         struct btrfs_key key;
2584         int slot;
2585         struct btrfs_ref ref = { 0 };
2586         u64 ref_offset;
2587         int ret;
2588
2589         if (clone_len == 0)
2590                 return 0;
2591
2592         if (clone_info->disk_offset == 0 &&
2593             btrfs_fs_incompat(fs_info, NO_HOLES))
2594                 return 0;
2595
2596         key.objectid = btrfs_ino(BTRFS_I(inode));
2597         key.type = BTRFS_EXTENT_DATA_KEY;
2598         key.offset = clone_info->file_offset;
2599         ret = btrfs_insert_empty_item(trans, root, path, &key,
2600                                       clone_info->item_size);
2601         if (ret)
2602                 return ret;
2603         leaf = path->nodes[0];
2604         slot = path->slots[0];
2605         write_extent_buffer(leaf, clone_info->extent_buf,
2606                             btrfs_item_ptr_offset(leaf, slot),
2607                             clone_info->item_size);
2608         extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2609         btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2610         btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2611         btrfs_mark_buffer_dirty(leaf);
2612         btrfs_release_path(path);
2613
2614         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
2615                         clone_info->file_offset, clone_len);
2616         if (ret)
2617                 return ret;
2618
2619         /* If it's a hole, nothing more needs to be done. */
2620         if (clone_info->disk_offset == 0)
2621                 return 0;
2622
2623         inode_add_bytes(inode, clone_len);
2624         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2625                                clone_info->disk_offset,
2626                                clone_info->disk_len, 0);
2627         ref_offset = clone_info->file_offset - clone_info->data_offset;
2628         btrfs_init_data_ref(&ref, root->root_key.objectid,
2629                             btrfs_ino(BTRFS_I(inode)), ref_offset);
2630         ret = btrfs_inc_extent_ref(trans, &ref);
2631
2632         return ret;
2633 }
2634
2635 /*
2636  * The respective range must have been previously locked, as well as the inode.
2637  * The end offset is inclusive (last byte of the range).
2638  * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2639  * cloning.
2640  * When cloning, we don't want to end up in a state where we dropped extents
2641  * without inserting a new one, so we must abort the transaction to avoid a
2642  * corruption.
2643  */
2644 int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2645                            const u64 start, const u64 end,
2646                            struct btrfs_clone_extent_info *clone_info,
2647                            struct btrfs_trans_handle **trans_out)
2648 {
2649         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2650         u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2651         u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2652         struct btrfs_root *root = BTRFS_I(inode)->root;
2653         struct btrfs_trans_handle *trans = NULL;
2654         struct btrfs_block_rsv *rsv;
2655         unsigned int rsv_count;
2656         u64 cur_offset;
2657         u64 drop_end;
2658         u64 len = end - start;
2659         int ret = 0;
2660
2661         if (end <= start)
2662                 return -EINVAL;
2663
2664         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2665         if (!rsv) {
2666                 ret = -ENOMEM;
2667                 goto out;
2668         }
2669         rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2670         rsv->failfast = 1;
2671
2672         /*
2673          * 1 - update the inode
2674          * 1 - removing the extents in the range
2675          * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2676          *     an extent
2677          */
2678         if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2679                 rsv_count = 3;
2680         else
2681                 rsv_count = 2;
2682
2683         trans = btrfs_start_transaction(root, rsv_count);
2684         if (IS_ERR(trans)) {
2685                 ret = PTR_ERR(trans);
2686                 trans = NULL;
2687                 goto out_free;
2688         }
2689
2690         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2691                                       min_size, false);
2692         BUG_ON(ret);
2693         trans->block_rsv = rsv;
2694
2695         cur_offset = start;
2696         while (cur_offset < end) {
2697                 ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path,
2698                                            cur_offset, end + 1, &drop_end,
2699                                            1, 0, 0, NULL);
2700                 if (ret != -ENOSPC) {
2701                         /*
2702                          * When cloning we want to avoid transaction aborts when
2703                          * nothing was done and we are attempting to clone parts
2704                          * of inline extents, in such cases -EOPNOTSUPP is
2705                          * returned by __btrfs_drop_extents() without having
2706                          * changed anything in the file.
2707                          */
2708                         if (clone_info && ret && ret != -EOPNOTSUPP)
2709                                 btrfs_abort_transaction(trans, ret);
2710                         break;
2711                 }
2712
2713                 trans->block_rsv = &fs_info->trans_block_rsv;
2714
2715                 if (!clone_info && cur_offset < drop_end &&
2716                     cur_offset < ino_size) {
2717                         ret = fill_holes(trans, BTRFS_I(inode), path,
2718                                         cur_offset, drop_end);
2719                         if (ret) {
2720                                 /*
2721                                  * If we failed then we didn't insert our hole
2722                                  * entries for the area we dropped, so now the
2723                                  * fs is corrupted, so we must abort the
2724                                  * transaction.
2725                                  */
2726                                 btrfs_abort_transaction(trans, ret);
2727                                 break;
2728                         }
2729                 } else if (!clone_info && cur_offset < drop_end) {
2730                         /*
2731                          * We are past the i_size here, but since we didn't
2732                          * insert holes we need to clear the mapped area so we
2733                          * know to not set disk_i_size in this area until a new
2734                          * file extent is inserted here.
2735                          */
2736                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2737                                         cur_offset, drop_end - cur_offset);
2738                         if (ret) {
2739                                 /*
2740                                  * We couldn't clear our area, so we could
2741                                  * presumably adjust up and corrupt the fs, so
2742                                  * we need to abort.
2743                                  */
2744                                 btrfs_abort_transaction(trans, ret);
2745                                 break;
2746                         }
2747                 }
2748
2749                 if (clone_info && drop_end > clone_info->file_offset) {
2750                         u64 clone_len = drop_end - clone_info->file_offset;
2751
2752                         ret = btrfs_insert_clone_extent(trans, inode, path,
2753                                                         clone_info, clone_len);
2754                         if (ret) {
2755                                 btrfs_abort_transaction(trans, ret);
2756                                 break;
2757                         }
2758                         clone_info->data_len -= clone_len;
2759                         clone_info->data_offset += clone_len;
2760                         clone_info->file_offset += clone_len;
2761                 }
2762
2763                 cur_offset = drop_end;
2764
2765                 ret = btrfs_update_inode(trans, root, inode);
2766                 if (ret)
2767                         break;
2768
2769                 btrfs_end_transaction(trans);
2770                 btrfs_btree_balance_dirty(fs_info);
2771
2772                 trans = btrfs_start_transaction(root, rsv_count);
2773                 if (IS_ERR(trans)) {
2774                         ret = PTR_ERR(trans);
2775                         trans = NULL;
2776                         break;
2777                 }
2778
2779                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2780                                               rsv, min_size, false);
2781                 BUG_ON(ret);    /* shouldn't happen */
2782                 trans->block_rsv = rsv;
2783
2784                 if (!clone_info) {
2785                         ret = find_first_non_hole(inode, &cur_offset, &len);
2786                         if (unlikely(ret < 0))
2787                                 break;
2788                         if (ret && !len) {
2789                                 ret = 0;
2790                                 break;
2791                         }
2792                 }
2793         }
2794
2795         /*
2796          * If we were cloning, force the next fsync to be a full one since we
2797          * we replaced (or just dropped in the case of cloning holes when
2798          * NO_HOLES is enabled) extents and extent maps.
2799          * This is for the sake of simplicity, and cloning into files larger
2800          * than 16Mb would force the full fsync any way (when
2801          * try_release_extent_mapping() is invoked during page cache truncation.
2802          */
2803         if (clone_info)
2804                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2805                         &BTRFS_I(inode)->runtime_flags);
2806
2807         if (ret)
2808                 goto out_trans;
2809
2810         trans->block_rsv = &fs_info->trans_block_rsv;
2811         /*
2812          * If we are using the NO_HOLES feature we might have had already an
2813          * hole that overlaps a part of the region [lockstart, lockend] and
2814          * ends at (or beyond) lockend. Since we have no file extent items to
2815          * represent holes, drop_end can be less than lockend and so we must
2816          * make sure we have an extent map representing the existing hole (the
2817          * call to __btrfs_drop_extents() might have dropped the existing extent
2818          * map representing the existing hole), otherwise the fast fsync path
2819          * will not record the existence of the hole region
2820          * [existing_hole_start, lockend].
2821          */
2822         if (drop_end <= end)
2823                 drop_end = end + 1;
2824         /*
2825          * Don't insert file hole extent item if it's for a range beyond eof
2826          * (because it's useless) or if it represents a 0 bytes range (when
2827          * cur_offset == drop_end).
2828          */
2829         if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2830                 ret = fill_holes(trans, BTRFS_I(inode), path,
2831                                 cur_offset, drop_end);
2832                 if (ret) {
2833                         /* Same comment as above. */
2834                         btrfs_abort_transaction(trans, ret);
2835                         goto out_trans;
2836                 }
2837         } else if (!clone_info && cur_offset < drop_end) {
2838                 /* See the comment in the loop above for the reasoning here. */
2839                 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2840                                         cur_offset, drop_end - cur_offset);
2841                 if (ret) {
2842                         btrfs_abort_transaction(trans, ret);
2843                         goto out_trans;
2844                 }
2845
2846         }
2847         if (clone_info) {
2848                 ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2849                                                 clone_info->data_len);
2850                 if (ret) {
2851                         btrfs_abort_transaction(trans, ret);
2852                         goto out_trans;
2853                 }
2854         }
2855
2856 out_trans:
2857         if (!trans)
2858                 goto out_free;
2859
2860         trans->block_rsv = &fs_info->trans_block_rsv;
2861         if (ret)
2862                 btrfs_end_transaction(trans);
2863         else
2864                 *trans_out = trans;
2865 out_free:
2866         btrfs_free_block_rsv(fs_info, rsv);
2867 out:
2868         return ret;
2869 }
2870
2871 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2872 {
2873         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2874         struct btrfs_root *root = BTRFS_I(inode)->root;
2875         struct extent_state *cached_state = NULL;
2876         struct btrfs_path *path;
2877         struct btrfs_trans_handle *trans = NULL;
2878         u64 lockstart;
2879         u64 lockend;
2880         u64 tail_start;
2881         u64 tail_len;
2882         u64 orig_start = offset;
2883         int ret = 0;
2884         bool same_block;
2885         u64 ino_size;
2886         bool truncated_block = false;
2887         bool updated_inode = false;
2888
2889         ret = btrfs_wait_ordered_range(inode, offset, len);
2890         if (ret)
2891                 return ret;
2892
2893         inode_lock(inode);
2894         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2895         ret = find_first_non_hole(inode, &offset, &len);
2896         if (ret < 0)
2897                 goto out_only_mutex;
2898         if (ret && !len) {
2899                 /* Already in a large hole */
2900                 ret = 0;
2901                 goto out_only_mutex;
2902         }
2903
2904         lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2905         lockend = round_down(offset + len,
2906                              btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2907         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2908                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2909         /*
2910          * We needn't truncate any block which is beyond the end of the file
2911          * because we are sure there is no data there.
2912          */
2913         /*
2914          * Only do this if we are in the same block and we aren't doing the
2915          * entire block.
2916          */
2917         if (same_block && len < fs_info->sectorsize) {
2918                 if (offset < ino_size) {
2919                         truncated_block = true;
2920                         ret = btrfs_truncate_block(inode, offset, len, 0);
2921                 } else {
2922                         ret = 0;
2923                 }
2924                 goto out_only_mutex;
2925         }
2926
2927         /* zero back part of the first block */
2928         if (offset < ino_size) {
2929                 truncated_block = true;
2930                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2931                 if (ret) {
2932                         inode_unlock(inode);
2933                         return ret;
2934                 }
2935         }
2936
2937         /* Check the aligned pages after the first unaligned page,
2938          * if offset != orig_start, which means the first unaligned page
2939          * including several following pages are already in holes,
2940          * the extra check can be skipped */
2941         if (offset == orig_start) {
2942                 /* after truncate page, check hole again */
2943                 len = offset + len - lockstart;
2944                 offset = lockstart;
2945                 ret = find_first_non_hole(inode, &offset, &len);
2946                 if (ret < 0)
2947                         goto out_only_mutex;
2948                 if (ret && !len) {
2949                         ret = 0;
2950                         goto out_only_mutex;
2951                 }
2952                 lockstart = offset;
2953         }
2954
2955         /* Check the tail unaligned part is in a hole */
2956         tail_start = lockend + 1;
2957         tail_len = offset + len - tail_start;
2958         if (tail_len) {
2959                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2960                 if (unlikely(ret < 0))
2961                         goto out_only_mutex;
2962                 if (!ret) {
2963                         /* zero the front end of the last page */
2964                         if (tail_start + tail_len < ino_size) {
2965                                 truncated_block = true;
2966                                 ret = btrfs_truncate_block(inode,
2967                                                         tail_start + tail_len,
2968                                                         0, 1);
2969                                 if (ret)
2970                                         goto out_only_mutex;
2971                         }
2972                 }
2973         }
2974
2975         if (lockend < lockstart) {
2976                 ret = 0;
2977                 goto out_only_mutex;
2978         }
2979
2980         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2981                                           &cached_state);
2982         if (ret)
2983                 goto out_only_mutex;
2984
2985         path = btrfs_alloc_path();
2986         if (!path) {
2987                 ret = -ENOMEM;
2988                 goto out;
2989         }
2990
2991         ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2992                                      &trans);
2993         btrfs_free_path(path);
2994         if (ret)
2995                 goto out;
2996
2997         ASSERT(trans != NULL);
2998         inode_inc_iversion(inode);
2999         inode->i_mtime = inode->i_ctime = current_time(inode);
3000         ret = btrfs_update_inode(trans, root, inode);
3001         updated_inode = true;
3002         btrfs_end_transaction(trans);
3003         btrfs_btree_balance_dirty(fs_info);
3004 out:
3005         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3006                              &cached_state);
3007 out_only_mutex:
3008         if (!updated_inode && truncated_block && !ret) {
3009                 /*
3010                  * If we only end up zeroing part of a page, we still need to
3011                  * update the inode item, so that all the time fields are
3012                  * updated as well as the necessary btrfs inode in memory fields
3013                  * for detecting, at fsync time, if the inode isn't yet in the
3014                  * log tree or it's there but not up to date.
3015                  */
3016                 struct timespec64 now = current_time(inode);
3017
3018                 inode_inc_iversion(inode);
3019                 inode->i_mtime = now;
3020                 inode->i_ctime = now;
3021                 trans = btrfs_start_transaction(root, 1);
3022                 if (IS_ERR(trans)) {
3023                         ret = PTR_ERR(trans);
3024                 } else {
3025                         int ret2;
3026
3027                         ret = btrfs_update_inode(trans, root, inode);
3028                         ret2 = btrfs_end_transaction(trans);
3029                         if (!ret)
3030                                 ret = ret2;
3031                 }
3032         }
3033         inode_unlock(inode);
3034         return ret;
3035 }
3036
3037 /* Helper structure to record which range is already reserved */
3038 struct falloc_range {
3039         struct list_head list;
3040         u64 start;
3041         u64 len;
3042 };
3043
3044 /*
3045  * Helper function to add falloc range
3046  *
3047  * Caller should have locked the larger range of extent containing
3048  * [start, len)
3049  */
3050 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3051 {
3052         struct falloc_range *prev = NULL;
3053         struct falloc_range *range = NULL;
3054
3055         if (list_empty(head))
3056                 goto insert;
3057
3058         /*
3059          * As fallocate iterate by bytenr order, we only need to check
3060          * the last range.
3061          */
3062         prev = list_entry(head->prev, struct falloc_range, list);
3063         if (prev->start + prev->len == start) {
3064                 prev->len += len;
3065                 return 0;
3066         }
3067 insert:
3068         range = kmalloc(sizeof(*range), GFP_KERNEL);
3069         if (!range)
3070                 return -ENOMEM;
3071         range->start = start;
3072         range->len = len;
3073         list_add_tail(&range->list, head);
3074         return 0;
3075 }
3076
3077 static int btrfs_fallocate_update_isize(struct inode *inode,
3078                                         const u64 end,
3079                                         const int mode)
3080 {
3081         struct btrfs_trans_handle *trans;
3082         struct btrfs_root *root = BTRFS_I(inode)->root;
3083         int ret;
3084         int ret2;
3085
3086         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3087                 return 0;
3088
3089         trans = btrfs_start_transaction(root, 1);
3090         if (IS_ERR(trans))
3091                 return PTR_ERR(trans);
3092
3093         inode->i_ctime = current_time(inode);
3094         i_size_write(inode, end);
3095         btrfs_inode_safe_disk_i_size_write(inode, 0);
3096         ret = btrfs_update_inode(trans, root, inode);
3097         ret2 = btrfs_end_transaction(trans);
3098
3099         return ret ? ret : ret2;
3100 }
3101
3102 enum {
3103         RANGE_BOUNDARY_WRITTEN_EXTENT,
3104         RANGE_BOUNDARY_PREALLOC_EXTENT,
3105         RANGE_BOUNDARY_HOLE,
3106 };
3107
3108 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3109                                                  u64 offset)
3110 {
3111         const u64 sectorsize = btrfs_inode_sectorsize(inode);
3112         struct extent_map *em;
3113         int ret;
3114
3115         offset = round_down(offset, sectorsize);
3116         em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3117         if (IS_ERR(em))
3118                 return PTR_ERR(em);
3119
3120         if (em->block_start == EXTENT_MAP_HOLE)
3121                 ret = RANGE_BOUNDARY_HOLE;
3122         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3123                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3124         else
3125                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3126
3127         free_extent_map(em);
3128         return ret;
3129 }
3130
3131 static int btrfs_zero_range(struct inode *inode,
3132                             loff_t offset,
3133                             loff_t len,
3134                             const int mode)
3135 {
3136         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3137         struct extent_map *em;
3138         struct extent_changeset *data_reserved = NULL;
3139         int ret;
3140         u64 alloc_hint = 0;
3141         const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3142         u64 alloc_start = round_down(offset, sectorsize);
3143         u64 alloc_end = round_up(offset + len, sectorsize);
3144         u64 bytes_to_reserve = 0;
3145         bool space_reserved = false;
3146
3147         inode_dio_wait(inode);
3148
3149         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3150                               alloc_end - alloc_start);
3151         if (IS_ERR(em)) {
3152                 ret = PTR_ERR(em);
3153                 goto out;
3154         }
3155
3156         /*
3157          * Avoid hole punching and extent allocation for some cases. More cases
3158          * could be considered, but these are unlikely common and we keep things
3159          * as simple as possible for now. Also, intentionally, if the target
3160          * range contains one or more prealloc extents together with regular
3161          * extents and holes, we drop all the existing extents and allocate a
3162          * new prealloc extent, so that we get a larger contiguous disk extent.
3163          */
3164         if (em->start <= alloc_start &&
3165             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3166                 const u64 em_end = em->start + em->len;
3167
3168                 if (em_end >= offset + len) {
3169                         /*
3170                          * The whole range is already a prealloc extent,
3171                          * do nothing except updating the inode's i_size if
3172                          * needed.
3173                          */
3174                         free_extent_map(em);
3175                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3176                                                            mode);
3177                         goto out;
3178                 }
3179                 /*
3180                  * Part of the range is already a prealloc extent, so operate
3181                  * only on the remaining part of the range.
3182                  */
3183                 alloc_start = em_end;
3184                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3185                 len = offset + len - alloc_start;
3186                 offset = alloc_start;
3187                 alloc_hint = em->block_start + em->len;
3188         }
3189         free_extent_map(em);
3190
3191         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3192             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3193                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3194                                       sectorsize);
3195                 if (IS_ERR(em)) {
3196                         ret = PTR_ERR(em);
3197                         goto out;
3198                 }
3199
3200                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3201                         free_extent_map(em);
3202                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3203                                                            mode);
3204                         goto out;
3205                 }
3206                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3207                         free_extent_map(em);
3208                         ret = btrfs_truncate_block(inode, offset, len, 0);
3209                         if (!ret)
3210                                 ret = btrfs_fallocate_update_isize(inode,
3211                                                                    offset + len,
3212                                                                    mode);
3213                         return ret;
3214                 }
3215                 free_extent_map(em);
3216                 alloc_start = round_down(offset, sectorsize);
3217                 alloc_end = alloc_start + sectorsize;
3218                 goto reserve_space;
3219         }
3220
3221         alloc_start = round_up(offset, sectorsize);
3222         alloc_end = round_down(offset + len, sectorsize);
3223
3224         /*
3225          * For unaligned ranges, check the pages at the boundaries, they might
3226          * map to an extent, in which case we need to partially zero them, or
3227          * they might map to a hole, in which case we need our allocation range
3228          * to cover them.
3229          */
3230         if (!IS_ALIGNED(offset, sectorsize)) {
3231                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3232                                                             offset);
3233                 if (ret < 0)
3234                         goto out;
3235                 if (ret == RANGE_BOUNDARY_HOLE) {
3236                         alloc_start = round_down(offset, sectorsize);
3237                         ret = 0;
3238                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3239                         ret = btrfs_truncate_block(inode, offset, 0, 0);
3240                         if (ret)
3241                                 goto out;
3242                 } else {
3243                         ret = 0;
3244                 }
3245         }
3246
3247         if (!IS_ALIGNED(offset + len, sectorsize)) {
3248                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3249                                                             offset + len);
3250                 if (ret < 0)
3251                         goto out;
3252                 if (ret == RANGE_BOUNDARY_HOLE) {
3253                         alloc_end = round_up(offset + len, sectorsize);
3254                         ret = 0;
3255                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3256                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3257                         if (ret)
3258                                 goto out;
3259                 } else {
3260                         ret = 0;
3261                 }
3262         }
3263
3264 reserve_space:
3265         if (alloc_start < alloc_end) {
3266                 struct extent_state *cached_state = NULL;
3267                 const u64 lockstart = alloc_start;
3268                 const u64 lockend = alloc_end - 1;
3269
3270                 bytes_to_reserve = alloc_end - alloc_start;
3271                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3272                                                       bytes_to_reserve);
3273                 if (ret < 0)
3274                         goto out;
3275                 space_reserved = true;
3276                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3277                                                   &cached_state);
3278                 if (ret)
3279                         goto out;
3280                 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3281                                                 alloc_start, bytes_to_reserve);
3282                 if (ret)
3283                         goto out;
3284                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3285                                                 alloc_end - alloc_start,
3286                                                 i_blocksize(inode),
3287                                                 offset + len, &alloc_hint);
3288                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3289                                      lockend, &cached_state);
3290                 /* btrfs_prealloc_file_range releases reserved space on error */
3291                 if (ret) {
3292                         space_reserved = false;
3293                         goto out;
3294                 }
3295         }
3296         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3297  out:
3298         if (ret && space_reserved)
3299                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3300                                                alloc_start, bytes_to_reserve);
3301         extent_changeset_free(data_reserved);
3302
3303         return ret;
3304 }
3305
3306 static long btrfs_fallocate(struct file *file, int mode,
3307                             loff_t offset, loff_t len)
3308 {
3309         struct inode *inode = file_inode(file);
3310         struct extent_state *cached_state = NULL;
3311         struct extent_changeset *data_reserved = NULL;
3312         struct falloc_range *range;
3313         struct falloc_range *tmp;
3314         struct list_head reserve_list;
3315         u64 cur_offset;
3316         u64 last_byte;
3317         u64 alloc_start;
3318         u64 alloc_end;
3319         u64 alloc_hint = 0;
3320         u64 locked_end;
3321         u64 actual_end = 0;
3322         struct extent_map *em;
3323         int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3324         int ret;
3325
3326         alloc_start = round_down(offset, blocksize);
3327         alloc_end = round_up(offset + len, blocksize);
3328         cur_offset = alloc_start;
3329
3330         /* Make sure we aren't being give some crap mode */
3331         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3332                      FALLOC_FL_ZERO_RANGE))
3333                 return -EOPNOTSUPP;
3334
3335         if (mode & FALLOC_FL_PUNCH_HOLE)
3336                 return btrfs_punch_hole(inode, offset, len);
3337
3338         /*
3339          * Only trigger disk allocation, don't trigger qgroup reserve
3340          *
3341          * For qgroup space, it will be checked later.
3342          */
3343         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3344                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3345                                                       alloc_end - alloc_start);
3346                 if (ret < 0)
3347                         return ret;
3348         }
3349
3350         inode_lock(inode);
3351
3352         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3353                 ret = inode_newsize_ok(inode, offset + len);
3354                 if (ret)
3355                         goto out;
3356         }
3357
3358         /*
3359          * TODO: Move these two operations after we have checked
3360          * accurate reserved space, or fallocate can still fail but
3361          * with page truncated or size expanded.
3362          *
3363          * But that's a minor problem and won't do much harm BTW.
3364          */
3365         if (alloc_start > inode->i_size) {
3366                 ret = btrfs_cont_expand(inode, i_size_read(inode),
3367                                         alloc_start);
3368                 if (ret)
3369                         goto out;
3370         } else if (offset + len > inode->i_size) {
3371                 /*
3372                  * If we are fallocating from the end of the file onward we
3373                  * need to zero out the end of the block if i_size lands in the
3374                  * middle of a block.
3375                  */
3376                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3377                 if (ret)
3378                         goto out;
3379         }
3380
3381         /*
3382          * wait for ordered IO before we have any locks.  We'll loop again
3383          * below with the locks held.
3384          */
3385         ret = btrfs_wait_ordered_range(inode, alloc_start,
3386                                        alloc_end - alloc_start);
3387         if (ret)
3388                 goto out;
3389
3390         if (mode & FALLOC_FL_ZERO_RANGE) {
3391                 ret = btrfs_zero_range(inode, offset, len, mode);
3392                 inode_unlock(inode);
3393                 return ret;
3394         }
3395
3396         locked_end = alloc_end - 1;
3397         while (1) {
3398                 struct btrfs_ordered_extent *ordered;
3399
3400                 /* the extent lock is ordered inside the running
3401                  * transaction
3402                  */
3403                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3404                                  locked_end, &cached_state);
3405                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3406                                                             locked_end);
3407
3408                 if (ordered &&
3409                     ordered->file_offset + ordered->num_bytes > alloc_start &&
3410                     ordered->file_offset < alloc_end) {
3411                         btrfs_put_ordered_extent(ordered);
3412                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3413                                              alloc_start, locked_end,
3414                                              &cached_state);
3415                         /*
3416                          * we can't wait on the range with the transaction
3417                          * running or with the extent lock held
3418                          */
3419                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3420                                                        alloc_end - alloc_start);
3421                         if (ret)
3422                                 goto out;
3423                 } else {
3424                         if (ordered)
3425                                 btrfs_put_ordered_extent(ordered);
3426                         break;
3427                 }
3428         }
3429
3430         /* First, check if we exceed the qgroup limit */
3431         INIT_LIST_HEAD(&reserve_list);
3432         while (cur_offset < alloc_end) {
3433                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3434                                       alloc_end - cur_offset);
3435                 if (IS_ERR(em)) {
3436                         ret = PTR_ERR(em);
3437                         break;
3438                 }
3439                 last_byte = min(extent_map_end(em), alloc_end);
3440                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3441                 last_byte = ALIGN(last_byte, blocksize);
3442                 if (em->block_start == EXTENT_MAP_HOLE ||
3443                     (cur_offset >= inode->i_size &&
3444                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3445                         ret = add_falloc_range(&reserve_list, cur_offset,
3446                                                last_byte - cur_offset);
3447                         if (ret < 0) {
3448                                 free_extent_map(em);
3449                                 break;
3450                         }
3451                         ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3452                                         &data_reserved, cur_offset,
3453                                         last_byte - cur_offset);
3454                         if (ret < 0) {
3455                                 cur_offset = last_byte;
3456                                 free_extent_map(em);
3457                                 break;
3458                         }
3459                 } else {
3460                         /*
3461                          * Do not need to reserve unwritten extent for this
3462                          * range, free reserved data space first, otherwise
3463                          * it'll result in false ENOSPC error.
3464                          */
3465                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3466                                 data_reserved, cur_offset,
3467                                 last_byte - cur_offset);
3468                 }
3469                 free_extent_map(em);
3470                 cur_offset = last_byte;
3471         }
3472
3473         /*
3474          * If ret is still 0, means we're OK to fallocate.
3475          * Or just cleanup the list and exit.
3476          */
3477         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3478                 if (!ret)
3479                         ret = btrfs_prealloc_file_range(inode, mode,
3480                                         range->start,
3481                                         range->len, i_blocksize(inode),
3482                                         offset + len, &alloc_hint);
3483                 else
3484                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3485                                         data_reserved, range->start,
3486                                         range->len);
3487                 list_del(&range->list);
3488                 kfree(range);
3489         }
3490         if (ret < 0)
3491                 goto out_unlock;
3492
3493         /*
3494          * We didn't need to allocate any more space, but we still extended the
3495          * size of the file so we need to update i_size and the inode item.
3496          */
3497         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3498 out_unlock:
3499         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3500                              &cached_state);
3501 out:
3502         inode_unlock(inode);
3503         /* Let go of our reservation. */
3504         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3505                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3506                                 cur_offset, alloc_end - cur_offset);
3507         extent_changeset_free(data_reserved);
3508         return ret;
3509 }
3510
3511 static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3512                                   int whence)
3513 {
3514         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3515         struct extent_map *em = NULL;
3516         struct extent_state *cached_state = NULL;
3517         loff_t i_size = inode->i_size;
3518         u64 lockstart;
3519         u64 lockend;
3520         u64 start;
3521         u64 len;
3522         int ret = 0;
3523
3524         if (i_size == 0 || offset >= i_size)
3525                 return -ENXIO;
3526
3527         /*
3528          * offset can be negative, in this case we start finding DATA/HOLE from
3529          * the very start of the file.
3530          */
3531         start = max_t(loff_t, 0, offset);
3532
3533         lockstart = round_down(start, fs_info->sectorsize);
3534         lockend = round_up(i_size, fs_info->sectorsize);
3535         if (lockend <= lockstart)
3536                 lockend = lockstart + fs_info->sectorsize;
3537         lockend--;
3538         len = lockend - lockstart + 1;
3539
3540         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3541                          &cached_state);
3542
3543         while (start < i_size) {
3544                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3545                 if (IS_ERR(em)) {
3546                         ret = PTR_ERR(em);
3547                         em = NULL;
3548                         break;
3549                 }
3550
3551                 if (whence == SEEK_HOLE &&
3552                     (em->block_start == EXTENT_MAP_HOLE ||
3553                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3554                         break;
3555                 else if (whence == SEEK_DATA &&
3556                            (em->block_start != EXTENT_MAP_HOLE &&
3557                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3558                         break;
3559
3560                 start = em->start + em->len;
3561                 free_extent_map(em);
3562                 em = NULL;
3563                 cond_resched();
3564         }
3565         free_extent_map(em);
3566         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3567                              &cached_state);
3568         if (ret) {
3569                 offset = ret;
3570         } else {
3571                 if (whence == SEEK_DATA && start >= i_size)
3572                         offset = -ENXIO;
3573                 else
3574                         offset = min_t(loff_t, start, i_size);
3575         }
3576
3577         return offset;
3578 }
3579
3580 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3581 {
3582         struct inode *inode = file->f_mapping->host;
3583
3584         switch (whence) {
3585         default:
3586                 return generic_file_llseek(file, offset, whence);
3587         case SEEK_DATA:
3588         case SEEK_HOLE:
3589                 inode_lock_shared(inode);
3590                 offset = find_desired_extent(inode, offset, whence);
3591                 inode_unlock_shared(inode);
3592                 break;
3593         }
3594
3595         if (offset < 0)
3596                 return offset;
3597
3598         return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3599 }
3600
3601 static int btrfs_file_open(struct inode *inode, struct file *filp)
3602 {
3603         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3604         return generic_file_open(inode, filp);
3605 }
3606
3607 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3608 {
3609         ssize_t ret = 0;
3610
3611         if (iocb->ki_flags & IOCB_DIRECT) {
3612                 struct inode *inode = file_inode(iocb->ki_filp);
3613
3614                 inode_lock_shared(inode);
3615                 ret = btrfs_direct_IO(iocb, to);
3616                 inode_unlock_shared(inode);
3617                 if (ret < 0)
3618                         return ret;
3619         }
3620
3621         return generic_file_buffered_read(iocb, to, ret);
3622 }
3623
3624 const struct file_operations btrfs_file_operations = {
3625         .llseek         = btrfs_file_llseek,
3626         .read_iter      = btrfs_file_read_iter,
3627         .splice_read    = generic_file_splice_read,
3628         .write_iter     = btrfs_file_write_iter,
3629         .splice_write   = iter_file_splice_write,
3630         .mmap           = btrfs_file_mmap,
3631         .open           = btrfs_file_open,
3632         .release        = btrfs_release_file,
3633         .fsync          = btrfs_sync_file,
3634         .fallocate      = btrfs_fallocate,
3635         .unlocked_ioctl = btrfs_ioctl,
3636 #ifdef CONFIG_COMPAT
3637         .compat_ioctl   = btrfs_compat_ioctl,
3638 #endif
3639         .remap_file_range = btrfs_remap_file_range,
3640 };
3641
3642 void __cold btrfs_auto_defrag_exit(void)
3643 {
3644         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3645 }
3646
3647 int __init btrfs_auto_defrag_init(void)
3648 {
3649         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3650                                         sizeof(struct inode_defrag), 0,
3651                                         SLAB_MEM_SPREAD,
3652                                         NULL);
3653         if (!btrfs_inode_defrag_cachep)
3654                 return -ENOMEM;
3655
3656         return 0;
3657 }
3658
3659 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3660 {
3661         int ret;
3662
3663         /*
3664          * So with compression we will find and lock a dirty page and clear the
3665          * first one as dirty, setup an async extent, and immediately return
3666          * with the entire range locked but with nobody actually marked with
3667          * writeback.  So we can't just filemap_write_and_wait_range() and
3668          * expect it to work since it will just kick off a thread to do the
3669          * actual work.  So we need to call filemap_fdatawrite_range _again_
3670          * since it will wait on the page lock, which won't be unlocked until
3671          * after the pages have been marked as writeback and so we're good to go
3672          * from there.  We have to do this otherwise we'll miss the ordered
3673          * extents and that results in badness.  Please Josef, do not think you
3674          * know better and pull this out at some point in the future, it is
3675          * right and you are wrong.
3676          */
3677         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3678         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3679                              &BTRFS_I(inode)->runtime_flags))
3680                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3681
3682         return ret;
3683 }