2 * linux/fs/ext4/inode.c
4 * Copyright (C) 1992, 1993, 1994, 1995
5 * Remy Card (card@masi.ibp.fr)
6 * Laboratoire MASI - Institut Blaise Pascal
7 * Universite Pierre et Marie Curie (Paris VI)
11 * linux/fs/minix/inode.c
13 * Copyright (C) 1991, 1992 Linus Torvalds
15 * Goal-directed block allocation by Stephen Tweedie
16 * (sct@redhat.com), 1993, 1998
17 * Big-endian to little-endian byte-swapping/bitmaps by
18 * David S. Miller (davem@caip.rutgers.edu), 1995
19 * 64-bit file support on 64-bit platforms by Jakub Jelinek
20 * (jj@sunsite.ms.mff.cuni.cz)
22 * Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
25 #include <linux/module.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41 #include <linux/kernel.h>
42 #include <linux/slab.h>
44 #include "ext4_jbd2.h"
47 #include "ext4_extents.h"
49 #include <trace/events/ext4.h>
51 #define MPAGE_DA_EXTENT_TAIL 0x01
53 static inline int ext4_begin_ordered_truncate(struct inode *inode,
56 trace_ext4_begin_ordered_truncate(inode, new_size);
57 return jbd2_journal_begin_ordered_truncate(
58 EXT4_SB(inode->i_sb)->s_journal,
59 &EXT4_I(inode)->jinode,
63 static void ext4_invalidatepage(struct page *page, unsigned long offset);
64 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
65 struct buffer_head *bh_result, int create);
66 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
67 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
68 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
69 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
72 * Test whether an inode is a fast symlink.
74 static int ext4_inode_is_fast_symlink(struct inode *inode)
76 int ea_blocks = EXT4_I(inode)->i_file_acl ?
77 (inode->i_sb->s_blocksize >> 9) : 0;
79 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
83 * Work out how many blocks we need to proceed with the next chunk of a
84 * truncate transaction.
86 static unsigned long blocks_for_truncate(struct inode *inode)
90 needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
92 /* Give ourselves just enough room to cope with inodes in which
93 * i_blocks is corrupt: we've seen disk corruptions in the past
94 * which resulted in random data in an inode which looked enough
95 * like a regular file for ext4 to try to delete it. Things
96 * will go a bit crazy if that happens, but at least we should
97 * try not to panic the whole kernel. */
101 /* But we need to bound the transaction so we don't overflow the
103 if (needed > EXT4_MAX_TRANS_DATA)
104 needed = EXT4_MAX_TRANS_DATA;
106 return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
110 * Truncate transactions can be complex and absolutely huge. So we need to
111 * be able to restart the transaction at a conventient checkpoint to make
112 * sure we don't overflow the journal.
114 * start_transaction gets us a new handle for a truncate transaction,
115 * and extend_transaction tries to extend the existing one a bit. If
116 * extend fails, we need to propagate the failure up and restart the
117 * transaction in the top-level truncate loop. --sct
119 static handle_t *start_transaction(struct inode *inode)
123 result = ext4_journal_start(inode, blocks_for_truncate(inode));
127 ext4_std_error(inode->i_sb, PTR_ERR(result));
132 * Try to extend this transaction for the purposes of truncation.
134 * Returns 0 if we managed to create more room. If we can't create more
135 * room, and the transaction must be restarted we return 1.
137 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
139 if (!ext4_handle_valid(handle))
141 if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
143 if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
149 * Restart the transaction associated with *handle. This does a commit,
150 * so before we call here everything must be consistently dirtied against
153 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
159 * Drop i_data_sem to avoid deadlock with ext4_map_blocks. At this
160 * moment, get_block can be called only for blocks inside i_size since
161 * page cache has been already dropped and writes are blocked by
162 * i_mutex. So we can safely drop the i_data_sem here.
164 BUG_ON(EXT4_JOURNAL(inode) == NULL);
165 jbd_debug(2, "restarting handle %p\n", handle);
166 up_write(&EXT4_I(inode)->i_data_sem);
167 ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
168 down_write(&EXT4_I(inode)->i_data_sem);
169 ext4_discard_preallocations(inode);
175 * Called at the last iput() if i_nlink is zero.
177 void ext4_evict_inode(struct inode *inode)
182 trace_ext4_evict_inode(inode);
183 if (inode->i_nlink) {
184 truncate_inode_pages(&inode->i_data, 0);
188 if (!is_bad_inode(inode))
189 dquot_initialize(inode);
191 if (ext4_should_order_data(inode))
192 ext4_begin_ordered_truncate(inode, 0);
193 truncate_inode_pages(&inode->i_data, 0);
195 if (is_bad_inode(inode))
198 handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
199 if (IS_ERR(handle)) {
200 ext4_std_error(inode->i_sb, PTR_ERR(handle));
202 * If we're going to skip the normal cleanup, we still need to
203 * make sure that the in-core orphan linked list is properly
206 ext4_orphan_del(NULL, inode);
211 ext4_handle_sync(handle);
213 err = ext4_mark_inode_dirty(handle, inode);
215 ext4_warning(inode->i_sb,
216 "couldn't mark inode dirty (err %d)", err);
220 ext4_truncate(inode);
223 * ext4_ext_truncate() doesn't reserve any slop when it
224 * restarts journal transactions; therefore there may not be
225 * enough credits left in the handle to remove the inode from
226 * the orphan list and set the dtime field.
228 if (!ext4_handle_has_enough_credits(handle, 3)) {
229 err = ext4_journal_extend(handle, 3);
231 err = ext4_journal_restart(handle, 3);
233 ext4_warning(inode->i_sb,
234 "couldn't extend journal (err %d)", err);
236 ext4_journal_stop(handle);
237 ext4_orphan_del(NULL, inode);
243 * Kill off the orphan record which ext4_truncate created.
244 * AKPM: I think this can be inside the above `if'.
245 * Note that ext4_orphan_del() has to be able to cope with the
246 * deletion of a non-existent orphan - this is because we don't
247 * know if ext4_truncate() actually created an orphan record.
248 * (Well, we could do this if we need to, but heck - it works)
250 ext4_orphan_del(handle, inode);
251 EXT4_I(inode)->i_dtime = get_seconds();
254 * One subtle ordering requirement: if anything has gone wrong
255 * (transaction abort, IO errors, whatever), then we can still
256 * do these next steps (the fs will already have been marked as
257 * having errors), but we can't free the inode if the mark_dirty
260 if (ext4_mark_inode_dirty(handle, inode))
261 /* If that failed, just do the required in-core inode clear. */
262 ext4_clear_inode(inode);
264 ext4_free_inode(handle, inode);
265 ext4_journal_stop(handle);
268 ext4_clear_inode(inode); /* We must guarantee clearing of inode... */
274 struct buffer_head *bh;
277 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
279 p->key = *(p->p = v);
284 * ext4_block_to_path - parse the block number into array of offsets
285 * @inode: inode in question (we are only interested in its superblock)
286 * @i_block: block number to be parsed
287 * @offsets: array to store the offsets in
288 * @boundary: set this non-zero if the referred-to block is likely to be
289 * followed (on disk) by an indirect block.
291 * To store the locations of file's data ext4 uses a data structure common
292 * for UNIX filesystems - tree of pointers anchored in the inode, with
293 * data blocks at leaves and indirect blocks in intermediate nodes.
294 * This function translates the block number into path in that tree -
295 * return value is the path length and @offsets[n] is the offset of
296 * pointer to (n+1)th node in the nth one. If @block is out of range
297 * (negative or too large) warning is printed and zero returned.
299 * Note: function doesn't find node addresses, so no IO is needed. All
300 * we need to know is the capacity of indirect blocks (taken from the
305 * Portability note: the last comparison (check that we fit into triple
306 * indirect block) is spelled differently, because otherwise on an
307 * architecture with 32-bit longs and 8Kb pages we might get into trouble
308 * if our filesystem had 8Kb blocks. We might use long long, but that would
309 * kill us on x86. Oh, well, at least the sign propagation does not matter -
310 * i_block would have to be negative in the very beginning, so we would not
314 static int ext4_block_to_path(struct inode *inode,
316 ext4_lblk_t offsets[4], int *boundary)
318 int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
319 int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
320 const long direct_blocks = EXT4_NDIR_BLOCKS,
321 indirect_blocks = ptrs,
322 double_blocks = (1 << (ptrs_bits * 2));
326 if (i_block < direct_blocks) {
327 offsets[n++] = i_block;
328 final = direct_blocks;
329 } else if ((i_block -= direct_blocks) < indirect_blocks) {
330 offsets[n++] = EXT4_IND_BLOCK;
331 offsets[n++] = i_block;
333 } else if ((i_block -= indirect_blocks) < double_blocks) {
334 offsets[n++] = EXT4_DIND_BLOCK;
335 offsets[n++] = i_block >> ptrs_bits;
336 offsets[n++] = i_block & (ptrs - 1);
338 } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
339 offsets[n++] = EXT4_TIND_BLOCK;
340 offsets[n++] = i_block >> (ptrs_bits * 2);
341 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
342 offsets[n++] = i_block & (ptrs - 1);
345 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
346 i_block + direct_blocks +
347 indirect_blocks + double_blocks, inode->i_ino);
350 *boundary = final - 1 - (i_block & (ptrs - 1));
354 static int __ext4_check_blockref(const char *function, unsigned int line,
356 __le32 *p, unsigned int max)
358 struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
362 while (bref < p+max) {
363 blk = le32_to_cpu(*bref++);
365 unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
367 es->s_last_error_block = cpu_to_le64(blk);
368 ext4_error_inode(inode, function, line, blk,
377 #define ext4_check_indirect_blockref(inode, bh) \
378 __ext4_check_blockref(__func__, __LINE__, inode, \
379 (__le32 *)(bh)->b_data, \
380 EXT4_ADDR_PER_BLOCK((inode)->i_sb))
382 #define ext4_check_inode_blockref(inode) \
383 __ext4_check_blockref(__func__, __LINE__, inode, \
384 EXT4_I(inode)->i_data, \
388 * ext4_get_branch - read the chain of indirect blocks leading to data
389 * @inode: inode in question
390 * @depth: depth of the chain (1 - direct pointer, etc.)
391 * @offsets: offsets of pointers in inode/indirect blocks
392 * @chain: place to store the result
393 * @err: here we store the error value
395 * Function fills the array of triples <key, p, bh> and returns %NULL
396 * if everything went OK or the pointer to the last filled triple
397 * (incomplete one) otherwise. Upon the return chain[i].key contains
398 * the number of (i+1)-th block in the chain (as it is stored in memory,
399 * i.e. little-endian 32-bit), chain[i].p contains the address of that
400 * number (it points into struct inode for i==0 and into the bh->b_data
401 * for i>0) and chain[i].bh points to the buffer_head of i-th indirect
402 * block for i>0 and NULL for i==0. In other words, it holds the block
403 * numbers of the chain, addresses they were taken from (and where we can
404 * verify that chain did not change) and buffer_heads hosting these
407 * Function stops when it stumbles upon zero pointer (absent block)
408 * (pointer to last triple returned, *@err == 0)
409 * or when it gets an IO error reading an indirect block
410 * (ditto, *@err == -EIO)
411 * or when it reads all @depth-1 indirect blocks successfully and finds
412 * the whole chain, all way to the data (returns %NULL, *err == 0).
414 * Need to be called with
415 * down_read(&EXT4_I(inode)->i_data_sem)
417 static Indirect *ext4_get_branch(struct inode *inode, int depth,
418 ext4_lblk_t *offsets,
419 Indirect chain[4], int *err)
421 struct super_block *sb = inode->i_sb;
423 struct buffer_head *bh;
426 /* i_data is not going away, no lock needed */
427 add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
431 bh = sb_getblk(sb, le32_to_cpu(p->key));
435 if (!bh_uptodate_or_lock(bh)) {
436 if (bh_submit_read(bh) < 0) {
440 /* validate block references */
441 if (ext4_check_indirect_blockref(inode, bh)) {
447 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
461 * ext4_find_near - find a place for allocation with sufficient locality
463 * @ind: descriptor of indirect block.
465 * This function returns the preferred place for block allocation.
466 * It is used when heuristic for sequential allocation fails.
468 * + if there is a block to the left of our position - allocate near it.
469 * + if pointer will live in indirect block - allocate near that block.
470 * + if pointer will live in inode - allocate in the same
473 * In the latter case we colour the starting block by the callers PID to
474 * prevent it from clashing with concurrent allocations for a different inode
475 * in the same block group. The PID is used here so that functionally related
476 * files will be close-by on-disk.
478 * Caller must make sure that @ind is valid and will stay that way.
480 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
482 struct ext4_inode_info *ei = EXT4_I(inode);
483 __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
485 ext4_fsblk_t bg_start;
486 ext4_fsblk_t last_block;
487 ext4_grpblk_t colour;
488 ext4_group_t block_group;
489 int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
491 /* Try to find previous block */
492 for (p = ind->p - 1; p >= start; p--) {
494 return le32_to_cpu(*p);
497 /* No such thing, so let's try location of indirect block */
499 return ind->bh->b_blocknr;
502 * It is going to be referred to from the inode itself? OK, just put it
503 * into the same cylinder group then.
505 block_group = ei->i_block_group;
506 if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
507 block_group &= ~(flex_size-1);
508 if (S_ISREG(inode->i_mode))
511 bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
512 last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
515 * If we are doing delayed allocation, we don't need take
516 * colour into account.
518 if (test_opt(inode->i_sb, DELALLOC))
521 if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
522 colour = (current->pid % 16) *
523 (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
525 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
526 return bg_start + colour;
530 * ext4_find_goal - find a preferred place for allocation.
532 * @block: block we want
533 * @partial: pointer to the last triple within a chain
535 * Normally this function find the preferred place for block allocation,
537 * Because this is only used for non-extent files, we limit the block nr
540 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
546 * XXX need to get goal block from mballoc's data structures
549 goal = ext4_find_near(inode, partial);
550 goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
555 * ext4_blks_to_allocate: Look up the block map and count the number
556 * of direct blocks need to be allocated for the given branch.
558 * @branch: chain of indirect blocks
559 * @k: number of blocks need for indirect blocks
560 * @blks: number of data blocks to be mapped.
561 * @blocks_to_boundary: the offset in the indirect block
563 * return the total number of blocks to be allocate, including the
564 * direct and indirect blocks.
566 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
567 int blocks_to_boundary)
569 unsigned int count = 0;
572 * Simple case, [t,d]Indirect block(s) has not allocated yet
573 * then it's clear blocks on that path have not allocated
576 /* right now we don't handle cross boundary allocation */
577 if (blks < blocks_to_boundary + 1)
580 count += blocks_to_boundary + 1;
585 while (count < blks && count <= blocks_to_boundary &&
586 le32_to_cpu(*(branch[0].p + count)) == 0) {
593 * ext4_alloc_blocks: multiple allocate blocks needed for a branch
594 * @indirect_blks: the number of blocks need to allocate for indirect
597 * @new_blocks: on return it will store the new block numbers for
598 * the indirect blocks(if needed) and the first direct block,
599 * @blks: on return it will store the total number of allocated
602 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
603 ext4_lblk_t iblock, ext4_fsblk_t goal,
604 int indirect_blks, int blks,
605 ext4_fsblk_t new_blocks[4], int *err)
607 struct ext4_allocation_request ar;
609 unsigned long count = 0, blk_allocated = 0;
611 ext4_fsblk_t current_block = 0;
615 * Here we try to allocate the requested multiple blocks at once,
616 * on a best-effort basis.
617 * To build a branch, we should allocate blocks for
618 * the indirect blocks(if not allocated yet), and at least
619 * the first direct block of this branch. That's the
620 * minimum number of blocks need to allocate(required)
622 /* first we try to allocate the indirect blocks */
623 target = indirect_blks;
626 /* allocating blocks for indirect blocks and direct blocks */
627 current_block = ext4_new_meta_blocks(handle, inode,
632 if (unlikely(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS)) {
633 EXT4_ERROR_INODE(inode,
634 "current_block %llu + count %lu > %d!",
635 current_block, count,
636 EXT4_MAX_BLOCK_FILE_PHYS);
642 /* allocate blocks for indirect blocks */
643 while (index < indirect_blks && count) {
644 new_blocks[index++] = current_block++;
649 * save the new block number
650 * for the first direct block
652 new_blocks[index] = current_block;
653 printk(KERN_INFO "%s returned more blocks than "
654 "requested\n", __func__);
660 target = blks - count ;
661 blk_allocated = count;
664 /* Now allocate data blocks */
665 memset(&ar, 0, sizeof(ar));
670 if (S_ISREG(inode->i_mode))
671 /* enable in-core preallocation only for regular files */
672 ar.flags = EXT4_MB_HINT_DATA;
674 current_block = ext4_mb_new_blocks(handle, &ar, err);
675 if (unlikely(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS)) {
676 EXT4_ERROR_INODE(inode,
677 "current_block %llu + ar.len %d > %d!",
678 current_block, ar.len,
679 EXT4_MAX_BLOCK_FILE_PHYS);
684 if (*err && (target == blks)) {
686 * if the allocation failed and we didn't allocate
692 if (target == blks) {
694 * save the new block number
695 * for the first direct block
697 new_blocks[index] = current_block;
699 blk_allocated += ar.len;
702 /* total number of blocks allocated for direct blocks */
707 for (i = 0; i < index; i++)
708 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
713 * ext4_alloc_branch - allocate and set up a chain of blocks.
715 * @indirect_blks: number of allocated indirect blocks
716 * @blks: number of allocated direct blocks
717 * @offsets: offsets (in the blocks) to store the pointers to next.
718 * @branch: place to store the chain in.
720 * This function allocates blocks, zeroes out all but the last one,
721 * links them into chain and (if we are synchronous) writes them to disk.
722 * In other words, it prepares a branch that can be spliced onto the
723 * inode. It stores the information about that chain in the branch[], in
724 * the same format as ext4_get_branch() would do. We are calling it after
725 * we had read the existing part of chain and partial points to the last
726 * triple of that (one with zero ->key). Upon the exit we have the same
727 * picture as after the successful ext4_get_block(), except that in one
728 * place chain is disconnected - *branch->p is still zero (we did not
729 * set the last link), but branch->key contains the number that should
730 * be placed into *branch->p to fill that gap.
732 * If allocation fails we free all blocks we've allocated (and forget
733 * their buffer_heads) and return the error value the from failed
734 * ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
735 * as described above and return 0.
737 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
738 ext4_lblk_t iblock, int indirect_blks,
739 int *blks, ext4_fsblk_t goal,
740 ext4_lblk_t *offsets, Indirect *branch)
742 int blocksize = inode->i_sb->s_blocksize;
745 struct buffer_head *bh;
747 ext4_fsblk_t new_blocks[4];
748 ext4_fsblk_t current_block;
750 num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
751 *blks, new_blocks, &err);
755 branch[0].key = cpu_to_le32(new_blocks[0]);
757 * metadata blocks and data blocks are allocated.
759 for (n = 1; n <= indirect_blks; n++) {
761 * Get buffer_head for parent block, zero it out
762 * and set the pointer to new one, then send
765 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
773 BUFFER_TRACE(bh, "call get_create_access");
774 err = ext4_journal_get_create_access(handle, bh);
776 /* Don't brelse(bh) here; it's done in
777 * ext4_journal_forget() below */
782 memset(bh->b_data, 0, blocksize);
783 branch[n].p = (__le32 *) bh->b_data + offsets[n];
784 branch[n].key = cpu_to_le32(new_blocks[n]);
785 *branch[n].p = branch[n].key;
786 if (n == indirect_blks) {
787 current_block = new_blocks[n];
789 * End of chain, update the last new metablock of
790 * the chain to point to the new allocated
791 * data blocks numbers
793 for (i = 1; i < num; i++)
794 *(branch[n].p + i) = cpu_to_le32(++current_block);
796 BUFFER_TRACE(bh, "marking uptodate");
797 set_buffer_uptodate(bh);
800 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
801 err = ext4_handle_dirty_metadata(handle, inode, bh);
808 /* Allocation failed, free what we already allocated */
809 ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
810 for (i = 1; i <= n ; i++) {
812 * branch[i].bh is newly allocated, so there is no
813 * need to revoke the block, which is why we don't
814 * need to set EXT4_FREE_BLOCKS_METADATA.
816 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
817 EXT4_FREE_BLOCKS_FORGET);
819 for (i = n+1; i < indirect_blks; i++)
820 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
822 ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
828 * ext4_splice_branch - splice the allocated branch onto inode.
830 * @block: (logical) number of block we are adding
831 * @chain: chain of indirect blocks (with a missing link - see
833 * @where: location of missing link
834 * @num: number of indirect blocks we are adding
835 * @blks: number of direct blocks we are adding
837 * This function fills the missing link and does all housekeeping needed in
838 * inode (->i_blocks, etc.). In case of success we end up with the full
839 * chain to new block and return 0.
841 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
842 ext4_lblk_t block, Indirect *where, int num,
847 ext4_fsblk_t current_block;
850 * If we're splicing into a [td]indirect block (as opposed to the
851 * inode) then we need to get write access to the [td]indirect block
855 BUFFER_TRACE(where->bh, "get_write_access");
856 err = ext4_journal_get_write_access(handle, where->bh);
862 *where->p = where->key;
865 * Update the host buffer_head or inode to point to more just allocated
866 * direct blocks blocks
868 if (num == 0 && blks > 1) {
869 current_block = le32_to_cpu(where->key) + 1;
870 for (i = 1; i < blks; i++)
871 *(where->p + i) = cpu_to_le32(current_block++);
874 /* We are done with atomic stuff, now do the rest of housekeeping */
875 /* had we spliced it onto indirect block? */
878 * If we spliced it onto an indirect block, we haven't
879 * altered the inode. Note however that if it is being spliced
880 * onto an indirect block at the very end of the file (the
881 * file is growing) then we *will* alter the inode to reflect
882 * the new i_size. But that is not done here - it is done in
883 * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
885 jbd_debug(5, "splicing indirect only\n");
886 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
887 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
892 * OK, we spliced it into the inode itself on a direct block.
894 ext4_mark_inode_dirty(handle, inode);
895 jbd_debug(5, "splicing direct\n");
900 for (i = 1; i <= num; i++) {
902 * branch[i].bh is newly allocated, so there is no
903 * need to revoke the block, which is why we don't
904 * need to set EXT4_FREE_BLOCKS_METADATA.
906 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
907 EXT4_FREE_BLOCKS_FORGET);
909 ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
916 * The ext4_ind_map_blocks() function handles non-extents inodes
917 * (i.e., using the traditional indirect/double-indirect i_blocks
918 * scheme) for ext4_map_blocks().
920 * Allocation strategy is simple: if we have to allocate something, we will
921 * have to go the whole way to leaf. So let's do it before attaching anything
922 * to tree, set linkage between the newborn blocks, write them if sync is
923 * required, recheck the path, free and repeat if check fails, otherwise
924 * set the last missing link (that will protect us from any truncate-generated
925 * removals - all blocks on the path are immune now) and possibly force the
926 * write on the parent block.
927 * That has a nice additional property: no special recovery from the failed
928 * allocations is needed - we simply release blocks and do not touch anything
929 * reachable from inode.
931 * `handle' can be NULL if create == 0.
933 * return > 0, # of blocks mapped or allocated.
934 * return = 0, if plain lookup failed.
935 * return < 0, error case.
937 * The ext4_ind_get_blocks() function should be called with
938 * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
939 * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
940 * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
943 static int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
944 struct ext4_map_blocks *map,
948 ext4_lblk_t offsets[4];
953 int blocks_to_boundary = 0;
956 ext4_fsblk_t first_block = 0;
958 J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
959 J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
960 depth = ext4_block_to_path(inode, map->m_lblk, offsets,
961 &blocks_to_boundary);
966 partial = ext4_get_branch(inode, depth, offsets, chain, &err);
968 /* Simplest case - block found, no allocation needed */
970 first_block = le32_to_cpu(chain[depth - 1].key);
973 while (count < map->m_len && count <= blocks_to_boundary) {
976 blk = le32_to_cpu(*(chain[depth-1].p + count));
978 if (blk == first_block + count)
986 /* Next simple case - plain lookup or failed read of indirect block */
987 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
991 * Okay, we need to do block allocation.
993 goal = ext4_find_goal(inode, map->m_lblk, partial);
995 /* the number of blocks need to allocate for [d,t]indirect blocks */
996 indirect_blks = (chain + depth) - partial - 1;
999 * Next look up the indirect map to count the totoal number of
1000 * direct blocks to allocate for this branch.
1002 count = ext4_blks_to_allocate(partial, indirect_blks,
1003 map->m_len, blocks_to_boundary);
1005 * Block out ext4_truncate while we alter the tree
1007 err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
1009 offsets + (partial - chain), partial);
1012 * The ext4_splice_branch call will free and forget any buffers
1013 * on the new chain if there is a failure, but that risks using
1014 * up transaction credits, especially for bitmaps where the
1015 * credits cannot be returned. Can we handle this somehow? We
1016 * may need to return -EAGAIN upwards in the worst case. --sct
1019 err = ext4_splice_branch(handle, inode, map->m_lblk,
1020 partial, indirect_blks, count);
1024 map->m_flags |= EXT4_MAP_NEW;
1026 ext4_update_inode_fsync_trans(handle, inode, 1);
1028 map->m_flags |= EXT4_MAP_MAPPED;
1029 map->m_pblk = le32_to_cpu(chain[depth-1].key);
1031 if (count > blocks_to_boundary)
1032 map->m_flags |= EXT4_MAP_BOUNDARY;
1034 /* Clean up and exit */
1035 partial = chain + depth - 1; /* the whole chain */
1037 while (partial > chain) {
1038 BUFFER_TRACE(partial->bh, "call brelse");
1039 brelse(partial->bh);
1047 qsize_t *ext4_get_reserved_space(struct inode *inode)
1049 return &EXT4_I(inode)->i_reserved_quota;
1054 * Calculate the number of metadata blocks need to reserve
1055 * to allocate a new block at @lblocks for non extent file based file
1057 static int ext4_indirect_calc_metadata_amount(struct inode *inode,
1060 struct ext4_inode_info *ei = EXT4_I(inode);
1061 sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
1064 if (lblock < EXT4_NDIR_BLOCKS)
1067 lblock -= EXT4_NDIR_BLOCKS;
1069 if (ei->i_da_metadata_calc_len &&
1070 (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
1071 ei->i_da_metadata_calc_len++;
1074 ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
1075 ei->i_da_metadata_calc_len = 1;
1076 blk_bits = order_base_2(lblock);
1077 return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
1081 * Calculate the number of metadata blocks need to reserve
1082 * to allocate a block located at @lblock
1084 static int ext4_calc_metadata_amount(struct inode *inode, sector_t lblock)
1086 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1087 return ext4_ext_calc_metadata_amount(inode, lblock);
1089 return ext4_indirect_calc_metadata_amount(inode, lblock);
1093 * Called with i_data_sem down, which is important since we can call
1094 * ext4_discard_preallocations() from here.
1096 void ext4_da_update_reserve_space(struct inode *inode,
1097 int used, int quota_claim)
1099 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1100 struct ext4_inode_info *ei = EXT4_I(inode);
1102 spin_lock(&ei->i_block_reservation_lock);
1103 trace_ext4_da_update_reserve_space(inode, used);
1104 if (unlikely(used > ei->i_reserved_data_blocks)) {
1105 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
1106 "with only %d reserved data blocks\n",
1107 __func__, inode->i_ino, used,
1108 ei->i_reserved_data_blocks);
1110 used = ei->i_reserved_data_blocks;
1113 /* Update per-inode reservations */
1114 ei->i_reserved_data_blocks -= used;
1115 ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
1116 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1117 used + ei->i_allocated_meta_blocks);
1118 ei->i_allocated_meta_blocks = 0;
1120 if (ei->i_reserved_data_blocks == 0) {
1122 * We can release all of the reserved metadata blocks
1123 * only when we have written all of the delayed
1124 * allocation blocks.
1126 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1127 ei->i_reserved_meta_blocks);
1128 ei->i_reserved_meta_blocks = 0;
1129 ei->i_da_metadata_calc_len = 0;
1131 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1133 /* Update quota subsystem for data blocks */
1135 dquot_claim_block(inode, used);
1138 * We did fallocate with an offset that is already delayed
1139 * allocated. So on delayed allocated writeback we should
1140 * not re-claim the quota for fallocated blocks.
1142 dquot_release_reservation_block(inode, used);
1146 * If we have done all the pending block allocations and if
1147 * there aren't any writers on the inode, we can discard the
1148 * inode's preallocations.
1150 if ((ei->i_reserved_data_blocks == 0) &&
1151 (atomic_read(&inode->i_writecount) == 0))
1152 ext4_discard_preallocations(inode);
1155 static int __check_block_validity(struct inode *inode, const char *func,
1157 struct ext4_map_blocks *map)
1159 if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
1161 ext4_error_inode(inode, func, line, map->m_pblk,
1162 "lblock %lu mapped to illegal pblock "
1163 "(length %d)", (unsigned long) map->m_lblk,
1170 #define check_block_validity(inode, map) \
1171 __check_block_validity((inode), __func__, __LINE__, (map))
1174 * Return the number of contiguous dirty pages in a given inode
1175 * starting at page frame idx.
1177 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1178 unsigned int max_pages)
1180 struct address_space *mapping = inode->i_mapping;
1182 struct pagevec pvec;
1184 int i, nr_pages, done = 0;
1188 pagevec_init(&pvec, 0);
1191 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1192 PAGECACHE_TAG_DIRTY,
1193 (pgoff_t)PAGEVEC_SIZE);
1196 for (i = 0; i < nr_pages; i++) {
1197 struct page *page = pvec.pages[i];
1198 struct buffer_head *bh, *head;
1201 if (unlikely(page->mapping != mapping) ||
1203 PageWriteback(page) ||
1204 page->index != idx) {
1209 if (page_has_buffers(page)) {
1210 bh = head = page_buffers(page);
1212 if (!buffer_delay(bh) &&
1213 !buffer_unwritten(bh))
1215 bh = bh->b_this_page;
1216 } while (!done && (bh != head));
1223 if (num >= max_pages) {
1228 pagevec_release(&pvec);
1234 * The ext4_map_blocks() function tries to look up the requested blocks,
1235 * and returns if the blocks are already mapped.
1237 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1238 * and store the allocated blocks in the result buffer head and mark it
1241 * If file type is extents based, it will call ext4_ext_map_blocks(),
1242 * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
1245 * On success, it returns the number of blocks being mapped or allocate.
1246 * if create==0 and the blocks are pre-allocated and uninitialized block,
1247 * the result buffer head is unmapped. If the create ==1, it will make sure
1248 * the buffer head is mapped.
1250 * It returns 0 if plain look up failed (blocks have not been allocated), in
1251 * that casem, buffer head is unmapped
1253 * It returns the error in case of allocation failure.
1255 int ext4_map_blocks(handle_t *handle, struct inode *inode,
1256 struct ext4_map_blocks *map, int flags)
1261 ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
1262 "logical block %lu\n", inode->i_ino, flags, map->m_len,
1263 (unsigned long) map->m_lblk);
1265 * Try to see if we can get the block without requesting a new
1266 * file system block.
1268 down_read((&EXT4_I(inode)->i_data_sem));
1269 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1270 retval = ext4_ext_map_blocks(handle, inode, map, 0);
1272 retval = ext4_ind_map_blocks(handle, inode, map, 0);
1274 up_read((&EXT4_I(inode)->i_data_sem));
1276 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1277 int ret = check_block_validity(inode, map);
1282 /* If it is only a block(s) look up */
1283 if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1287 * Returns if the blocks have already allocated
1289 * Note that if blocks have been preallocated
1290 * ext4_ext_get_block() returns th create = 0
1291 * with buffer head unmapped.
1293 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
1297 * When we call get_blocks without the create flag, the
1298 * BH_Unwritten flag could have gotten set if the blocks
1299 * requested were part of a uninitialized extent. We need to
1300 * clear this flag now that we are committed to convert all or
1301 * part of the uninitialized extent to be an initialized
1302 * extent. This is because we need to avoid the combination
1303 * of BH_Unwritten and BH_Mapped flags being simultaneously
1304 * set on the buffer_head.
1306 map->m_flags &= ~EXT4_MAP_UNWRITTEN;
1309 * New blocks allocate and/or writing to uninitialized extent
1310 * will possibly result in updating i_data, so we take
1311 * the write lock of i_data_sem, and call get_blocks()
1312 * with create == 1 flag.
1314 down_write((&EXT4_I(inode)->i_data_sem));
1317 * if the caller is from delayed allocation writeout path
1318 * we have already reserved fs blocks for allocation
1319 * let the underlying get_block() function know to
1320 * avoid double accounting
1322 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1323 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1325 * We need to check for EXT4 here because migrate
1326 * could have changed the inode type in between
1328 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
1329 retval = ext4_ext_map_blocks(handle, inode, map, flags);
1331 retval = ext4_ind_map_blocks(handle, inode, map, flags);
1333 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
1335 * We allocated new blocks which will result in
1336 * i_data's format changing. Force the migrate
1337 * to fail by clearing migrate flags
1339 ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
1343 * Update reserved blocks/metadata blocks after successful
1344 * block allocation which had been deferred till now. We don't
1345 * support fallocate for non extent files. So we can update
1346 * reserve space here.
1349 (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
1350 ext4_da_update_reserve_space(inode, retval, 1);
1352 if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1353 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1355 up_write((&EXT4_I(inode)->i_data_sem));
1356 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
1357 int ret = check_block_validity(inode, map);
1364 /* Maximum number of blocks we map for direct IO at once. */
1365 #define DIO_MAX_BLOCKS 4096
1367 static int _ext4_get_block(struct inode *inode, sector_t iblock,
1368 struct buffer_head *bh, int flags)
1370 handle_t *handle = ext4_journal_current_handle();
1371 struct ext4_map_blocks map;
1372 int ret = 0, started = 0;
1375 map.m_lblk = iblock;
1376 map.m_len = bh->b_size >> inode->i_blkbits;
1378 if (flags && !handle) {
1379 /* Direct IO write... */
1380 if (map.m_len > DIO_MAX_BLOCKS)
1381 map.m_len = DIO_MAX_BLOCKS;
1382 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
1383 handle = ext4_journal_start(inode, dio_credits);
1384 if (IS_ERR(handle)) {
1385 ret = PTR_ERR(handle);
1391 ret = ext4_map_blocks(handle, inode, &map, flags);
1393 map_bh(bh, inode->i_sb, map.m_pblk);
1394 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1395 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
1399 ext4_journal_stop(handle);
1403 int ext4_get_block(struct inode *inode, sector_t iblock,
1404 struct buffer_head *bh, int create)
1406 return _ext4_get_block(inode, iblock, bh,
1407 create ? EXT4_GET_BLOCKS_CREATE : 0);
1411 * `handle' can be NULL if create is zero
1413 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1414 ext4_lblk_t block, int create, int *errp)
1416 struct ext4_map_blocks map;
1417 struct buffer_head *bh;
1420 J_ASSERT(handle != NULL || create == 0);
1424 err = ext4_map_blocks(handle, inode, &map,
1425 create ? EXT4_GET_BLOCKS_CREATE : 0);
1433 bh = sb_getblk(inode->i_sb, map.m_pblk);
1438 if (map.m_flags & EXT4_MAP_NEW) {
1439 J_ASSERT(create != 0);
1440 J_ASSERT(handle != NULL);
1443 * Now that we do not always journal data, we should
1444 * keep in mind whether this should always journal the
1445 * new buffer as metadata. For now, regular file
1446 * writes use ext4_get_block instead, so it's not a
1450 BUFFER_TRACE(bh, "call get_create_access");
1451 fatal = ext4_journal_get_create_access(handle, bh);
1452 if (!fatal && !buffer_uptodate(bh)) {
1453 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1454 set_buffer_uptodate(bh);
1457 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1458 err = ext4_handle_dirty_metadata(handle, inode, bh);
1462 BUFFER_TRACE(bh, "not a new buffer");
1472 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1473 ext4_lblk_t block, int create, int *err)
1475 struct buffer_head *bh;
1477 bh = ext4_getblk(handle, inode, block, create, err);
1480 if (buffer_uptodate(bh))
1482 ll_rw_block(READ_META, 1, &bh);
1484 if (buffer_uptodate(bh))
1491 static int walk_page_buffers(handle_t *handle,
1492 struct buffer_head *head,
1496 int (*fn)(handle_t *handle,
1497 struct buffer_head *bh))
1499 struct buffer_head *bh;
1500 unsigned block_start, block_end;
1501 unsigned blocksize = head->b_size;
1503 struct buffer_head *next;
1505 for (bh = head, block_start = 0;
1506 ret == 0 && (bh != head || !block_start);
1507 block_start = block_end, bh = next) {
1508 next = bh->b_this_page;
1509 block_end = block_start + blocksize;
1510 if (block_end <= from || block_start >= to) {
1511 if (partial && !buffer_uptodate(bh))
1515 err = (*fn)(handle, bh);
1523 * To preserve ordering, it is essential that the hole instantiation and
1524 * the data write be encapsulated in a single transaction. We cannot
1525 * close off a transaction and start a new one between the ext4_get_block()
1526 * and the commit_write(). So doing the jbd2_journal_start at the start of
1527 * prepare_write() is the right place.
1529 * Also, this function can nest inside ext4_writepage() ->
1530 * block_write_full_page(). In that case, we *know* that ext4_writepage()
1531 * has generated enough buffer credits to do the whole page. So we won't
1532 * block on the journal in that case, which is good, because the caller may
1535 * By accident, ext4 can be reentered when a transaction is open via
1536 * quota file writes. If we were to commit the transaction while thus
1537 * reentered, there can be a deadlock - we would be holding a quota
1538 * lock, and the commit would never complete if another thread had a
1539 * transaction open and was blocking on the quota lock - a ranking
1542 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1543 * will _not_ run commit under these circumstances because handle->h_ref
1544 * is elevated. We'll still have enough credits for the tiny quotafile
1547 static int do_journal_get_write_access(handle_t *handle,
1548 struct buffer_head *bh)
1550 int dirty = buffer_dirty(bh);
1553 if (!buffer_mapped(bh) || buffer_freed(bh))
1556 * __block_write_begin() could have dirtied some buffers. Clean
1557 * the dirty bit as jbd2_journal_get_write_access() could complain
1558 * otherwise about fs integrity issues. Setting of the dirty bit
1559 * by __block_write_begin() isn't a real problem here as we clear
1560 * the bit before releasing a page lock and thus writeback cannot
1561 * ever write the buffer.
1564 clear_buffer_dirty(bh);
1565 ret = ext4_journal_get_write_access(handle, bh);
1567 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1572 * Truncate blocks that were not used by write. We have to truncate the
1573 * pagecache as well so that corresponding buffers get properly unmapped.
1575 static void ext4_truncate_failed_write(struct inode *inode)
1577 truncate_inode_pages(inode->i_mapping, inode->i_size);
1578 ext4_truncate(inode);
1581 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
1582 struct buffer_head *bh_result, int create);
1583 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1584 loff_t pos, unsigned len, unsigned flags,
1585 struct page **pagep, void **fsdata)
1587 struct inode *inode = mapping->host;
1588 int ret, needed_blocks;
1595 trace_ext4_write_begin(inode, pos, len, flags);
1597 * Reserve one block more for addition to orphan list in case
1598 * we allocate blocks but write fails for some reason
1600 needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1601 index = pos >> PAGE_CACHE_SHIFT;
1602 from = pos & (PAGE_CACHE_SIZE - 1);
1606 handle = ext4_journal_start(inode, needed_blocks);
1607 if (IS_ERR(handle)) {
1608 ret = PTR_ERR(handle);
1612 /* We cannot recurse into the filesystem as the transaction is already
1614 flags |= AOP_FLAG_NOFS;
1616 page = grab_cache_page_write_begin(mapping, index, flags);
1618 ext4_journal_stop(handle);
1624 if (ext4_should_dioread_nolock(inode))
1625 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
1627 ret = __block_write_begin(page, pos, len, ext4_get_block);
1629 if (!ret && ext4_should_journal_data(inode)) {
1630 ret = walk_page_buffers(handle, page_buffers(page),
1631 from, to, NULL, do_journal_get_write_access);
1636 page_cache_release(page);
1638 * __block_write_begin may have instantiated a few blocks
1639 * outside i_size. Trim these off again. Don't need
1640 * i_size_read because we hold i_mutex.
1642 * Add inode to orphan list in case we crash before
1645 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1646 ext4_orphan_add(handle, inode);
1648 ext4_journal_stop(handle);
1649 if (pos + len > inode->i_size) {
1650 ext4_truncate_failed_write(inode);
1652 * If truncate failed early the inode might
1653 * still be on the orphan list; we need to
1654 * make sure the inode is removed from the
1655 * orphan list in that case.
1658 ext4_orphan_del(NULL, inode);
1662 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1668 /* For write_end() in data=journal mode */
1669 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1671 if (!buffer_mapped(bh) || buffer_freed(bh))
1673 set_buffer_uptodate(bh);
1674 return ext4_handle_dirty_metadata(handle, NULL, bh);
1677 static int ext4_generic_write_end(struct file *file,
1678 struct address_space *mapping,
1679 loff_t pos, unsigned len, unsigned copied,
1680 struct page *page, void *fsdata)
1682 int i_size_changed = 0;
1683 struct inode *inode = mapping->host;
1684 handle_t *handle = ext4_journal_current_handle();
1686 copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1689 * No need to use i_size_read() here, the i_size
1690 * cannot change under us because we hold i_mutex.
1692 * But it's important to update i_size while still holding page lock:
1693 * page writeout could otherwise come in and zero beyond i_size.
1695 if (pos + copied > inode->i_size) {
1696 i_size_write(inode, pos + copied);
1700 if (pos + copied > EXT4_I(inode)->i_disksize) {
1701 /* We need to mark inode dirty even if
1702 * new_i_size is less that inode->i_size
1703 * bu greater than i_disksize.(hint delalloc)
1705 ext4_update_i_disksize(inode, (pos + copied));
1709 page_cache_release(page);
1712 * Don't mark the inode dirty under page lock. First, it unnecessarily
1713 * makes the holding time of page lock longer. Second, it forces lock
1714 * ordering of page lock and transaction start for journaling
1718 ext4_mark_inode_dirty(handle, inode);
1724 * We need to pick up the new inode size which generic_commit_write gave us
1725 * `file' can be NULL - eg, when called from page_symlink().
1727 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1728 * buffers are managed internally.
1730 static int ext4_ordered_write_end(struct file *file,
1731 struct address_space *mapping,
1732 loff_t pos, unsigned len, unsigned copied,
1733 struct page *page, void *fsdata)
1735 handle_t *handle = ext4_journal_current_handle();
1736 struct inode *inode = mapping->host;
1739 trace_ext4_ordered_write_end(inode, pos, len, copied);
1740 ret = ext4_jbd2_file_inode(handle, inode);
1743 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1746 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1747 /* if we have allocated more blocks and copied
1748 * less. We will have blocks allocated outside
1749 * inode->i_size. So truncate them
1751 ext4_orphan_add(handle, inode);
1755 ret2 = ext4_journal_stop(handle);
1759 if (pos + len > inode->i_size) {
1760 ext4_truncate_failed_write(inode);
1762 * If truncate failed early the inode might still be
1763 * on the orphan list; we need to make sure the inode
1764 * is removed from the orphan list in that case.
1767 ext4_orphan_del(NULL, inode);
1771 return ret ? ret : copied;
1774 static int ext4_writeback_write_end(struct file *file,
1775 struct address_space *mapping,
1776 loff_t pos, unsigned len, unsigned copied,
1777 struct page *page, void *fsdata)
1779 handle_t *handle = ext4_journal_current_handle();
1780 struct inode *inode = mapping->host;
1783 trace_ext4_writeback_write_end(inode, pos, len, copied);
1784 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1787 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1788 /* if we have allocated more blocks and copied
1789 * less. We will have blocks allocated outside
1790 * inode->i_size. So truncate them
1792 ext4_orphan_add(handle, inode);
1797 ret2 = ext4_journal_stop(handle);
1801 if (pos + len > inode->i_size) {
1802 ext4_truncate_failed_write(inode);
1804 * If truncate failed early the inode might still be
1805 * on the orphan list; we need to make sure the inode
1806 * is removed from the orphan list in that case.
1809 ext4_orphan_del(NULL, inode);
1812 return ret ? ret : copied;
1815 static int ext4_journalled_write_end(struct file *file,
1816 struct address_space *mapping,
1817 loff_t pos, unsigned len, unsigned copied,
1818 struct page *page, void *fsdata)
1820 handle_t *handle = ext4_journal_current_handle();
1821 struct inode *inode = mapping->host;
1827 trace_ext4_journalled_write_end(inode, pos, len, copied);
1828 from = pos & (PAGE_CACHE_SIZE - 1);
1832 if (!PageUptodate(page))
1834 page_zero_new_buffers(page, from+copied, to);
1837 ret = walk_page_buffers(handle, page_buffers(page), from,
1838 to, &partial, write_end_fn);
1840 SetPageUptodate(page);
1841 new_i_size = pos + copied;
1842 if (new_i_size > inode->i_size)
1843 i_size_write(inode, pos+copied);
1844 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1845 if (new_i_size > EXT4_I(inode)->i_disksize) {
1846 ext4_update_i_disksize(inode, new_i_size);
1847 ret2 = ext4_mark_inode_dirty(handle, inode);
1853 page_cache_release(page);
1854 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1855 /* if we have allocated more blocks and copied
1856 * less. We will have blocks allocated outside
1857 * inode->i_size. So truncate them
1859 ext4_orphan_add(handle, inode);
1861 ret2 = ext4_journal_stop(handle);
1864 if (pos + len > inode->i_size) {
1865 ext4_truncate_failed_write(inode);
1867 * If truncate failed early the inode might still be
1868 * on the orphan list; we need to make sure the inode
1869 * is removed from the orphan list in that case.
1872 ext4_orphan_del(NULL, inode);
1875 return ret ? ret : copied;
1879 * Reserve a single block located at lblock
1881 static int ext4_da_reserve_space(struct inode *inode, sector_t lblock)
1884 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1885 struct ext4_inode_info *ei = EXT4_I(inode);
1886 unsigned long md_needed;
1890 * recalculate the amount of metadata blocks to reserve
1891 * in order to allocate nrblocks
1892 * worse case is one extent per block
1895 spin_lock(&ei->i_block_reservation_lock);
1896 md_needed = ext4_calc_metadata_amount(inode, lblock);
1897 trace_ext4_da_reserve_space(inode, md_needed);
1898 spin_unlock(&ei->i_block_reservation_lock);
1901 * We will charge metadata quota at writeout time; this saves
1902 * us from metadata over-estimation, though we may go over by
1903 * a small amount in the end. Here we just reserve for data.
1905 ret = dquot_reserve_block(inode, 1);
1909 * We do still charge estimated metadata to the sb though;
1910 * we cannot afford to run out of free blocks.
1912 if (ext4_claim_free_blocks(sbi, md_needed + 1)) {
1913 dquot_release_reservation_block(inode, 1);
1914 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1920 spin_lock(&ei->i_block_reservation_lock);
1921 ei->i_reserved_data_blocks++;
1922 ei->i_reserved_meta_blocks += md_needed;
1923 spin_unlock(&ei->i_block_reservation_lock);
1925 return 0; /* success */
1928 static void ext4_da_release_space(struct inode *inode, int to_free)
1930 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1931 struct ext4_inode_info *ei = EXT4_I(inode);
1934 return; /* Nothing to release, exit */
1936 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1938 trace_ext4_da_release_space(inode, to_free);
1939 if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1941 * if there aren't enough reserved blocks, then the
1942 * counter is messed up somewhere. Since this
1943 * function is called from invalidate page, it's
1944 * harmless to return without any action.
1946 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1947 "ino %lu, to_free %d with only %d reserved "
1948 "data blocks\n", inode->i_ino, to_free,
1949 ei->i_reserved_data_blocks);
1951 to_free = ei->i_reserved_data_blocks;
1953 ei->i_reserved_data_blocks -= to_free;
1955 if (ei->i_reserved_data_blocks == 0) {
1957 * We can release all of the reserved metadata blocks
1958 * only when we have written all of the delayed
1959 * allocation blocks.
1961 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1962 ei->i_reserved_meta_blocks);
1963 ei->i_reserved_meta_blocks = 0;
1964 ei->i_da_metadata_calc_len = 0;
1967 /* update fs dirty data blocks counter */
1968 percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1970 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1972 dquot_release_reservation_block(inode, to_free);
1975 static void ext4_da_page_release_reservation(struct page *page,
1976 unsigned long offset)
1979 struct buffer_head *head, *bh;
1980 unsigned int curr_off = 0;
1982 head = page_buffers(page);
1985 unsigned int next_off = curr_off + bh->b_size;
1987 if ((offset <= curr_off) && (buffer_delay(bh))) {
1989 clear_buffer_delay(bh);
1991 curr_off = next_off;
1992 } while ((bh = bh->b_this_page) != head);
1993 ext4_da_release_space(page->mapping->host, to_release);
1997 * Delayed allocation stuff
2001 * mpage_da_submit_io - walks through extent of pages and try to write
2002 * them with writepage() call back
2004 * @mpd->inode: inode
2005 * @mpd->first_page: first page of the extent
2006 * @mpd->next_page: page after the last page of the extent
2008 * By the time mpage_da_submit_io() is called we expect all blocks
2009 * to be allocated. this may be wrong if allocation failed.
2011 * As pages are already locked by write_cache_pages(), we can't use it
2013 static int mpage_da_submit_io(struct mpage_da_data *mpd,
2014 struct ext4_map_blocks *map)
2016 struct pagevec pvec;
2017 unsigned long index, end;
2018 int ret = 0, err, nr_pages, i;
2019 struct inode *inode = mpd->inode;
2020 struct address_space *mapping = inode->i_mapping;
2021 loff_t size = i_size_read(inode);
2022 unsigned int len, block_start;
2023 struct buffer_head *bh, *page_bufs = NULL;
2024 int journal_data = ext4_should_journal_data(inode);
2025 sector_t pblock = 0, cur_logical = 0;
2026 struct ext4_io_submit io_submit;
2028 BUG_ON(mpd->next_page <= mpd->first_page);
2029 memset(&io_submit, 0, sizeof(io_submit));
2031 * We need to start from the first_page to the next_page - 1
2032 * to make sure we also write the mapped dirty buffer_heads.
2033 * If we look at mpd->b_blocknr we would only be looking
2034 * at the currently mapped buffer_heads.
2036 index = mpd->first_page;
2037 end = mpd->next_page - 1;
2039 pagevec_init(&pvec, 0);
2040 while (index <= end) {
2041 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2044 for (i = 0; i < nr_pages; i++) {
2045 int commit_write = 0, redirty_page = 0;
2046 struct page *page = pvec.pages[i];
2048 index = page->index;
2052 if (index == size >> PAGE_CACHE_SHIFT)
2053 len = size & ~PAGE_CACHE_MASK;
2055 len = PAGE_CACHE_SIZE;
2057 cur_logical = index << (PAGE_CACHE_SHIFT -
2059 pblock = map->m_pblk + (cur_logical -
2064 BUG_ON(!PageLocked(page));
2065 BUG_ON(PageWriteback(page));
2068 * If the page does not have buffers (for
2069 * whatever reason), try to create them using
2070 * __block_write_begin. If this fails,
2071 * redirty the page and move on.
2073 if (!page_has_buffers(page)) {
2074 if (__block_write_begin(page, 0, len,
2075 noalloc_get_block_write)) {
2077 redirty_page_for_writepage(mpd->wbc,
2085 bh = page_bufs = page_buffers(page);
2090 if (map && (cur_logical >= map->m_lblk) &&
2091 (cur_logical <= (map->m_lblk +
2092 (map->m_len - 1)))) {
2093 if (buffer_delay(bh)) {
2094 clear_buffer_delay(bh);
2095 bh->b_blocknr = pblock;
2097 if (buffer_unwritten(bh) ||
2099 BUG_ON(bh->b_blocknr != pblock);
2100 if (map->m_flags & EXT4_MAP_UNINIT)
2101 set_buffer_uninit(bh);
2102 clear_buffer_unwritten(bh);
2105 /* redirty page if block allocation undone */
2106 if (buffer_delay(bh) || buffer_unwritten(bh))
2108 bh = bh->b_this_page;
2109 block_start += bh->b_size;
2112 } while (bh != page_bufs);
2118 /* mark the buffer_heads as dirty & uptodate */
2119 block_commit_write(page, 0, len);
2122 * Delalloc doesn't support data journalling,
2123 * but eventually maybe we'll lift this
2126 if (unlikely(journal_data && PageChecked(page)))
2127 err = __ext4_journalled_writepage(page, len);
2129 err = ext4_bio_write_page(&io_submit, page,
2133 mpd->pages_written++;
2135 * In error case, we have to continue because
2136 * remaining pages are still locked
2141 pagevec_release(&pvec);
2143 ext4_io_submit(&io_submit);
2147 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2148 sector_t logical, long blk_cnt)
2152 struct pagevec pvec;
2153 struct inode *inode = mpd->inode;
2154 struct address_space *mapping = inode->i_mapping;
2156 index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2157 end = (logical + blk_cnt - 1) >>
2158 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2159 while (index <= end) {
2160 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2163 for (i = 0; i < nr_pages; i++) {
2164 struct page *page = pvec.pages[i];
2165 if (page->index > end)
2167 BUG_ON(!PageLocked(page));
2168 BUG_ON(PageWriteback(page));
2169 block_invalidatepage(page, 0);
2170 ClearPageUptodate(page);
2173 index = pvec.pages[nr_pages - 1]->index + 1;
2174 pagevec_release(&pvec);
2179 static void ext4_print_free_blocks(struct inode *inode)
2181 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2182 printk(KERN_CRIT "Total free blocks count %lld\n",
2183 ext4_count_free_blocks(inode->i_sb));
2184 printk(KERN_CRIT "Free/Dirty block details\n");
2185 printk(KERN_CRIT "free_blocks=%lld\n",
2186 (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2187 printk(KERN_CRIT "dirty_blocks=%lld\n",
2188 (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2189 printk(KERN_CRIT "Block reservation details\n");
2190 printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2191 EXT4_I(inode)->i_reserved_data_blocks);
2192 printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2193 EXT4_I(inode)->i_reserved_meta_blocks);
2198 * mpage_da_map_and_submit - go through given space, map them
2199 * if necessary, and then submit them for I/O
2201 * @mpd - bh describing space
2203 * The function skips space we know is already mapped to disk blocks.
2206 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
2208 int err, blks, get_blocks_flags;
2209 struct ext4_map_blocks map, *mapp = NULL;
2210 sector_t next = mpd->b_blocknr;
2211 unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2212 loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2213 handle_t *handle = NULL;
2216 * If the blocks are mapped already, or we couldn't accumulate
2217 * any blocks, then proceed immediately to the submission stage.
2219 if ((mpd->b_size == 0) ||
2220 ((mpd->b_state & (1 << BH_Mapped)) &&
2221 !(mpd->b_state & (1 << BH_Delay)) &&
2222 !(mpd->b_state & (1 << BH_Unwritten))))
2225 handle = ext4_journal_current_handle();
2229 * Call ext4_map_blocks() to allocate any delayed allocation
2230 * blocks, or to convert an uninitialized extent to be
2231 * initialized (in the case where we have written into
2232 * one or more preallocated blocks).
2234 * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2235 * indicate that we are on the delayed allocation path. This
2236 * affects functions in many different parts of the allocation
2237 * call path. This flag exists primarily because we don't
2238 * want to change *many* call functions, so ext4_map_blocks()
2239 * will set the magic i_delalloc_reserved_flag once the
2240 * inode's allocation semaphore is taken.
2242 * If the blocks in questions were delalloc blocks, set
2243 * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2244 * variables are updated after the blocks have been allocated.
2247 map.m_len = max_blocks;
2248 get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
2249 if (ext4_should_dioread_nolock(mpd->inode))
2250 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2251 if (mpd->b_state & (1 << BH_Delay))
2252 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2254 blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
2256 struct super_block *sb = mpd->inode->i_sb;
2260 * If get block returns EAGAIN or ENOSPC and there
2261 * appears to be free blocks we will call
2262 * ext4_writepage() for all of the pages which will
2263 * just redirty the pages.
2268 if (err == -ENOSPC &&
2269 ext4_count_free_blocks(sb)) {
2275 * get block failure will cause us to loop in
2276 * writepages, because a_ops->writepage won't be able
2277 * to make progress. The page will be redirtied by
2278 * writepage and writepages will again try to write
2281 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2282 ext4_msg(sb, KERN_CRIT,
2283 "delayed block allocation failed for inode %lu "
2284 "at logical offset %llu with max blocks %zd "
2285 "with error %d", mpd->inode->i_ino,
2286 (unsigned long long) next,
2287 mpd->b_size >> mpd->inode->i_blkbits, err);
2288 ext4_msg(sb, KERN_CRIT,
2289 "This should not happen!! Data will be lost\n");
2291 ext4_print_free_blocks(mpd->inode);
2293 /* invalidate all the pages */
2294 ext4_da_block_invalidatepages(mpd, next,
2295 mpd->b_size >> mpd->inode->i_blkbits);
2301 if (map.m_flags & EXT4_MAP_NEW) {
2302 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
2305 for (i = 0; i < map.m_len; i++)
2306 unmap_underlying_metadata(bdev, map.m_pblk + i);
2309 if (ext4_should_order_data(mpd->inode)) {
2310 err = ext4_jbd2_file_inode(handle, mpd->inode);
2312 /* This only happens if the journal is aborted */
2317 * Update on-disk size along with block allocation.
2319 disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2320 if (disksize > i_size_read(mpd->inode))
2321 disksize = i_size_read(mpd->inode);
2322 if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2323 ext4_update_i_disksize(mpd->inode, disksize);
2324 err = ext4_mark_inode_dirty(handle, mpd->inode);
2326 ext4_error(mpd->inode->i_sb,
2327 "Failed to mark inode %lu dirty",
2332 mpage_da_submit_io(mpd, mapp);
2336 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2337 (1 << BH_Delay) | (1 << BH_Unwritten))
2340 * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2342 * @mpd->lbh - extent of blocks
2343 * @logical - logical number of the block in the file
2344 * @bh - bh of the block (used to access block's state)
2346 * the function is used to collect contig. blocks in same state
2348 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2349 sector_t logical, size_t b_size,
2350 unsigned long b_state)
2353 int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2356 * XXX Don't go larger than mballoc is willing to allocate
2357 * This is a stopgap solution. We eventually need to fold
2358 * mpage_da_submit_io() into this function and then call
2359 * ext4_map_blocks() multiple times in a loop
2361 if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
2364 /* check if thereserved journal credits might overflow */
2365 if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
2366 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2368 * With non-extent format we are limited by the journal
2369 * credit available. Total credit needed to insert
2370 * nrblocks contiguous blocks is dependent on the
2371 * nrblocks. So limit nrblocks.
2374 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2375 EXT4_MAX_TRANS_DATA) {
2377 * Adding the new buffer_head would make it cross the
2378 * allowed limit for which we have journal credit
2379 * reserved. So limit the new bh->b_size
2381 b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2382 mpd->inode->i_blkbits;
2383 /* we will do mpage_da_submit_io in the next loop */
2387 * First block in the extent
2389 if (mpd->b_size == 0) {
2390 mpd->b_blocknr = logical;
2391 mpd->b_size = b_size;
2392 mpd->b_state = b_state & BH_FLAGS;
2396 next = mpd->b_blocknr + nrblocks;
2398 * Can we merge the block to our big extent?
2400 if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2401 mpd->b_size += b_size;
2407 * We couldn't merge the block to our extent, so we
2408 * need to flush current extent and start new one
2410 mpage_da_map_and_submit(mpd);
2414 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2416 return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2420 * __mpage_da_writepage - finds extent of pages and blocks
2422 * @page: page to consider
2423 * @wbc: not used, we just follow rules
2426 * The function finds extents of pages and scan them for all blocks.
2428 static int __mpage_da_writepage(struct page *page,
2429 struct writeback_control *wbc,
2430 struct mpage_da_data *mpd)
2432 struct inode *inode = mpd->inode;
2433 struct buffer_head *bh, *head;
2437 * Can we merge this page to current extent?
2439 if (mpd->next_page != page->index) {
2441 * Nope, we can't. So, we map non-allocated blocks
2442 * and start IO on them
2444 if (mpd->next_page != mpd->first_page) {
2445 mpage_da_map_and_submit(mpd);
2447 * skip rest of the page in the page_vec
2449 redirty_page_for_writepage(wbc, page);
2451 return MPAGE_DA_EXTENT_TAIL;
2455 * Start next extent of pages ...
2457 mpd->first_page = page->index;
2467 mpd->next_page = page->index + 1;
2468 logical = (sector_t) page->index <<
2469 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2471 if (!page_has_buffers(page)) {
2472 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2473 (1 << BH_Dirty) | (1 << BH_Uptodate));
2475 return MPAGE_DA_EXTENT_TAIL;
2478 * Page with regular buffer heads, just add all dirty ones
2480 head = page_buffers(page);
2483 BUG_ON(buffer_locked(bh));
2485 * We need to try to allocate
2486 * unmapped blocks in the same page.
2487 * Otherwise we won't make progress
2488 * with the page in ext4_writepage
2490 if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2491 mpage_add_bh_to_extent(mpd, logical,
2495 return MPAGE_DA_EXTENT_TAIL;
2496 } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2498 * mapped dirty buffer. We need to update
2499 * the b_state because we look at
2500 * b_state in mpage_da_map_blocks. We don't
2501 * update b_size because if we find an
2502 * unmapped buffer_head later we need to
2503 * use the b_state flag of that buffer_head.
2505 if (mpd->b_size == 0)
2506 mpd->b_state = bh->b_state & BH_FLAGS;
2509 } while ((bh = bh->b_this_page) != head);
2516 * This is a special get_blocks_t callback which is used by
2517 * ext4_da_write_begin(). It will either return mapped block or
2518 * reserve space for a single block.
2520 * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2521 * We also have b_blocknr = -1 and b_bdev initialized properly
2523 * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2524 * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2525 * initialized properly.
2527 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2528 struct buffer_head *bh, int create)
2530 struct ext4_map_blocks map;
2532 sector_t invalid_block = ~((sector_t) 0xffff);
2534 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2537 BUG_ON(create == 0);
2538 BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
2540 map.m_lblk = iblock;
2544 * first, we need to know whether the block is allocated already
2545 * preallocated blocks are unmapped but should treated
2546 * the same as allocated blocks.
2548 ret = ext4_map_blocks(NULL, inode, &map, 0);
2552 if (buffer_delay(bh))
2553 return 0; /* Not sure this could or should happen */
2555 * XXX: __block_write_begin() unmaps passed block, is it OK?
2557 ret = ext4_da_reserve_space(inode, iblock);
2559 /* not enough space to reserve */
2562 map_bh(bh, inode->i_sb, invalid_block);
2564 set_buffer_delay(bh);
2568 map_bh(bh, inode->i_sb, map.m_pblk);
2569 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
2571 if (buffer_unwritten(bh)) {
2572 /* A delayed write to unwritten bh should be marked
2573 * new and mapped. Mapped ensures that we don't do
2574 * get_block multiple times when we write to the same
2575 * offset and new ensures that we do proper zero out
2576 * for partial write.
2579 set_buffer_mapped(bh);
2585 * This function is used as a standard get_block_t calback function
2586 * when there is no desire to allocate any blocks. It is used as a
2587 * callback function for block_write_begin() and block_write_full_page().
2588 * These functions should only try to map a single block at a time.
2590 * Since this function doesn't do block allocations even if the caller
2591 * requests it by passing in create=1, it is critically important that
2592 * any caller checks to make sure that any buffer heads are returned
2593 * by this function are either all already mapped or marked for
2594 * delayed allocation before calling block_write_full_page(). Otherwise,
2595 * b_blocknr could be left unitialized, and the page write functions will
2596 * be taken by surprise.
2598 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2599 struct buffer_head *bh_result, int create)
2601 BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2602 return _ext4_get_block(inode, iblock, bh_result, 0);
2605 static int bget_one(handle_t *handle, struct buffer_head *bh)
2611 static int bput_one(handle_t *handle, struct buffer_head *bh)
2617 static int __ext4_journalled_writepage(struct page *page,
2620 struct address_space *mapping = page->mapping;
2621 struct inode *inode = mapping->host;
2622 struct buffer_head *page_bufs;
2623 handle_t *handle = NULL;
2627 ClearPageChecked(page);
2628 page_bufs = page_buffers(page);
2630 walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2631 /* As soon as we unlock the page, it can go away, but we have
2632 * references to buffers so we are safe */
2635 handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2636 if (IS_ERR(handle)) {
2637 ret = PTR_ERR(handle);
2641 ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2642 do_journal_get_write_access);
2644 err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2648 err = ext4_journal_stop(handle);
2652 walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2653 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2658 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
2659 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
2662 * Note that we don't need to start a transaction unless we're journaling data
2663 * because we should have holes filled from ext4_page_mkwrite(). We even don't
2664 * need to file the inode to the transaction's list in ordered mode because if
2665 * we are writing back data added by write(), the inode is already there and if
2666 * we are writing back data modified via mmap(), noone guarantees in which
2667 * transaction the data will hit the disk. In case we are journaling data, we
2668 * cannot start transaction directly because transaction start ranks above page
2669 * lock so we have to do some magic.
2671 * This function can get called via...
2672 * - ext4_da_writepages after taking page lock (have journal handle)
2673 * - journal_submit_inode_data_buffers (no journal handle)
2674 * - shrink_page_list via pdflush (no journal handle)
2675 * - grab_page_cache when doing write_begin (have journal handle)
2677 * We don't do any block allocation in this function. If we have page with
2678 * multiple blocks we need to write those buffer_heads that are mapped. This
2679 * is important for mmaped based write. So if we do with blocksize 1K
2680 * truncate(f, 1024);
2681 * a = mmap(f, 0, 4096);
2683 * truncate(f, 4096);
2684 * we have in the page first buffer_head mapped via page_mkwrite call back
2685 * but other bufer_heads would be unmapped but dirty(dirty done via the
2686 * do_wp_page). So writepage should write the first block. If we modify
2687 * the mmap area beyond 1024 we will again get a page_fault and the
2688 * page_mkwrite callback will do the block allocation and mark the
2689 * buffer_heads mapped.
2691 * We redirty the page if we have any buffer_heads that is either delay or
2692 * unwritten in the page.
2694 * We can get recursively called as show below.
2696 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2699 * But since we don't do any block allocation we should not deadlock.
2700 * Page also have the dirty flag cleared so we don't get recurive page_lock.
2702 static int ext4_writepage(struct page *page,
2703 struct writeback_control *wbc)
2705 int ret = 0, commit_write = 0;
2708 struct buffer_head *page_bufs = NULL;
2709 struct inode *inode = page->mapping->host;
2711 trace_ext4_writepage(inode, page);
2712 size = i_size_read(inode);
2713 if (page->index == size >> PAGE_CACHE_SHIFT)
2714 len = size & ~PAGE_CACHE_MASK;
2716 len = PAGE_CACHE_SIZE;
2719 * If the page does not have buffers (for whatever reason),
2720 * try to create them using __block_write_begin. If this
2721 * fails, redirty the page and move on.
2723 if (!page_has_buffers(page)) {
2724 if (__block_write_begin(page, 0, len,
2725 noalloc_get_block_write)) {
2727 redirty_page_for_writepage(wbc, page);
2733 page_bufs = page_buffers(page);
2734 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2735 ext4_bh_delay_or_unwritten)) {
2737 * We don't want to do block allocation, so redirty
2738 * the page and return. We may reach here when we do
2739 * a journal commit via journal_submit_inode_data_buffers.
2740 * We can also reach here via shrink_page_list
2745 /* now mark the buffer_heads as dirty and uptodate */
2746 block_commit_write(page, 0, len);
2748 if (PageChecked(page) && ext4_should_journal_data(inode))
2750 * It's mmapped pagecache. Add buffers and journal it. There
2751 * doesn't seem much point in redirtying the page here.
2753 return __ext4_journalled_writepage(page, len);
2755 if (buffer_uninit(page_bufs)) {
2756 ext4_set_bh_endio(page_bufs, inode);
2757 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2758 wbc, ext4_end_io_buffer_write);
2760 ret = block_write_full_page(page, noalloc_get_block_write,
2767 * This is called via ext4_da_writepages() to
2768 * calulate the total number of credits to reserve to fit
2769 * a single extent allocation into a single transaction,
2770 * ext4_da_writpeages() will loop calling this before
2771 * the block allocation.
2774 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2776 int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2779 * With non-extent format the journal credit needed to
2780 * insert nrblocks contiguous block is dependent on
2781 * number of contiguous block. So we will limit
2782 * number of contiguous block to a sane value
2784 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2785 (max_blocks > EXT4_MAX_TRANS_DATA))
2786 max_blocks = EXT4_MAX_TRANS_DATA;
2788 return ext4_chunk_trans_blocks(inode, max_blocks);
2792 * write_cache_pages_da - walk the list of dirty pages of the given
2793 * address space and call the callback function (which usually writes
2796 * This is a forked version of write_cache_pages(). Differences:
2797 * Range cyclic is ignored.
2798 * no_nrwrite_index_update is always presumed true
2800 static int write_cache_pages_da(struct address_space *mapping,
2801 struct writeback_control *wbc,
2802 struct mpage_da_data *mpd,
2803 pgoff_t *done_index)
2807 struct pagevec pvec;
2810 pgoff_t end; /* Inclusive */
2811 long nr_to_write = wbc->nr_to_write;
2814 pagevec_init(&pvec, 0);
2815 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2816 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2818 if (wbc->sync_mode == WB_SYNC_ALL)
2819 tag = PAGECACHE_TAG_TOWRITE;
2821 tag = PAGECACHE_TAG_DIRTY;
2823 *done_index = index;
2824 while (!done && (index <= end)) {
2827 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2828 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2832 for (i = 0; i < nr_pages; i++) {
2833 struct page *page = pvec.pages[i];
2836 * At this point, the page may be truncated or
2837 * invalidated (changing page->mapping to NULL), or
2838 * even swizzled back from swapper_space to tmpfs file
2839 * mapping. However, page->index will not change
2840 * because we have a reference on the page.
2842 if (page->index > end) {
2847 *done_index = page->index + 1;
2852 * Page truncated or invalidated. We can freely skip it
2853 * then, even for data integrity operations: the page
2854 * has disappeared concurrently, so there could be no
2855 * real expectation of this data interity operation
2856 * even if there is now a new, dirty page at the same
2857 * pagecache address.
2859 if (unlikely(page->mapping != mapping)) {
2865 if (!PageDirty(page)) {
2866 /* someone wrote it for us */
2867 goto continue_unlock;
2870 if (PageWriteback(page)) {
2871 if (wbc->sync_mode != WB_SYNC_NONE)
2872 wait_on_page_writeback(page);
2874 goto continue_unlock;
2877 BUG_ON(PageWriteback(page));
2878 if (!clear_page_dirty_for_io(page))
2879 goto continue_unlock;
2881 ret = __mpage_da_writepage(page, wbc, mpd);
2882 if (unlikely(ret)) {
2883 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2892 if (nr_to_write > 0) {
2894 if (nr_to_write == 0 &&
2895 wbc->sync_mode == WB_SYNC_NONE) {
2897 * We stop writing back only if we are
2898 * not doing integrity sync. In case of
2899 * integrity sync we have to keep going
2900 * because someone may be concurrently
2901 * dirtying pages, and we might have
2902 * synced a lot of newly appeared dirty
2903 * pages, but have not synced all of the
2911 pagevec_release(&pvec);
2918 static int ext4_da_writepages(struct address_space *mapping,
2919 struct writeback_control *wbc)
2922 int range_whole = 0;
2923 handle_t *handle = NULL;
2924 struct mpage_da_data mpd;
2925 struct inode *inode = mapping->host;
2926 int pages_written = 0;
2928 unsigned int max_pages;
2929 int range_cyclic, cycled = 1, io_done = 0;
2930 int needed_blocks, ret = 0;
2931 long desired_nr_to_write, nr_to_writebump = 0;
2932 loff_t range_start = wbc->range_start;
2933 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2934 pgoff_t done_index = 0;
2937 trace_ext4_da_writepages(inode, wbc);
2940 * No pages to write? This is mainly a kludge to avoid starting
2941 * a transaction for special inodes like journal inode on last iput()
2942 * because that could violate lock ordering on umount
2944 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2948 * If the filesystem has aborted, it is read-only, so return
2949 * right away instead of dumping stack traces later on that
2950 * will obscure the real source of the problem. We test
2951 * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2952 * the latter could be true if the filesystem is mounted
2953 * read-only, and in that case, ext4_da_writepages should
2954 * *never* be called, so if that ever happens, we would want
2957 if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2960 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2963 range_cyclic = wbc->range_cyclic;
2964 if (wbc->range_cyclic) {
2965 index = mapping->writeback_index;
2968 wbc->range_start = index << PAGE_CACHE_SHIFT;
2969 wbc->range_end = LLONG_MAX;
2970 wbc->range_cyclic = 0;
2973 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2974 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2978 * This works around two forms of stupidity. The first is in
2979 * the writeback code, which caps the maximum number of pages
2980 * written to be 1024 pages. This is wrong on multiple
2981 * levels; different architectues have a different page size,
2982 * which changes the maximum amount of data which gets
2983 * written. Secondly, 4 megabytes is way too small. XFS
2984 * forces this value to be 16 megabytes by multiplying
2985 * nr_to_write parameter by four, and then relies on its
2986 * allocator to allocate larger extents to make them
2987 * contiguous. Unfortunately this brings us to the second
2988 * stupidity, which is that ext4's mballoc code only allocates
2989 * at most 2048 blocks. So we force contiguous writes up to
2990 * the number of dirty blocks in the inode, or
2991 * sbi->max_writeback_mb_bump whichever is smaller.
2993 max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2994 if (!range_cyclic && range_whole) {
2995 if (wbc->nr_to_write == LONG_MAX)
2996 desired_nr_to_write = wbc->nr_to_write;
2998 desired_nr_to_write = wbc->nr_to_write * 8;
3000 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
3002 if (desired_nr_to_write > max_pages)
3003 desired_nr_to_write = max_pages;
3005 if (wbc->nr_to_write < desired_nr_to_write) {
3006 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
3007 wbc->nr_to_write = desired_nr_to_write;
3011 mpd.inode = mapping->host;
3013 pages_skipped = wbc->pages_skipped;
3016 if (wbc->sync_mode == WB_SYNC_ALL)
3017 tag_pages_for_writeback(mapping, index, end);
3019 while (!ret && wbc->nr_to_write > 0) {
3022 * we insert one extent at a time. So we need
3023 * credit needed for single extent allocation.
3024 * journalled mode is currently not supported
3027 BUG_ON(ext4_should_journal_data(inode));
3028 needed_blocks = ext4_da_writepages_trans_blocks(inode);
3030 /* start a new transaction*/
3031 handle = ext4_journal_start(inode, needed_blocks);
3032 if (IS_ERR(handle)) {
3033 ret = PTR_ERR(handle);
3034 ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
3035 "%ld pages, ino %lu; err %d", __func__,
3036 wbc->nr_to_write, inode->i_ino, ret);
3037 goto out_writepages;
3041 * Now call __mpage_da_writepage to find the next
3042 * contiguous region of logical blocks that need
3043 * blocks to be allocated by ext4. We don't actually
3044 * submit the blocks for I/O here, even though
3045 * write_cache_pages thinks it will, and will set the
3046 * pages as clean for write before calling
3047 * __mpage_da_writepage().
3055 mpd.pages_written = 0;
3057 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
3059 * If we have a contiguous extent of pages and we
3060 * haven't done the I/O yet, map the blocks and submit
3063 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
3064 mpage_da_map_and_submit(&mpd);
3065 ret = MPAGE_DA_EXTENT_TAIL;
3067 trace_ext4_da_write_pages(inode, &mpd);
3068 wbc->nr_to_write -= mpd.pages_written;
3070 ext4_journal_stop(handle);
3072 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
3073 /* commit the transaction which would
3074 * free blocks released in the transaction
3077 jbd2_journal_force_commit_nested(sbi->s_journal);
3078 wbc->pages_skipped = pages_skipped;
3080 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
3082 * got one extent now try with
3085 pages_written += mpd.pages_written;
3086 wbc->pages_skipped = pages_skipped;