2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_format.h"
21 #include "xfs_log_format.h"
22 #include "xfs_trans_resv.h"
24 #include "xfs_mount.h"
25 #include "xfs_inode.h"
26 #include "xfs_error.h"
27 #include "xfs_trans.h"
28 #include "xfs_trans_priv.h"
29 #include "xfs_inode_item.h"
30 #include "xfs_quota.h"
31 #include "xfs_trace.h"
32 #include "xfs_icache.h"
33 #include "xfs_bmap_util.h"
34 #include "xfs_dquot_item.h"
35 #include "xfs_dquot.h"
37 #include <linux/kthread.h>
38 #include <linux/freezer.h>
40 STATIC
void __xfs_inode_clear_reclaim_tag(struct xfs_mount
*mp
,
41 struct xfs_perag
*pag
, struct xfs_inode
*ip
);
44 * Allocate and initialise an xfs_inode.
54 * if this didn't occur in transactions, we could use
55 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
56 * code up to do this anyway.
58 ip
= kmem_zone_alloc(xfs_inode_zone
, KM_SLEEP
);
61 if (inode_init_always(mp
->m_super
, VFS_I(ip
))) {
62 kmem_zone_free(xfs_inode_zone
, ip
);
66 XFS_STATS_INC(mp
, vn_active
);
67 ASSERT(atomic_read(&ip
->i_pincount
) == 0);
68 ASSERT(!spin_is_locked(&ip
->i_flags_lock
));
69 ASSERT(!xfs_isiflocked(ip
));
70 ASSERT(ip
->i_ino
== 0);
72 mrlock_init(&ip
->i_iolock
, MRLOCK_BARRIER
, "xfsio", ip
->i_ino
);
74 /* initialise the xfs inode */
77 memset(&ip
->i_imap
, 0, sizeof(struct xfs_imap
));
79 memset(&ip
->i_df
, 0, sizeof(xfs_ifork_t
));
81 ip
->i_delayed_blks
= 0;
82 memset(&ip
->i_d
, 0, sizeof(ip
->i_d
));
88 xfs_inode_free_callback(
89 struct rcu_head
*head
)
91 struct inode
*inode
= container_of(head
, struct inode
, i_rcu
);
92 struct xfs_inode
*ip
= XFS_I(inode
);
94 kmem_zone_free(xfs_inode_zone
, ip
);
101 switch (ip
->i_d
.di_mode
& S_IFMT
) {
105 xfs_idestroy_fork(ip
, XFS_DATA_FORK
);
110 xfs_idestroy_fork(ip
, XFS_ATTR_FORK
);
113 ASSERT(!(ip
->i_itemp
->ili_item
.li_flags
& XFS_LI_IN_AIL
));
114 xfs_inode_item_destroy(ip
);
119 * Because we use RCU freeing we need to ensure the inode always
120 * appears to be reclaimed with an invalid inode number when in the
121 * free state. The ip->i_flags_lock provides the barrier against lookup
124 spin_lock(&ip
->i_flags_lock
);
125 ip
->i_flags
= XFS_IRECLAIM
;
127 spin_unlock(&ip
->i_flags_lock
);
129 /* asserts to verify all state is correct here */
130 ASSERT(atomic_read(&ip
->i_pincount
) == 0);
131 ASSERT(!xfs_isiflocked(ip
));
132 XFS_STATS_DEC(ip
->i_mount
, vn_active
);
134 call_rcu(&VFS_I(ip
)->i_rcu
, xfs_inode_free_callback
);
138 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
139 * part of the structure. This is made more complex by the fact we store
140 * information about the on-disk values in the VFS inode and so we can't just
141 * overwrite the values unconditionally. Hence we save the parameters we
142 * need to retain across reinitialisation, and rewrite them into the VFS inode
143 * after reinitialisation even if it fails.
147 struct xfs_mount
*mp
,
151 uint32_t nlink
= inode
->i_nlink
;
152 uint32_t generation
= inode
->i_generation
;
153 uint64_t version
= inode
->i_version
;
155 error
= inode_init_always(mp
->m_super
, inode
);
157 set_nlink(inode
, nlink
);
158 inode
->i_generation
= generation
;
159 inode
->i_version
= version
;
164 * Check the validity of the inode we just found it the cache
168 struct xfs_perag
*pag
,
169 struct xfs_inode
*ip
,
172 int lock_flags
) __releases(RCU
)
174 struct inode
*inode
= VFS_I(ip
);
175 struct xfs_mount
*mp
= ip
->i_mount
;
179 * check for re-use of an inode within an RCU grace period due to the
180 * radix tree nodes not being updated yet. We monitor for this by
181 * setting the inode number to zero before freeing the inode structure.
182 * If the inode has been reallocated and set up, then the inode number
183 * will not match, so check for that, too.
185 spin_lock(&ip
->i_flags_lock
);
186 if (ip
->i_ino
!= ino
) {
187 trace_xfs_iget_skip(ip
);
188 XFS_STATS_INC(mp
, xs_ig_frecycle
);
195 * If we are racing with another cache hit that is currently
196 * instantiating this inode or currently recycling it out of
197 * reclaimabe state, wait for the initialisation to complete
200 * XXX(hch): eventually we should do something equivalent to
201 * wait_on_inode to wait for these flags to be cleared
202 * instead of polling for it.
204 if (ip
->i_flags
& (XFS_INEW
|XFS_IRECLAIM
)) {
205 trace_xfs_iget_skip(ip
);
206 XFS_STATS_INC(mp
, xs_ig_frecycle
);
212 * If lookup is racing with unlink return an error immediately.
214 if (ip
->i_d
.di_mode
== 0 && !(flags
& XFS_IGET_CREATE
)) {
220 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
221 * Need to carefully get it back into useable state.
223 if (ip
->i_flags
& XFS_IRECLAIMABLE
) {
224 trace_xfs_iget_reclaim(ip
);
227 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
228 * from stomping over us while we recycle the inode. We can't
229 * clear the radix tree reclaimable tag yet as it requires
230 * pag_ici_lock to be held exclusive.
232 ip
->i_flags
|= XFS_IRECLAIM
;
234 spin_unlock(&ip
->i_flags_lock
);
237 error
= xfs_reinit_inode(mp
, inode
);
240 * Re-initializing the inode failed, and we are in deep
241 * trouble. Try to re-add it to the reclaim list.
244 spin_lock(&ip
->i_flags_lock
);
246 ip
->i_flags
&= ~(XFS_INEW
| XFS_IRECLAIM
);
247 ASSERT(ip
->i_flags
& XFS_IRECLAIMABLE
);
248 trace_xfs_iget_reclaim_fail(ip
);
252 spin_lock(&pag
->pag_ici_lock
);
253 spin_lock(&ip
->i_flags_lock
);
256 * Clear the per-lifetime state in the inode as we are now
257 * effectively a new inode and need to return to the initial
258 * state before reuse occurs.
260 ip
->i_flags
&= ~XFS_IRECLAIM_RESET_FLAGS
;
261 ip
->i_flags
|= XFS_INEW
;
262 __xfs_inode_clear_reclaim_tag(mp
, pag
, ip
);
263 inode
->i_state
= I_NEW
;
265 ASSERT(!rwsem_is_locked(&ip
->i_iolock
.mr_lock
));
266 mrlock_init(&ip
->i_iolock
, MRLOCK_BARRIER
, "xfsio", ip
->i_ino
);
268 spin_unlock(&ip
->i_flags_lock
);
269 spin_unlock(&pag
->pag_ici_lock
);
271 /* If the VFS inode is being torn down, pause and try again. */
273 trace_xfs_iget_skip(ip
);
278 /* We've got a live one. */
279 spin_unlock(&ip
->i_flags_lock
);
281 trace_xfs_iget_hit(ip
);
285 xfs_ilock(ip
, lock_flags
);
287 xfs_iflags_clear(ip
, XFS_ISTALE
| XFS_IDONTCACHE
);
288 XFS_STATS_INC(mp
, xs_ig_found
);
293 spin_unlock(&ip
->i_flags_lock
);
301 struct xfs_mount
*mp
,
302 struct xfs_perag
*pag
,
305 struct xfs_inode
**ipp
,
309 struct xfs_inode
*ip
;
311 xfs_agino_t agino
= XFS_INO_TO_AGINO(mp
, ino
);
314 ip
= xfs_inode_alloc(mp
, ino
);
318 error
= xfs_iread(mp
, tp
, ip
, flags
);
322 trace_xfs_iget_miss(ip
);
324 if ((ip
->i_d
.di_mode
== 0) && !(flags
& XFS_IGET_CREATE
)) {
330 * Preload the radix tree so we can insert safely under the
331 * write spinlock. Note that we cannot sleep inside the preload
332 * region. Since we can be called from transaction context, don't
333 * recurse into the file system.
335 if (radix_tree_preload(GFP_NOFS
)) {
341 * Because the inode hasn't been added to the radix-tree yet it can't
342 * be found by another thread, so we can do the non-sleeping lock here.
345 if (!xfs_ilock_nowait(ip
, lock_flags
))
350 * These values must be set before inserting the inode into the radix
351 * tree as the moment it is inserted a concurrent lookup (allowed by the
352 * RCU locking mechanism) can find it and that lookup must see that this
353 * is an inode currently under construction (i.e. that XFS_INEW is set).
354 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
355 * memory barrier that ensures this detection works correctly at lookup
359 if (flags
& XFS_IGET_DONTCACHE
)
360 iflags
|= XFS_IDONTCACHE
;
364 xfs_iflags_set(ip
, iflags
);
366 /* insert the new inode */
367 spin_lock(&pag
->pag_ici_lock
);
368 error
= radix_tree_insert(&pag
->pag_ici_root
, agino
, ip
);
369 if (unlikely(error
)) {
370 WARN_ON(error
!= -EEXIST
);
371 XFS_STATS_INC(mp
, xs_ig_dup
);
373 goto out_preload_end
;
375 spin_unlock(&pag
->pag_ici_lock
);
376 radix_tree_preload_end();
382 spin_unlock(&pag
->pag_ici_lock
);
383 radix_tree_preload_end();
385 xfs_iunlock(ip
, lock_flags
);
387 __destroy_inode(VFS_I(ip
));
393 * Look up an inode by number in the given file system.
394 * The inode is looked up in the cache held in each AG.
395 * If the inode is found in the cache, initialise the vfs inode
398 * If it is not in core, read it in from the file system's device,
399 * add it to the cache and initialise the vfs inode.
401 * The inode is locked according to the value of the lock_flags parameter.
402 * This flag parameter indicates how and if the inode's IO lock and inode lock
405 * mp -- the mount point structure for the current file system. It points
406 * to the inode hash table.
407 * tp -- a pointer to the current transaction if there is one. This is
408 * simply passed through to the xfs_iread() call.
409 * ino -- the number of the inode desired. This is the unique identifier
410 * within the file system for the inode being requested.
411 * lock_flags -- flags indicating how to lock the inode. See the comment
412 * for xfs_ilock() for a list of valid values.
429 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
430 * doesn't get freed while it's being referenced during a
431 * radix tree traversal here. It assumes this function
432 * aqcuires only the ILOCK (and therefore it has no need to
433 * involve the IOLOCK in this synchronization).
435 ASSERT((lock_flags
& (XFS_IOLOCK_EXCL
| XFS_IOLOCK_SHARED
)) == 0);
437 /* reject inode numbers outside existing AGs */
438 if (!ino
|| XFS_INO_TO_AGNO(mp
, ino
) >= mp
->m_sb
.sb_agcount
)
441 XFS_STATS_INC(mp
, xs_ig_attempts
);
443 /* get the perag structure and ensure that it's inode capable */
444 pag
= xfs_perag_get(mp
, XFS_INO_TO_AGNO(mp
, ino
));
445 agino
= XFS_INO_TO_AGINO(mp
, ino
);
450 ip
= radix_tree_lookup(&pag
->pag_ici_root
, agino
);
453 error
= xfs_iget_cache_hit(pag
, ip
, ino
, flags
, lock_flags
);
455 goto out_error_or_again
;
458 XFS_STATS_INC(mp
, xs_ig_missed
);
460 error
= xfs_iget_cache_miss(mp
, pag
, tp
, ino
, &ip
,
463 goto out_error_or_again
;
470 * If we have a real type for an on-disk inode, we can setup the inode
471 * now. If it's a new inode being created, xfs_ialloc will handle it.
473 if (xfs_iflags_test(ip
, XFS_INEW
) && ip
->i_d
.di_mode
!= 0)
474 xfs_setup_existing_inode(ip
);
478 if (error
== -EAGAIN
) {
487 * The inode lookup is done in batches to keep the amount of lock traffic and
488 * radix tree lookups to a minimum. The batch size is a trade off between
489 * lookup reduction and stack usage. This is in the reclaim path, so we can't
492 #define XFS_LOOKUP_BATCH 32
495 xfs_inode_ag_walk_grab(
496 struct xfs_inode
*ip
)
498 struct inode
*inode
= VFS_I(ip
);
500 ASSERT(rcu_read_lock_held());
503 * check for stale RCU freed inode
505 * If the inode has been reallocated, it doesn't matter if it's not in
506 * the AG we are walking - we are walking for writeback, so if it
507 * passes all the "valid inode" checks and is dirty, then we'll write
508 * it back anyway. If it has been reallocated and still being
509 * initialised, the XFS_INEW check below will catch it.
511 spin_lock(&ip
->i_flags_lock
);
513 goto out_unlock_noent
;
515 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
516 if (__xfs_iflags_test(ip
, XFS_INEW
| XFS_IRECLAIMABLE
| XFS_IRECLAIM
))
517 goto out_unlock_noent
;
518 spin_unlock(&ip
->i_flags_lock
);
520 /* nothing to sync during shutdown */
521 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
))
522 return -EFSCORRUPTED
;
524 /* If we can't grab the inode, it must on it's way to reclaim. */
532 spin_unlock(&ip
->i_flags_lock
);
538 struct xfs_mount
*mp
,
539 struct xfs_perag
*pag
,
540 int (*execute
)(struct xfs_inode
*ip
, int flags
,
546 uint32_t first_index
;
558 struct xfs_inode
*batch
[XFS_LOOKUP_BATCH
];
565 nr_found
= radix_tree_gang_lookup(&pag
->pag_ici_root
,
566 (void **)batch
, first_index
,
569 nr_found
= radix_tree_gang_lookup_tag(
571 (void **) batch
, first_index
,
572 XFS_LOOKUP_BATCH
, tag
);
580 * Grab the inodes before we drop the lock. if we found
581 * nothing, nr == 0 and the loop will be skipped.
583 for (i
= 0; i
< nr_found
; i
++) {
584 struct xfs_inode
*ip
= batch
[i
];
586 if (done
|| xfs_inode_ag_walk_grab(ip
))
590 * Update the index for the next lookup. Catch
591 * overflows into the next AG range which can occur if
592 * we have inodes in the last block of the AG and we
593 * are currently pointing to the last inode.
595 * Because we may see inodes that are from the wrong AG
596 * due to RCU freeing and reallocation, only update the
597 * index if it lies in this AG. It was a race that lead
598 * us to see this inode, so another lookup from the
599 * same index will not find it again.
601 if (XFS_INO_TO_AGNO(mp
, ip
->i_ino
) != pag
->pag_agno
)
603 first_index
= XFS_INO_TO_AGINO(mp
, ip
->i_ino
+ 1);
604 if (first_index
< XFS_INO_TO_AGINO(mp
, ip
->i_ino
))
608 /* unlock now we've grabbed the inodes. */
611 for (i
= 0; i
< nr_found
; i
++) {
614 error
= execute(batch
[i
], flags
, args
);
616 if (error
== -EAGAIN
) {
620 if (error
&& last_error
!= -EFSCORRUPTED
)
624 /* bail out if the filesystem is corrupted. */
625 if (error
== -EFSCORRUPTED
)
630 } while (nr_found
&& !done
);
640 * Background scanning to trim post-EOF preallocated space. This is queued
641 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
645 struct xfs_mount
*mp
)
648 if (radix_tree_tagged(&mp
->m_perag_tree
, XFS_ICI_EOFBLOCKS_TAG
))
649 queue_delayed_work(mp
->m_eofblocks_workqueue
,
650 &mp
->m_eofblocks_work
,
651 msecs_to_jiffies(xfs_eofb_secs
* 1000));
656 xfs_eofblocks_worker(
657 struct work_struct
*work
)
659 struct xfs_mount
*mp
= container_of(to_delayed_work(work
),
660 struct xfs_mount
, m_eofblocks_work
);
661 xfs_icache_free_eofblocks(mp
, NULL
);
662 xfs_queue_eofblocks(mp
);
666 xfs_inode_ag_iterator(
667 struct xfs_mount
*mp
,
668 int (*execute
)(struct xfs_inode
*ip
, int flags
,
673 struct xfs_perag
*pag
;
679 while ((pag
= xfs_perag_get(mp
, ag
))) {
680 ag
= pag
->pag_agno
+ 1;
681 error
= xfs_inode_ag_walk(mp
, pag
, execute
, flags
, args
, -1);
685 if (error
== -EFSCORRUPTED
)
693 xfs_inode_ag_iterator_tag(
694 struct xfs_mount
*mp
,
695 int (*execute
)(struct xfs_inode
*ip
, int flags
,
701 struct xfs_perag
*pag
;
707 while ((pag
= xfs_perag_get_tag(mp
, ag
, tag
))) {
708 ag
= pag
->pag_agno
+ 1;
709 error
= xfs_inode_ag_walk(mp
, pag
, execute
, flags
, args
, tag
);
713 if (error
== -EFSCORRUPTED
)
721 * Queue a new inode reclaim pass if there are reclaimable inodes and there
722 * isn't a reclaim pass already in progress. By default it runs every 5s based
723 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
724 * tunable, but that can be done if this method proves to be ineffective or too
728 xfs_reclaim_work_queue(
729 struct xfs_mount
*mp
)
733 if (radix_tree_tagged(&mp
->m_perag_tree
, XFS_ICI_RECLAIM_TAG
)) {
734 queue_delayed_work(mp
->m_reclaim_workqueue
, &mp
->m_reclaim_work
,
735 msecs_to_jiffies(xfs_syncd_centisecs
/ 6 * 10));
741 * This is a fast pass over the inode cache to try to get reclaim moving on as
742 * many inodes as possible in a short period of time. It kicks itself every few
743 * seconds, as well as being kicked by the inode cache shrinker when memory
744 * goes low. It scans as quickly as possible avoiding locked inodes or those
745 * already being flushed, and once done schedules a future pass.
749 struct work_struct
*work
)
751 struct xfs_mount
*mp
= container_of(to_delayed_work(work
),
752 struct xfs_mount
, m_reclaim_work
);
754 xfs_reclaim_inodes(mp
, SYNC_TRYLOCK
);
755 xfs_reclaim_work_queue(mp
);
759 __xfs_inode_set_reclaim_tag(
760 struct xfs_perag
*pag
,
761 struct xfs_inode
*ip
)
763 radix_tree_tag_set(&pag
->pag_ici_root
,
764 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
),
765 XFS_ICI_RECLAIM_TAG
);
767 if (!pag
->pag_ici_reclaimable
) {
768 /* propagate the reclaim tag up into the perag radix tree */
769 spin_lock(&ip
->i_mount
->m_perag_lock
);
770 radix_tree_tag_set(&ip
->i_mount
->m_perag_tree
,
771 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
772 XFS_ICI_RECLAIM_TAG
);
773 spin_unlock(&ip
->i_mount
->m_perag_lock
);
775 /* schedule periodic background inode reclaim */
776 xfs_reclaim_work_queue(ip
->i_mount
);
778 trace_xfs_perag_set_reclaim(ip
->i_mount
, pag
->pag_agno
,
781 pag
->pag_ici_reclaimable
++;
785 * We set the inode flag atomically with the radix tree tag.
786 * Once we get tag lookups on the radix tree, this inode flag
790 xfs_inode_set_reclaim_tag(
793 struct xfs_mount
*mp
= ip
->i_mount
;
794 struct xfs_perag
*pag
;
796 pag
= xfs_perag_get(mp
, XFS_INO_TO_AGNO(mp
, ip
->i_ino
));
797 spin_lock(&pag
->pag_ici_lock
);
798 spin_lock(&ip
->i_flags_lock
);
799 __xfs_inode_set_reclaim_tag(pag
, ip
);
800 __xfs_iflags_set(ip
, XFS_IRECLAIMABLE
);
801 spin_unlock(&ip
->i_flags_lock
);
802 spin_unlock(&pag
->pag_ici_lock
);
807 __xfs_inode_clear_reclaim(
811 pag
->pag_ici_reclaimable
--;
812 if (!pag
->pag_ici_reclaimable
) {
813 /* clear the reclaim tag from the perag radix tree */
814 spin_lock(&ip
->i_mount
->m_perag_lock
);
815 radix_tree_tag_clear(&ip
->i_mount
->m_perag_tree
,
816 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
817 XFS_ICI_RECLAIM_TAG
);
818 spin_unlock(&ip
->i_mount
->m_perag_lock
);
819 trace_xfs_perag_clear_reclaim(ip
->i_mount
, pag
->pag_agno
,
825 __xfs_inode_clear_reclaim_tag(
830 radix_tree_tag_clear(&pag
->pag_ici_root
,
831 XFS_INO_TO_AGINO(mp
, ip
->i_ino
), XFS_ICI_RECLAIM_TAG
);
832 __xfs_inode_clear_reclaim(pag
, ip
);
836 * Grab the inode for reclaim exclusively.
837 * Return 0 if we grabbed it, non-zero otherwise.
840 xfs_reclaim_inode_grab(
841 struct xfs_inode
*ip
,
844 ASSERT(rcu_read_lock_held());
846 /* quick check for stale RCU freed inode */
851 * If we are asked for non-blocking operation, do unlocked checks to
852 * see if the inode already is being flushed or in reclaim to avoid
855 if ((flags
& SYNC_TRYLOCK
) &&
856 __xfs_iflags_test(ip
, XFS_IFLOCK
| XFS_IRECLAIM
))
860 * The radix tree lock here protects a thread in xfs_iget from racing
861 * with us starting reclaim on the inode. Once we have the
862 * XFS_IRECLAIM flag set it will not touch us.
864 * Due to RCU lookup, we may find inodes that have been freed and only
865 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
866 * aren't candidates for reclaim at all, so we must check the
867 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
869 spin_lock(&ip
->i_flags_lock
);
870 if (!__xfs_iflags_test(ip
, XFS_IRECLAIMABLE
) ||
871 __xfs_iflags_test(ip
, XFS_IRECLAIM
)) {
872 /* not a reclaim candidate. */
873 spin_unlock(&ip
->i_flags_lock
);
876 __xfs_iflags_set(ip
, XFS_IRECLAIM
);
877 spin_unlock(&ip
->i_flags_lock
);
882 * Inodes in different states need to be treated differently. The following
883 * table lists the inode states and the reclaim actions necessary:
885 * inode state iflush ret required action
886 * --------------- ---------- ---------------
888 * shutdown EIO unpin and reclaim
889 * clean, unpinned 0 reclaim
890 * stale, unpinned 0 reclaim
891 * clean, pinned(*) 0 requeue
892 * stale, pinned EAGAIN requeue
893 * dirty, async - requeue
894 * dirty, sync 0 reclaim
896 * (*) dgc: I don't think the clean, pinned state is possible but it gets
897 * handled anyway given the order of checks implemented.
899 * Also, because we get the flush lock first, we know that any inode that has
900 * been flushed delwri has had the flush completed by the time we check that
901 * the inode is clean.
903 * Note that because the inode is flushed delayed write by AIL pushing, the
904 * flush lock may already be held here and waiting on it can result in very
905 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
906 * the caller should push the AIL first before trying to reclaim inodes to
907 * minimise the amount of time spent waiting. For background relaim, we only
908 * bother to reclaim clean inodes anyway.
910 * Hence the order of actions after gaining the locks should be:
912 * shutdown => unpin and reclaim
913 * pinned, async => requeue
914 * pinned, sync => unpin
917 * dirty, async => requeue
918 * dirty, sync => flush, wait and reclaim
922 struct xfs_inode
*ip
,
923 struct xfs_perag
*pag
,
926 struct xfs_buf
*bp
= NULL
;
931 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
932 if (!xfs_iflock_nowait(ip
)) {
933 if (!(sync_mode
& SYNC_WAIT
))
938 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
)) {
940 xfs_iflush_abort(ip
, false);
943 if (xfs_ipincount(ip
)) {
944 if (!(sync_mode
& SYNC_WAIT
))
948 if (xfs_iflags_test(ip
, XFS_ISTALE
))
950 if (xfs_inode_clean(ip
))
954 * Never flush out dirty data during non-blocking reclaim, as it would
955 * just contend with AIL pushing trying to do the same job.
957 if (!(sync_mode
& SYNC_WAIT
))
961 * Now we have an inode that needs flushing.
963 * Note that xfs_iflush will never block on the inode buffer lock, as
964 * xfs_ifree_cluster() can lock the inode buffer before it locks the
965 * ip->i_lock, and we are doing the exact opposite here. As a result,
966 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
967 * result in an ABBA deadlock with xfs_ifree_cluster().
969 * As xfs_ifree_cluser() must gather all inodes that are active in the
970 * cache to mark them stale, if we hit this case we don't actually want
971 * to do IO here - we want the inode marked stale so we can simply
972 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
973 * inode, back off and try again. Hopefully the next pass through will
974 * see the stale flag set on the inode.
976 error
= xfs_iflush(ip
, &bp
);
977 if (error
== -EAGAIN
) {
978 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
979 /* backoff longer than in xfs_ifree_cluster */
985 error
= xfs_bwrite(bp
);
992 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
994 XFS_STATS_INC(ip
->i_mount
, xs_ig_reclaims
);
996 * Remove the inode from the per-AG radix tree.
998 * Because radix_tree_delete won't complain even if the item was never
999 * added to the tree assert that it's been there before to catch
1000 * problems with the inode life time early on.
1002 spin_lock(&pag
->pag_ici_lock
);
1003 if (!radix_tree_delete(&pag
->pag_ici_root
,
1004 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
)))
1006 __xfs_inode_clear_reclaim(pag
, ip
);
1007 spin_unlock(&pag
->pag_ici_lock
);
1010 * Here we do an (almost) spurious inode lock in order to coordinate
1011 * with inode cache radix tree lookups. This is because the lookup
1012 * can reference the inodes in the cache without taking references.
1014 * We make that OK here by ensuring that we wait until the inode is
1015 * unlocked after the lookup before we go ahead and free it.
1017 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
1018 xfs_qm_dqdetach(ip
);
1019 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
1027 xfs_iflags_clear(ip
, XFS_IRECLAIM
);
1028 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
1030 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1031 * a short while. However, this just burns CPU time scanning the tree
1032 * waiting for IO to complete and the reclaim work never goes back to
1033 * the idle state. Instead, return 0 to let the next scheduled
1034 * background reclaim attempt to reclaim the inode again.
1040 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1041 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1042 * then a shut down during filesystem unmount reclaim walk leak all the
1043 * unreclaimed inodes.
1046 xfs_reclaim_inodes_ag(
1047 struct xfs_mount
*mp
,
1051 struct xfs_perag
*pag
;
1055 int trylock
= flags
& SYNC_TRYLOCK
;
1061 while ((pag
= xfs_perag_get_tag(mp
, ag
, XFS_ICI_RECLAIM_TAG
))) {
1062 unsigned long first_index
= 0;
1066 ag
= pag
->pag_agno
+ 1;
1069 if (!mutex_trylock(&pag
->pag_ici_reclaim_lock
)) {
1074 first_index
= pag
->pag_ici_reclaim_cursor
;
1076 mutex_lock(&pag
->pag_ici_reclaim_lock
);
1079 struct xfs_inode
*batch
[XFS_LOOKUP_BATCH
];
1083 nr_found
= radix_tree_gang_lookup_tag(
1085 (void **)batch
, first_index
,
1087 XFS_ICI_RECLAIM_TAG
);
1095 * Grab the inodes before we drop the lock. if we found
1096 * nothing, nr == 0 and the loop will be skipped.
1098 for (i
= 0; i
< nr_found
; i
++) {
1099 struct xfs_inode
*ip
= batch
[i
];
1101 if (done
|| xfs_reclaim_inode_grab(ip
, flags
))
1105 * Update the index for the next lookup. Catch
1106 * overflows into the next AG range which can
1107 * occur if we have inodes in the last block of
1108 * the AG and we are currently pointing to the
1111 * Because we may see inodes that are from the
1112 * wrong AG due to RCU freeing and
1113 * reallocation, only update the index if it
1114 * lies in this AG. It was a race that lead us
1115 * to see this inode, so another lookup from
1116 * the same index will not find it again.
1118 if (XFS_INO_TO_AGNO(mp
, ip
->i_ino
) !=
1121 first_index
= XFS_INO_TO_AGINO(mp
, ip
->i_ino
+ 1);
1122 if (first_index
< XFS_INO_TO_AGINO(mp
, ip
->i_ino
))
1126 /* unlock now we've grabbed the inodes. */
1129 for (i
= 0; i
< nr_found
; i
++) {
1132 error
= xfs_reclaim_inode(batch
[i
], pag
, flags
);
1133 if (error
&& last_error
!= -EFSCORRUPTED
)
1137 *nr_to_scan
-= XFS_LOOKUP_BATCH
;
1141 } while (nr_found
&& !done
&& *nr_to_scan
> 0);
1143 if (trylock
&& !done
)
1144 pag
->pag_ici_reclaim_cursor
= first_index
;
1146 pag
->pag_ici_reclaim_cursor
= 0;
1147 mutex_unlock(&pag
->pag_ici_reclaim_lock
);
1152 * if we skipped any AG, and we still have scan count remaining, do
1153 * another pass this time using blocking reclaim semantics (i.e
1154 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1155 * ensure that when we get more reclaimers than AGs we block rather
1156 * than spin trying to execute reclaim.
1158 if (skipped
&& (flags
& SYNC_WAIT
) && *nr_to_scan
> 0) {
1170 int nr_to_scan
= INT_MAX
;
1172 return xfs_reclaim_inodes_ag(mp
, mode
, &nr_to_scan
);
1176 * Scan a certain number of inodes for reclaim.
1178 * When called we make sure that there is a background (fast) inode reclaim in
1179 * progress, while we will throttle the speed of reclaim via doing synchronous
1180 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1181 * them to be cleaned, which we hope will not be very long due to the
1182 * background walker having already kicked the IO off on those dirty inodes.
1185 xfs_reclaim_inodes_nr(
1186 struct xfs_mount
*mp
,
1189 /* kick background reclaimer and push the AIL */
1190 xfs_reclaim_work_queue(mp
);
1191 xfs_ail_push_all(mp
->m_ail
);
1193 return xfs_reclaim_inodes_ag(mp
, SYNC_TRYLOCK
| SYNC_WAIT
, &nr_to_scan
);
1197 * Return the number of reclaimable inodes in the filesystem for
1198 * the shrinker to determine how much to reclaim.
1201 xfs_reclaim_inodes_count(
1202 struct xfs_mount
*mp
)
1204 struct xfs_perag
*pag
;
1205 xfs_agnumber_t ag
= 0;
1206 int reclaimable
= 0;
1208 while ((pag
= xfs_perag_get_tag(mp
, ag
, XFS_ICI_RECLAIM_TAG
))) {
1209 ag
= pag
->pag_agno
+ 1;
1210 reclaimable
+= pag
->pag_ici_reclaimable
;
1218 struct xfs_inode
*ip
,
1219 struct xfs_eofblocks
*eofb
)
1221 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_UID
) &&
1222 !uid_eq(VFS_I(ip
)->i_uid
, eofb
->eof_uid
))
1225 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_GID
) &&
1226 !gid_eq(VFS_I(ip
)->i_gid
, eofb
->eof_gid
))
1229 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_PRID
) &&
1230 xfs_get_projid(ip
) != eofb
->eof_prid
)
1237 * A union-based inode filtering algorithm. Process the inode if any of the
1238 * criteria match. This is for global/internal scans only.
1241 xfs_inode_match_id_union(
1242 struct xfs_inode
*ip
,
1243 struct xfs_eofblocks
*eofb
)
1245 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_UID
) &&
1246 uid_eq(VFS_I(ip
)->i_uid
, eofb
->eof_uid
))
1249 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_GID
) &&
1250 gid_eq(VFS_I(ip
)->i_gid
, eofb
->eof_gid
))
1253 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_PRID
) &&
1254 xfs_get_projid(ip
) == eofb
->eof_prid
)
1261 xfs_inode_free_eofblocks(
1262 struct xfs_inode
*ip
,
1267 struct xfs_eofblocks
*eofb
= args
;
1268 bool need_iolock
= true;
1271 ASSERT(!eofb
|| (eofb
&& eofb
->eof_scan_owner
!= 0));
1273 if (!xfs_can_free_eofblocks(ip
, false)) {
1274 /* inode could be preallocated or append-only */
1275 trace_xfs_inode_free_eofblocks_invalid(ip
);
1276 xfs_inode_clear_eofblocks_tag(ip
);
1281 * If the mapping is dirty the operation can block and wait for some
1282 * time. Unless we are waiting, skip it.
1284 if (!(flags
& SYNC_WAIT
) &&
1285 mapping_tagged(VFS_I(ip
)->i_mapping
, PAGECACHE_TAG_DIRTY
))
1289 if (eofb
->eof_flags
& XFS_EOF_FLAGS_UNION
)
1290 match
= xfs_inode_match_id_union(ip
, eofb
);
1292 match
= xfs_inode_match_id(ip
, eofb
);
1296 /* skip the inode if the file size is too small */
1297 if (eofb
->eof_flags
& XFS_EOF_FLAGS_MINFILESIZE
&&
1298 XFS_ISIZE(ip
) < eofb
->eof_min_file_size
)
1302 * A scan owner implies we already hold the iolock. Skip it in
1303 * xfs_free_eofblocks() to avoid deadlock. This also eliminates
1304 * the possibility of EAGAIN being returned.
1306 if (eofb
->eof_scan_owner
== ip
->i_ino
)
1307 need_iolock
= false;
1310 ret
= xfs_free_eofblocks(ip
->i_mount
, ip
, need_iolock
);
1312 /* don't revisit the inode if we're not waiting */
1313 if (ret
== -EAGAIN
&& !(flags
& SYNC_WAIT
))
1320 xfs_icache_free_eofblocks(
1321 struct xfs_mount
*mp
,
1322 struct xfs_eofblocks
*eofb
)
1324 int flags
= SYNC_TRYLOCK
;
1326 if (eofb
&& (eofb
->eof_flags
& XFS_EOF_FLAGS_SYNC
))
1329 return xfs_inode_ag_iterator_tag(mp
, xfs_inode_free_eofblocks
, flags
,
1330 eofb
, XFS_ICI_EOFBLOCKS_TAG
);
1334 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1335 * multiple quotas, we don't know exactly which quota caused an allocation
1336 * failure. We make a best effort by including each quota under low free space
1337 * conditions (less than 1% free space) in the scan.
1340 xfs_inode_free_quota_eofblocks(
1341 struct xfs_inode
*ip
)
1344 struct xfs_eofblocks eofb
= {0};
1345 struct xfs_dquot
*dq
;
1347 ASSERT(xfs_isilocked(ip
, XFS_IOLOCK_EXCL
));
1350 * Set the scan owner to avoid a potential livelock. Otherwise, the scan
1351 * can repeatedly trylock on the inode we're currently processing. We
1352 * run a sync scan to increase effectiveness and use the union filter to
1353 * cover all applicable quotas in a single scan.
1355 eofb
.eof_scan_owner
= ip
->i_ino
;
1356 eofb
.eof_flags
= XFS_EOF_FLAGS_UNION
|XFS_EOF_FLAGS_SYNC
;
1358 if (XFS_IS_UQUOTA_ENFORCED(ip
->i_mount
)) {
1359 dq
= xfs_inode_dquot(ip
, XFS_DQ_USER
);
1360 if (dq
&& xfs_dquot_lowsp(dq
)) {
1361 eofb
.eof_uid
= VFS_I(ip
)->i_uid
;
1362 eofb
.eof_flags
|= XFS_EOF_FLAGS_UID
;
1367 if (XFS_IS_GQUOTA_ENFORCED(ip
->i_mount
)) {
1368 dq
= xfs_inode_dquot(ip
, XFS_DQ_GROUP
);
1369 if (dq
&& xfs_dquot_lowsp(dq
)) {
1370 eofb
.eof_gid
= VFS_I(ip
)->i_gid
;
1371 eofb
.eof_flags
|= XFS_EOF_FLAGS_GID
;
1377 xfs_icache_free_eofblocks(ip
->i_mount
, &eofb
);
1383 xfs_inode_set_eofblocks_tag(
1386 struct xfs_mount
*mp
= ip
->i_mount
;
1387 struct xfs_perag
*pag
;
1390 pag
= xfs_perag_get(mp
, XFS_INO_TO_AGNO(mp
, ip
->i_ino
));
1391 spin_lock(&pag
->pag_ici_lock
);
1392 trace_xfs_inode_set_eofblocks_tag(ip
);
1394 tagged
= radix_tree_tagged(&pag
->pag_ici_root
,
1395 XFS_ICI_EOFBLOCKS_TAG
);
1396 radix_tree_tag_set(&pag
->pag_ici_root
,
1397 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
),
1398 XFS_ICI_EOFBLOCKS_TAG
);
1400 /* propagate the eofblocks tag up into the perag radix tree */
1401 spin_lock(&ip
->i_mount
->m_perag_lock
);
1402 radix_tree_tag_set(&ip
->i_mount
->m_perag_tree
,
1403 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
1404 XFS_ICI_EOFBLOCKS_TAG
);
1405 spin_unlock(&ip
->i_mount
->m_perag_lock
);
1407 /* kick off background trimming */
1408 xfs_queue_eofblocks(ip
->i_mount
);
1410 trace_xfs_perag_set_eofblocks(ip
->i_mount
, pag
->pag_agno
,
1414 spin_unlock(&pag
->pag_ici_lock
);
1419 xfs_inode_clear_eofblocks_tag(
1422 struct xfs_mount
*mp
= ip
->i_mount
;
1423 struct xfs_perag
*pag
;
1425 pag
= xfs_perag_get(mp
, XFS_INO_TO_AGNO(mp
, ip
->i_ino
));
1426 spin_lock(&pag
->pag_ici_lock
);
1427 trace_xfs_inode_clear_eofblocks_tag(ip
);
1429 radix_tree_tag_clear(&pag
->pag_ici_root
,
1430 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
),
1431 XFS_ICI_EOFBLOCKS_TAG
);
1432 if (!radix_tree_tagged(&pag
->pag_ici_root
, XFS_ICI_EOFBLOCKS_TAG
)) {
1433 /* clear the eofblocks tag from the perag radix tree */
1434 spin_lock(&ip
->i_mount
->m_perag_lock
);
1435 radix_tree_tag_clear(&ip
->i_mount
->m_perag_tree
,
1436 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
1437 XFS_ICI_EOFBLOCKS_TAG
);
1438 spin_unlock(&ip
->i_mount
->m_perag_lock
);
1439 trace_xfs_perag_clear_eofblocks(ip
->i_mount
, pag
->pag_agno
,
1443 spin_unlock(&pag
->pag_ici_lock
);