1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * File open, close, extend, truncate
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
26 #include <linux/capability.h>
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/highmem.h>
31 #include <linux/pagemap.h>
32 #include <linux/uio.h>
33 #include <linux/sched.h>
34 #include <linux/splice.h>
35 #include <linux/mount.h>
36 #include <linux/writeback.h>
38 #define MLOG_MASK_PREFIX ML_INODE
39 #include <cluster/masklog.h>
47 #include "extent_map.h"
57 #include "buffer_head_io.h"
59 static int ocfs2_sync_inode(struct inode
*inode
)
61 filemap_fdatawrite(inode
->i_mapping
);
62 return sync_mapping_buffers(inode
->i_mapping
);
65 static int ocfs2_file_open(struct inode
*inode
, struct file
*file
)
68 int mode
= file
->f_flags
;
69 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
71 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode
, file
,
72 file
->f_path
.dentry
->d_name
.len
, file
->f_path
.dentry
->d_name
.name
);
74 spin_lock(&oi
->ip_lock
);
76 /* Check that the inode hasn't been wiped from disk by another
77 * node. If it hasn't then we're safe as long as we hold the
78 * spin lock until our increment of open count. */
79 if (OCFS2_I(inode
)->ip_flags
& OCFS2_INODE_DELETED
) {
80 spin_unlock(&oi
->ip_lock
);
87 oi
->ip_flags
|= OCFS2_INODE_OPEN_DIRECT
;
90 spin_unlock(&oi
->ip_lock
);
97 static int ocfs2_file_release(struct inode
*inode
, struct file
*file
)
99 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
101 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode
, file
,
102 file
->f_path
.dentry
->d_name
.len
,
103 file
->f_path
.dentry
->d_name
.name
);
105 spin_lock(&oi
->ip_lock
);
106 if (!--oi
->ip_open_count
)
107 oi
->ip_flags
&= ~OCFS2_INODE_OPEN_DIRECT
;
108 spin_unlock(&oi
->ip_lock
);
115 static int ocfs2_sync_file(struct file
*file
,
116 struct dentry
*dentry
,
121 struct inode
*inode
= dentry
->d_inode
;
122 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
124 mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file
, dentry
, datasync
,
125 dentry
->d_name
.len
, dentry
->d_name
.name
);
127 err
= ocfs2_sync_inode(dentry
->d_inode
);
131 journal
= osb
->journal
->j_journal
;
132 err
= journal_force_commit(journal
);
137 return (err
< 0) ? -EIO
: 0;
140 int ocfs2_should_update_atime(struct inode
*inode
,
141 struct vfsmount
*vfsmnt
)
144 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
146 if (ocfs2_is_hard_readonly(osb
) || ocfs2_is_soft_readonly(osb
))
149 if ((inode
->i_flags
& S_NOATIME
) ||
150 ((inode
->i_sb
->s_flags
& MS_NODIRATIME
) && S_ISDIR(inode
->i_mode
)))
154 * We can be called with no vfsmnt structure - NFSD will
157 * Note that our action here is different than touch_atime() -
158 * if we can't tell whether this is a noatime mount, then we
159 * don't know whether to trust the value of s_atime_quantum.
164 if ((vfsmnt
->mnt_flags
& MNT_NOATIME
) ||
165 ((vfsmnt
->mnt_flags
& MNT_NODIRATIME
) && S_ISDIR(inode
->i_mode
)))
168 if (vfsmnt
->mnt_flags
& MNT_RELATIME
) {
169 if ((timespec_compare(&inode
->i_atime
, &inode
->i_mtime
) <= 0) ||
170 (timespec_compare(&inode
->i_atime
, &inode
->i_ctime
) <= 0))
177 if ((now
.tv_sec
- inode
->i_atime
.tv_sec
<= osb
->s_atime_quantum
))
183 int ocfs2_update_inode_atime(struct inode
*inode
,
184 struct buffer_head
*bh
)
187 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
192 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
193 if (handle
== NULL
) {
199 inode
->i_atime
= CURRENT_TIME
;
200 ret
= ocfs2_mark_inode_dirty(handle
, inode
, bh
);
204 ocfs2_commit_trans(OCFS2_SB(inode
->i_sb
), handle
);
210 static int ocfs2_set_inode_size(handle_t
*handle
,
212 struct buffer_head
*fe_bh
,
218 i_size_write(inode
, new_i_size
);
219 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
220 inode
->i_ctime
= inode
->i_mtime
= CURRENT_TIME
;
222 status
= ocfs2_mark_inode_dirty(handle
, inode
, fe_bh
);
233 static int ocfs2_simple_size_update(struct inode
*inode
,
234 struct buffer_head
*di_bh
,
238 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
239 handle_t
*handle
= NULL
;
241 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
242 if (handle
== NULL
) {
248 ret
= ocfs2_set_inode_size(handle
, inode
, di_bh
,
253 ocfs2_commit_trans(osb
, handle
);
258 static int ocfs2_orphan_for_truncate(struct ocfs2_super
*osb
,
260 struct buffer_head
*fe_bh
,
265 struct ocfs2_dinode
*di
;
269 /* TODO: This needs to actually orphan the inode in this
272 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
273 if (IS_ERR(handle
)) {
274 status
= PTR_ERR(handle
);
279 status
= ocfs2_journal_access(handle
, inode
, fe_bh
,
280 OCFS2_JOURNAL_ACCESS_WRITE
);
287 * Do this before setting i_size.
289 status
= ocfs2_zero_tail_for_truncate(inode
, handle
, new_i_size
);
295 i_size_write(inode
, new_i_size
);
296 inode
->i_blocks
= ocfs2_align_bytes_to_sectors(new_i_size
);
297 inode
->i_ctime
= inode
->i_mtime
= CURRENT_TIME
;
299 di
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
300 di
->i_size
= cpu_to_le64(new_i_size
);
301 di
->i_ctime
= di
->i_mtime
= cpu_to_le64(inode
->i_ctime
.tv_sec
);
302 di
->i_ctime_nsec
= di
->i_mtime_nsec
= cpu_to_le32(inode
->i_ctime
.tv_nsec
);
304 status
= ocfs2_journal_dirty(handle
, fe_bh
);
309 ocfs2_commit_trans(osb
, handle
);
316 static int ocfs2_truncate_file(struct inode
*inode
,
317 struct buffer_head
*di_bh
,
321 struct ocfs2_dinode
*fe
= NULL
;
322 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
323 struct ocfs2_truncate_context
*tc
= NULL
;
325 mlog_entry("(inode = %llu, new_i_size = %llu\n",
326 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
327 (unsigned long long)new_i_size
);
329 fe
= (struct ocfs2_dinode
*) di_bh
->b_data
;
330 if (!OCFS2_IS_VALID_DINODE(fe
)) {
331 OCFS2_RO_ON_INVALID_DINODE(inode
->i_sb
, fe
);
336 mlog_bug_on_msg(le64_to_cpu(fe
->i_size
) != i_size_read(inode
),
337 "Inode %llu, inode i_size = %lld != di "
338 "i_size = %llu, i_flags = 0x%x\n",
339 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
341 (unsigned long long)le64_to_cpu(fe
->i_size
),
342 le32_to_cpu(fe
->i_flags
));
344 if (new_i_size
> le64_to_cpu(fe
->i_size
)) {
345 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
346 (unsigned long long)le64_to_cpu(fe
->i_size
),
347 (unsigned long long)new_i_size
);
353 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
354 (unsigned long long)le64_to_cpu(fe
->i_blkno
),
355 (unsigned long long)le64_to_cpu(fe
->i_size
),
356 (unsigned long long)new_i_size
);
358 /* lets handle the simple truncate cases before doing any more
359 * cluster locking. */
360 if (new_i_size
== le64_to_cpu(fe
->i_size
))
363 down_write(&OCFS2_I(inode
)->ip_alloc_sem
);
365 /* This forces other nodes to sync and drop their pages. Do
366 * this even if we have a truncate without allocation change -
367 * ocfs2 cluster sizes can be much greater than page size, so
368 * we have to truncate them anyway. */
369 status
= ocfs2_data_lock(inode
, 1);
371 up_write(&OCFS2_I(inode
)->ip_alloc_sem
);
377 unmap_mapping_range(inode
->i_mapping
, new_i_size
+ PAGE_SIZE
- 1, 0, 1);
378 truncate_inode_pages(inode
->i_mapping
, new_i_size
);
380 /* alright, we're going to need to do a full blown alloc size
381 * change. Orphan the inode so that recovery can complete the
382 * truncate if necessary. This does the task of marking
384 status
= ocfs2_orphan_for_truncate(osb
, inode
, di_bh
, new_i_size
);
387 goto bail_unlock_data
;
390 status
= ocfs2_prepare_truncate(osb
, inode
, di_bh
, &tc
);
393 goto bail_unlock_data
;
396 status
= ocfs2_commit_truncate(osb
, inode
, di_bh
, tc
);
399 goto bail_unlock_data
;
402 /* TODO: orphan dir cleanup here. */
404 ocfs2_data_unlock(inode
, 1);
406 up_write(&OCFS2_I(inode
)->ip_alloc_sem
);
415 * extend allocation only here.
416 * we'll update all the disk stuff, and oip->alloc_size
418 * expect stuff to be locked, a transaction started and enough data /
419 * metadata reservations in the contexts.
421 * Will return -EAGAIN, and a reason if a restart is needed.
422 * If passed in, *reason will always be set, even in error.
424 int ocfs2_do_extend_allocation(struct ocfs2_super
*osb
,
428 struct buffer_head
*fe_bh
,
430 struct ocfs2_alloc_context
*data_ac
,
431 struct ocfs2_alloc_context
*meta_ac
,
432 enum ocfs2_alloc_restarted
*reason_ret
)
436 struct ocfs2_dinode
*fe
= (struct ocfs2_dinode
*) fe_bh
->b_data
;
437 enum ocfs2_alloc_restarted reason
= RESTART_NONE
;
438 u32 bit_off
, num_bits
;
441 BUG_ON(!clusters_to_add
);
443 free_extents
= ocfs2_num_free_extents(osb
, inode
, fe
);
444 if (free_extents
< 0) {
445 status
= free_extents
;
450 /* there are two cases which could cause us to EAGAIN in the
451 * we-need-more-metadata case:
452 * 1) we haven't reserved *any*
453 * 2) we are so fragmented, we've needed to add metadata too
455 if (!free_extents
&& !meta_ac
) {
456 mlog(0, "we haven't reserved any metadata!\n");
458 reason
= RESTART_META
;
460 } else if ((!free_extents
)
461 && (ocfs2_alloc_context_bits_left(meta_ac
)
462 < ocfs2_extend_meta_needed(fe
))) {
463 mlog(0, "filesystem is really fragmented...\n");
465 reason
= RESTART_META
;
469 status
= ocfs2_claim_clusters(osb
, handle
, data_ac
, 1,
470 &bit_off
, &num_bits
);
472 if (status
!= -ENOSPC
)
477 BUG_ON(num_bits
> clusters_to_add
);
479 /* reserve our write early -- insert_extent may update the inode */
480 status
= ocfs2_journal_access(handle
, inode
, fe_bh
,
481 OCFS2_JOURNAL_ACCESS_WRITE
);
487 block
= ocfs2_clusters_to_blocks(osb
->sb
, bit_off
);
488 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
489 num_bits
, bit_off
, (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
490 status
= ocfs2_insert_extent(osb
, handle
, inode
, fe_bh
,
491 *logical_offset
, block
, num_bits
,
498 status
= ocfs2_journal_dirty(handle
, fe_bh
);
504 clusters_to_add
-= num_bits
;
505 *logical_offset
+= num_bits
;
507 if (clusters_to_add
) {
508 mlog(0, "need to alloc once more, clusters = %u, wanted = "
509 "%u\n", fe
->i_clusters
, clusters_to_add
);
511 reason
= RESTART_TRANS
;
517 *reason_ret
= reason
;
522 * For a given allocation, determine which allocators will need to be
523 * accessed, and lock them, reserving the appropriate number of bits.
525 * Called from ocfs2_extend_allocation() for file systems which don't
526 * support holes, and from ocfs2_write() for file systems which
527 * understand sparse inodes.
529 int ocfs2_lock_allocators(struct inode
*inode
, struct ocfs2_dinode
*di
,
530 u32 clusters_to_add
, u32 extents_to_split
,
531 struct ocfs2_alloc_context
**data_ac
,
532 struct ocfs2_alloc_context
**meta_ac
)
534 int ret
, num_free_extents
;
535 unsigned int max_recs_needed
= clusters_to_add
+ 2 * extents_to_split
;
536 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
541 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
542 "clusters_to_add = %u, extents_to_split = %u\n",
543 (unsigned long long)OCFS2_I(inode
)->ip_blkno
, i_size_read(inode
),
544 le32_to_cpu(di
->i_clusters
), clusters_to_add
, extents_to_split
);
546 num_free_extents
= ocfs2_num_free_extents(osb
, inode
, di
);
547 if (num_free_extents
< 0) {
548 ret
= num_free_extents
;
554 * Sparse allocation file systems need to be more conservative
555 * with reserving room for expansion - the actual allocation
556 * happens while we've got a journal handle open so re-taking
557 * a cluster lock (because we ran out of room for another
558 * extent) will violate ordering rules.
560 * Most of the time we'll only be seeing this 1 cluster at a time
563 * Always lock for any unwritten extents - we might want to
564 * add blocks during a split.
566 if (!num_free_extents
||
567 (ocfs2_sparse_alloc(osb
) && num_free_extents
< max_recs_needed
)) {
568 ret
= ocfs2_reserve_new_metadata(osb
, di
, meta_ac
);
576 ret
= ocfs2_reserve_clusters(osb
, clusters_to_add
, data_ac
);
586 ocfs2_free_alloc_context(*meta_ac
);
591 * We cannot have an error and a non null *data_ac.
598 static int ocfs2_extend_allocation(struct inode
*inode
,
602 int restart_func
= 0;
603 int drop_alloc_sem
= 0;
605 u32 prev_clusters
, logical_start
;
606 struct buffer_head
*bh
= NULL
;
607 struct ocfs2_dinode
*fe
= NULL
;
608 handle_t
*handle
= NULL
;
609 struct ocfs2_alloc_context
*data_ac
= NULL
;
610 struct ocfs2_alloc_context
*meta_ac
= NULL
;
611 enum ocfs2_alloc_restarted why
;
612 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
614 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add
);
617 * This function only exists for file systems which don't
620 BUG_ON(ocfs2_sparse_alloc(osb
));
622 status
= ocfs2_read_block(osb
, OCFS2_I(inode
)->ip_blkno
, &bh
,
623 OCFS2_BH_CACHED
, inode
);
629 fe
= (struct ocfs2_dinode
*) bh
->b_data
;
630 if (!OCFS2_IS_VALID_DINODE(fe
)) {
631 OCFS2_RO_ON_INVALID_DINODE(inode
->i_sb
, fe
);
636 logical_start
= OCFS2_I(inode
)->ip_clusters
;
639 BUG_ON(le32_to_cpu(fe
->i_clusters
) != OCFS2_I(inode
)->ip_clusters
);
641 /* blocks peope in read/write from reading our allocation
642 * until we're done changing it. We depend on i_mutex to block
643 * other extend/truncate calls while we're here. Ordering wrt
644 * start_trans is important here -- always do it before! */
645 down_write(&OCFS2_I(inode
)->ip_alloc_sem
);
648 status
= ocfs2_lock_allocators(inode
, fe
, clusters_to_add
, 0, &data_ac
,
655 credits
= ocfs2_calc_extend_credits(osb
->sb
, fe
, clusters_to_add
);
656 handle
= ocfs2_start_trans(osb
, credits
);
657 if (IS_ERR(handle
)) {
658 status
= PTR_ERR(handle
);
664 restarted_transaction
:
665 /* reserve a write to the file entry early on - that we if we
666 * run out of credits in the allocation path, we can still
668 status
= ocfs2_journal_access(handle
, inode
, bh
,
669 OCFS2_JOURNAL_ACCESS_WRITE
);
675 prev_clusters
= OCFS2_I(inode
)->ip_clusters
;
677 status
= ocfs2_do_extend_allocation(osb
,
686 if ((status
< 0) && (status
!= -EAGAIN
)) {
687 if (status
!= -ENOSPC
)
692 status
= ocfs2_journal_dirty(handle
, bh
);
698 spin_lock(&OCFS2_I(inode
)->ip_lock
);
699 clusters_to_add
-= (OCFS2_I(inode
)->ip_clusters
- prev_clusters
);
700 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
702 if (why
!= RESTART_NONE
&& clusters_to_add
) {
703 if (why
== RESTART_META
) {
704 mlog(0, "restarting function.\n");
707 BUG_ON(why
!= RESTART_TRANS
);
709 mlog(0, "restarting transaction.\n");
710 /* TODO: This can be more intelligent. */
711 credits
= ocfs2_calc_extend_credits(osb
->sb
,
714 status
= ocfs2_extend_trans(handle
, credits
);
716 /* handle still has to be committed at
722 goto restarted_transaction
;
726 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
727 le32_to_cpu(fe
->i_clusters
),
728 (unsigned long long)le64_to_cpu(fe
->i_size
));
729 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
730 OCFS2_I(inode
)->ip_clusters
, i_size_read(inode
));
733 if (drop_alloc_sem
) {
734 up_write(&OCFS2_I(inode
)->ip_alloc_sem
);
738 ocfs2_commit_trans(osb
, handle
);
742 ocfs2_free_alloc_context(data_ac
);
746 ocfs2_free_alloc_context(meta_ac
);
749 if ((!status
) && restart_func
) {
762 /* Some parts of this taken from generic_cont_expand, which turned out
763 * to be too fragile to do exactly what we need without us having to
764 * worry about recursive locking in ->prepare_write() and
765 * ->commit_write(). */
766 static int ocfs2_write_zero_page(struct inode
*inode
,
769 struct address_space
*mapping
= inode
->i_mapping
;
773 handle_t
*handle
= NULL
;
776 offset
= (size
& (PAGE_CACHE_SIZE
-1)); /* Within page */
777 /* ugh. in prepare/commit_write, if from==to==start of block, we
778 ** skip the prepare. make sure we never send an offset for the start
781 if ((offset
& (inode
->i_sb
->s_blocksize
- 1)) == 0) {
784 index
= size
>> PAGE_CACHE_SHIFT
;
786 page
= grab_cache_page(mapping
, index
);
793 ret
= ocfs2_prepare_write_nolock(inode
, page
, offset
, offset
);
799 if (ocfs2_should_order_data(inode
)) {
800 handle
= ocfs2_start_walk_page_trans(inode
, page
, offset
,
802 if (IS_ERR(handle
)) {
803 ret
= PTR_ERR(handle
);
809 /* must not update i_size! */
810 ret
= block_commit_write(page
, offset
, offset
);
817 ocfs2_commit_trans(OCFS2_SB(inode
->i_sb
), handle
);
820 page_cache_release(page
);
825 static int ocfs2_zero_extend(struct inode
*inode
,
830 struct super_block
*sb
= inode
->i_sb
;
832 start_off
= ocfs2_align_bytes_to_blocks(sb
, i_size_read(inode
));
833 while (start_off
< zero_to_size
) {
834 ret
= ocfs2_write_zero_page(inode
, start_off
);
840 start_off
+= sb
->s_blocksize
;
843 * Very large extends have the potential to lock up
844 * the cpu for extended periods of time.
854 * A tail_to_skip value > 0 indicates that we're being called from
855 * ocfs2_file_aio_write(). This has the following implications:
857 * - we don't want to update i_size
858 * - di_bh will be NULL, which is fine because it's only used in the
859 * case where we want to update i_size.
860 * - ocfs2_zero_extend() will then only be filling the hole created
861 * between i_size and the start of the write.
863 static int ocfs2_extend_file(struct inode
*inode
,
864 struct buffer_head
*di_bh
,
869 u32 clusters_to_add
= 0;
871 BUG_ON(!tail_to_skip
&& !di_bh
);
873 /* setattr sometimes calls us like this. */
877 if (i_size_read(inode
) == new_i_size
)
879 BUG_ON(new_i_size
< i_size_read(inode
));
881 if (ocfs2_sparse_alloc(OCFS2_SB(inode
->i_sb
))) {
882 BUG_ON(tail_to_skip
!= 0);
883 goto out_update_size
;
886 clusters_to_add
= ocfs2_clusters_for_bytes(inode
->i_sb
, new_i_size
) -
887 OCFS2_I(inode
)->ip_clusters
;
890 * protect the pages that ocfs2_zero_extend is going to be
891 * pulling into the page cache.. we do this before the
892 * metadata extend so that we don't get into the situation
893 * where we've extended the metadata but can't get the data
896 ret
= ocfs2_data_lock(inode
, 1);
902 if (clusters_to_add
) {
903 ret
= ocfs2_extend_allocation(inode
, clusters_to_add
);
911 * Call this even if we don't add any clusters to the tree. We
912 * still need to zero the area between the old i_size and the
915 ret
= ocfs2_zero_extend(inode
, (u64
)new_i_size
- tail_to_skip
);
923 /* We're being called from ocfs2_setattr() which wants
924 * us to update i_size */
925 ret
= ocfs2_simple_size_update(inode
, di_bh
, new_i_size
);
931 if (!ocfs2_sparse_alloc(OCFS2_SB(inode
->i_sb
)))
932 ocfs2_data_unlock(inode
, 1);
938 int ocfs2_setattr(struct dentry
*dentry
, struct iattr
*attr
)
940 int status
= 0, size_change
;
941 struct inode
*inode
= dentry
->d_inode
;
942 struct super_block
*sb
= inode
->i_sb
;
943 struct ocfs2_super
*osb
= OCFS2_SB(sb
);
944 struct buffer_head
*bh
= NULL
;
945 handle_t
*handle
= NULL
;
947 mlog_entry("(0x%p, '%.*s')\n", dentry
,
948 dentry
->d_name
.len
, dentry
->d_name
.name
);
950 if (attr
->ia_valid
& ATTR_MODE
)
951 mlog(0, "mode change: %d\n", attr
->ia_mode
);
952 if (attr
->ia_valid
& ATTR_UID
)
953 mlog(0, "uid change: %d\n", attr
->ia_uid
);
954 if (attr
->ia_valid
& ATTR_GID
)
955 mlog(0, "gid change: %d\n", attr
->ia_gid
);
956 if (attr
->ia_valid
& ATTR_SIZE
)
957 mlog(0, "size change...\n");
958 if (attr
->ia_valid
& (ATTR_ATIME
| ATTR_MTIME
| ATTR_CTIME
))
959 mlog(0, "time change...\n");
961 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
962 | ATTR_GID | ATTR_UID | ATTR_MODE)
963 if (!(attr
->ia_valid
& OCFS2_VALID_ATTRS
)) {
964 mlog(0, "can't handle attrs: 0x%x\n", attr
->ia_valid
);
968 status
= inode_change_ok(inode
, attr
);
972 size_change
= S_ISREG(inode
->i_mode
) && attr
->ia_valid
& ATTR_SIZE
;
974 status
= ocfs2_rw_lock(inode
, 1);
981 status
= ocfs2_meta_lock(inode
, &bh
, 1);
983 if (status
!= -ENOENT
)
988 if (size_change
&& attr
->ia_size
!= i_size_read(inode
)) {
989 if (i_size_read(inode
) > attr
->ia_size
)
990 status
= ocfs2_truncate_file(inode
, bh
, attr
->ia_size
);
992 status
= ocfs2_extend_file(inode
, bh
, attr
->ia_size
, 0);
994 if (status
!= -ENOSPC
)
1001 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
1002 if (IS_ERR(handle
)) {
1003 status
= PTR_ERR(handle
);
1009 * This will intentionally not wind up calling vmtruncate(),
1010 * since all the work for a size change has been done above.
1011 * Otherwise, we could get into problems with truncate as
1012 * ip_alloc_sem is used there to protect against i_size
1015 status
= inode_setattr(inode
, attr
);
1021 status
= ocfs2_mark_inode_dirty(handle
, inode
, bh
);
1026 ocfs2_commit_trans(osb
, handle
);
1028 ocfs2_meta_unlock(inode
, 1);
1031 ocfs2_rw_unlock(inode
, 1);
1040 int ocfs2_getattr(struct vfsmount
*mnt
,
1041 struct dentry
*dentry
,
1044 struct inode
*inode
= dentry
->d_inode
;
1045 struct super_block
*sb
= dentry
->d_inode
->i_sb
;
1046 struct ocfs2_super
*osb
= sb
->s_fs_info
;
1051 err
= ocfs2_inode_revalidate(dentry
);
1058 generic_fillattr(inode
, stat
);
1060 /* We set the blksize from the cluster size for performance */
1061 stat
->blksize
= osb
->s_clustersize
;
1069 int ocfs2_permission(struct inode
*inode
, int mask
, struct nameidata
*nd
)
1075 ret
= ocfs2_meta_lock(inode
, NULL
, 0);
1082 ret
= generic_permission(inode
, mask
, NULL
);
1084 ocfs2_meta_unlock(inode
, 0);
1090 static int ocfs2_write_remove_suid(struct inode
*inode
)
1093 struct buffer_head
*bh
= NULL
;
1094 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1096 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1097 struct ocfs2_dinode
*di
;
1099 mlog_entry("(Inode %llu, mode 0%o)\n",
1100 (unsigned long long)oi
->ip_blkno
, inode
->i_mode
);
1102 handle
= ocfs2_start_trans(osb
, OCFS2_INODE_UPDATE_CREDITS
);
1103 if (handle
== NULL
) {
1109 ret
= ocfs2_read_block(osb
, oi
->ip_blkno
, &bh
, OCFS2_BH_CACHED
, inode
);
1115 ret
= ocfs2_journal_access(handle
, inode
, bh
,
1116 OCFS2_JOURNAL_ACCESS_WRITE
);
1122 inode
->i_mode
&= ~S_ISUID
;
1123 if ((inode
->i_mode
& S_ISGID
) && (inode
->i_mode
& S_IXGRP
))
1124 inode
->i_mode
&= ~S_ISGID
;
1126 di
= (struct ocfs2_dinode
*) bh
->b_data
;
1127 di
->i_mode
= cpu_to_le16(inode
->i_mode
);
1129 ret
= ocfs2_journal_dirty(handle
, bh
);
1135 ocfs2_commit_trans(osb
, handle
);
1142 * Will look for holes and unwritten extents in the range starting at
1143 * pos for count bytes (inclusive).
1145 static int ocfs2_check_range_for_holes(struct inode
*inode
, loff_t pos
,
1149 unsigned int extent_flags
;
1150 u32 cpos
, clusters
, extent_len
, phys_cpos
;
1151 struct super_block
*sb
= inode
->i_sb
;
1153 cpos
= pos
>> OCFS2_SB(sb
)->s_clustersize_bits
;
1154 clusters
= ocfs2_clusters_for_bytes(sb
, pos
+ count
) - cpos
;
1157 ret
= ocfs2_get_clusters(inode
, cpos
, &phys_cpos
, &extent_len
,
1164 if (phys_cpos
== 0 || (extent_flags
& OCFS2_EXT_UNWRITTEN
)) {
1169 if (extent_len
> clusters
)
1170 extent_len
= clusters
;
1172 clusters
-= extent_len
;
1179 static int ocfs2_prepare_inode_for_write(struct dentry
*dentry
,
1185 int ret
= 0, meta_level
= appending
;
1186 struct inode
*inode
= dentry
->d_inode
;
1188 loff_t newsize
, saved_pos
;
1191 * We sample i_size under a read level meta lock to see if our write
1192 * is extending the file, if it is we back off and get a write level
1196 ret
= ocfs2_meta_lock(inode
, NULL
, meta_level
);
1203 /* Clear suid / sgid if necessary. We do this here
1204 * instead of later in the write path because
1205 * remove_suid() calls ->setattr without any hint that
1206 * we may have already done our cluster locking. Since
1207 * ocfs2_setattr() *must* take cluster locks to
1208 * proceeed, this will lead us to recursively lock the
1209 * inode. There's also the dinode i_size state which
1210 * can be lost via setattr during extending writes (we
1211 * set inode->i_size at the end of a write. */
1212 if (should_remove_suid(dentry
)) {
1213 if (meta_level
== 0) {
1214 ocfs2_meta_unlock(inode
, meta_level
);
1219 ret
= ocfs2_write_remove_suid(inode
);
1226 /* work on a copy of ppos until we're sure that we won't have
1227 * to recalculate it due to relocking. */
1229 saved_pos
= i_size_read(inode
);
1230 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos
);
1235 if (ocfs2_sparse_alloc(OCFS2_SB(inode
->i_sb
))) {
1236 loff_t end
= saved_pos
+ count
;
1239 * Skip the O_DIRECT checks if we don't need
1242 if (!direct_io
|| !(*direct_io
))
1246 * Allowing concurrent direct writes means
1247 * i_size changes wouldn't be synchronized, so
1248 * one node could wind up truncating another
1251 if (end
> i_size_read(inode
)) {
1257 * We don't fill holes during direct io, so
1258 * check for them here. If any are found, the
1259 * caller will have to retake some cluster
1260 * locks and initiate the io as buffered.
1262 ret
= ocfs2_check_range_for_holes(inode
, saved_pos
,
1273 * The rest of this loop is concerned with legacy file
1274 * systems which don't support sparse files.
1277 newsize
= count
+ saved_pos
;
1279 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1280 (long long) saved_pos
, (long long) newsize
,
1281 (long long) i_size_read(inode
));
1283 /* No need for a higher level metadata lock if we're
1284 * never going past i_size. */
1285 if (newsize
<= i_size_read(inode
))
1288 if (meta_level
== 0) {
1289 ocfs2_meta_unlock(inode
, meta_level
);
1294 spin_lock(&OCFS2_I(inode
)->ip_lock
);
1295 clusters
= ocfs2_clusters_for_bytes(inode
->i_sb
, newsize
) -
1296 OCFS2_I(inode
)->ip_clusters
;
1297 spin_unlock(&OCFS2_I(inode
)->ip_lock
);
1299 mlog(0, "Writing at EOF, may need more allocation: "
1300 "i_size = %lld, newsize = %lld, need %u clusters\n",
1301 (long long) i_size_read(inode
), (long long) newsize
,
1304 /* We only want to continue the rest of this loop if
1305 * our extend will actually require more
1310 ret
= ocfs2_extend_file(inode
, NULL
, newsize
, count
);
1323 ocfs2_meta_unlock(inode
, meta_level
);
1330 ocfs2_set_next_iovec(const struct iovec
**iovp
, size_t *basep
, size_t bytes
)
1332 const struct iovec
*iov
= *iovp
;
1333 size_t base
= *basep
;
1336 int copy
= min(bytes
, iov
->iov_len
- base
);
1340 if (iov
->iov_len
== base
) {
1349 static struct page
* ocfs2_get_write_source(char **ret_src_buf
,
1350 const struct iovec
*cur_iov
,
1354 char *buf
= cur_iov
->iov_base
+ iov_offset
;
1355 struct page
*src_page
= NULL
;
1358 off
= (unsigned long)(buf
) & ~PAGE_CACHE_MASK
;
1360 if (!segment_eq(get_fs(), KERNEL_DS
)) {
1362 * Pull in the user page. We want to do this outside
1363 * of the meta data locks in order to preserve locking
1364 * order in case of page fault.
1366 ret
= get_user_pages(current
, current
->mm
,
1367 (unsigned long)buf
& PAGE_CACHE_MASK
, 1,
1368 0, 0, &src_page
, NULL
);
1370 *ret_src_buf
= kmap(src_page
) + off
;
1372 src_page
= ERR_PTR(-EFAULT
);
1380 static void ocfs2_put_write_source(struct page
*page
)
1384 page_cache_release(page
);
1388 static ssize_t
ocfs2_file_buffered_write(struct file
*file
, loff_t
*ppos
,
1389 const struct iovec
*iov
,
1390 unsigned long nr_segs
,
1392 ssize_t o_direct_written
)
1395 ssize_t copied
, total
= 0;
1396 size_t iov_offset
= 0, bytes
;
1398 const struct iovec
*cur_iov
= iov
;
1399 struct page
*user_page
, *page
;
1404 * handle partial DIO write. Adjust cur_iov if needed.
1406 ocfs2_set_next_iovec(&cur_iov
, &iov_offset
, o_direct_written
);
1411 user_page
= ocfs2_get_write_source(&buf
, cur_iov
, iov_offset
);
1412 if (IS_ERR(user_page
)) {
1413 ret
= PTR_ERR(user_page
);
1417 /* Stay within our page boundaries */
1418 bytes
= min((PAGE_CACHE_SIZE
- ((unsigned long)pos
& ~PAGE_CACHE_MASK
)),
1419 (PAGE_CACHE_SIZE
- ((unsigned long)buf
& ~PAGE_CACHE_MASK
)));
1420 /* Stay within the vector boundary */
1421 bytes
= min_t(size_t, bytes
, cur_iov
->iov_len
- iov_offset
);
1422 /* Stay within count */
1423 bytes
= min(bytes
, count
);
1426 ret
= ocfs2_write_begin(file
, file
->f_mapping
, pos
, bytes
, 0,
1433 dst
= kmap_atomic(page
, KM_USER0
);
1434 memcpy(dst
+ (pos
& (PAGE_CACHE_SIZE
- 1)), buf
, bytes
);
1435 kunmap_atomic(dst
, KM_USER0
);
1436 flush_dcache_page(page
);
1437 ocfs2_put_write_source(user_page
);
1439 copied
= ocfs2_write_end(file
, file
->f_mapping
, pos
, bytes
,
1440 bytes
, page
, fsdata
);
1448 *ppos
= pos
+ copied
;
1451 ocfs2_set_next_iovec(&cur_iov
, &iov_offset
, copied
);
1455 return total
? total
: ret
;
1458 static ssize_t
ocfs2_file_aio_write(struct kiocb
*iocb
,
1459 const struct iovec
*iov
,
1460 unsigned long nr_segs
,
1463 int ret
, direct_io
, appending
, rw_level
, have_alloc_sem
= 0;
1464 int can_do_direct
, sync
= 0;
1465 ssize_t written
= 0;
1466 size_t ocount
; /* original count */
1467 size_t count
; /* after file limit checks */
1468 loff_t
*ppos
= &iocb
->ki_pos
;
1469 struct file
*file
= iocb
->ki_filp
;
1470 struct inode
*inode
= file
->f_path
.dentry
->d_inode
;
1472 mlog_entry("(0x%p, %u, '%.*s')\n", file
,
1473 (unsigned int)nr_segs
,
1474 file
->f_path
.dentry
->d_name
.len
,
1475 file
->f_path
.dentry
->d_name
.name
);
1477 if (iocb
->ki_left
== 0)
1480 ret
= generic_segment_checks(iov
, &nr_segs
, &ocount
, VERIFY_READ
);
1486 vfs_check_frozen(inode
->i_sb
, SB_FREEZE_WRITE
);
1488 appending
= file
->f_flags
& O_APPEND
? 1 : 0;
1489 direct_io
= file
->f_flags
& O_DIRECT
? 1 : 0;
1491 mutex_lock(&inode
->i_mutex
);
1494 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1496 down_read(&inode
->i_alloc_sem
);
1500 /* concurrent O_DIRECT writes are allowed */
1501 rw_level
= !direct_io
;
1502 ret
= ocfs2_rw_lock(inode
, rw_level
);
1508 can_do_direct
= direct_io
;
1509 ret
= ocfs2_prepare_inode_for_write(file
->f_path
.dentry
, ppos
,
1510 iocb
->ki_left
, appending
,
1518 * We can't complete the direct I/O as requested, fall back to
1521 if (direct_io
&& !can_do_direct
) {
1522 ocfs2_rw_unlock(inode
, rw_level
);
1523 up_read(&inode
->i_alloc_sem
);
1533 if (!sync
&& ((file
->f_flags
& O_SYNC
) || IS_SYNC(inode
)))
1537 * XXX: Is it ok to execute these checks a second time?
1539 ret
= generic_write_checks(file
, ppos
, &count
, S_ISBLK(inode
->i_mode
));
1544 * Set pos so that sync_page_range_nolock() below understands
1545 * where to start from. We might've moved it around via the
1546 * calls above. The range we want to actually sync starts from
1552 /* communicate with ocfs2_dio_end_io */
1553 ocfs2_iocb_set_rw_locked(iocb
, rw_level
);
1556 written
= generic_file_direct_write(iocb
, iov
, &nr_segs
, *ppos
,
1557 ppos
, count
, ocount
);
1563 written
= ocfs2_file_buffered_write(file
, ppos
, iov
, nr_segs
,
1567 if (ret
!= -EFAULT
|| ret
!= -ENOSPC
)
1574 /* buffered aio wouldn't have proper lock coverage today */
1575 BUG_ON(ret
== -EIOCBQUEUED
&& !(file
->f_flags
& O_DIRECT
));
1578 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1579 * function pointer which is called when o_direct io completes so that
1580 * it can unlock our rw lock. (it's the clustered equivalent of
1581 * i_alloc_sem; protects truncate from racing with pending ios).
1582 * Unfortunately there are error cases which call end_io and others
1583 * that don't. so we don't have to unlock the rw_lock if either an
1584 * async dio is going to do it in the future or an end_io after an
1585 * error has already done it.
1587 if (ret
== -EIOCBQUEUED
|| !ocfs2_iocb_is_rw_locked(iocb
)) {
1594 ocfs2_rw_unlock(inode
, rw_level
);
1598 up_read(&inode
->i_alloc_sem
);
1600 if (written
> 0 && sync
) {
1603 err
= sync_page_range_nolock(inode
, file
->f_mapping
, pos
, count
);
1608 mutex_unlock(&inode
->i_mutex
);
1611 return written
? written
: ret
;
1614 static int ocfs2_splice_write_actor(struct pipe_inode_info
*pipe
,
1615 struct pipe_buffer
*buf
,
1616 struct splice_desc
*sd
)
1620 struct file
*file
= sd
->u
.file
;
1621 unsigned int offset
;
1622 struct page
*page
= NULL
;
1626 ret
= buf
->ops
->confirm(pipe
, buf
);
1630 offset
= sd
->pos
& ~PAGE_CACHE_MASK
;
1632 if (count
+ offset
> PAGE_CACHE_SIZE
)
1633 count
= PAGE_CACHE_SIZE
- offset
;
1635 ret
= ocfs2_write_begin(file
, file
->f_mapping
, sd
->pos
, count
, 0,
1642 src
= buf
->ops
->map(pipe
, buf
, 1);
1643 dst
= kmap_atomic(page
, KM_USER1
);
1644 memcpy(dst
+ offset
, src
+ buf
->offset
, count
);
1645 kunmap_atomic(page
, KM_USER1
);
1646 buf
->ops
->unmap(pipe
, buf
, src
);
1648 copied
= ocfs2_write_end(file
, file
->f_mapping
, sd
->pos
, count
, count
,
1657 return copied
? copied
: ret
;
1660 static ssize_t
__ocfs2_file_splice_write(struct pipe_inode_info
*pipe
,
1667 struct address_space
*mapping
= out
->f_mapping
;
1668 struct inode
*inode
= mapping
->host
;
1669 struct splice_desc sd
= {
1676 ret
= __splice_from_pipe(pipe
, &sd
, ocfs2_splice_write_actor
);
1680 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(inode
))) {
1681 err
= generic_osync_inode(inode
, mapping
,
1682 OSYNC_METADATA
|OSYNC_DATA
);
1691 static ssize_t
ocfs2_file_splice_write(struct pipe_inode_info
*pipe
,
1698 struct inode
*inode
= out
->f_path
.dentry
->d_inode
;
1700 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out
, pipe
,
1702 out
->f_path
.dentry
->d_name
.len
,
1703 out
->f_path
.dentry
->d_name
.name
);
1705 inode_double_lock(inode
, pipe
->inode
);
1707 ret
= ocfs2_rw_lock(inode
, 1);
1713 ret
= ocfs2_prepare_inode_for_write(out
->f_path
.dentry
, ppos
, len
, 0,
1720 /* ok, we're done with i_size and alloc work */
1721 ret
= __ocfs2_file_splice_write(pipe
, out
, ppos
, len
, flags
);
1724 ocfs2_rw_unlock(inode
, 1);
1726 inode_double_unlock(inode
, pipe
->inode
);
1732 static ssize_t
ocfs2_file_splice_read(struct file
*in
,
1734 struct pipe_inode_info
*pipe
,
1739 struct inode
*inode
= in
->f_path
.dentry
->d_inode
;
1741 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in
, pipe
,
1743 in
->f_path
.dentry
->d_name
.len
,
1744 in
->f_path
.dentry
->d_name
.name
);
1747 * See the comment in ocfs2_file_aio_read()
1749 ret
= ocfs2_meta_lock(inode
, NULL
, 0);
1754 ocfs2_meta_unlock(inode
, 0);
1756 ret
= generic_file_splice_read(in
, ppos
, pipe
, len
, flags
);
1763 static ssize_t
ocfs2_file_aio_read(struct kiocb
*iocb
,
1764 const struct iovec
*iov
,
1765 unsigned long nr_segs
,
1768 int ret
= 0, rw_level
= -1, have_alloc_sem
= 0, lock_level
= 0;
1769 struct file
*filp
= iocb
->ki_filp
;
1770 struct inode
*inode
= filp
->f_path
.dentry
->d_inode
;
1772 mlog_entry("(0x%p, %u, '%.*s')\n", filp
,
1773 (unsigned int)nr_segs
,
1774 filp
->f_path
.dentry
->d_name
.len
,
1775 filp
->f_path
.dentry
->d_name
.name
);
1784 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
1785 * need locks to protect pending reads from racing with truncate.
1787 if (filp
->f_flags
& O_DIRECT
) {
1788 down_read(&inode
->i_alloc_sem
);
1791 ret
= ocfs2_rw_lock(inode
, 0);
1797 /* communicate with ocfs2_dio_end_io */
1798 ocfs2_iocb_set_rw_locked(iocb
, rw_level
);
1802 * We're fine letting folks race truncates and extending
1803 * writes with read across the cluster, just like they can
1804 * locally. Hence no rw_lock during read.
1806 * Take and drop the meta data lock to update inode fields
1807 * like i_size. This allows the checks down below
1808 * generic_file_aio_read() a chance of actually working.
1810 ret
= ocfs2_meta_lock_atime(inode
, filp
->f_vfsmnt
, &lock_level
);
1815 ocfs2_meta_unlock(inode
, lock_level
);
1817 ret
= generic_file_aio_read(iocb
, iov
, nr_segs
, iocb
->ki_pos
);
1819 mlog(ML_ERROR
, "generic_file_aio_read returned -EINVAL\n");
1821 /* buffered aio wouldn't have proper lock coverage today */
1822 BUG_ON(ret
== -EIOCBQUEUED
&& !(filp
->f_flags
& O_DIRECT
));
1824 /* see ocfs2_file_aio_write */
1825 if (ret
== -EIOCBQUEUED
|| !ocfs2_iocb_is_rw_locked(iocb
)) {
1832 up_read(&inode
->i_alloc_sem
);
1834 ocfs2_rw_unlock(inode
, rw_level
);
1840 const struct inode_operations ocfs2_file_iops
= {
1841 .setattr
= ocfs2_setattr
,
1842 .getattr
= ocfs2_getattr
,
1843 .permission
= ocfs2_permission
,
1846 const struct inode_operations ocfs2_special_file_iops
= {
1847 .setattr
= ocfs2_setattr
,
1848 .getattr
= ocfs2_getattr
,
1849 .permission
= ocfs2_permission
,
1852 const struct file_operations ocfs2_fops
= {
1853 .read
= do_sync_read
,
1854 .write
= do_sync_write
,
1856 .fsync
= ocfs2_sync_file
,
1857 .release
= ocfs2_file_release
,
1858 .open
= ocfs2_file_open
,
1859 .aio_read
= ocfs2_file_aio_read
,
1860 .aio_write
= ocfs2_file_aio_write
,
1861 .ioctl
= ocfs2_ioctl
,
1862 #ifdef CONFIG_COMPAT
1863 .compat_ioctl
= ocfs2_compat_ioctl
,
1865 .splice_read
= ocfs2_file_splice_read
,
1866 .splice_write
= ocfs2_file_splice_write
,
1869 const struct file_operations ocfs2_dops
= {
1870 .read
= generic_read_dir
,
1871 .readdir
= ocfs2_readdir
,
1872 .fsync
= ocfs2_sync_file
,
1873 .ioctl
= ocfs2_ioctl
,
1874 #ifdef CONFIG_COMPAT
1875 .compat_ioctl
= ocfs2_compat_ioctl
,