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 "ext4_jbd2.h"
43 #include "ext4_extents.h"
45 #define MPAGE_DA_EXTENT_TAIL 0x01
47 static inline int ext4_begin_ordered_truncate(struct inode *inode,
50 return jbd2_journal_begin_ordered_truncate(
51 EXT4_SB(inode->i_sb)->s_journal,
52 &EXT4_I(inode)->jinode,
56 static void ext4_invalidatepage(struct page *page, unsigned long offset);
59 * Test whether an inode is a fast symlink.
61 static int ext4_inode_is_fast_symlink(struct inode *inode)
63 int ea_blocks = EXT4_I(inode)->i_file_acl ?
64 (inode->i_sb->s_blocksize >> 9) : 0;
66 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
70 * The ext4 forget function must perform a revoke if we are freeing data
71 * which has been journaled. Metadata (eg. indirect blocks) must be
72 * revoked in all cases.
74 * "bh" may be NULL: a metadata block may have been freed from memory
75 * but there may still be a record of it in the journal, and that record
76 * still needs to be revoked.
78 * If the handle isn't valid we're not journaling so there's nothing to do.
80 int ext4_forget(handle_t *handle, int is_metadata, struct inode *inode,
81 struct buffer_head *bh, ext4_fsblk_t blocknr)
85 if (!ext4_handle_valid(handle))
90 BUFFER_TRACE(bh, "enter");
92 jbd_debug(4, "forgetting bh %p: is_metadata = %d, mode %o, "
94 bh, is_metadata, inode->i_mode,
95 test_opt(inode->i_sb, DATA_FLAGS));
97 /* Never use the revoke function if we are doing full data
98 * journaling: there is no need to, and a V1 superblock won't
99 * support it. Otherwise, only skip the revoke on un-journaled
102 if (test_opt(inode->i_sb, DATA_FLAGS) == EXT4_MOUNT_JOURNAL_DATA ||
103 (!is_metadata && !ext4_should_journal_data(inode))) {
105 BUFFER_TRACE(bh, "call jbd2_journal_forget");
106 return ext4_journal_forget(handle, bh);
112 * data!=journal && (is_metadata || should_journal_data(inode))
114 BUFFER_TRACE(bh, "call ext4_journal_revoke");
115 err = ext4_journal_revoke(handle, blocknr, bh);
117 ext4_abort(inode->i_sb, __func__,
118 "error %d when attempting revoke", err);
119 BUFFER_TRACE(bh, "exit");
124 * Work out how many blocks we need to proceed with the next chunk of a
125 * truncate transaction.
127 static unsigned long blocks_for_truncate(struct inode *inode)
131 needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
133 /* Give ourselves just enough room to cope with inodes in which
134 * i_blocks is corrupt: we've seen disk corruptions in the past
135 * which resulted in random data in an inode which looked enough
136 * like a regular file for ext4 to try to delete it. Things
137 * will go a bit crazy if that happens, but at least we should
138 * try not to panic the whole kernel. */
142 /* But we need to bound the transaction so we don't overflow the
144 if (needed > EXT4_MAX_TRANS_DATA)
145 needed = EXT4_MAX_TRANS_DATA;
147 return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
151 * Truncate transactions can be complex and absolutely huge. So we need to
152 * be able to restart the transaction at a conventient checkpoint to make
153 * sure we don't overflow the journal.
155 * start_transaction gets us a new handle for a truncate transaction,
156 * and extend_transaction tries to extend the existing one a bit. If
157 * extend fails, we need to propagate the failure up and restart the
158 * transaction in the top-level truncate loop. --sct
160 static handle_t *start_transaction(struct inode *inode)
164 result = ext4_journal_start(inode, blocks_for_truncate(inode));
168 ext4_std_error(inode->i_sb, PTR_ERR(result));
173 * Try to extend this transaction for the purposes of truncation.
175 * Returns 0 if we managed to create more room. If we can't create more
176 * room, and the transaction must be restarted we return 1.
178 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
180 if (!ext4_handle_valid(handle))
182 if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
184 if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
190 * Restart the transaction associated with *handle. This does a commit,
191 * so before we call here everything must be consistently dirtied against
194 static int ext4_journal_test_restart(handle_t *handle, struct inode *inode)
196 BUG_ON(EXT4_JOURNAL(inode) == NULL);
197 jbd_debug(2, "restarting handle %p\n", handle);
198 return ext4_journal_restart(handle, blocks_for_truncate(inode));
202 * Called at the last iput() if i_nlink is zero.
204 void ext4_delete_inode(struct inode *inode)
209 if (ext4_should_order_data(inode))
210 ext4_begin_ordered_truncate(inode, 0);
211 truncate_inode_pages(&inode->i_data, 0);
213 if (is_bad_inode(inode))
216 handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
217 if (IS_ERR(handle)) {
218 ext4_std_error(inode->i_sb, PTR_ERR(handle));
220 * If we're going to skip the normal cleanup, we still need to
221 * make sure that the in-core orphan linked list is properly
224 ext4_orphan_del(NULL, inode);
229 ext4_handle_sync(handle);
231 err = ext4_mark_inode_dirty(handle, inode);
233 ext4_warning(inode->i_sb, __func__,
234 "couldn't mark inode dirty (err %d)", err);
238 ext4_truncate(inode);
241 * ext4_ext_truncate() doesn't reserve any slop when it
242 * restarts journal transactions; therefore there may not be
243 * enough credits left in the handle to remove the inode from
244 * the orphan list and set the dtime field.
246 if (!ext4_handle_has_enough_credits(handle, 3)) {
247 err = ext4_journal_extend(handle, 3);
249 err = ext4_journal_restart(handle, 3);
251 ext4_warning(inode->i_sb, __func__,
252 "couldn't extend journal (err %d)", err);
254 ext4_journal_stop(handle);
260 * Kill off the orphan record which ext4_truncate created.
261 * AKPM: I think this can be inside the above `if'.
262 * Note that ext4_orphan_del() has to be able to cope with the
263 * deletion of a non-existent orphan - this is because we don't
264 * know if ext4_truncate() actually created an orphan record.
265 * (Well, we could do this if we need to, but heck - it works)
267 ext4_orphan_del(handle, inode);
268 EXT4_I(inode)->i_dtime = get_seconds();
271 * One subtle ordering requirement: if anything has gone wrong
272 * (transaction abort, IO errors, whatever), then we can still
273 * do these next steps (the fs will already have been marked as
274 * having errors), but we can't free the inode if the mark_dirty
277 if (ext4_mark_inode_dirty(handle, inode))
278 /* If that failed, just do the required in-core inode clear. */
281 ext4_free_inode(handle, inode);
282 ext4_journal_stop(handle);
285 clear_inode(inode); /* We must guarantee clearing of inode... */
291 struct buffer_head *bh;
294 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
296 p->key = *(p->p = v);
301 * ext4_block_to_path - parse the block number into array of offsets
302 * @inode: inode in question (we are only interested in its superblock)
303 * @i_block: block number to be parsed
304 * @offsets: array to store the offsets in
305 * @boundary: set this non-zero if the referred-to block is likely to be
306 * followed (on disk) by an indirect block.
308 * To store the locations of file's data ext4 uses a data structure common
309 * for UNIX filesystems - tree of pointers anchored in the inode, with
310 * data blocks at leaves and indirect blocks in intermediate nodes.
311 * This function translates the block number into path in that tree -
312 * return value is the path length and @offsets[n] is the offset of
313 * pointer to (n+1)th node in the nth one. If @block is out of range
314 * (negative or too large) warning is printed and zero returned.
316 * Note: function doesn't find node addresses, so no IO is needed. All
317 * we need to know is the capacity of indirect blocks (taken from the
322 * Portability note: the last comparison (check that we fit into triple
323 * indirect block) is spelled differently, because otherwise on an
324 * architecture with 32-bit longs and 8Kb pages we might get into trouble
325 * if our filesystem had 8Kb blocks. We might use long long, but that would
326 * kill us on x86. Oh, well, at least the sign propagation does not matter -
327 * i_block would have to be negative in the very beginning, so we would not
331 static int ext4_block_to_path(struct inode *inode,
333 ext4_lblk_t offsets[4], int *boundary)
335 int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
336 int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
337 const long direct_blocks = EXT4_NDIR_BLOCKS,
338 indirect_blocks = ptrs,
339 double_blocks = (1 << (ptrs_bits * 2));
344 ext4_warning(inode->i_sb, "ext4_block_to_path", "block < 0");
345 } else if (i_block < direct_blocks) {
346 offsets[n++] = i_block;
347 final = direct_blocks;
348 } else if ((i_block -= direct_blocks) < indirect_blocks) {
349 offsets[n++] = EXT4_IND_BLOCK;
350 offsets[n++] = i_block;
352 } else if ((i_block -= indirect_blocks) < double_blocks) {
353 offsets[n++] = EXT4_DIND_BLOCK;
354 offsets[n++] = i_block >> ptrs_bits;
355 offsets[n++] = i_block & (ptrs - 1);
357 } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
358 offsets[n++] = EXT4_TIND_BLOCK;
359 offsets[n++] = i_block >> (ptrs_bits * 2);
360 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
361 offsets[n++] = i_block & (ptrs - 1);
364 ext4_warning(inode->i_sb, "ext4_block_to_path",
365 "block %lu > max in inode %lu",
366 i_block + direct_blocks +
367 indirect_blocks + double_blocks, inode->i_ino);
370 *boundary = final - 1 - (i_block & (ptrs - 1));
374 static int __ext4_check_blockref(const char *function, struct inode *inode,
375 __le32 *p, unsigned int max) {
377 unsigned int maxblocks = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es);
379 while (bref < p+max) {
380 if (unlikely(le32_to_cpu(*bref) >= maxblocks)) {
381 ext4_error(inode->i_sb, function,
382 "block reference %u >= max (%u) "
383 "in inode #%lu, offset=%d",
384 le32_to_cpu(*bref), maxblocks,
385 inode->i_ino, (int)(bref-p));
394 #define ext4_check_indirect_blockref(inode, bh) \
395 __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data, \
396 EXT4_ADDR_PER_BLOCK((inode)->i_sb))
398 #define ext4_check_inode_blockref(inode) \
399 __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data, \
403 * ext4_get_branch - read the chain of indirect blocks leading to data
404 * @inode: inode in question
405 * @depth: depth of the chain (1 - direct pointer, etc.)
406 * @offsets: offsets of pointers in inode/indirect blocks
407 * @chain: place to store the result
408 * @err: here we store the error value
410 * Function fills the array of triples <key, p, bh> and returns %NULL
411 * if everything went OK or the pointer to the last filled triple
412 * (incomplete one) otherwise. Upon the return chain[i].key contains
413 * the number of (i+1)-th block in the chain (as it is stored in memory,
414 * i.e. little-endian 32-bit), chain[i].p contains the address of that
415 * number (it points into struct inode for i==0 and into the bh->b_data
416 * for i>0) and chain[i].bh points to the buffer_head of i-th indirect
417 * block for i>0 and NULL for i==0. In other words, it holds the block
418 * numbers of the chain, addresses they were taken from (and where we can
419 * verify that chain did not change) and buffer_heads hosting these
422 * Function stops when it stumbles upon zero pointer (absent block)
423 * (pointer to last triple returned, *@err == 0)
424 * or when it gets an IO error reading an indirect block
425 * (ditto, *@err == -EIO)
426 * or when it reads all @depth-1 indirect blocks successfully and finds
427 * the whole chain, all way to the data (returns %NULL, *err == 0).
429 * Need to be called with
430 * down_read(&EXT4_I(inode)->i_data_sem)
432 static Indirect *ext4_get_branch(struct inode *inode, int depth,
433 ext4_lblk_t *offsets,
434 Indirect chain[4], int *err)
436 struct super_block *sb = inode->i_sb;
438 struct buffer_head *bh;
441 /* i_data is not going away, no lock needed */
442 add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
446 bh = sb_getblk(sb, le32_to_cpu(p->key));
450 if (!bh_uptodate_or_lock(bh)) {
451 if (bh_submit_read(bh) < 0) {
455 /* validate block references */
456 if (ext4_check_indirect_blockref(inode, bh)) {
462 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
476 * ext4_find_near - find a place for allocation with sufficient locality
478 * @ind: descriptor of indirect block.
480 * This function returns the preferred place for block allocation.
481 * It is used when heuristic for sequential allocation fails.
483 * + if there is a block to the left of our position - allocate near it.
484 * + if pointer will live in indirect block - allocate near that block.
485 * + if pointer will live in inode - allocate in the same
488 * In the latter case we colour the starting block by the callers PID to
489 * prevent it from clashing with concurrent allocations for a different inode
490 * in the same block group. The PID is used here so that functionally related
491 * files will be close-by on-disk.
493 * Caller must make sure that @ind is valid and will stay that way.
495 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
497 struct ext4_inode_info *ei = EXT4_I(inode);
498 __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
500 ext4_fsblk_t bg_start;
501 ext4_fsblk_t last_block;
502 ext4_grpblk_t colour;
503 ext4_group_t block_group;
504 int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
506 /* Try to find previous block */
507 for (p = ind->p - 1; p >= start; p--) {
509 return le32_to_cpu(*p);
512 /* No such thing, so let's try location of indirect block */
514 return ind->bh->b_blocknr;
517 * It is going to be referred to from the inode itself? OK, just put it
518 * into the same cylinder group then.
520 block_group = ei->i_block_group;
521 if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
522 block_group &= ~(flex_size-1);
523 if (S_ISREG(inode->i_mode))
526 bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
527 last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
530 * If we are doing delayed allocation, we don't need take
531 * colour into account.
533 if (test_opt(inode->i_sb, DELALLOC))
536 if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
537 colour = (current->pid % 16) *
538 (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
540 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
541 return bg_start + colour;
545 * ext4_find_goal - find a preferred place for allocation.
547 * @block: block we want
548 * @partial: pointer to the last triple within a chain
550 * Normally this function find the preferred place for block allocation,
553 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
557 * XXX need to get goal block from mballoc's data structures
560 return ext4_find_near(inode, partial);
564 * ext4_blks_to_allocate: Look up the block map and count the number
565 * of direct blocks need to be allocated for the given branch.
567 * @branch: chain of indirect blocks
568 * @k: number of blocks need for indirect blocks
569 * @blks: number of data blocks to be mapped.
570 * @blocks_to_boundary: the offset in the indirect block
572 * return the total number of blocks to be allocate, including the
573 * direct and indirect blocks.
575 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
576 int blocks_to_boundary)
578 unsigned int count = 0;
581 * Simple case, [t,d]Indirect block(s) has not allocated yet
582 * then it's clear blocks on that path have not allocated
585 /* right now we don't handle cross boundary allocation */
586 if (blks < blocks_to_boundary + 1)
589 count += blocks_to_boundary + 1;
594 while (count < blks && count <= blocks_to_boundary &&
595 le32_to_cpu(*(branch[0].p + count)) == 0) {
602 * ext4_alloc_blocks: multiple allocate blocks needed for a branch
603 * @indirect_blks: the number of blocks need to allocate for indirect
606 * @new_blocks: on return it will store the new block numbers for
607 * the indirect blocks(if needed) and the first direct block,
608 * @blks: on return it will store the total number of allocated
611 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
612 ext4_lblk_t iblock, ext4_fsblk_t goal,
613 int indirect_blks, int blks,
614 ext4_fsblk_t new_blocks[4], int *err)
616 struct ext4_allocation_request ar;
618 unsigned long count = 0, blk_allocated = 0;
620 ext4_fsblk_t current_block = 0;
624 * Here we try to allocate the requested multiple blocks at once,
625 * on a best-effort basis.
626 * To build a branch, we should allocate blocks for
627 * the indirect blocks(if not allocated yet), and at least
628 * the first direct block of this branch. That's the
629 * minimum number of blocks need to allocate(required)
631 /* first we try to allocate the indirect blocks */
632 target = indirect_blks;
635 /* allocating blocks for indirect blocks and direct blocks */
636 current_block = ext4_new_meta_blocks(handle, inode,
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);
676 if (*err && (target == blks)) {
678 * if the allocation failed and we didn't allocate
684 if (target == blks) {
686 * save the new block number
687 * for the first direct block
689 new_blocks[index] = current_block;
691 blk_allocated += ar.len;
694 /* total number of blocks allocated for direct blocks */
699 for (i = 0; i < index; i++)
700 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
705 * ext4_alloc_branch - allocate and set up a chain of blocks.
707 * @indirect_blks: number of allocated indirect blocks
708 * @blks: number of allocated direct blocks
709 * @offsets: offsets (in the blocks) to store the pointers to next.
710 * @branch: place to store the chain in.
712 * This function allocates blocks, zeroes out all but the last one,
713 * links them into chain and (if we are synchronous) writes them to disk.
714 * In other words, it prepares a branch that can be spliced onto the
715 * inode. It stores the information about that chain in the branch[], in
716 * the same format as ext4_get_branch() would do. We are calling it after
717 * we had read the existing part of chain and partial points to the last
718 * triple of that (one with zero ->key). Upon the exit we have the same
719 * picture as after the successful ext4_get_block(), except that in one
720 * place chain is disconnected - *branch->p is still zero (we did not
721 * set the last link), but branch->key contains the number that should
722 * be placed into *branch->p to fill that gap.
724 * If allocation fails we free all blocks we've allocated (and forget
725 * their buffer_heads) and return the error value the from failed
726 * ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
727 * as described above and return 0.
729 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
730 ext4_lblk_t iblock, int indirect_blks,
731 int *blks, ext4_fsblk_t goal,
732 ext4_lblk_t *offsets, Indirect *branch)
734 int blocksize = inode->i_sb->s_blocksize;
737 struct buffer_head *bh;
739 ext4_fsblk_t new_blocks[4];
740 ext4_fsblk_t current_block;
742 num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
743 *blks, new_blocks, &err);
747 branch[0].key = cpu_to_le32(new_blocks[0]);
749 * metadata blocks and data blocks are allocated.
751 for (n = 1; n <= indirect_blks; n++) {
753 * Get buffer_head for parent block, zero it out
754 * and set the pointer to new one, then send
757 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
760 BUFFER_TRACE(bh, "call get_create_access");
761 err = ext4_journal_get_create_access(handle, bh);
768 memset(bh->b_data, 0, blocksize);
769 branch[n].p = (__le32 *) bh->b_data + offsets[n];
770 branch[n].key = cpu_to_le32(new_blocks[n]);
771 *branch[n].p = branch[n].key;
772 if (n == indirect_blks) {
773 current_block = new_blocks[n];
775 * End of chain, update the last new metablock of
776 * the chain to point to the new allocated
777 * data blocks numbers
779 for (i=1; i < num; i++)
780 *(branch[n].p + i) = cpu_to_le32(++current_block);
782 BUFFER_TRACE(bh, "marking uptodate");
783 set_buffer_uptodate(bh);
786 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
787 err = ext4_handle_dirty_metadata(handle, inode, bh);
794 /* Allocation failed, free what we already allocated */
795 for (i = 1; i <= n ; i++) {
796 BUFFER_TRACE(branch[i].bh, "call jbd2_journal_forget");
797 ext4_journal_forget(handle, branch[i].bh);
799 for (i = 0; i < indirect_blks; i++)
800 ext4_free_blocks(handle, inode, new_blocks[i], 1, 0);
802 ext4_free_blocks(handle, inode, new_blocks[i], num, 0);
808 * ext4_splice_branch - splice the allocated branch onto inode.
810 * @block: (logical) number of block we are adding
811 * @chain: chain of indirect blocks (with a missing link - see
813 * @where: location of missing link
814 * @num: number of indirect blocks we are adding
815 * @blks: number of direct blocks we are adding
817 * This function fills the missing link and does all housekeeping needed in
818 * inode (->i_blocks, etc.). In case of success we end up with the full
819 * chain to new block and return 0.
821 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
822 ext4_lblk_t block, Indirect *where, int num, int blks)
826 ext4_fsblk_t current_block;
829 * If we're splicing into a [td]indirect block (as opposed to the
830 * inode) then we need to get write access to the [td]indirect block
834 BUFFER_TRACE(where->bh, "get_write_access");
835 err = ext4_journal_get_write_access(handle, where->bh);
841 *where->p = where->key;
844 * Update the host buffer_head or inode to point to more just allocated
845 * direct blocks blocks
847 if (num == 0 && blks > 1) {
848 current_block = le32_to_cpu(where->key) + 1;
849 for (i = 1; i < blks; i++)
850 *(where->p + i) = cpu_to_le32(current_block++);
853 /* We are done with atomic stuff, now do the rest of housekeeping */
855 inode->i_ctime = ext4_current_time(inode);
856 ext4_mark_inode_dirty(handle, inode);
858 /* had we spliced it onto indirect block? */
861 * If we spliced it onto an indirect block, we haven't
862 * altered the inode. Note however that if it is being spliced
863 * onto an indirect block at the very end of the file (the
864 * file is growing) then we *will* alter the inode to reflect
865 * the new i_size. But that is not done here - it is done in
866 * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
868 jbd_debug(5, "splicing indirect only\n");
869 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
870 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
875 * OK, we spliced it into the inode itself on a direct block.
876 * Inode was dirtied above.
878 jbd_debug(5, "splicing direct\n");
883 for (i = 1; i <= num; i++) {
884 BUFFER_TRACE(where[i].bh, "call jbd2_journal_forget");
885 ext4_journal_forget(handle, where[i].bh);
886 ext4_free_blocks(handle, inode,
887 le32_to_cpu(where[i-1].key), 1, 0);
889 ext4_free_blocks(handle, inode, le32_to_cpu(where[num].key), blks, 0);
895 * Allocation strategy is simple: if we have to allocate something, we will
896 * have to go the whole way to leaf. So let's do it before attaching anything
897 * to tree, set linkage between the newborn blocks, write them if sync is
898 * required, recheck the path, free and repeat if check fails, otherwise
899 * set the last missing link (that will protect us from any truncate-generated
900 * removals - all blocks on the path are immune now) and possibly force the
901 * write on the parent block.
902 * That has a nice additional property: no special recovery from the failed
903 * allocations is needed - we simply release blocks and do not touch anything
904 * reachable from inode.
906 * `handle' can be NULL if create == 0.
908 * return > 0, # of blocks mapped or allocated.
909 * return = 0, if plain lookup failed.
910 * return < 0, error case.
913 * Need to be called with
914 * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
915 * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
917 static int ext4_get_blocks_handle(handle_t *handle, struct inode *inode,
918 ext4_lblk_t iblock, unsigned int maxblocks,
919 struct buffer_head *bh_result,
920 int create, int extend_disksize)
923 ext4_lblk_t offsets[4];
928 int blocks_to_boundary = 0;
930 struct ext4_inode_info *ei = EXT4_I(inode);
932 ext4_fsblk_t first_block = 0;
936 J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
937 J_ASSERT(handle != NULL || create == 0);
938 depth = ext4_block_to_path(inode, iblock, offsets,
939 &blocks_to_boundary);
944 partial = ext4_get_branch(inode, depth, offsets, chain, &err);
946 /* Simplest case - block found, no allocation needed */
948 first_block = le32_to_cpu(chain[depth - 1].key);
949 clear_buffer_new(bh_result);
952 while (count < maxblocks && count <= blocks_to_boundary) {
955 blk = le32_to_cpu(*(chain[depth-1].p + count));
957 if (blk == first_block + count)
965 /* Next simple case - plain lookup or failed read of indirect block */
966 if (!create || err == -EIO)
970 * Okay, we need to do block allocation.
972 goal = ext4_find_goal(inode, iblock, partial);
974 /* the number of blocks need to allocate for [d,t]indirect blocks */
975 indirect_blks = (chain + depth) - partial - 1;
978 * Next look up the indirect map to count the totoal number of
979 * direct blocks to allocate for this branch.
981 count = ext4_blks_to_allocate(partial, indirect_blks,
982 maxblocks, blocks_to_boundary);
984 * Block out ext4_truncate while we alter the tree
986 err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
988 offsets + (partial - chain), partial);
991 * The ext4_splice_branch call will free and forget any buffers
992 * on the new chain if there is a failure, but that risks using
993 * up transaction credits, especially for bitmaps where the
994 * credits cannot be returned. Can we handle this somehow? We
995 * may need to return -EAGAIN upwards in the worst case. --sct
998 err = ext4_splice_branch(handle, inode, iblock,
999 partial, indirect_blks, count);
1001 * i_disksize growing is protected by i_data_sem. Don't forget to
1002 * protect it if you're about to implement concurrent
1003 * ext4_get_block() -bzzz
1005 if (!err && extend_disksize) {
1006 disksize = ((loff_t) iblock + count) << inode->i_blkbits;
1007 if (disksize > i_size_read(inode))
1008 disksize = i_size_read(inode);
1009 if (disksize > ei->i_disksize)
1010 ei->i_disksize = disksize;
1015 set_buffer_new(bh_result);
1017 map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
1018 if (count > blocks_to_boundary)
1019 set_buffer_boundary(bh_result);
1021 /* Clean up and exit */
1022 partial = chain + depth - 1; /* the whole chain */
1024 while (partial > chain) {
1025 BUFFER_TRACE(partial->bh, "call brelse");
1026 brelse(partial->bh);
1029 BUFFER_TRACE(bh_result, "returned");
1034 qsize_t ext4_get_reserved_space(struct inode *inode)
1036 unsigned long long total;
1038 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1039 total = EXT4_I(inode)->i_reserved_data_blocks +
1040 EXT4_I(inode)->i_reserved_meta_blocks;
1041 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1046 * Calculate the number of metadata blocks need to reserve
1047 * to allocate @blocks for non extent file based file
1049 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
1051 int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1052 int ind_blks, dind_blks, tind_blks;
1054 /* number of new indirect blocks needed */
1055 ind_blks = (blocks + icap - 1) / icap;
1057 dind_blks = (ind_blks + icap - 1) / icap;
1061 return ind_blks + dind_blks + tind_blks;
1065 * Calculate the number of metadata blocks need to reserve
1066 * to allocate given number of blocks
1068 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1073 if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1074 return ext4_ext_calc_metadata_amount(inode, blocks);
1076 return ext4_indirect_calc_metadata_amount(inode, blocks);
1079 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1081 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1082 int total, mdb, mdb_free;
1084 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1085 /* recalculate the number of metablocks still need to be reserved */
1086 total = EXT4_I(inode)->i_reserved_data_blocks - used;
1087 mdb = ext4_calc_metadata_amount(inode, total);
1089 /* figure out how many metablocks to release */
1090 BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1091 mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1094 /* Account for allocated meta_blocks */
1095 mdb_free -= EXT4_I(inode)->i_allocated_meta_blocks;
1097 /* update fs dirty blocks counter */
1098 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1099 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1100 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1103 /* update per-inode reservations */
1104 BUG_ON(used > EXT4_I(inode)->i_reserved_data_blocks);
1105 EXT4_I(inode)->i_reserved_data_blocks -= used;
1106 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1109 * free those over-booking quota for metadata blocks
1112 vfs_dq_release_reservation_block(inode, mdb_free);
1115 * If we have done all the pending block allocations and if
1116 * there aren't any writers on the inode, we can discard the
1117 * inode's preallocations.
1119 if (!total && (atomic_read(&inode->i_writecount) == 0))
1120 ext4_discard_preallocations(inode);
1124 * The ext4_get_blocks_wrap() function try to look up the requested blocks,
1125 * and returns if the blocks are already mapped.
1127 * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1128 * and store the allocated blocks in the result buffer head and mark it
1131 * If file type is extents based, it will call ext4_ext_get_blocks(),
1132 * Otherwise, call with ext4_get_blocks_handle() to handle indirect mapping
1135 * On success, it returns the number of blocks being mapped or allocate.
1136 * if create==0 and the blocks are pre-allocated and uninitialized block,
1137 * the result buffer head is unmapped. If the create ==1, it will make sure
1138 * the buffer head is mapped.
1140 * It returns 0 if plain look up failed (blocks have not been allocated), in
1141 * that casem, buffer head is unmapped
1143 * It returns the error in case of allocation failure.
1145 int ext4_get_blocks_wrap(handle_t *handle, struct inode *inode, sector_t block,
1146 unsigned int max_blocks, struct buffer_head *bh,
1147 int create, int extend_disksize, int flag)
1151 clear_buffer_mapped(bh);
1152 clear_buffer_unwritten(bh);
1155 * Try to see if we can get the block without requesting
1156 * for new file system block.
1158 down_read((&EXT4_I(inode)->i_data_sem));
1159 if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1160 retval = ext4_ext_get_blocks(handle, inode, block, max_blocks,
1163 retval = ext4_get_blocks_handle(handle,
1164 inode, block, max_blocks, bh, 0, 0);
1166 up_read((&EXT4_I(inode)->i_data_sem));
1168 /* If it is only a block(s) look up */
1173 * Returns if the blocks have already allocated
1175 * Note that if blocks have been preallocated
1176 * ext4_ext_get_block() returns th create = 0
1177 * with buffer head unmapped.
1179 if (retval > 0 && buffer_mapped(bh))
1183 * When we call get_blocks without the create flag, the
1184 * BH_Unwritten flag could have gotten set if the blocks
1185 * requested were part of a uninitialized extent. We need to
1186 * clear this flag now that we are committed to convert all or
1187 * part of the uninitialized extent to be an initialized
1188 * extent. This is because we need to avoid the combination
1189 * of BH_Unwritten and BH_Mapped flags being simultaneously
1190 * set on the buffer_head.
1192 clear_buffer_unwritten(bh);
1195 * New blocks allocate and/or writing to uninitialized extent
1196 * will possibly result in updating i_data, so we take
1197 * the write lock of i_data_sem, and call get_blocks()
1198 * with create == 1 flag.
1200 down_write((&EXT4_I(inode)->i_data_sem));
1203 * if the caller is from delayed allocation writeout path
1204 * we have already reserved fs blocks for allocation
1205 * let the underlying get_block() function know to
1206 * avoid double accounting
1209 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1211 * We need to check for EXT4 here because migrate
1212 * could have changed the inode type in between
1214 if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1215 retval = ext4_ext_get_blocks(handle, inode, block, max_blocks,
1216 bh, create, extend_disksize);
1218 retval = ext4_get_blocks_handle(handle, inode, block,
1219 max_blocks, bh, create, extend_disksize);
1221 if (retval > 0 && buffer_new(bh)) {
1223 * We allocated new blocks which will result in
1224 * i_data's format changing. Force the migrate
1225 * to fail by clearing migrate flags
1227 EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
1233 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1235 * Update reserved blocks/metadata blocks
1236 * after successful block allocation
1237 * which were deferred till now
1239 if ((retval > 0) && buffer_delay(bh))
1240 ext4_da_update_reserve_space(inode, retval);
1243 up_write((&EXT4_I(inode)->i_data_sem));
1247 /* Maximum number of blocks we map for direct IO at once. */
1248 #define DIO_MAX_BLOCKS 4096
1250 int ext4_get_block(struct inode *inode, sector_t iblock,
1251 struct buffer_head *bh_result, int create)
1253 handle_t *handle = ext4_journal_current_handle();
1254 int ret = 0, started = 0;
1255 unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1258 if (create && !handle) {
1259 /* Direct IO write... */
1260 if (max_blocks > DIO_MAX_BLOCKS)
1261 max_blocks = DIO_MAX_BLOCKS;
1262 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1263 handle = ext4_journal_start(inode, dio_credits);
1264 if (IS_ERR(handle)) {
1265 ret = PTR_ERR(handle);
1271 ret = ext4_get_blocks_wrap(handle, inode, iblock,
1272 max_blocks, bh_result, create, 0, 0);
1274 bh_result->b_size = (ret << inode->i_blkbits);
1278 ext4_journal_stop(handle);
1284 * `handle' can be NULL if create is zero
1286 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1287 ext4_lblk_t block, int create, int *errp)
1289 struct buffer_head dummy;
1292 J_ASSERT(handle != NULL || create == 0);
1295 dummy.b_blocknr = -1000;
1296 buffer_trace_init(&dummy.b_history);
1297 err = ext4_get_blocks_wrap(handle, inode, block, 1,
1298 &dummy, create, 1, 0);
1300 * ext4_get_blocks_handle() returns number of blocks
1301 * mapped. 0 in case of a HOLE.
1309 if (!err && buffer_mapped(&dummy)) {
1310 struct buffer_head *bh;
1311 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1316 if (buffer_new(&dummy)) {
1317 J_ASSERT(create != 0);
1318 J_ASSERT(handle != NULL);
1321 * Now that we do not always journal data, we should
1322 * keep in mind whether this should always journal the
1323 * new buffer as metadata. For now, regular file
1324 * writes use ext4_get_block instead, so it's not a
1328 BUFFER_TRACE(bh, "call get_create_access");
1329 fatal = ext4_journal_get_create_access(handle, bh);
1330 if (!fatal && !buffer_uptodate(bh)) {
1331 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1332 set_buffer_uptodate(bh);
1335 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1336 err = ext4_handle_dirty_metadata(handle, inode, bh);
1340 BUFFER_TRACE(bh, "not a new buffer");
1353 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1354 ext4_lblk_t block, int create, int *err)
1356 struct buffer_head *bh;
1358 bh = ext4_getblk(handle, inode, block, create, err);
1361 if (buffer_uptodate(bh))
1363 ll_rw_block(READ_META, 1, &bh);
1365 if (buffer_uptodate(bh))
1372 static int walk_page_buffers(handle_t *handle,
1373 struct buffer_head *head,
1377 int (*fn)(handle_t *handle,
1378 struct buffer_head *bh))
1380 struct buffer_head *bh;
1381 unsigned block_start, block_end;
1382 unsigned blocksize = head->b_size;
1384 struct buffer_head *next;
1386 for (bh = head, block_start = 0;
1387 ret == 0 && (bh != head || !block_start);
1388 block_start = block_end, bh = next)
1390 next = bh->b_this_page;
1391 block_end = block_start + blocksize;
1392 if (block_end <= from || block_start >= to) {
1393 if (partial && !buffer_uptodate(bh))
1397 err = (*fn)(handle, bh);
1405 * To preserve ordering, it is essential that the hole instantiation and
1406 * the data write be encapsulated in a single transaction. We cannot
1407 * close off a transaction and start a new one between the ext4_get_block()
1408 * and the commit_write(). So doing the jbd2_journal_start at the start of
1409 * prepare_write() is the right place.
1411 * Also, this function can nest inside ext4_writepage() ->
1412 * block_write_full_page(). In that case, we *know* that ext4_writepage()
1413 * has generated enough buffer credits to do the whole page. So we won't
1414 * block on the journal in that case, which is good, because the caller may
1417 * By accident, ext4 can be reentered when a transaction is open via
1418 * quota file writes. If we were to commit the transaction while thus
1419 * reentered, there can be a deadlock - we would be holding a quota
1420 * lock, and the commit would never complete if another thread had a
1421 * transaction open and was blocking on the quota lock - a ranking
1424 * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1425 * will _not_ run commit under these circumstances because handle->h_ref
1426 * is elevated. We'll still have enough credits for the tiny quotafile
1429 static int do_journal_get_write_access(handle_t *handle,
1430 struct buffer_head *bh)
1432 if (!buffer_mapped(bh) || buffer_freed(bh))
1434 return ext4_journal_get_write_access(handle, bh);
1437 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1438 loff_t pos, unsigned len, unsigned flags,
1439 struct page **pagep, void **fsdata)
1441 struct inode *inode = mapping->host;
1442 int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
1449 trace_mark(ext4_write_begin,
1450 "dev %s ino %lu pos %llu len %u flags %u",
1451 inode->i_sb->s_id, inode->i_ino,
1452 (unsigned long long) pos, len, flags);
1453 index = pos >> PAGE_CACHE_SHIFT;
1454 from = pos & (PAGE_CACHE_SIZE - 1);
1458 handle = ext4_journal_start(inode, needed_blocks);
1459 if (IS_ERR(handle)) {
1460 ret = PTR_ERR(handle);
1464 /* We cannot recurse into the filesystem as the transaction is already
1466 flags |= AOP_FLAG_NOFS;
1468 page = grab_cache_page_write_begin(mapping, index, flags);
1470 ext4_journal_stop(handle);
1476 ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1479 if (!ret && ext4_should_journal_data(inode)) {
1480 ret = walk_page_buffers(handle, page_buffers(page),
1481 from, to, NULL, do_journal_get_write_access);
1486 ext4_journal_stop(handle);
1487 page_cache_release(page);
1489 * block_write_begin may have instantiated a few blocks
1490 * outside i_size. Trim these off again. Don't need
1491 * i_size_read because we hold i_mutex.
1493 if (pos + len > inode->i_size)
1494 vmtruncate(inode, inode->i_size);
1497 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1503 /* For write_end() in data=journal mode */
1504 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1506 if (!buffer_mapped(bh) || buffer_freed(bh))
1508 set_buffer_uptodate(bh);
1509 return ext4_handle_dirty_metadata(handle, NULL, bh);
1513 * We need to pick up the new inode size which generic_commit_write gave us
1514 * `file' can be NULL - eg, when called from page_symlink().
1516 * ext4 never places buffers on inode->i_mapping->private_list. metadata
1517 * buffers are managed internally.
1519 static int ext4_ordered_write_end(struct file *file,
1520 struct address_space *mapping,
1521 loff_t pos, unsigned len, unsigned copied,
1522 struct page *page, void *fsdata)
1524 handle_t *handle = ext4_journal_current_handle();
1525 struct inode *inode = mapping->host;
1528 trace_mark(ext4_ordered_write_end,
1529 "dev %s ino %lu pos %llu len %u copied %u",
1530 inode->i_sb->s_id, inode->i_ino,
1531 (unsigned long long) pos, len, copied);
1532 ret = ext4_jbd2_file_inode(handle, inode);
1537 new_i_size = pos + copied;
1538 if (new_i_size > EXT4_I(inode)->i_disksize) {
1539 ext4_update_i_disksize(inode, new_i_size);
1540 /* We need to mark inode dirty even if
1541 * new_i_size is less that inode->i_size
1542 * bu greater than i_disksize.(hint delalloc)
1544 ext4_mark_inode_dirty(handle, inode);
1547 ret2 = generic_write_end(file, mapping, pos, len, copied,
1553 ret2 = ext4_journal_stop(handle);
1557 return ret ? ret : copied;
1560 static int ext4_writeback_write_end(struct file *file,
1561 struct address_space *mapping,
1562 loff_t pos, unsigned len, unsigned copied,
1563 struct page *page, void *fsdata)
1565 handle_t *handle = ext4_journal_current_handle();
1566 struct inode *inode = mapping->host;
1570 trace_mark(ext4_writeback_write_end,
1571 "dev %s ino %lu pos %llu len %u copied %u",
1572 inode->i_sb->s_id, inode->i_ino,
1573 (unsigned long long) pos, len, copied);
1574 new_i_size = pos + copied;
1575 if (new_i_size > EXT4_I(inode)->i_disksize) {
1576 ext4_update_i_disksize(inode, new_i_size);
1577 /* We need to mark inode dirty even if
1578 * new_i_size is less that inode->i_size
1579 * bu greater than i_disksize.(hint delalloc)
1581 ext4_mark_inode_dirty(handle, inode);
1584 ret2 = generic_write_end(file, mapping, pos, len, copied,
1590 ret2 = ext4_journal_stop(handle);
1594 return ret ? ret : copied;
1597 static int ext4_journalled_write_end(struct file *file,
1598 struct address_space *mapping,
1599 loff_t pos, unsigned len, unsigned copied,
1600 struct page *page, void *fsdata)
1602 handle_t *handle = ext4_journal_current_handle();
1603 struct inode *inode = mapping->host;
1609 trace_mark(ext4_journalled_write_end,
1610 "dev %s ino %lu pos %llu len %u copied %u",
1611 inode->i_sb->s_id, inode->i_ino,
1612 (unsigned long long) pos, len, copied);
1613 from = pos & (PAGE_CACHE_SIZE - 1);
1617 if (!PageUptodate(page))
1619 page_zero_new_buffers(page, from+copied, to);
1622 ret = walk_page_buffers(handle, page_buffers(page), from,
1623 to, &partial, write_end_fn);
1625 SetPageUptodate(page);
1626 new_i_size = pos + copied;
1627 if (new_i_size > inode->i_size)
1628 i_size_write(inode, pos+copied);
1629 EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1630 if (new_i_size > EXT4_I(inode)->i_disksize) {
1631 ext4_update_i_disksize(inode, new_i_size);
1632 ret2 = ext4_mark_inode_dirty(handle, inode);
1638 ret2 = ext4_journal_stop(handle);
1641 page_cache_release(page);
1643 return ret ? ret : copied;
1646 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1649 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1650 unsigned long md_needed, mdblocks, total = 0;
1653 * recalculate the amount of metadata blocks to reserve
1654 * in order to allocate nrblocks
1655 * worse case is one extent per block
1658 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1659 total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1660 mdblocks = ext4_calc_metadata_amount(inode, total);
1661 BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1663 md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1664 total = md_needed + nrblocks;
1667 * Make quota reservation here to prevent quota overflow
1668 * later. Real quota accounting is done at pages writeout
1671 if (vfs_dq_reserve_block(inode, total)) {
1672 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1676 if (ext4_claim_free_blocks(sbi, total)) {
1677 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1678 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1682 vfs_dq_release_reservation_block(inode, total);
1685 EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1686 EXT4_I(inode)->i_reserved_meta_blocks = mdblocks;
1688 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1689 return 0; /* success */
1692 static void ext4_da_release_space(struct inode *inode, int to_free)
1694 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1695 int total, mdb, mdb_free, release;
1698 return; /* Nothing to release, exit */
1700 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1702 if (!EXT4_I(inode)->i_reserved_data_blocks) {
1704 * if there is no reserved blocks, but we try to free some
1705 * then the counter is messed up somewhere.
1706 * but since this function is called from invalidate
1707 * page, it's harmless to return without any action
1709 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1710 "blocks for inode %lu, but there is no reserved "
1711 "data blocks\n", to_free, inode->i_ino);
1712 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1716 /* recalculate the number of metablocks still need to be reserved */
1717 total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1718 mdb = ext4_calc_metadata_amount(inode, total);
1720 /* figure out how many metablocks to release */
1721 BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1722 mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1724 release = to_free + mdb_free;
1726 /* update fs dirty blocks counter for truncate case */
1727 percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1729 /* update per-inode reservations */
1730 BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1731 EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1733 BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1734 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1735 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1737 vfs_dq_release_reservation_block(inode, release);
1740 static void ext4_da_page_release_reservation(struct page *page,
1741 unsigned long offset)
1744 struct buffer_head *head, *bh;
1745 unsigned int curr_off = 0;
1747 head = page_buffers(page);
1750 unsigned int next_off = curr_off + bh->b_size;
1752 if ((offset <= curr_off) && (buffer_delay(bh))) {
1754 clear_buffer_delay(bh);
1756 curr_off = next_off;
1757 } while ((bh = bh->b_this_page) != head);
1758 ext4_da_release_space(page->mapping->host, to_release);
1762 * Delayed allocation stuff
1765 struct mpage_da_data {
1766 struct inode *inode;
1767 sector_t b_blocknr; /* start block number of extent */
1768 size_t b_size; /* size of extent */
1769 unsigned long b_state; /* state of the extent */
1770 unsigned long first_page, next_page; /* extent of pages */
1771 struct writeback_control *wbc;
1778 * mpage_da_submit_io - walks through extent of pages and try to write
1779 * them with writepage() call back
1781 * @mpd->inode: inode
1782 * @mpd->first_page: first page of the extent
1783 * @mpd->next_page: page after the last page of the extent
1785 * By the time mpage_da_submit_io() is called we expect all blocks
1786 * to be allocated. this may be wrong if allocation failed.
1788 * As pages are already locked by write_cache_pages(), we can't use it
1790 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1793 struct pagevec pvec;
1794 unsigned long index, end;
1795 int ret = 0, err, nr_pages, i;
1796 struct inode *inode = mpd->inode;
1797 struct address_space *mapping = inode->i_mapping;
1799 BUG_ON(mpd->next_page <= mpd->first_page);
1801 * We need to start from the first_page to the next_page - 1
1802 * to make sure we also write the mapped dirty buffer_heads.
1803 * If we look at mpd->b_blocknr we would only be looking
1804 * at the currently mapped buffer_heads.
1806 index = mpd->first_page;
1807 end = mpd->next_page - 1;
1809 pagevec_init(&pvec, 0);
1810 while (index <= end) {
1811 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1814 for (i = 0; i < nr_pages; i++) {
1815 struct page *page = pvec.pages[i];
1817 index = page->index;
1822 BUG_ON(!PageLocked(page));
1823 BUG_ON(PageWriteback(page));
1825 pages_skipped = mpd->wbc->pages_skipped;
1826 err = mapping->a_ops->writepage(page, mpd->wbc);
1827 if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1829 * have successfully written the page
1830 * without skipping the same
1832 mpd->pages_written++;
1834 * In error case, we have to continue because
1835 * remaining pages are still locked
1836 * XXX: unlock and re-dirty them?
1841 pagevec_release(&pvec);
1847 * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1849 * @mpd->inode - inode to walk through
1850 * @exbh->b_blocknr - first block on a disk
1851 * @exbh->b_size - amount of space in bytes
1852 * @logical - first logical block to start assignment with
1854 * the function goes through all passed space and put actual disk
1855 * block numbers into buffer heads, dropping BH_Delay
1857 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1858 struct buffer_head *exbh)
1860 struct inode *inode = mpd->inode;
1861 struct address_space *mapping = inode->i_mapping;
1862 int blocks = exbh->b_size >> inode->i_blkbits;
1863 sector_t pblock = exbh->b_blocknr, cur_logical;
1864 struct buffer_head *head, *bh;
1866 struct pagevec pvec;
1869 index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1870 end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1871 cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1873 pagevec_init(&pvec, 0);
1875 while (index <= end) {
1876 /* XXX: optimize tail */
1877 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1880 for (i = 0; i < nr_pages; i++) {
1881 struct page *page = pvec.pages[i];
1883 index = page->index;
1888 BUG_ON(!PageLocked(page));
1889 BUG_ON(PageWriteback(page));
1890 BUG_ON(!page_has_buffers(page));
1892 bh = page_buffers(page);
1895 /* skip blocks out of the range */
1897 if (cur_logical >= logical)
1900 } while ((bh = bh->b_this_page) != head);
1903 if (cur_logical >= logical + blocks)
1905 if (buffer_delay(bh)) {
1906 bh->b_blocknr = pblock;
1907 clear_buffer_delay(bh);
1908 bh->b_bdev = inode->i_sb->s_bdev;
1909 } else if (buffer_unwritten(bh)) {
1910 bh->b_blocknr = pblock;
1911 clear_buffer_unwritten(bh);
1912 set_buffer_mapped(bh);
1914 bh->b_bdev = inode->i_sb->s_bdev;
1915 } else if (buffer_mapped(bh))
1916 BUG_ON(bh->b_blocknr != pblock);
1920 } while ((bh = bh->b_this_page) != head);
1922 pagevec_release(&pvec);
1928 * __unmap_underlying_blocks - just a helper function to unmap
1929 * set of blocks described by @bh
1931 static inline void __unmap_underlying_blocks(struct inode *inode,
1932 struct buffer_head *bh)
1934 struct block_device *bdev = inode->i_sb->s_bdev;
1937 blocks = bh->b_size >> inode->i_blkbits;
1938 for (i = 0; i < blocks; i++)
1939 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
1942 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
1943 sector_t logical, long blk_cnt)
1947 struct pagevec pvec;
1948 struct inode *inode = mpd->inode;
1949 struct address_space *mapping = inode->i_mapping;
1951 index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
1952 end = (logical + blk_cnt - 1) >>
1953 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1954 while (index <= end) {
1955 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1958 for (i = 0; i < nr_pages; i++) {
1959 struct page *page = pvec.pages[i];
1960 index = page->index;
1965 BUG_ON(!PageLocked(page));
1966 BUG_ON(PageWriteback(page));
1967 block_invalidatepage(page, 0);
1968 ClearPageUptodate(page);
1975 static void ext4_print_free_blocks(struct inode *inode)
1977 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1978 printk(KERN_EMERG "Total free blocks count %lld\n",
1979 ext4_count_free_blocks(inode->i_sb));
1980 printk(KERN_EMERG "Free/Dirty block details\n");
1981 printk(KERN_EMERG "free_blocks=%lld\n",
1982 (long long)percpu_counter_sum(&sbi->s_freeblocks_counter));
1983 printk(KERN_EMERG "dirty_blocks=%lld\n",
1984 (long long)percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1985 printk(KERN_EMERG "Block reservation details\n");
1986 printk(KERN_EMERG "i_reserved_data_blocks=%u\n",
1987 EXT4_I(inode)->i_reserved_data_blocks);
1988 printk(KERN_EMERG "i_reserved_meta_blocks=%u\n",
1989 EXT4_I(inode)->i_reserved_meta_blocks);
1993 #define EXT4_DELALLOC_RSVED 1
1994 static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
1995 struct buffer_head *bh_result, int create)
1998 unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1999 loff_t disksize = EXT4_I(inode)->i_disksize;
2000 handle_t *handle = NULL;
2002 handle = ext4_journal_current_handle();
2004 ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
2005 bh_result, create, 0, EXT4_DELALLOC_RSVED);
2009 bh_result->b_size = (ret << inode->i_blkbits);
2011 if (ext4_should_order_data(inode)) {
2013 retval = ext4_jbd2_file_inode(handle, inode);
2016 * Failed to add inode for ordered mode. Don't
2023 * Update on-disk size along with block allocation we don't
2024 * use 'extend_disksize' as size may change within already
2025 * allocated block -bzzz
2027 disksize = ((loff_t) iblock + ret) << inode->i_blkbits;
2028 if (disksize > i_size_read(inode))
2029 disksize = i_size_read(inode);
2030 if (disksize > EXT4_I(inode)->i_disksize) {
2031 ext4_update_i_disksize(inode, disksize);
2032 ret = ext4_mark_inode_dirty(handle, inode);
2039 * mpage_da_map_blocks - go through given space
2041 * @mpd - bh describing space
2043 * The function skips space we know is already mapped to disk blocks.
2046 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2049 struct buffer_head new;
2053 * We consider only non-mapped and non-allocated blocks
2055 if ((mpd->b_state & (1 << BH_Mapped)) &&
2056 !(mpd->b_state & (1 << BH_Delay)))
2059 * We need to make sure the BH_Delay flag is passed down to
2060 * ext4_da_get_block_write(), since it calls
2061 * ext4_get_blocks_wrap() with the EXT4_DELALLOC_RSVED flag.
2062 * This flag causes ext4_get_blocks_wrap() to call
2063 * ext4_da_update_reserve_space() if the passed buffer head
2064 * has the BH_Delay flag set. In the future, once we clean up
2065 * the interfaces to ext4_get_blocks_wrap(), we should pass in
2066 * a separate flag which requests that the delayed allocation
2067 * statistics should be updated, instead of depending on the
2068 * state information getting passed down via the map_bh's
2069 * state bitmasks plus the magic EXT4_DELALLOC_RSVED flag.
2071 new.b_state = mpd->b_state & (1 << BH_Delay);
2073 new.b_size = mpd->b_size;
2074 next = mpd->b_blocknr;
2076 * If we didn't accumulate anything
2077 * to write simply return
2082 err = ext4_da_get_block_write(mpd->inode, next, &new, 1);
2085 * If get block returns with error we simply
2086 * return. Later writepage will redirty the page and
2087 * writepages will find the dirty page again
2092 if (err == -ENOSPC &&
2093 ext4_count_free_blocks(mpd->inode->i_sb)) {
2099 * get block failure will cause us to loop in
2100 * writepages, because a_ops->writepage won't be able
2101 * to make progress. The page will be redirtied by
2102 * writepage and writepages will again try to write
2105 printk(KERN_EMERG "%s block allocation failed for inode %lu "
2106 "at logical offset %llu with max blocks "
2107 "%zd with error %d\n",
2108 __func__, mpd->inode->i_ino,
2109 (unsigned long long)next,
2110 mpd->b_size >> mpd->inode->i_blkbits, err);
2111 printk(KERN_EMERG "This should not happen.!! "
2112 "Data will be lost\n");
2113 if (err == -ENOSPC) {
2114 ext4_print_free_blocks(mpd->inode);
2116 /* invlaidate all the pages */
2117 ext4_da_block_invalidatepages(mpd, next,
2118 mpd->b_size >> mpd->inode->i_blkbits);
2121 BUG_ON(new.b_size == 0);
2123 if (buffer_new(&new))
2124 __unmap_underlying_blocks(mpd->inode, &new);
2127 * If blocks are delayed marked, we need to
2128 * put actual blocknr and drop delayed bit
2130 if ((mpd->b_state & (1 << BH_Delay)) ||
2131 (mpd->b_state & (1 << BH_Unwritten)))
2132 mpage_put_bnr_to_bhs(mpd, next, &new);
2137 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2138 (1 << BH_Delay) | (1 << BH_Unwritten))
2141 * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2143 * @mpd->lbh - extent of blocks
2144 * @logical - logical number of the block in the file
2145 * @bh - bh of the block (used to access block's state)
2147 * the function is used to collect contig. blocks in same state
2149 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2150 sector_t logical, size_t b_size,
2151 unsigned long b_state)
2154 int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2156 /* check if thereserved journal credits might overflow */
2157 if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2158 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2160 * With non-extent format we are limited by the journal
2161 * credit available. Total credit needed to insert
2162 * nrblocks contiguous blocks is dependent on the
2163 * nrblocks. So limit nrblocks.
2166 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2167 EXT4_MAX_TRANS_DATA) {
2169 * Adding the new buffer_head would make it cross the
2170 * allowed limit for which we have journal credit
2171 * reserved. So limit the new bh->b_size
2173 b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2174 mpd->inode->i_blkbits;
2175 /* we will do mpage_da_submit_io in the next loop */
2179 * First block in the extent
2181 if (mpd->b_size == 0) {
2182 mpd->b_blocknr = logical;
2183 mpd->b_size = b_size;
2184 mpd->b_state = b_state & BH_FLAGS;
2188 next = mpd->b_blocknr + nrblocks;
2190 * Can we merge the block to our big extent?
2192 if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2193 mpd->b_size += b_size;
2199 * We couldn't merge the block to our extent, so we
2200 * need to flush current extent and start new one
2202 if (mpage_da_map_blocks(mpd) == 0)
2203 mpage_da_submit_io(mpd);
2209 * __mpage_da_writepage - finds extent of pages and blocks
2211 * @page: page to consider
2212 * @wbc: not used, we just follow rules
2215 * The function finds extents of pages and scan them for all blocks.
2217 static int __mpage_da_writepage(struct page *page,
2218 struct writeback_control *wbc, void *data)
2220 struct mpage_da_data *mpd = data;
2221 struct inode *inode = mpd->inode;
2222 struct buffer_head *bh, *head;
2227 * Rest of the page in the page_vec
2228 * redirty then and skip then. We will
2229 * try to to write them again after
2230 * starting a new transaction
2232 redirty_page_for_writepage(wbc, page);
2234 return MPAGE_DA_EXTENT_TAIL;
2237 * Can we merge this page to current extent?
2239 if (mpd->next_page != page->index) {
2241 * Nope, we can't. So, we map non-allocated blocks
2242 * and start IO on them using writepage()
2244 if (mpd->next_page != mpd->first_page) {
2245 if (mpage_da_map_blocks(mpd) == 0)
2246 mpage_da_submit_io(mpd);
2248 * skip rest of the page in the page_vec
2251 redirty_page_for_writepage(wbc, page);
2253 return MPAGE_DA_EXTENT_TAIL;
2257 * Start next extent of pages ...
2259 mpd->first_page = page->index;
2269 mpd->next_page = page->index + 1;
2270 logical = (sector_t) page->index <<
2271 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2273 if (!page_has_buffers(page)) {
2274 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2275 (1 << BH_Dirty) | (1 << BH_Uptodate));
2277 return MPAGE_DA_EXTENT_TAIL;
2280 * Page with regular buffer heads, just add all dirty ones
2282 head = page_buffers(page);
2285 BUG_ON(buffer_locked(bh));
2287 * We need to try to allocate
2288 * unmapped blocks in the same page.
2289 * Otherwise we won't make progress
2290 * with the page in ext4_da_writepage
2292 if (buffer_dirty(bh) &&
2293 (!buffer_mapped(bh) || buffer_delay(bh))) {
2294 mpage_add_bh_to_extent(mpd, logical,
2298 return MPAGE_DA_EXTENT_TAIL;
2299 } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2301 * mapped dirty buffer. We need to update
2302 * the b_state because we look at
2303 * b_state in mpage_da_map_blocks. We don't
2304 * update b_size because if we find an
2305 * unmapped buffer_head later we need to
2306 * use the b_state flag of that buffer_head.
2308 if (mpd->b_size == 0)
2309 mpd->b_state = bh->b_state & BH_FLAGS;
2312 } while ((bh = bh->b_this_page) != head);
2319 * this is a special callback for ->write_begin() only
2320 * it's intention is to return mapped block or reserve space
2322 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2323 struct buffer_head *bh_result, int create)
2326 sector_t invalid_block = ~((sector_t) 0xffff);
2328 if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2331 BUG_ON(create == 0);
2332 BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2335 * first, we need to know whether the block is allocated already
2336 * preallocated blocks are unmapped but should treated
2337 * the same as allocated blocks.
2339 ret = ext4_get_blocks_wrap(NULL, inode, iblock, 1, bh_result, 0, 0, 0);
2340 if ((ret == 0) && !buffer_delay(bh_result)) {
2341 /* the block isn't (pre)allocated yet, let's reserve space */
2343 * XXX: __block_prepare_write() unmaps passed block,
2346 ret = ext4_da_reserve_space(inode, 1);
2348 /* not enough space to reserve */
2351 map_bh(bh_result, inode->i_sb, invalid_block);
2352 set_buffer_new(bh_result);
2353 set_buffer_delay(bh_result);
2354 } else if (ret > 0) {
2355 bh_result->b_size = (ret << inode->i_blkbits);
2357 * With sub-block writes into unwritten extents
2358 * we also need to mark the buffer as new so that
2359 * the unwritten parts of the buffer gets correctly zeroed.
2361 if (buffer_unwritten(bh_result))
2362 set_buffer_new(bh_result);
2369 static int ext4_bh_unmapped_or_delay(handle_t *handle, struct buffer_head *bh)
2372 * unmapped buffer is possible for holes.
2373 * delay buffer is possible with delayed allocation
2375 return ((!buffer_mapped(bh) || buffer_delay(bh)) && buffer_dirty(bh));
2378 static int ext4_normal_get_block_write(struct inode *inode, sector_t iblock,
2379 struct buffer_head *bh_result, int create)
2382 unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2385 * we don't want to do block allocation in writepage
2386 * so call get_block_wrap with create = 0
2388 ret = ext4_get_blocks_wrap(NULL, inode, iblock, max_blocks,
2389 bh_result, 0, 0, 0);
2391 bh_result->b_size = (ret << inode->i_blkbits);
2398 * get called vi ext4_da_writepages after taking page lock (have journal handle)
2399 * get called via journal_submit_inode_data_buffers (no journal handle)
2400 * get called via shrink_page_list via pdflush (no journal handle)
2401 * or grab_page_cache when doing write_begin (have journal handle)
2403 static int ext4_da_writepage(struct page *page,
2404 struct writeback_control *wbc)
2409 struct buffer_head *page_bufs;
2410 struct inode *inode = page->mapping->host;
2412 trace_mark(ext4_da_writepage,
2413 "dev %s ino %lu page_index %lu",
2414 inode->i_sb->s_id, inode->i_ino, page->index);
2415 size = i_size_read(inode);
2416 if (page->index == size >> PAGE_CACHE_SHIFT)
2417 len = size & ~PAGE_CACHE_MASK;
2419 len = PAGE_CACHE_SIZE;
2421 if (page_has_buffers(page)) {
2422 page_bufs = page_buffers(page);
2423 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2424 ext4_bh_unmapped_or_delay)) {
2426 * We don't want to do block allocation
2427 * So redirty the page and return
2428 * We may reach here when we do a journal commit
2429 * via journal_submit_inode_data_buffers.
2430 * If we don't have mapping block we just ignore
2431 * them. We can also reach here via shrink_page_list
2433 redirty_page_for_writepage(wbc, page);
2439 * The test for page_has_buffers() is subtle:
2440 * We know the page is dirty but it lost buffers. That means
2441 * that at some moment in time after write_begin()/write_end()
2442 * has been called all buffers have been clean and thus they
2443 * must have been written at least once. So they are all
2444 * mapped and we can happily proceed with mapping them
2445 * and writing the page.
2447 * Try to initialize the buffer_heads and check whether
2448 * all are mapped and non delay. We don't want to
2449 * do block allocation here.
2451 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
2452 ext4_normal_get_block_write);
2454 page_bufs = page_buffers(page);
2455 /* check whether all are mapped and non delay */
2456 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2457 ext4_bh_unmapped_or_delay)) {
2458 redirty_page_for_writepage(wbc, page);
2464 * We can't do block allocation here
2465 * so just redity the page and unlock
2468 redirty_page_for_writepage(wbc, page);
2472 /* now mark the buffer_heads as dirty and uptodate */
2473 block_commit_write(page, 0, PAGE_CACHE_SIZE);
2476 if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2477 ret = nobh_writepage(page, ext4_normal_get_block_write, wbc);
2479 ret = block_write_full_page(page,
2480 ext4_normal_get_block_write,
2487 * This is called via ext4_da_writepages() to
2488 * calulate the total number of credits to reserve to fit
2489 * a single extent allocation into a single transaction,
2490 * ext4_da_writpeages() will loop calling this before
2491 * the block allocation.
2494 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2496 int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2499 * With non-extent format the journal credit needed to
2500 * insert nrblocks contiguous block is dependent on
2501 * number of contiguous block. So we will limit
2502 * number of contiguous block to a sane value
2504 if (!(inode->i_flags & EXT4_EXTENTS_FL) &&
2505 (max_blocks > EXT4_MAX_TRANS_DATA))
2506 max_blocks = EXT4_MAX_TRANS_DATA;
2508 return ext4_chunk_trans_blocks(inode, max_blocks);
2511 static int ext4_da_writepages(struct address_space *mapping,
2512 struct writeback_control *wbc)
2515 int range_whole = 0;
2516 handle_t *handle = NULL;
2517 struct mpage_da_data mpd;
2518 struct inode *inode = mapping->host;
2519 int no_nrwrite_index_update;
2520 int pages_written = 0;
2522 int range_cyclic, cycled = 1, io_done = 0;
2523 int needed_blocks, ret = 0, nr_to_writebump = 0;
2524 struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2526 trace_mark(ext4_da_writepages,
2527 "dev %s ino %lu nr_t_write %ld "
2528 "pages_skipped %ld range_start %llu "
2529 "range_end %llu nonblocking %d "
2530 "for_kupdate %d for_reclaim %d "
2531 "for_writepages %d range_cyclic %d",
2532 inode->i_sb->s_id, inode->i_ino,
2533 wbc->nr_to_write, wbc->pages_skipped,
2534 (unsigned long long) wbc->range_start,
2535 (unsigned long long) wbc->range_end,
2536 wbc->nonblocking, wbc->for_kupdate,
2537 wbc->for_reclaim, wbc->for_writepages,
2541 * No pages to write? This is mainly a kludge to avoid starting
2542 * a transaction for special inodes like journal inode on last iput()
2543 * because that could violate lock ordering on umount
2545 if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2549 * If the filesystem has aborted, it is read-only, so return
2550 * right away instead of dumping stack traces later on that
2551 * will obscure the real source of the problem. We test
2552 * EXT4_MOUNT_ABORT instead of sb->s_flag's MS_RDONLY because
2553 * the latter could be true if the filesystem is mounted
2554 * read-only, and in that case, ext4_da_writepages should
2555 * *never* be called, so if that ever happens, we would want
2558 if (unlikely(sbi->s_mount_opt & EXT4_MOUNT_ABORT))
2562 * Make sure nr_to_write is >= sbi->s_mb_stream_request
2563 * This make sure small files blocks are allocated in
2564 * single attempt. This ensure that small files
2565 * get less fragmented.
2567 if (wbc->nr_to_write < sbi->s_mb_stream_request) {
2568 nr_to_writebump = sbi->s_mb_stream_request - wbc->nr_to_write;
2569 wbc->nr_to_write = sbi->s_mb_stream_request;
2571 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2574 range_cyclic = wbc->range_cyclic;
2575 if (wbc->range_cyclic) {
2576 index = mapping->writeback_index;
2579 wbc->range_start = index << PAGE_CACHE_SHIFT;
2580 wbc->range_end = LLONG_MAX;
2581 wbc->range_cyclic = 0;
2583 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2586 mpd.inode = mapping->host;
2589 * we don't want write_cache_pages to update
2590 * nr_to_write and writeback_index
2592 no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2593 wbc->no_nrwrite_index_update = 1;
2594 pages_skipped = wbc->pages_skipped;
2597 while (!ret && wbc->nr_to_write > 0) {
2600 * we insert one extent at a time. So we need
2601 * credit needed for single extent allocation.
2602 * journalled mode is currently not supported
2605 BUG_ON(ext4_should_journal_data(inode));
2606 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2608 /* start a new transaction*/
2609 handle = ext4_journal_start(inode, needed_blocks);
2610 if (IS_ERR(handle)) {
2611 ret = PTR_ERR(handle);
2612 printk(KERN_CRIT "%s: jbd2_start: "
2613 "%ld pages, ino %lu; err %d\n", __func__,
2614 wbc->nr_to_write, inode->i_ino, ret);
2616 goto out_writepages;
2620 * Now call __mpage_da_writepage to find the next
2621 * contiguous region of logical blocks that need
2622 * blocks to be allocated by ext4. We don't actually
2623 * submit the blocks for I/O here, even though
2624 * write_cache_pages thinks it will, and will set the
2625 * pages as clean for write before calling
2626 * __mpage_da_writepage().
2634 mpd.pages_written = 0;
2636 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2639 * If we have a contigous extent of pages and we
2640 * haven't done the I/O yet, map the blocks and submit
2643 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2644 if (mpage_da_map_blocks(&mpd) == 0)
2645 mpage_da_submit_io(&mpd);
2647 ret = MPAGE_DA_EXTENT_TAIL;
2649 wbc->nr_to_write -= mpd.pages_written;
2651 ext4_journal_stop(handle);
2653 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2654 /* commit the transaction which would
2655 * free blocks released in the transaction
2658 jbd2_journal_force_commit_nested(sbi->s_journal);
2659 wbc->pages_skipped = pages_skipped;
2661 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2663 * got one extent now try with
2666 pages_written += mpd.pages_written;
2667 wbc->pages_skipped = pages_skipped;
2670 } else if (wbc->nr_to_write)
2672 * There is no more writeout needed
2673 * or we requested for a noblocking writeout
2674 * and we found the device congested
2678 if (!io_done && !cycled) {
2681 wbc->range_start = index << PAGE_CACHE_SHIFT;
2682 wbc->range_end = mapping->writeback_index - 1;
2685 if (pages_skipped != wbc->pages_skipped)
2686 printk(KERN_EMERG "This should not happen leaving %s "
2687 "with nr_to_write = %ld ret = %d\n",
2688 __func__, wbc->nr_to_write, ret);
2691 index += pages_written;
2692 wbc->range_cyclic = range_cyclic;
2693 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2695 * set the writeback_index so that range_cyclic
2696 * mode will write it back later
2698 mapping->writeback_index = index;
2701 if (!no_nrwrite_index_update)
2702 wbc->no_nrwrite_index_update = 0;
2703 wbc->nr_to_write -= nr_to_writebump;
2704 trace_mark(ext4_da_writepage_result,
2705 "dev %s ino %lu ret %d pages_written %d "
2706 "pages_skipped %ld congestion %d "
2707 "more_io %d no_nrwrite_index_update %d",
2708 inode->i_sb->s_id, inode->i_ino, ret,
2709 pages_written, wbc->pages_skipped,
2710 wbc->encountered_congestion, wbc->more_io,
2711 wbc->no_nrwrite_index_update);
2715 #define FALL_BACK_TO_NONDELALLOC 1
2716 static int ext4_nonda_switch(struct super_block *sb)
2718 s64 free_blocks, dirty_blocks;
2719 struct ext4_sb_info *sbi = EXT4_SB(sb);
2722 * switch to non delalloc mode if we are running low
2723 * on free block. The free block accounting via percpu
2724 * counters can get slightly wrong with percpu_counter_batch getting
2725 * accumulated on each CPU without updating global counters
2726 * Delalloc need an accurate free block accounting. So switch
2727 * to non delalloc when we are near to error range.
2729 free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2730 dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2731 if (2 * free_blocks < 3 * dirty_blocks ||
2732 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2734 * free block count is less that 150% of dirty blocks
2735 * or free blocks is less that watermark
2742 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2743 loff_t pos, unsigned len, unsigned flags,
2744 struct page **pagep, void **fsdata)
2746 int ret, retries = 0;
2750 struct inode *inode = mapping->host;
2753 index = pos >> PAGE_CACHE_SHIFT;
2754 from = pos & (PAGE_CACHE_SIZE - 1);
2757 if (ext4_nonda_switch(inode->i_sb)) {
2758 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2759 return ext4_write_begin(file, mapping, pos,
2760 len, flags, pagep, fsdata);
2762 *fsdata = (void *)0;
2764 trace_mark(ext4_da_write_begin,
2765 "dev %s ino %lu pos %llu len %u flags %u",
2766 inode->i_sb->s_id, inode->i_ino,
2767 (unsigned long long) pos, len, flags);
2770 * With delayed allocation, we don't log the i_disksize update
2771 * if there is delayed block allocation. But we still need
2772 * to journalling the i_disksize update if writes to the end
2773 * of file which has an already mapped buffer.
2775 handle = ext4_journal_start(inode, 1);
2776 if (IS_ERR(handle)) {
2777 ret = PTR_ERR(handle);
2780 /* We cannot recurse into the filesystem as the transaction is already
2782 flags |= AOP_FLAG_NOFS;
2784 page = grab_cache_page_write_begin(mapping, index, flags);
2786 ext4_journal_stop(handle);
2792 ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
2793 ext4_da_get_block_prep);
2796 ext4_journal_stop(handle);
2797 page_cache_release(page);
2799 * block_write_begin may have instantiated a few blocks
2800 * outside i_size. Trim these off again. Don't need
2801 * i_size_read because we hold i_mutex.
2803 if (pos + len > inode->i_size)
2804 vmtruncate(inode, inode->i_size);
2807 if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2814 * Check if we should update i_disksize
2815 * when write to the end of file but not require block allocation
2817 static int ext4_da_should_update_i_disksize(struct page *page,
2818 unsigned long offset)
2820 struct buffer_head *bh;
2821 struct inode *inode = page->mapping->host;
2825 bh = page_buffers(page);
2826 idx = offset >> inode->i_blkbits;
2828 for (i = 0; i < idx; i++)
2829 bh = bh->b_this_page;
2831 if (!buffer_mapped(bh) || (buffer_delay(bh)))
2836 static int ext4_da_write_end(struct file *file,
2837 struct address_space *mapping,
2838 loff_t pos, unsigned len, unsigned copied,
2839 struct page *page, void *fsdata)
2841 struct inode *inode = mapping->host;
2843 handle_t *handle = ext4_journal_current_handle();
2845 unsigned long start, end;
2846 int write_mode = (int)(unsigned long)fsdata;
2848 if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2849 if (ext4_should_order_data(inode)) {
2850 return ext4_ordered_write_end(file, mapping, pos,
2851 len, copied, page, fsdata);
2852 } else if (ext4_should_writeback_data(inode)) {
2853 return ext4_writeback_write_end(file, mapping, pos,
2854 len, copied, page, fsdata);
2860 trace_mark(ext4_da_write_end,
2861 "dev %s ino %lu pos %llu len %u copied %u",
2862 inode->i_sb->s_id, inode->i_ino,
2863 (unsigned long long) pos, len, copied);
2864 start = pos & (PAGE_CACHE_SIZE - 1);
2865 end = start + copied - 1;
2868 * generic_write_end() will run mark_inode_dirty() if i_size
2869 * changes. So let's piggyback the i_disksize mark_inode_dirty
2873 new_i_size = pos + copied;
2874 if (new_i_size > EXT4_I(inode)->i_disksize) {
2875 if (ext4_da_should_update_i_disksize(page, end)) {
2876 down_write(&EXT4_I(inode)->i_data_sem);
2877 if (new_i_size > EXT4_I(inode)->i_disksize) {
2879 * Updating i_disksize when extending file
2880 * without needing block allocation
2882 if (ext4_should_order_data(inode))
2883 ret = ext4_jbd2_file_inode(handle,
2886 EXT4_I(inode)->i_disksize = new_i_size;
2888 up_write(&EXT4_I(inode)->i_data_sem);
2889 /* We need to mark inode dirty even if
2890 * new_i_size is less that inode->i_size
2891 * bu greater than i_disksize.(hint delalloc)
2893 ext4_mark_inode_dirty(handle, inode);
2896 ret2 = generic_write_end(file, mapping, pos, len, copied,
2901 ret2 = ext4_journal_stop(handle);
2905 return ret ? ret : copied;
2908 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2911 * Drop reserved blocks
2913 BUG_ON(!PageLocked(page));
2914 if (!page_has_buffers(page))
2917 ext4_da_page_release_reservation(page, offset);
2920 ext4_invalidatepage(page, offset);
2926 * Force all delayed allocation blocks to be allocated for a given inode.
2928 int ext4_alloc_da_blocks(struct inode *inode)
2930 if (!EXT4_I(inode)->i_reserved_data_blocks &&
2931 !EXT4_I(inode)->i_reserved_meta_blocks)
2935 * We do something simple for now. The filemap_flush() will
2936 * also start triggering a write of the data blocks, which is
2937 * not strictly speaking necessary (and for users of
2938 * laptop_mode, not even desirable). However, to do otherwise
2939 * would require replicating code paths in:
2941 * ext4_da_writepages() ->
2942 * write_cache_pages() ---> (via passed in callback function)
2943 * __mpage_da_writepage() -->
2944 * mpage_add_bh_to_extent()
2945 * mpage_da_map_blocks()
2947 * The problem is that write_cache_pages(), located in
2948 * mm/page-writeback.c, marks pages clean in preparation for
2949 * doing I/O, which is not desirable if we're not planning on
2952 * We could call write_cache_pages(), and then redirty all of
2953 * the pages by calling redirty_page_for_writeback() but that
2954 * would be ugly in the extreme. So instead we would need to
2955 * replicate parts of the code in the above functions,
2956 * simplifying them becuase we wouldn't actually intend to
2957 * write out the pages, but rather only collect contiguous
2958 * logical block extents, call the multi-block allocator, and
2959 * then update the buffer heads with the block allocations.
2961 * For now, though, we'll cheat by calling filemap_flush(),
2962 * which will map the blocks, and start the I/O, but not
2963 * actually wait for the I/O to complete.
2965 return filemap_flush(inode->i_mapping);
2969 * bmap() is special. It gets used by applications such as lilo and by
2970 * the swapper to find the on-disk block of a specific piece of data.
2972 * Naturally, this is dangerous if the block concerned is still in the
2973 * journal. If somebody makes a swapfile on an ext4 data-journaling
2974 * filesystem and enables swap, then they may get a nasty shock when the
2975 * data getting swapped to that swapfile suddenly gets overwritten by
2976 * the original zero's written out previously to the journal and
2977 * awaiting writeback in the kernel's buffer cache.
2979 * So, if we see any bmap calls here on a modified, data-journaled file,
2980 * take extra steps to flush any blocks which might be in the cache.
2982 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2984 struct inode *inode = mapping->host;
2988 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2989 test_opt(inode->i_sb, DELALLOC)) {
2991 * With delalloc we want to sync the file
2992 * so that we can make sure we allocate
2995 filemap_write_and_wait(mapping);
2998 if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
3000 * This is a REALLY heavyweight approach, but the use of
3001 * bmap on dirty files is expected to be extremely rare:
3002 * only if we run lilo or swapon on a freshly made file
3003 * do we expect this to happen.
3005 * (bmap requires CAP_SYS_RAWIO so this does not
3006 * represent an unprivileged user DOS attack --- we'd be
3007 * in trouble if mortal users could trigger this path at
3010 * NB. EXT4_STATE_JDATA is not set on files other than
3011 * regular files. If somebody wants to bmap a directory
3012 * or symlink and gets confused because the buffer
3013 * hasn't yet been flushed to disk, they deserve
3014 * everything they get.
3017 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
3018 journal = EXT4_JOURNAL(inode);
3019 jbd2_journal_lock_updates(journal);
3020 err = jbd2_journal_flush(journal);
3021 jbd2_journal_unlock_updates(journal);
3027 return generic_block_bmap(mapping, block, ext4_get_block);
3030 static int bget_one(handle_t *handle, struct buffer_head *bh)
3036 static int bput_one(handle_t *handle, struct buffer_head *bh)
3043 * Note that we don't need to start a transaction unless we're journaling data
3044 * because we should have holes filled from ext4_page_mkwrite(). We even don't
3045 * need to file the inode to the transaction's list in ordered mode because if
3046 * we are writing back data added by write(), the inode is already there and if
3047 * we are writing back data modified via mmap(), noone guarantees in which
3048 * transaction the data will hit the disk. In case we are journaling data, we
3049 * cannot start transaction directly because transaction start ranks above page
3050 * lock so we have to do some magic.
3052 * In all journaling modes block_write_full_page() will start the I/O.
3056 * ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
3061 * ext4_file_write() -> generic_file_write() -> __alloc_pages() -> ...
3063 * Same applies to ext4_get_block(). We will deadlock on various things like
3064 * lock_journal and i_data_sem
3066 * Setting PF_MEMALLOC here doesn't work - too many internal memory
3069 * 16May01: If we're reentered then journal_current_handle() will be
3070 * non-zero. We simply *return*.
3072 * 1 July 2001: @@@ FIXME:
3073 * In journalled data mode, a data buffer may be metadata against the
3074 * current transaction. But the same file is part of a shared mapping
3075 * and someone does a writepage() on it.
3077 * We will move the buffer onto the async_data list, but *after* it has
3078 * been dirtied. So there's a small window where we have dirty data on
3081 * Note that this only applies to the last partial page in the file. The
3082 * bit which block_write_full_page() uses prepare/commit for. (That's
3083 * broken code anyway: it's wrong for msync()).
3085 * It's a rare case: affects the final partial page, for journalled data
3086 * where the file is subject to bith write() and writepage() in the same
3087 * transction. To fix it we'll need a custom block_write_full_page().
3088 * We'll probably need that anyway for journalling writepage() output.
3090 * We don't honour synchronous mounts for writepage(). That would be
3091 * disastrous. Any write() or metadata operation will sync the fs for
3095 static int __ext4_normal_writepage(struct page *page,
3096 struct writeback_control *wbc)
3098 struct inode *inode = page->mapping->host;
3100 if (test_opt(inode->i_sb, NOBH))
3101 return nobh_writepage(page,
3102 ext4_normal_get_block_write, wbc);
3104 return block_write_full_page(page,
3105 ext4_normal_get_block_write,
3109 static int ext4_normal_writepage(struct page *page,
3110 struct writeback_control *wbc)
3112 struct inode *inode = page->mapping->host;
3113 loff_t size = i_size_read(inode);
3116 trace_mark(ext4_normal_writepage,
3117 "dev %s ino %lu page_index %lu",
3118 inode->i_sb->s_id, inode->i_ino, page->index);
3119 J_ASSERT(PageLocked(page));
3120 if (page->index == size >> PAGE_CACHE_SHIFT)
3121 len = size & ~PAGE_CACHE_MASK;
3123 len = PAGE_CACHE_SIZE;
3125 if (page_has_buffers(page)) {
3126 /* if page has buffers it should all be mapped
3127 * and allocated. If there are not buffers attached
3128 * to the page we know the page is dirty but it lost
3129 * buffers. That means that at some moment in time
3130 * after write_begin() / write_end() has been called
3131 * all buffers have been clean and thus they must have been
3132 * written at least once. So they are all mapped and we can
3133 * happily proceed with mapping them and writing the page.
3135 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3136 ext4_bh_unmapped_or_delay));
3139 if (!ext4_journal_current_handle())
3140 return __ext4_normal_writepage(page, wbc);
3142 redirty_page_for_writepage(wbc, page);
3147 static int __ext4_journalled_writepage(struct page *page,
3148 struct writeback_control *wbc)
3150 struct address_space *mapping = page->mapping;
3151 struct inode *inode = mapping->host;
3152 struct buffer_head *page_bufs;
3153 handle_t *handle = NULL;
3157 ret = block_prepare_write(page, 0, PAGE_CACHE_SIZE,
3158 ext4_normal_get_block_write);
3162 page_bufs = page_buffers(page);
3163 walk_page_buffers(handle, page_bufs, 0, PAGE_CACHE_SIZE, NULL,
3165 /* As soon as we unlock the page, it can go away, but we have
3166 * references to buffers so we are safe */
3169 handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
3170 if (IS_ERR(handle)) {
3171 ret = PTR_ERR(handle);
3175 ret = walk_page_buffers(handle, page_bufs, 0,
3176 PAGE_CACHE_SIZE, NULL, do_journal_get_write_access);
3178 err = walk_page_buffers(handle, page_bufs, 0,
3179 PAGE_CACHE_SIZE, NULL, write_end_fn);
3182 err = ext4_journal_stop(handle);
3186 walk_page_buffers(handle, page_bufs, 0,
3187 PAGE_CACHE_SIZE, NULL, bput_one);
3188 EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
3197 static int ext4_journalled_writepage(struct page *page,
3198 struct writeback_control *wbc)
3200 struct inode *inode = page->mapping->host;
3201 loff_t size = i_size_read(inode);
3204 trace_mark(ext4_journalled_writepage,
3205 "dev %s ino %lu page_index %lu",
3206 inode->i_sb->s_id, inode->i_ino, page->index);
3207 J_ASSERT(PageLocked(page));
3208 if (page->index == size >> PAGE_CACHE_SHIFT)
3209 len = size & ~PAGE_CACHE_MASK;
3211 len = PAGE_CACHE_SIZE;
3213 if (page_has_buffers(page)) {
3214 /* if page has buffers it should all be mapped
3215 * and allocated. If there are not buffers attached
3216 * to the page we know the page is dirty but it lost
3217 * buffers. That means that at some moment in time
3218 * after write_begin() / write_end() has been called
3219 * all buffers have been clean and thus they must have been
3220 * written at least once. So they are all mapped and we can
3221 * happily proceed with mapping them and writing the page.
3223 BUG_ON(walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
3224 ext4_bh_unmapped_or_delay));
3227 if (ext4_journal_current_handle())
3230 if (PageChecked(page)) {
3232 * It's mmapped pagecache. Add buffers and journal it. There
3233 * doesn't seem much point in redirtying the page here.
3235 ClearPageChecked(page);
3236 return __ext4_journalled_writepage(page, wbc);
3239 * It may be a page full of checkpoint-mode buffers. We don't
3240 * really know unless we go poke around in the buffer_heads.
3241 * But block_write_full_page will do the right thing.
3243 return block_write_full_page(page,
3244 ext4_normal_get_block_write,
3248 redirty_page_for_writepage(wbc, page);
3253 static int ext4_readpage(struct file *file, struct page *page)
3255 return mpage_readpage(page, ext4_get_block);
3259 ext4_readpages(struct file *file, struct address_space *mapping,
3260 struct list_head *pages, unsigned nr_pages)
3262 return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3265 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3267 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3270 * If it's a full truncate we just forget about the pending dirtying
3273 ClearPageChecked(page);
3276 jbd2_journal_invalidatepage(journal, page, offset);
3278 block_invalidatepage(page, offset);
3281 static int ext4_releasepage(struct page *page, gfp_t wait)
3283 journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3285 WARN_ON(PageChecked(page));
3286 if (!page_has_buffers(page))
3289 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3291 return try_to_free_buffers(page);
3295 * If the O_DIRECT write will extend the file then add this inode to the
3296 * orphan list. So recovery will truncate it back to the original size
3297 * if the machine crashes during the write.
3299 * If the O_DIRECT write is intantiating holes inside i_size and the machine
3300 * crashes then stale disk data _may_ be exposed inside the file. But current
3301 * VFS code falls back into buffered path in that case so we are safe.
3303 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3304 const struct iovec *iov, loff_t offset,
3305 unsigned long nr_segs)
3307 struct file *file = iocb->ki_filp;
3308 struct inode *inode = file->f_mapping->host;
3309 struct ext4_inode_info *ei = EXT4_I(inode);
3313 size_t count = iov_length(iov, nr_segs);
3316 loff_t final_size = offset + count;
3318 if (final_size > inode->i_size) {
3319 /* Credits for sb + inode write */
3320 handle = ext4_journal_start(inode, 2);
3321 if (IS_ERR(handle)) {
3322 ret = PTR_ERR(handle);
3325 ret = ext4_orphan_add(handle, inode);
3327 ext4_journal_stop(handle);
3331 ei->i_disksize = inode->i_size;
3332 ext4_journal_stop(handle);
3336 ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3338 ext4_get_block, NULL);
3343 /* Credits for sb + inode write */
3344 handle = ext4_journal_start(inode, 2);
3345 if (IS_ERR(handle)) {
3346 /* This is really bad luck. We've written the data
3347 * but cannot extend i_size. Bail out and pretend
3348 * the write failed... */
3349 ret = PTR_ERR(handle);
3353 ext4_orphan_del(handle, inode);
3355 loff_t end = offset + ret;
3356 if (end > inode->i_size) {
3357 ei->i_disksize = end;
3358 i_size_write(inode, end);
3360 * We're going to return a positive `ret'
3361 * here due to non-zero-length I/O, so there's
3362 * no way of reporting error returns from
3363 * ext4_mark_inode_dirty() to userspace. So
3366 ext4_mark_inode_dirty(handle, inode);
3369 err = ext4_journal_stop(handle);
3378 * Pages can be marked dirty completely asynchronously from ext4's journalling
3379 * activity. By filemap_sync_pte(), try_to_unmap_one(), etc. We cannot do
3380 * much here because ->set_page_dirty is called under VFS locks. The page is
3381 * not necessarily locked.
3383 * We cannot just dirty the page and leave attached buffers clean, because the
3384 * buffers' dirty state is "definitive". We cannot just set the buffers dirty
3385 * or jbddirty because all the journalling code will explode.
3387 * So what we do is to mark the page "pending dirty" and next time writepage
3388 * is called, propagate that into the buffers appropriately.
3390 static int ext4_journalled_set_page_dirty(struct page *page)
3392 SetPageChecked(page);
3393 return __set_page_dirty_nobuffers(page);
3396 static const struct address_space_operations ext4_ordered_aops = {
3397 .readpage = ext4_readpage,
3398 .readpages = ext4_readpages,
3399 .writepage = ext4_normal_writepage,
3400 .sync_page = block_sync_page,
3401 .write_begin = ext4_write_begin,
3402 .write_end = ext4_ordered_write_end,
3404 .invalidatepage = ext4_invalidatepage,
3405 .releasepage = ext4_releasepage,
3406 .direct_IO = ext4_direct_IO,
3407 .migratepage = buffer_migrate_page,
3408 .is_partially_uptodate = block_is_partially_uptodate,
3411 static const struct address_space_operations ext4_writeback_aops = {
3412 .readpage = ext4_readpage,
3413 .readpages = ext4_readpages,
3414 .writepage = ext4_normal_writepage,
3415 .sync_page = block_sync_page,
3416 .write_begin = ext4_write_begin,
3417 .write_end = ext4_writeback_write_end,
3419 .invalidatepage = ext4_invalidatepage,
3420 .releasepage = ext4_releasepage,
3421 .direct_IO = ext4_direct_IO,
3422 .migratepage = buffer_migrate_page,
3423 .is_partially_uptodate = block_is_partially_uptodate,
3426 static const struct address_space_operations ext4_journalled_aops = {
3427 .readpage = ext4_readpage,
3428 .readpages = ext4_readpages,
3429 .writepage = ext4_journalled_writepage,
3430 .sync_page = block_sync_page,
3431 .write_begin = ext4_write_begin,
3432 .write_end = ext4_journalled_write_end,
3433 .set_page_dirty = ext4_journalled_set_page_dirty,
3435 .invalidatepage = ext4_invalidatepage,
3436 .releasepage = ext4_releasepage,
3437 .is_partially_uptodate = block_is_partially_uptodate,
3440 static const struct address_space_operations ext4_da_aops = {
3441 .readpage = ext4_readpage,
3442 .readpages = ext4_readpages,
3443 .writepage = ext4_da_writepage,
3444 .writepages = ext4_da_writepages,
3445 .sync_page = block_sync_page,
3446 .write_begin = ext4_da_write_begin,
3447 .write_end = ext4_da_write_end,
3449 .invalidatepage = ext4_da_invalidatepage,
3450 .releasepage = ext4_releasepage,
3451 .direct_IO = ext4_direct_IO,
3452 .migratepage = buffer_migrate_page,
3453 .is_partially_uptodate = block_is_partially_uptodate,
3456 void ext4_set_aops(struct inode *inode)
3458 if (ext4_should_order_data(inode) &&
3459 test_opt(inode->i_sb, DELALLOC))
3460 inode->i_mapping->a_ops = &ext4_da_aops;
3461 else if (ext4_should_order_data(inode))
3462 inode->i_mapping->a_ops = &ext4_ordered_aops;
3463 else if (ext4_should_writeback_data(inode) &&
3464 test_opt(inode->i_sb, DELALLOC))
3465 inode->i_mapping->a_ops = &ext4_da_aops;
3466 else if (ext4_should_writeback_data(inode))
3467 inode->i_mapping->a_ops = &ext4_writeback_aops;
3469 inode->i_mapping->a_ops = &ext4_journalled_aops;
3473 * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3474 * up to the end of the block which corresponds to `from'.
3475 * This required during truncate. We need to physically zero the tail end
3476 * of that block so it doesn't yield old data if the file is later grown.
3478 int ext4_block_truncate_page(handle_t *handle,
3479 struct address_space *mapping, loff_t from)
3481 ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3482 unsigned offset = from & (PAGE_CACHE_SIZE-1);
3483 unsigned blocksize, length, pos;
3485 struct inode *inode = mapping->host;
3486 struct buffer_head *bh;
3490 page = grab_cache_page(mapping, from >> PAGE_CACHE_SHIFT);
3494 blocksize = inode->i_sb->s_blocksize;
3495 length = blocksize - (offset & (blocksize - 1));
3496 iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3499 * For "nobh" option, we can only work if we don't need to
3500 * read-in the page - otherwise we create buffers to do the IO.
3502 if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3503 ext4_should_writeback_data(inode) && PageUptodate(page)) {
3504 zero_user(page, offset, length);
3505 set_page_dirty(page);
3509 if (!page_has_buffers(page))
3510 create_empty_buffers(page, blocksize, 0);
3512 /* Find the buffer that contains "offset" */
3513 bh = page_buffers(page);
3515 while (offset >= pos) {
3516 bh = bh->b_this_page;
3522 if (buffer_freed(bh)) {
3523 BUFFER_TRACE(bh, "freed: skip");
3527 if (!buffer_mapped(bh)) {
3528 BUFFER_TRACE(bh, "unmapped");
3529 ext4_get_block(inode, iblock, bh, 0);
3530 /* unmapped? It's a hole - nothing to do */
3531 if (!buffer_mapped(bh)) {
3532 BUFFER_TRACE(bh, "still unmapped");
3537 /* Ok, it's mapped. Make sure it's up-to-date */
3538 if (PageUptodate(page))
3539 set_buffer_uptodate(bh);
3541 if (!buffer_uptodate(bh)) {
3543 ll_rw_block(READ, 1, &bh);
3545 /* Uhhuh. Read error. Complain and punt. */
3546 if (!buffer_uptodate(bh))
3550 if (ext4_should_journal_data(inode)) {
3551 BUFFER_TRACE(bh, "get write access");
3552 err = ext4_journal_get_write_access(handle, bh);
3557 zero_user(page, offset, length);
3559 BUFFER_TRACE(bh, "zeroed end of block");
3562 if (ext4_should_journal_data(inode)) {
3563 err = ext4_handle_dirty_metadata(handle, inode, bh);
3565 if (ext4_should_order_data(inode))
3566 err = ext4_jbd2_file_inode(handle, inode);
3567 mark_buffer_dirty(bh);
3572 page_cache_release(page);
3577 * Probably it should be a library function... search for first non-zero word
3578 * or memcmp with zero_page, whatever is better for particular architecture.
3581 static inline int all_zeroes(__le32 *p, __le32 *q)
3590 * ext4_find_shared - find the indirect blocks for partial truncation.
3591 * @inode: inode in question
3592 * @depth: depth of the affected branch
3593 * @offsets: offsets of pointers in that branch (see ext4_block_to_path)
3594 * @chain: place to store the pointers to partial indirect blocks
3595 * @top: place to the (detached) top of branch
3597 * This is a helper function used by ext4_truncate().
3599 * When we do truncate() we may have to clean the ends of several
3600 * indirect blocks but leave the blocks themselves alive. Block is
3601 * partially truncated if some data below the new i_size is refered
3602 * from it (and it is on the path to the first completely truncated
3603 * data block, indeed). We have to free the top of that path along
3604 * with everything to the right of the path. Since no allocation
3605 * past the truncation point is possible until ext4_truncate()
3606 * finishes, we may safely do the latter, but top of branch may
3607 * require special attention - pageout below the truncation point
3608 * might try to populate it.
3610 * We atomically detach the top of branch from the tree, store the
3611 * block number of its root in *@top, pointers to buffer_heads of
3612 * partially truncated blocks - in @chain[].bh and pointers to
3613 * their last elements that should not be removed - in
3614 * @chain[].p. Return value is the pointer to last filled element
3617 * The work left to caller to do the actual freeing of subtrees:
3618 * a) free the subtree starting from *@top
3619 * b) free the subtrees whose roots are stored in
3620 * (@chain[i].p+1 .. end of @chain[i].bh->b_data)
3621 * c) free the subtrees growing from the inode past the @chain[0].
3622 * (no partially truncated stuff there). */
3624 static Indirect *ext4_find_shared(struct inode *inode, int depth,
3625 ext4_lblk_t offsets[4], Indirect chain[4], __le32 *top)
3627 Indirect *partial, *p;
3631 /* Make k index the deepest non-null offest + 1 */
3632 for (k = depth; k > 1 && !offsets[k-1]; k--)
3634 partial = ext4_get_branch(inode, k, offsets, chain, &err);
3635 /* Writer: pointers */
3637 partial = chain + k-1;
3639 * If the branch acquired continuation since we've looked at it -
3640 * fine, it should all survive and (new) top doesn't belong to us.
3642 if (!partial->key && *partial->p)
3645 for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
3648 * OK, we've found the last block that must survive. The rest of our
3649 * branch should be detached before unlocking. However, if that rest
3650 * of branch is all ours and does not grow immediately from the inode
3651 * it's easier to cheat and just decrement partial->p.
3653 if (p == chain + k - 1 && p > chain) {
3657 /* Nope, don't do this in ext4. Must leave the tree intact */
3664 while (partial > p) {
3665 brelse(partial->bh);
3673 * Zero a number of block pointers in either an inode or an indirect block.
3674 * If we restart the transaction we must again get write access to the
3675 * indirect block for further modification.
3677 * We release `count' blocks on disk, but (last - first) may be greater
3678 * than `count' because there can be holes in there.
3680 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
3681 struct buffer_head *bh, ext4_fsblk_t block_to_free,
3682 unsigned long count, __le32 *first, __le32 *last)
3685 if (try_to_extend_transaction(handle, inode)) {
3687 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3688 ext4_handle_dirty_metadata(handle, inode, bh);
3690 ext4_mark_inode_dirty(handle, inode);
3691 ext4_journal_test_restart(handle, inode);
3693 BUFFER_TRACE(bh, "retaking write access");
3694 ext4_journal_get_write_access(handle, bh);
3699 * Any buffers which are on the journal will be in memory. We find
3700 * them on the hash table so jbd2_journal_revoke() will run jbd2_journal_forget()
3701 * on them. We've already detached each block from the file, so
3702 * bforget() in jbd2_journal_forget() should be safe.
3704 * AKPM: turn on bforget in jbd2_journal_forget()!!!
3706 for (p = first; p < last; p++) {
3707 u32 nr = le32_to_cpu(*p);
3709 struct buffer_head *tbh;
3712 tbh = sb_find_get_block(inode->i_sb, nr);
3713 ext4_forget(handle, 0, inode, tbh, nr);
3717 ext4_free_blocks(handle, inode, block_to_free, count, 0);
3721 * ext4_free_data - free a list of data blocks
3722 * @handle: handle for this transaction
3723 * @inode: inode we are dealing with
3724 * @this_bh: indirect buffer_head which contains *@first and *@last
3725 * @first: array of block numbers
3726 * @last: points immediately past the end of array
3728 * We are freeing all blocks refered from that array (numbers are stored as
3729 * little-endian 32-bit) and updating @inode->i_blocks appropriately.
3731 * We accumulate contiguous runs of blocks to free. Conveniently, if these
3732 * blocks are contiguous then releasing them at one time will only affect one
3733 * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
3734 * actually use a lot of journal space.
3736 * @this_bh will be %NULL if @first and @last point into the inode's direct
3739 static void ext4_free_data(handle_t *handle, struct inode *inode,
3740 struct buffer_head *this_bh,
3741 __le32 *first, __le32 *last)
3743 ext4_fsblk_t block_to_free = 0; /* Starting block # of a run */
3744 unsigned long count = 0; /* Number of blocks in the run */
3745 __le32 *block_to_free_p = NULL; /* Pointer into inode/ind
3748 ext4_fsblk_t nr; /* Current block # */
3749 __le32 *p; /* Pointer into inode/ind
3750 for current block */
3753 if (this_bh) { /* For indirect block */
3754 BUFFER_TRACE(this_bh, "get_write_access");
3755 err = ext4_journal_get_write_access(handle, this_bh);
3756 /* Important: if we can't update the indirect pointers
3757 * to the blocks, we can't free them. */
3762 for (p = first; p < last; p++) {
3763 nr = le32_to_cpu(*p);
3765 /* accumulate blocks to free if they're contiguous */
3768 block_to_free_p = p;
3770 } else if (nr == block_to_free + count) {
3773 ext4_clear_blocks(handle, inode, this_bh,
3775 count, block_to_free_p, p);
3777 block_to_free_p = p;
3784 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
3785 count, block_to_free_p, p);
3788 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
3791 * The buffer head should have an attached journal head at this
3792 * point. However, if the data is corrupted and an indirect
3793 * block pointed to itself, it would have been detached when
3794 * the block was cleared. Check for this instead of OOPSing.
3796 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
3797 ext4_handle_dirty_metadata(handle, inode, this_bh);
3799 ext4_error(inode->i_sb, __func__,
3800 "circular indirect block detected, "
3801 "inode=%lu, block=%llu",
3803 (unsigned long long) this_bh->b_blocknr);
3808 * ext4_free_branches - free an array of branches
3809 * @handle: JBD handle for this transaction
3810 * @inode: inode we are dealing with
3811 * @parent_bh: the buffer_head which contains *@first and *@last
3812 * @first: array of block numbers
3813 * @last: pointer immediately past the end of array
3814 * @depth: depth of the branches to free
3816 * We are freeing all blocks refered from these branches (numbers are
3817 * stored as little-endian 32-bit) and updating @inode->i_blocks
3820 static void ext4_free_branches(handle_t *handle, struct inode *inode,
3821 struct buffer_head *parent_bh,
3822 __le32 *first, __le32 *last, int depth)
3827 if (ext4_handle_is_aborted(handle))
3831 struct buffer_head *bh;
3832 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3834 while (--p >= first) {
3835 nr = le32_to_cpu(*p);
3837 continue; /* A hole */
3839 /* Go read the buffer for the next level down */
3840 bh = sb_bread(inode->i_sb, nr);
3843 * A read failure? Report error and clear slot
3847 ext4_error(inode->i_sb, "ext4_free_branches",
3848 "Read failure, inode=%lu, block=%llu",
3853 /* This zaps the entire block. Bottom up. */
3854 BUFFER_TRACE(bh, "free child branches");
3855 ext4_free_branches(handle, inode, bh,
3856 (__le32 *) bh->b_data,
3857 (__le32 *) bh->b_data + addr_per_block,
3861 * We've probably journalled the indirect block several
3862 * times during the truncate. But it's no longer
3863 * needed and we now drop it from the transaction via
3864 * jbd2_journal_revoke().
3866 * That's easy if it's exclusively part of this
3867 * transaction. But if it's part of the committing
3868 * transaction then jbd2_journal_forget() will simply
3869 * brelse() it. That means that if the underlying
3870 * block is reallocated in ext4_get_block(),
3871 * unmap_underlying_metadata() will find this block
3872 * and will try to get rid of it. damn, damn.
3874 * If this block has already been committed to the
3875 * journal, a revoke record will be written. And
3876 * revoke records must be emitted *before* clearing
3877 * this block's bit in the bitmaps.
3879 ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
3882 * Everything below this this pointer has been
3883 * released. Now let this top-of-subtree go.
3885 * We want the freeing of this indirect block to be
3886 * atomic in the journal with the updating of the
3887 * bitmap block which owns it. So make some room in
3890 * We zero the parent pointer *after* freeing its
3891 * pointee in the bitmaps, so if extend_transaction()
3892 * for some reason fails to put the bitmap changes and
3893 * the release into the same transaction, recovery
3894 * will merely complain about releasing a free block,
3895 * rather than leaking blocks.
3897 if (ext4_handle_is_aborted(handle))
3899 if (try_to_extend_transaction(handle, inode)) {
3900 ext4_mark_inode_dirty(handle, inode);
3901 ext4_journal_test_restart(handle, inode);
3904 ext4_free_blocks(handle, inode, nr, 1, 1);
3908 * The block which we have just freed is
3909 * pointed to by an indirect block: journal it
3911 BUFFER_TRACE(parent_bh, "get_write_access");
3912 if (!ext4_journal_get_write_access(handle,
3915 BUFFER_TRACE(parent_bh,
3916 "call ext4_handle_dirty_metadata");
3917 ext4_handle_dirty_metadata(handle,
3924 /* We have reached the bottom of the tree. */
3925 BUFFER_TRACE(parent_bh, "free data blocks");
3926 ext4_free_data(handle, inode, parent_bh, first, last);
3930 int ext4_can_truncate(struct inode *inode)
3932 if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
3934 if (S_ISREG(inode->i_mode))
3936 if (S_ISDIR(inode->i_mode))
3938 if (S_ISLNK(inode->i_mode))
3939 return !ext4_inode_is_fast_symlink(inode);
3946 * We block out ext4_get_block() block instantiations across the entire
3947 * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3948 * simultaneously on behalf of the same inode.
3950 * As we work through the truncate and commmit bits of it to the journal there
3951 * is one core, guiding principle: the file's tree must always be consistent on
3952 * disk. We must be able to restart the truncate after a crash.
3954 * The file's tree may be transiently inconsistent in memory (although it
3955 * probably isn't), but whenever we close off and commit a journal transaction,
3956 * the contents of (the filesystem + the journal) must be consistent and
3957 * restartable. It's pretty simple, really: bottom up, right to left (although
3958 * left-to-right works OK too).
3960 * Note that at recovery time, journal replay occurs *before* the restart of
3961 * truncate against the orphan inode list.
3963 * The committed inode has the new, desired i_size (which is the same as
3964 * i_disksize in this case). After a crash, ext4_orphan_cleanup() will see
3965 * that this inode's truncate did not complete and it will again call
3966 * ext4_truncate() to have another go. So there will be instantiated blocks
3967 * to the right of the truncation point in a crashed ext4 filesystem. But
3968 * that's fine - as long as they are linked from the inode, the post-crash
3969 * ext4_truncate() run will find them and release them.
3971 void ext4_truncate(struct inode *inode)
3974 struct ext4_inode_info *ei = EXT4_I(inode);
3975 __le32 *i_data = ei->i_data;
3976 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
3977 struct address_space *mapping = inode->i_mapping;
3978 ext4_lblk_t offsets[4];
3983 ext4_lblk_t last_block;
3984 unsigned blocksize = inode->i_sb->s_blocksize;
3986 if (!ext4_can_truncate(inode))
3989 if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
3990 ei->i_state |= EXT4_STATE_DA_ALLOC_CLOSE;
3992 if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
3993 ext4_ext_truncate(inode);
3997 handle = start_transaction(inode);
3999 return; /* AKPM: return what? */
4001 last_block = (inode->i_size + blocksize-1)
4002 >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4004 if (inode->i_size & (blocksize - 1))
4005 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4008 n = ext4_block_to_path(inode, last_block, offsets, NULL);
4010 goto out_stop; /* error */
4013 * OK. This truncate is going to happen. We add the inode to the
4014 * orphan list, so that if this truncate spans multiple transactions,
4015 * and we crash, we will resume the truncate when the filesystem
4016 * recovers. It also marks the inode dirty, to catch the new size.
4018 * Implication: the file must always be in a sane, consistent
4019 * truncatable state while each transaction commits.
4021 if (ext4_orphan_add(handle, inode))
4025 * From here we block out all ext4_get_block() callers who want to
4026 * modify the block allocation tree.
4028 down_write(&ei->i_data_sem);
4030 ext4_discard_preallocations(inode);
4033 * The orphan list entry will now protect us from any crash which
4034 * occurs before the truncate completes, so it is now safe to propagate
4035 * the new, shorter inode size (held for now in i_size) into the
4036 * on-disk inode. We do this via i_disksize, which is the value which
4037 * ext4 *really* writes onto the disk inode.
4039 ei->i_disksize = inode->i_size;
4041 if (n == 1) { /* direct blocks */
4042 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4043 i_data + EXT4_NDIR_BLOCKS);
4047 partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4048 /* Kill the top of shared branch (not detached) */
4050 if (partial == chain) {
4051 /* Shared branch grows from the inode */
4052 ext4_free_branches(handle, inode, NULL,
4053 &nr, &nr+1, (chain+n-1) - partial);
4056 * We mark the inode dirty prior to restart,
4057 * and prior to stop. No need for it here.
4060 /* Shared branch grows from an indirect block */
4061 BUFFER_TRACE(partial->bh, "get_write_access");
4062 ext4_free_branches(handle, inode, partial->bh,
4064 partial->p+1, (chain+n-1) - partial);
4067 /* Clear the ends of indirect blocks on the shared branch */
4068 while (partial > chain) {
4069 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4070 (__le32*)partial->bh->b_data+addr_per_block,
4071 (chain+n-1) - partial);
4072 BUFFER_TRACE(partial->bh, "call brelse");
4073 brelse (partial->bh);
4077 /* Kill the remaining (whole) subtrees */
4078 switch (offsets[0]) {
4080 nr = i_data[EXT4_IND_BLOCK];
4082 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4083 i_data[EXT4_IND_BLOCK] = 0;
4085 case EXT4_IND_BLOCK:
4086 nr = i_data[EXT4_DIND_BLOCK];
4088 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4089 i_data[EXT4_DIND_BLOCK] = 0;
4091 case EXT4_DIND_BLOCK:
4092 nr = i_data[EXT4_TIND_BLOCK];
4094 ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4095 i_data[EXT4_TIND_BLOCK] = 0;
4097 case EXT4_TIND_BLOCK:
4101 up_write(&ei->i_data_sem);
4102 inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4103 ext4_mark_inode_dirty(handle, inode);
4106 * In a multi-transaction truncate, we only make the final transaction
4110 ext4_handle_sync(handle);
4113 * If this was a simple ftruncate(), and the file will remain alive
4114 * then we need to clear up the orphan record which we created above.
4115 * However, if this was a real unlink then we were called by
4116 * ext4_delete_inode(), and we allow that function to clean up the
4117 * orphan info for us.
4120 ext4_orphan_del(handle, inode);
4122 ext4_journal_stop(handle);
4126 * ext4_get_inode_loc returns with an extra refcount against the inode's
4127 * underlying buffer_head on success. If 'in_mem' is true, we have all
4128 * data in memory that is needed to recreate the on-disk version of this
4131 static int __ext4_get_inode_loc(struct inode *inode,
4132 struct ext4_iloc *iloc, int in_mem)
4134 struct ext4_group_desc *gdp;
4135 struct buffer_head *bh;
4136 struct super_block *sb = inode->i_sb;
4138 int inodes_per_block, inode_offset;
4141 if (!ext4_valid_inum(sb, inode->i_ino))
4144 iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4145 gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4150 * Figure out the offset within the block group inode table
4152 inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4153 inode_offset = ((inode->i_ino - 1) %
4154 EXT4_INODES_PER_GROUP(sb));
4155 block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4156 iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4158 bh = sb_getblk(sb, block);
4160 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4161 "inode block - inode=%lu, block=%llu",
4162 inode->i_ino, block);
4165 if (!buffer_uptodate(bh)) {
4169 * If the buffer has the write error flag, we have failed
4170 * to write out another inode in the same block. In this
4171 * case, we don't have to read the block because we may
4172 * read the old inode data successfully.
4174 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4175 set_buffer_uptodate(bh);
4177 if (buffer_uptodate(bh)) {
4178 /* someone brought it uptodate while we waited */
4184 * If we have all information of the inode in memory and this
4185 * is the only valid inode in the block, we need not read the
4189 struct buffer_head *bitmap_bh;
4192 start = inode_offset & ~(inodes_per_block - 1);
4194 /* Is the inode bitmap in cache? */
4195 bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4200 * If the inode bitmap isn't in cache then the
4201 * optimisation may end up performing two reads instead
4202 * of one, so skip it.
4204 if (!buffer_uptodate(bitmap_bh)) {
4208 for (i = start; i < start + inodes_per_block; i++) {
4209 if (i == inode_offset)
4211 if (ext4_test_bit(i, bitmap_bh->b_data))
4215 if (i == start + inodes_per_block) {
4216 /* all other inodes are free, so skip I/O */
4217 memset(bh->b_data, 0, bh->b_size);
4218 set_buffer_uptodate(bh);
4226 * If we need to do any I/O, try to pre-readahead extra
4227 * blocks from the inode table.
4229 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4230 ext4_fsblk_t b, end, table;
4233 table = ext4_inode_table(sb, gdp);
4234 /* s_inode_readahead_blks is always a power of 2 */
4235 b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4238 end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4239 num = EXT4_INODES_PER_GROUP(sb);
4240 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4241 EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4242 num -= ext4_itable_unused_count(sb, gdp);
4243 table += num / inodes_per_block;
4247 sb_breadahead(sb, b++);
4251 * There are other valid inodes in the buffer, this inode
4252 * has in-inode xattrs, or we don't have this inode in memory.
4253 * Read the block from disk.
4256 bh->b_end_io = end_buffer_read_sync;
4257 submit_bh(READ_META, bh);
4259 if (!buffer_uptodate(bh)) {
4260 ext4_error(sb, __func__,
4261 "unable to read inode block - inode=%lu, "
4262 "block=%llu", inode->i_ino, block);
4272 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4274 /* We have all inode data except xattrs in memory here. */
4275 return __ext4_get_inode_loc(inode, iloc,
4276 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4279 void ext4_set_inode_flags(struct inode *inode)
4281 unsigned int flags = EXT4_I(inode)->i_flags;
4283 inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4284 if (flags & EXT4_SYNC_FL)
4285 inode->i_flags |= S_SYNC;
4286 if (flags & EXT4_APPEND_FL)
4287 inode->i_flags |= S_APPEND;
4288 if (flags & EXT4_IMMUTABLE_FL)
4289 inode->i_flags |= S_IMMUTABLE;
4290 if (flags & EXT4_NOATIME_FL)
4291 inode->i_flags |= S_NOATIME;
4292 if (flags & EXT4_DIRSYNC_FL)
4293 inode->i_flags |= S_DIRSYNC;
4296 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4297 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4299 unsigned int flags = ei->vfs_inode.i_flags;
4301 ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4302 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4304 ei->i_flags |= EXT4_SYNC_FL;
4305 if (flags & S_APPEND)
4306 ei->i_flags |= EXT4_APPEND_FL;
4307 if (flags & S_IMMUTABLE)
4308 ei->i_flags |= EXT4_IMMUTABLE_FL;
4309 if (flags & S_NOATIME)
4310 ei->i_flags |= EXT4_NOATIME_FL;
4311 if (flags & S_DIRSYNC)
4312 ei->i_flags |= EXT4_DIRSYNC_FL;
4314 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4315 struct ext4_inode_info *ei)
4318 struct inode *inode = &(ei->vfs_inode);
4319 struct super_block *sb = inode->i_sb;
4321 if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4322 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4323 /* we are using combined 48 bit field */
4324 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4325 le32_to_cpu(raw_inode->i_blocks_lo);
4326 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4327 /* i_blocks represent file system block size */
4328 return i_blocks << (inode->i_blkbits - 9);
4333 return le32_to_cpu(raw_inode->i_blocks_lo);
4337 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4339 struct ext4_iloc iloc;
4340 struct ext4_inode *raw_inode;
4341 struct ext4_inode_info *ei;
4342 struct buffer_head *bh;
4343 struct inode *inode;
4347 inode = iget_locked(sb, ino);
4349 return ERR_PTR(-ENOMEM);
4350 if (!(inode->i_state & I_NEW))
4354 #ifdef CONFIG_EXT4_FS_POSIX_ACL
4355 ei->i_acl = EXT4_ACL_NOT_CACHED;
4356 ei->i_default_acl = EXT4_ACL_NOT_CACHED;
4359 ret = __ext4_get_inode_loc(inode, &iloc, 0);
4363 raw_inode = ext4_raw_inode(&iloc);
4364 inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4365 inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4366 inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4367 if (!(test_opt(inode->i_sb, NO_UID32))) {
4368 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4369 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4371 inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4374 ei->i_dir_start_lookup = 0;
4375 ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4376 /* We now have enough fields to check if the inode was active or not.
4377 * This is needed because nfsd might try to access dead inodes
4378 * the test is that same one that e2fsck uses
4379 * NeilBrown 1999oct15
4381 if (inode->i_nlink == 0) {
4382 if (inode->i_mode == 0 ||
4383 !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4384 /* this inode is deleted */
4389 /* The only unlinked inodes we let through here have
4390 * valid i_mode and are being read by the orphan
4391 * recovery code: that's fine, we're about to complete
4392 * the process of deleting those. */
4394 ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4395 inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4396 ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4397 if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4399 ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4400 inode->i_size = ext4_isize(raw_inode);
4401 ei->i_disksize = inode->i_size;
4402 inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4403 ei->i_block_group = iloc.block_group;
4404 ei->i_last_alloc_group = ~0;
4406 * NOTE! The in-memory inode i_data array is in little-endian order
4407 * even on big-endian machines: we do NOT byteswap the block numbers!
4409 for (block = 0; block < EXT4_N_BLOCKS; block++)
4410 ei->i_data[block] = raw_inode->i_block[block];
4411 INIT_LIST_HEAD(&ei->i_orphan);
4413 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4414 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4415 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4416 EXT4_INODE_SIZE(inode->i_sb)) {
4421 if (ei->i_extra_isize == 0) {
4422 /* The extra space is currently unused. Use it. */
4423 ei->i_extra_isize = sizeof(struct ext4_inode) -
4424 EXT4_GOOD_OLD_INODE_SIZE;
4426 __le32 *magic = (void *)raw_inode +
4427 EXT4_GOOD_OLD_INODE_SIZE +
4429 if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4430 ei->i_state |= EXT4_STATE_XATTR;
4433 ei->i_extra_isize = 0;
4435 EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4436 EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4437 EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4438 EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4440 inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4441 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4442 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4444 (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4448 if (ei->i_file_acl &&
4450 (le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block) +
4451 EXT4_SB(sb)->s_gdb_count)) ||
4452 (ei->i_file_acl >= ext4_blocks_count(EXT4_SB(sb)->s_es)))) {
4453 ext4_error(sb, __func__,
4454 "bad extended attribute block %llu in inode #%lu",
4455 ei->i_file_acl, inode->i_ino);
4458 } else if (ei->i_flags & EXT4_EXTENTS_FL) {
4459 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4460 (S_ISLNK(inode->i_mode) &&
4461 !ext4_inode_is_fast_symlink(inode)))
4462 /* Validate extent which is part of inode */
4463 ret = ext4_ext_check_inode(inode);
4464 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4465 (S_ISLNK(inode->i_mode) &&
4466 !ext4_inode_is_fast_symlink(inode))) {
4467 /* Validate block references which are part of inode */
4468 ret = ext4_check_inode_blockref(inode);
4475 if (S_ISREG(inode->i_mode)) {
4476 inode->i_op = &ext4_file_inode_operations;
4477 inode->i_fop = &ext4_file_operations;
4478 ext4_set_aops(inode);
4479 } else if (S_ISDIR(inode->i_mode)) {
4480 inode->i_op = &ext4_dir_inode_operations;
4481 inode->i_fop = &ext4_dir_operations;
4482 } else if (S_ISLNK(inode->i_mode)) {
4483 if (ext4_inode_is_fast_symlink(inode)) {
4484 inode->i_op = &ext4_fast_symlink_inode_operations;
4485 nd_terminate_link(ei->i_data, inode->i_size,
4486 sizeof(ei->i_data) - 1);
4488 inode->i_op = &ext4_symlink_inode_operations;
4489 ext4_set_aops(inode);
4491 } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4492 S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4493 inode->i_op = &ext4_special_inode_operations;
4494 if (raw_inode->i_block[0])
4495 init_special_inode(inode, inode->i_mode,
4496 old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4498 init_special_inode(inode, inode->i_mode,
4499 new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4503 ext4_error(inode->i_sb, __func__,
4504 "bogus i_mode (%o) for inode=%lu",
4505 inode->i_mode, inode->i_ino);
4509 ext4_set_inode_flags(inode);
4510 unlock_new_inode(inode);
4515 return ERR_PTR(ret);
4518 static int ext4_inode_blocks_set(handle_t *handle,
4519 struct ext4_inode *raw_inode,
4520 struct ext4_inode_info *ei)
4522 struct inode *inode = &(ei->vfs_inode);
4523 u64 i_blocks = inode->i_blocks;
4524 struct super_block *sb = inode->i_sb;
4526 if (i_blocks <= ~0U) {
4528 * i_blocks can be represnted in a 32 bit variable
4529 * as multiple of 512 bytes
4531 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4532 raw_inode->i_blocks_high = 0;
4533 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4536 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4539 if (i_blocks <= 0xffffffffffffULL) {
4541 * i_blocks can be represented in a 48 bit variable
4542 * as multiple of 512 bytes
4544 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4545 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4546 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4548 ei->i_flags |= EXT4_HUGE_FILE_FL;
4549 /* i_block is stored in file system block size */
4550 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4551 raw_inode->i_blocks_lo = cpu_to_le32(i_blocks);
4552 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4558 * Post the struct inode info into an on-disk inode location in the
4559 * buffer-cache. This gobbles the caller's reference to the
4560 * buffer_head in the inode location struct.
4562 * The caller must have write access to iloc->bh.
4564 static int ext4_do_update_inode(handle_t *handle,
4565 struct inode *inode,
4566 struct ext4_iloc *iloc)
4568 struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4569 struct ext4_inode_info *ei = EXT4_I(inode);
4570 struct buffer_head *bh = iloc->bh;
4571 int err = 0, rc, block;
4573 /* For fields not not tracking in the in-memory inode,
4574 * initialise them to zero for new inodes. */
4575 if (ei->i_state & EXT4_STATE_NEW)
4576 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4578 ext4_get_inode_flags(ei);
4579 raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4580 if (!(test_opt(inode->i_sb, NO_UID32))) {
4581 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
4582 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
4584 * Fix up interoperability with old kernels. Otherwise, old inodes get
4585 * re-used with the upper 16 bits of the uid/gid intact
4588 raw_inode->i_uid_high =
4589 cpu_to_le16(high_16_bits(inode->i_uid));
4590 raw_inode->i_gid_high =
4591 cpu_to_le16(high_16_bits(inode->i_gid));
4593 raw_inode->i_uid_high = 0;
4594 raw_inode->i_gid_high = 0;
4597 raw_inode->i_uid_low =
4598 cpu_to_le16(fs_high2lowuid(inode->i_uid));
4599 raw_inode->i_gid_low =
4600 cpu_to_le16(fs_high2lowgid(inode->i_gid));
4601 raw_inode->i_uid_high = 0;
4602 raw_inode->i_gid_high = 0;
4604 raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4606 EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4607 EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4608 EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4609 EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4611 if (ext4_inode_blocks_set(handle, raw_inode, ei))
4613 raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4614 /* clear the migrate flag in the raw_inode */
4615 raw_inode->i_flags = cpu_to_le32(ei->i_flags & ~EXT4_EXT_MIGRATE);
4616 if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
4617 cpu_to_le32(EXT4_OS_HURD))
4618 raw_inode->i_file_acl_high =
4619 cpu_to_le16(ei->i_file_acl >> 32);
4620 raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4621 ext4_isize_set(raw_inode, ei->i_disksize);
4622 if (ei->i_disksize > 0x7fffffffULL) {
4623 struct super_block *sb = inode->i_sb;
4624 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4625 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4626 EXT4_SB(sb)->s_es->s_rev_level ==
4627 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
4628 /* If this is the first large file
4629 * created, add a flag to the superblock.
4631 err = ext4_journal_get_write_access(handle,
4632 EXT4_SB(sb)->s_sbh);
4635 ext4_update_dynamic_rev(sb);
4636 EXT4_SET_RO_COMPAT_FEATURE(sb,
4637 EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4639 ext4_handle_sync(handle);
4640 err = ext4_handle_dirty_metadata(handle, inode,
4641 EXT4_SB(sb)->s_sbh);
4644 raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4645 if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4646 if (old_valid_dev(inode->i_rdev)) {
4647 raw_inode->i_block[0] =
4648 cpu_to_le32(old_encode_dev(inode->i_rdev));
4649 raw_inode->i_block[1] = 0;
4651 raw_inode->i_block[0] = 0;
4652 raw_inode->i_block[1] =
4653 cpu_to_le32(new_encode_dev(inode->i_rdev));
4654 raw_inode->i_block[2] = 0;
4656 } else for (block = 0; block < EXT4_N_BLOCKS; block++)
4657 raw_inode->i_block[block] = ei->i_data[block];
4659 raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4660 if (ei->i_extra_isize) {
4661 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4662 raw_inode->i_version_hi =
4663 cpu_to_le32(inode->i_version >> 32);
4664 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
4667 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4668 rc = ext4_handle_dirty_metadata(handle, inode, bh);
4671 ei->i_state &= ~EXT4_STATE_NEW;
4675 ext4_std_error(inode->i_sb, err);
4680 * ext4_write_inode()
4682 * We are called from a few places:
4684 * - Within generic_file_write() for O_SYNC files.
4685 * Here, there will be no transaction running. We wait for any running
4686 * trasnaction to commit.
4688 * - Within sys_sync(), kupdate and such.
4689 * We wait on commit, if tol to.
4691 * - Within prune_icache() (PF_MEMALLOC == true)
4692 * Here we simply return. We can't afford to block kswapd on the
4695 * In all cases it is actually safe for us to return without doing anything,
4696 * because the inode has been copied into a raw inode buffer in
4697 * ext4_mark_inode_dirty(). This is a correctness thing for O_SYNC and for
4700 * Note that we are absolutely dependent upon all inode dirtiers doing the
4701 * right thing: they *must* call mark_inode_dirty() after dirtying info in
4702 * which we are interested.
4704 * It would be a bug for them to not do this. The code:
4706 * mark_inode_dirty(inode)
4708 * inode->i_size = expr;
4710 * is in error because a kswapd-driven write_inode() could occur while
4711 * `stuff()' is running, and the new i_size will be lost. Plus the inode
4712 * will no longer be on the superblock's dirty inode list.
4714 int ext4_write_inode(struct inode *inode, int wait)
4716 if (current->flags & PF_MEMALLOC)
4719 if (ext4_journal_current_handle()) {
4720 jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4728 return ext4_force_commit(inode->i_sb);
4731 int __ext4_write_dirty_metadata(struct inode *inode, struct buffer_head *bh)
4735 mark_buffer_dirty(bh);
4736 if (inode && inode_needs_sync(inode)) {
4737 sync_dirty_buffer(bh);
4738 if (buffer_req(bh) && !buffer_uptodate(bh)) {
4739 ext4_error(inode->i_sb, __func__,
4740 "IO error syncing inode, "
4741 "inode=%lu, block=%llu",
4743 (unsigned long long)bh->b_blocknr);
4753 * Called from notify_change.
4755 * We want to trap VFS attempts to truncate the file as soon as
4756 * possible. In particular, we want to make sure that when the VFS
4757 * shrinks i_size, we put the inode on the orphan list and modify
4758 * i_disksize immediately, so that during the subsequent flushing of
4759 * dirty pages and freeing of disk blocks, we can guarantee that any
4760 * commit will leave the blocks being flushed in an unused state on
4761 * disk. (On recovery, the inode will get truncated and the blocks will
4762 * be freed, so we have a strong guarantee that no future commit will
4763 * leave these blocks visible to the user.)
4765 * Another thing we have to assure is that if we are in ordered mode
4766 * and inode is still attached to the committing transaction, we must
4767 * we start writeout of all the dirty pages which are being truncated.
4768 * This way we are sure that all the data written in the previous
4769 * transaction are already on disk (truncate waits for pages under
4772 * Called with inode->i_mutex down.
4774 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4776 struct inode *inode = dentry->d_inode;
4778 const unsigned int ia_valid = attr->ia_valid;
4780 error = inode_change_ok(inode, attr);
4784 if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
4785 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
4788 /* (user+group)*(old+new) structure, inode write (sb,
4789 * inode block, ? - but truncate inode update has it) */
4790 handle = ext4_journal_start(inode, 2*(EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)+
4791 EXT4_QUOTA_DEL_BLOCKS(inode->i_sb))+3);
4792 if (IS_ERR(handle)) {
4793 error = PTR_ERR(handle);
4796 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
4798 ext4_journal_stop(handle);
4801 /* Update corresponding info in inode so that everything is in
4802 * one transaction */
4803 if (attr->ia_valid & ATTR_UID)
4804 inode->i_uid = attr->ia_uid;
4805 if (attr->ia_valid & ATTR_GID)
4806 inode->i_gid = attr->ia_gid;
4807 error = ext4_mark_inode_dirty(handle, inode);
4808 ext4_journal_stop(handle);
4811 if (attr->ia_valid & ATTR_SIZE) {
4812 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
4813 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4815 if (attr->ia_size > sbi->s_bitmap_maxbytes) {
4822 if (S_ISREG(inode->i_mode) &&
4823 attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
4826 handle = ext4_journal_start(inode, 3);
4827 if (IS_ERR(handle)) {
4828 error = PTR_ERR(handle);
4832 error = ext4_orphan_add(handle, inode);
4833 EXT4_I(inode)->i_disksize = attr->ia_size;
4834 rc = ext4_mark_inode_dirty(handle, inode);
4837 ext4_journal_stop(handle);
4839 if (ext4_should_order_data(inode)) {
4840 error = ext4_begin_ordered_truncate(inode,
4843 /* Do as much error cleanup as possible */
4844 handle = ext4_journal_start(inode, 3);
4845 if (IS_ERR(handle)) {
4846 ext4_orphan_del(NULL, inode);
4849 ext4_orphan_del(handle, inode);
4850 ext4_journal_stop(handle);
4856 rc = inode_setattr(inode, attr);
4858 /* If inode_setattr's call to ext4_truncate failed to get a
4859 * transaction handle at all, we need to clean up the in-core
4860 * orphan list manually. */
4862 ext4_orphan_del(NULL, inode);
4864 if (!rc && (ia_valid & ATTR_MODE))
4865 rc = ext4_acl_chmod(inode);
4868 ext4_std_error(inode->i_sb, error);
4874 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4877 struct inode *inode;
4878 unsigned long delalloc_blocks;
4880 inode = dentry->d_inode;
4881 generic_fillattr(inode, stat);
4884 * We can't update i_blocks if the block allocation is delayed
4885 * otherwise in the case of system crash before the real block
4886 * allocation is done, we will have i_blocks inconsistent with
4887 * on-disk file blocks.
4888 * We always keep i_blocks updated together with real
4889 * allocation. But to not confuse with user, stat
4890 * will return the blocks that include the delayed allocation
4891 * blocks for this file.
4893 spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
4894 delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
4895 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
4897 stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4901 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
4906 /* if nrblocks are contiguous */
4909 * With N contiguous data blocks, it need at most
4910 * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
4911 * 2 dindirect blocks
4914 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
4915 return indirects + 3;
4918 * if nrblocks are not contiguous, worse case, each block touch
4919 * a indirect block, and each indirect block touch a double indirect
4920 * block, plus a triple indirect block
4922 indirects = nrblocks * 2 + 1;
4926 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4928 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
4929 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
4930 return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
4934 * Account for index blocks, block groups bitmaps and block group
4935 * descriptor blocks if modify datablocks and index blocks
4936 * worse case, the indexs blocks spread over different block groups
4938 * If datablocks are discontiguous, they are possible to spread over
4939 * different block groups too. If they are contiugous, with flexbg,
4940 * they could still across block group boundary.
4942 * Also account for superblock, inode, quota and xattr blocks
4944 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4946 int groups, gdpblocks;
4951 * How many index blocks need to touch to modify nrblocks?
4952 * The "Chunk" flag indicating whether the nrblocks is
4953 * physically contiguous on disk
4955 * For Direct IO and fallocate, they calls get_block to allocate
4956 * one single extent at a time, so they could set the "Chunk" flag
4958 idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4963 * Now let's see how many group bitmaps and group descriptors need
4973 if (groups > EXT4_SB(inode->i_sb)->s_groups_count)
4974 groups = EXT4_SB(inode->i_sb)->s_groups_count;
4975 if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4976 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4978 /* bitmaps and block group descriptor blocks */
4979 ret += groups + gdpblocks;
4981 /* Blocks for super block, inode, quota and xattr blocks */
4982 ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4988 * Calulate the total number of credits to reserve to fit
4989 * the modification of a single pages into a single transaction,
4990 * which may include multiple chunks of block allocations.
4992 * This could be called via ext4_write_begin()
4994 * We need to consider the worse case, when
4995 * one new block per extent.
4997 int ext4_writepage_trans_blocks(struct inode *inode)
4999 int bpp = ext4_journal_blocks_per_page(inode);
5002 ret = ext4_meta_trans_blocks(inode, bpp, 0);
5004 /* Account for data blocks for journalled mode */
5005 if (ext4_should_journal_data(inode))
5011 * Calculate the journal credits for a chunk of data modification.
5013 * This is called from DIO, fallocate or whoever calling
5014 * ext4_get_blocks_wrap() to map/allocate a chunk of contigous disk blocks.
5016 * journal buffers for data blocks are not included here, as DIO
5017 * and fallocate do no need to journal data buffers.
5019 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5021 return ext4_meta_trans_blocks(inode, nrblocks, 1);
5025 * The caller must have previously called ext4_reserve_inode_write().
5026 * Give this, we know that the caller already has write access to iloc->bh.
5028 int ext4_mark_iloc_dirty(handle_t *handle,
5029 struct inode *inode, struct ext4_iloc *iloc)
5033 if (test_opt(inode->i_sb, I_VERSION))
5034 inode_inc_iversion(inode);
5036 /* the do_update_inode consumes one bh->b_count */
5039 /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5040 err = ext4_do_update_inode(handle, inode, iloc);
5046 * On success, We end up with an outstanding reference count against
5047 * iloc->bh. This _must_ be cleaned up later.
5051 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5052 struct ext4_iloc *iloc)
5056 err = ext4_get_inode_loc(inode, iloc);
5058 BUFFER_TRACE(iloc->bh, "get_write_access");
5059 err = ext4_journal_get_write_access(handle, iloc->bh);
5065 ext4_std_error(inode->i_sb, err);
5070 * Expand an inode by new_extra_isize bytes.
5071 * Returns 0 on success or negative error number on failure.
5073 static int ext4_expand_extra_isize(struct inode *inode,
5074 unsigned int new_extra_isize,
5075 struct ext4_iloc iloc,
5078 struct ext4_inode *raw_inode;
5079 struct ext4_xattr_ibody_header *header;
5080 struct ext4_xattr_entry *entry;
5082 if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5085 raw_inode = ext4_raw_inode(&iloc);
5087 header = IHDR(inode, raw_inode);
5088 entry = IFIRST(header);
5090 /* No extended attributes present */
5091 if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
5092 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5093 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5095 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5099 /* try to expand with EAs present */
5100 return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5105 * What we do here is to mark the in-core inode as clean with respect to inode
5106 * dirtiness (it may still be data-dirty).
5107 * This means that the in-core inode may be reaped by prune_icache
5108 * without having to perform any I/O. This is a very good thing,
5109 * because *any* task may call prune_icache - even ones which
5110 * have a transaction open against a different journal.
5112 * Is this cheating? Not really. Sure, we haven't written the
5113 * inode out, but prune_icache isn't a user-visible syncing function.
5114 * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5115 * we start and wait on commits.
5117 * Is this efficient/effective? Well, we're being nice to the system
5118 * by cleaning up our inodes proactively so they can be reaped
5119 * without I/O. But we are potentially leaving up to five seconds'
5120 * worth of inodes floating about which prune_icache wants us to
5121 * write out. One way to fix that would be to get prune_icache()
5122 * to do a write_super() to free up some memory. It has the desired
5125 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5127 struct ext4_iloc iloc;
5128 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5129 static unsigned int mnt_count;
5133 err = ext4_reserve_inode_write(handle, inode, &iloc);
5134 if (ext4_handle_valid(handle) &&
5135 EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
5136 !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
5138 * We need extra buffer credits since we may write into EA block
5139 * with this same handle. If journal_extend fails, then it will
5140 * only result in a minor loss of functionality for that inode.
5141 * If this is felt to be critical, then e2fsck should be run to
5142 * force a large enough s_min_extra_isize.
5144 if ((jbd2_journal_extend(handle,
5145 EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
5146 ret = ext4_expand_extra_isize(inode,
5147 sbi->s_want_extra_isize,
5150 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
5152 le16_to_cpu(sbi->s_es->s_mnt_count)) {
5153 ext4_warning(inode->i_sb, __func__,
5154 "Unable to expand inode %lu. Delete"
5155 " some EAs or run e2fsck.",
5158 le16_to_cpu(sbi->s_es->s_mnt_count);
5164 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
5169 * ext4_dirty_inode() is called from __mark_inode_dirty()
5171 * We're really interested in the case where a file is being extended.
5172 * i_size has been changed by generic_commit_write() and we thus need
5173 * to include the updated inode in the current transaction.
5175 * Also, vfs_dq_alloc_block() will always dirty the inode when blocks
5176 * are allocated to the file.
5178 * If the inode is marked synchronous, we don't honour that here - doing
5179 * so would cause a commit on atime updates, which we don't bother doing.
5180 * We handle synchronous inodes at the highest possible level.
5182 void ext4_dirty_inode(struct inode *inode)
5184 handle_t *current_handle = ext4_journal_current_handle();
5187 if (!ext4_handle_valid(current_handle)) {
5188 ext4_mark_inode_dirty(current_handle, inode);
5192 handle = ext4_journal_start(inode, 2);
5195 if (current_handle &&
5196 current_handle->h_transaction != handle->h_transaction) {
5197 /* This task has a transaction open against a different fs */
5198 printk(KERN_EMERG "%s: transactions do not match!\n",
5201 jbd_debug(5, "marking dirty. outer handle=%p\n",
5203 ext4_mark_inode_dirty(handle, inode);
5205 ext4_journal_stop(handle);
5212 * Bind an inode's backing buffer_head into this transaction, to prevent
5213 * it from being flushed to disk early. Unlike
5214 * ext4_reserve_inode_write, this leaves behind no bh reference and
5215 * returns no iloc structure, so the caller needs to repeat the iloc
5216 * lookup to mark the inode dirty later.
5218 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5220 struct ext4_iloc iloc;
5224 err = ext4_get_inode_loc(inode, &iloc);
5226 BUFFER_TRACE(iloc.bh, "get_write_access");
5227 err = jbd2_journal_get_write_access(handle, iloc.bh);
5229 err = ext4_handle_dirty_metadata(handle,
5235 ext4_std_error(inode->i_sb, err);
5240 int ext4_change_inode_journal_flag(struct inode *inode, int val)
5247 * We have to be very careful here: changing a data block's
5248 * journaling status dynamically is dangerous. If we write a
5249 * data block to the journal, change the status and then delete
5250 * that block, we risk forgetting to revoke the old log record
5251 * from the journal and so a subsequent replay can corrupt data.
5252 * So, first we make sure that the journal is empty and that
5253 * nobody is changing anything.
5256 journal = EXT4_JOURNAL(inode);
5259 if (is_journal_aborted(journal))
5262 jbd2_journal_lock_updates(journal);
5263 jbd2_journal_flush(journal);
5266 * OK, there are no updates running now, and all cached data is
5267 * synced to disk. We are now in a completely consistent state
5268 * which doesn't have anything in the journal, and we know that
5269 * no filesystem updates are running, so it is safe to modify
5270 * the inode's in-core data-journaling state flag now.
5274 EXT4_I(inode)->i_flags |= EXT4_JOURNAL_DATA_FL;
5276 EXT4_I(inode)->i_flags &= ~EXT4_JOURNAL_DATA_FL;
5277 ext4_set_aops(inode);
5279 jbd2_journal_unlock_updates(journal);
5281 /* Finally we can mark the inode as dirty. */
5283 handle = ext4_journal_start(inode, 1);
5285 return PTR_ERR(handle);
5287 err = ext4_mark_inode_dirty(handle, inode);
5288 ext4_handle_sync(handle);
5289 ext4_journal_stop(handle);
5290 ext4_std_error(inode->i_sb, err);
5295 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5297 return !buffer_mapped(bh);
5300 int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
5302 struct page *page = vmf->page;
5307 struct file *file = vma->vm_file;
5308 struct inode *inode = file->f_path.dentry->d_inode;
5309 struct address_space *mapping = inode->i_mapping;
5312 * Get i_alloc_sem to stop truncates messing with the inode. We cannot
5313 * get i_mutex because we are already holding mmap_sem.
5315 down_read(&inode->i_alloc_sem);
5316 size = i_size_read(inode);
5317 if (page->mapping != mapping || size <= page_offset(page)
5318 || !PageUptodate(page)) {
5319 /* page got truncated from under us? */
5323 if (PageMappedToDisk(page))
5326 if (page->index == size >> PAGE_CACHE_SHIFT)
5327 len = size & ~PAGE_CACHE_MASK;
5329 len = PAGE_CACHE_SIZE;
5331 if (page_has_buffers(page)) {
5332 /* return if we have all the buffers mapped */
5333 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
5338 * OK, we need to fill the hole... Do write_begin write_end
5339 * to do block allocation/reservation.We are not holding
5340 * inode.i__mutex here. That allow * parallel write_begin,
5341 * write_end call. lock_page prevent this from happening
5342 * on the same page though
5344 ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
5345 len, AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
5348 ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
5349 len, len, page, fsdata);
5355 ret = VM_FAULT_SIGBUS;
5356 up_read(&inode->i_alloc_sem);