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"
25 #include "xfs_mount.h"
26 #include "xfs_inode.h"
27 #include "xfs_error.h"
28 #include "xfs_trans.h"
29 #include "xfs_trans_priv.h"
30 #include "xfs_inode_item.h"
31 #include "xfs_quota.h"
32 #include "xfs_trace.h"
33 #include "xfs_icache.h"
34 #include "xfs_bmap_util.h"
35 #include "xfs_dquot_item.h"
36 #include "xfs_dquot.h"
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
41 STATIC
void __xfs_inode_clear_reclaim_tag(struct xfs_mount
*mp
,
42 struct xfs_perag
*pag
, struct xfs_inode
*ip
);
45 * Allocate and initialise an xfs_inode.
55 * if this didn't occur in transactions, we could use
56 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
57 * code up to do this anyway.
59 ip
= kmem_zone_alloc(xfs_inode_zone
, KM_SLEEP
);
62 if (inode_init_always(mp
->m_super
, VFS_I(ip
))) {
63 kmem_zone_free(xfs_inode_zone
, ip
);
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(xfs_icdinode_t
));
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
));
133 call_rcu(&VFS_I(ip
)->i_rcu
, xfs_inode_free_callback
);
137 * Check the validity of the inode we just found it the cache
141 struct xfs_perag
*pag
,
142 struct xfs_inode
*ip
,
145 int lock_flags
) __releases(RCU
)
147 struct inode
*inode
= VFS_I(ip
);
148 struct xfs_mount
*mp
= ip
->i_mount
;
152 * check for re-use of an inode within an RCU grace period due to the
153 * radix tree nodes not being updated yet. We monitor for this by
154 * setting the inode number to zero before freeing the inode structure.
155 * If the inode has been reallocated and set up, then the inode number
156 * will not match, so check for that, too.
158 spin_lock(&ip
->i_flags_lock
);
159 if (ip
->i_ino
!= ino
) {
160 trace_xfs_iget_skip(ip
);
161 XFS_STATS_INC(xs_ig_frecycle
);
168 * If we are racing with another cache hit that is currently
169 * instantiating this inode or currently recycling it out of
170 * reclaimabe state, wait for the initialisation to complete
173 * XXX(hch): eventually we should do something equivalent to
174 * wait_on_inode to wait for these flags to be cleared
175 * instead of polling for it.
177 if (ip
->i_flags
& (XFS_INEW
|XFS_IRECLAIM
)) {
178 trace_xfs_iget_skip(ip
);
179 XFS_STATS_INC(xs_ig_frecycle
);
185 * If lookup is racing with unlink return an error immediately.
187 if (ip
->i_d
.di_mode
== 0 && !(flags
& XFS_IGET_CREATE
)) {
193 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
194 * Need to carefully get it back into useable state.
196 if (ip
->i_flags
& XFS_IRECLAIMABLE
) {
197 trace_xfs_iget_reclaim(ip
);
200 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
201 * from stomping over us while we recycle the inode. We can't
202 * clear the radix tree reclaimable tag yet as it requires
203 * pag_ici_lock to be held exclusive.
205 ip
->i_flags
|= XFS_IRECLAIM
;
207 spin_unlock(&ip
->i_flags_lock
);
210 error
= inode_init_always(mp
->m_super
, inode
);
213 * Re-initializing the inode failed, and we are in deep
214 * trouble. Try to re-add it to the reclaim list.
217 spin_lock(&ip
->i_flags_lock
);
219 ip
->i_flags
&= ~(XFS_INEW
| XFS_IRECLAIM
);
220 ASSERT(ip
->i_flags
& XFS_IRECLAIMABLE
);
221 trace_xfs_iget_reclaim_fail(ip
);
225 spin_lock(&pag
->pag_ici_lock
);
226 spin_lock(&ip
->i_flags_lock
);
229 * Clear the per-lifetime state in the inode as we are now
230 * effectively a new inode and need to return to the initial
231 * state before reuse occurs.
233 ip
->i_flags
&= ~XFS_IRECLAIM_RESET_FLAGS
;
234 ip
->i_flags
|= XFS_INEW
;
235 __xfs_inode_clear_reclaim_tag(mp
, pag
, ip
);
236 inode
->i_state
= I_NEW
;
238 ASSERT(!rwsem_is_locked(&ip
->i_iolock
.mr_lock
));
239 mrlock_init(&ip
->i_iolock
, MRLOCK_BARRIER
, "xfsio", ip
->i_ino
);
241 spin_unlock(&ip
->i_flags_lock
);
242 spin_unlock(&pag
->pag_ici_lock
);
244 /* If the VFS inode is being torn down, pause and try again. */
246 trace_xfs_iget_skip(ip
);
251 /* We've got a live one. */
252 spin_unlock(&ip
->i_flags_lock
);
254 trace_xfs_iget_hit(ip
);
258 xfs_ilock(ip
, lock_flags
);
260 xfs_iflags_clear(ip
, XFS_ISTALE
| XFS_IDONTCACHE
);
261 XFS_STATS_INC(xs_ig_found
);
266 spin_unlock(&ip
->i_flags_lock
);
274 struct xfs_mount
*mp
,
275 struct xfs_perag
*pag
,
278 struct xfs_inode
**ipp
,
282 struct xfs_inode
*ip
;
284 xfs_agino_t agino
= XFS_INO_TO_AGINO(mp
, ino
);
287 ip
= xfs_inode_alloc(mp
, ino
);
291 error
= xfs_iread(mp
, tp
, ip
, flags
);
295 trace_xfs_iget_miss(ip
);
297 if ((ip
->i_d
.di_mode
== 0) && !(flags
& XFS_IGET_CREATE
)) {
303 * Preload the radix tree so we can insert safely under the
304 * write spinlock. Note that we cannot sleep inside the preload
305 * region. Since we can be called from transaction context, don't
306 * recurse into the file system.
308 if (radix_tree_preload(GFP_NOFS
)) {
314 * Because the inode hasn't been added to the radix-tree yet it can't
315 * be found by another thread, so we can do the non-sleeping lock here.
318 if (!xfs_ilock_nowait(ip
, lock_flags
))
323 * These values must be set before inserting the inode into the radix
324 * tree as the moment it is inserted a concurrent lookup (allowed by the
325 * RCU locking mechanism) can find it and that lookup must see that this
326 * is an inode currently under construction (i.e. that XFS_INEW is set).
327 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
328 * memory barrier that ensures this detection works correctly at lookup
332 if (flags
& XFS_IGET_DONTCACHE
)
333 iflags
|= XFS_IDONTCACHE
;
337 xfs_iflags_set(ip
, iflags
);
339 /* insert the new inode */
340 spin_lock(&pag
->pag_ici_lock
);
341 error
= radix_tree_insert(&pag
->pag_ici_root
, agino
, ip
);
342 if (unlikely(error
)) {
343 WARN_ON(error
!= -EEXIST
);
344 XFS_STATS_INC(xs_ig_dup
);
346 goto out_preload_end
;
348 spin_unlock(&pag
->pag_ici_lock
);
349 radix_tree_preload_end();
355 spin_unlock(&pag
->pag_ici_lock
);
356 radix_tree_preload_end();
358 xfs_iunlock(ip
, lock_flags
);
360 __destroy_inode(VFS_I(ip
));
366 * Look up an inode by number in the given file system.
367 * The inode is looked up in the cache held in each AG.
368 * If the inode is found in the cache, initialise the vfs inode
371 * If it is not in core, read it in from the file system's device,
372 * add it to the cache and initialise the vfs inode.
374 * The inode is locked according to the value of the lock_flags parameter.
375 * This flag parameter indicates how and if the inode's IO lock and inode lock
378 * mp -- the mount point structure for the current file system. It points
379 * to the inode hash table.
380 * tp -- a pointer to the current transaction if there is one. This is
381 * simply passed through to the xfs_iread() call.
382 * ino -- the number of the inode desired. This is the unique identifier
383 * within the file system for the inode being requested.
384 * lock_flags -- flags indicating how to lock the inode. See the comment
385 * for xfs_ilock() for a list of valid values.
402 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
403 * doesn't get freed while it's being referenced during a
404 * radix tree traversal here. It assumes this function
405 * aqcuires only the ILOCK (and therefore it has no need to
406 * involve the IOLOCK in this synchronization).
408 ASSERT((lock_flags
& (XFS_IOLOCK_EXCL
| XFS_IOLOCK_SHARED
)) == 0);
410 /* reject inode numbers outside existing AGs */
411 if (!ino
|| XFS_INO_TO_AGNO(mp
, ino
) >= mp
->m_sb
.sb_agcount
)
414 /* get the perag structure and ensure that it's inode capable */
415 pag
= xfs_perag_get(mp
, XFS_INO_TO_AGNO(mp
, ino
));
416 agino
= XFS_INO_TO_AGINO(mp
, ino
);
421 ip
= radix_tree_lookup(&pag
->pag_ici_root
, agino
);
424 error
= xfs_iget_cache_hit(pag
, ip
, ino
, flags
, lock_flags
);
426 goto out_error_or_again
;
429 XFS_STATS_INC(xs_ig_missed
);
431 error
= xfs_iget_cache_miss(mp
, pag
, tp
, ino
, &ip
,
434 goto out_error_or_again
;
441 * If we have a real type for an on-disk inode, we can set ops(&unlock)
442 * now. If it's a new inode being created, xfs_ialloc will handle it.
444 if (xfs_iflags_test(ip
, XFS_INEW
) && ip
->i_d
.di_mode
!= 0)
449 if (error
== -EAGAIN
) {
458 * The inode lookup is done in batches to keep the amount of lock traffic and
459 * radix tree lookups to a minimum. The batch size is a trade off between
460 * lookup reduction and stack usage. This is in the reclaim path, so we can't
463 #define XFS_LOOKUP_BATCH 32
466 xfs_inode_ag_walk_grab(
467 struct xfs_inode
*ip
)
469 struct inode
*inode
= VFS_I(ip
);
471 ASSERT(rcu_read_lock_held());
474 * check for stale RCU freed inode
476 * If the inode has been reallocated, it doesn't matter if it's not in
477 * the AG we are walking - we are walking for writeback, so if it
478 * passes all the "valid inode" checks and is dirty, then we'll write
479 * it back anyway. If it has been reallocated and still being
480 * initialised, the XFS_INEW check below will catch it.
482 spin_lock(&ip
->i_flags_lock
);
484 goto out_unlock_noent
;
486 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
487 if (__xfs_iflags_test(ip
, XFS_INEW
| XFS_IRECLAIMABLE
| XFS_IRECLAIM
))
488 goto out_unlock_noent
;
489 spin_unlock(&ip
->i_flags_lock
);
491 /* nothing to sync during shutdown */
492 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
))
493 return -EFSCORRUPTED
;
495 /* If we can't grab the inode, it must on it's way to reclaim. */
503 spin_unlock(&ip
->i_flags_lock
);
509 struct xfs_mount
*mp
,
510 struct xfs_perag
*pag
,
511 int (*execute
)(struct xfs_inode
*ip
, int flags
,
517 uint32_t first_index
;
529 struct xfs_inode
*batch
[XFS_LOOKUP_BATCH
];
536 nr_found
= radix_tree_gang_lookup(&pag
->pag_ici_root
,
537 (void **)batch
, first_index
,
540 nr_found
= radix_tree_gang_lookup_tag(
542 (void **) batch
, first_index
,
543 XFS_LOOKUP_BATCH
, tag
);
551 * Grab the inodes before we drop the lock. if we found
552 * nothing, nr == 0 and the loop will be skipped.
554 for (i
= 0; i
< nr_found
; i
++) {
555 struct xfs_inode
*ip
= batch
[i
];
557 if (done
|| xfs_inode_ag_walk_grab(ip
))
561 * Update the index for the next lookup. Catch
562 * overflows into the next AG range which can occur if
563 * we have inodes in the last block of the AG and we
564 * are currently pointing to the last inode.
566 * Because we may see inodes that are from the wrong AG
567 * due to RCU freeing and reallocation, only update the
568 * index if it lies in this AG. It was a race that lead
569 * us to see this inode, so another lookup from the
570 * same index will not find it again.
572 if (XFS_INO_TO_AGNO(mp
, ip
->i_ino
) != pag
->pag_agno
)
574 first_index
= XFS_INO_TO_AGINO(mp
, ip
->i_ino
+ 1);
575 if (first_index
< XFS_INO_TO_AGINO(mp
, ip
->i_ino
))
579 /* unlock now we've grabbed the inodes. */
582 for (i
= 0; i
< nr_found
; i
++) {
585 error
= execute(batch
[i
], flags
, args
);
587 if (error
== -EAGAIN
) {
591 if (error
&& last_error
!= -EFSCORRUPTED
)
595 /* bail out if the filesystem is corrupted. */
596 if (error
== -EFSCORRUPTED
)
601 } while (nr_found
&& !done
);
611 * Background scanning to trim post-EOF preallocated space. This is queued
612 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
616 struct xfs_mount
*mp
)
619 if (radix_tree_tagged(&mp
->m_perag_tree
, XFS_ICI_EOFBLOCKS_TAG
))
620 queue_delayed_work(mp
->m_eofblocks_workqueue
,
621 &mp
->m_eofblocks_work
,
622 msecs_to_jiffies(xfs_eofb_secs
* 1000));
627 xfs_eofblocks_worker(
628 struct work_struct
*work
)
630 struct xfs_mount
*mp
= container_of(to_delayed_work(work
),
631 struct xfs_mount
, m_eofblocks_work
);
632 xfs_icache_free_eofblocks(mp
, NULL
);
633 xfs_queue_eofblocks(mp
);
637 xfs_inode_ag_iterator(
638 struct xfs_mount
*mp
,
639 int (*execute
)(struct xfs_inode
*ip
, int flags
,
644 struct xfs_perag
*pag
;
650 while ((pag
= xfs_perag_get(mp
, ag
))) {
651 ag
= pag
->pag_agno
+ 1;
652 error
= xfs_inode_ag_walk(mp
, pag
, execute
, flags
, args
, -1);
656 if (error
== -EFSCORRUPTED
)
664 xfs_inode_ag_iterator_tag(
665 struct xfs_mount
*mp
,
666 int (*execute
)(struct xfs_inode
*ip
, int flags
,
672 struct xfs_perag
*pag
;
678 while ((pag
= xfs_perag_get_tag(mp
, ag
, tag
))) {
679 ag
= pag
->pag_agno
+ 1;
680 error
= xfs_inode_ag_walk(mp
, pag
, execute
, flags
, args
, tag
);
684 if (error
== -EFSCORRUPTED
)
692 * Queue a new inode reclaim pass if there are reclaimable inodes and there
693 * isn't a reclaim pass already in progress. By default it runs every 5s based
694 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
695 * tunable, but that can be done if this method proves to be ineffective or too
699 xfs_reclaim_work_queue(
700 struct xfs_mount
*mp
)
704 if (radix_tree_tagged(&mp
->m_perag_tree
, XFS_ICI_RECLAIM_TAG
)) {
705 queue_delayed_work(mp
->m_reclaim_workqueue
, &mp
->m_reclaim_work
,
706 msecs_to_jiffies(xfs_syncd_centisecs
/ 6 * 10));
712 * This is a fast pass over the inode cache to try to get reclaim moving on as
713 * many inodes as possible in a short period of time. It kicks itself every few
714 * seconds, as well as being kicked by the inode cache shrinker when memory
715 * goes low. It scans as quickly as possible avoiding locked inodes or those
716 * already being flushed, and once done schedules a future pass.
720 struct work_struct
*work
)
722 struct xfs_mount
*mp
= container_of(to_delayed_work(work
),
723 struct xfs_mount
, m_reclaim_work
);
725 xfs_reclaim_inodes(mp
, SYNC_TRYLOCK
);
726 xfs_reclaim_work_queue(mp
);
730 __xfs_inode_set_reclaim_tag(
731 struct xfs_perag
*pag
,
732 struct xfs_inode
*ip
)
734 radix_tree_tag_set(&pag
->pag_ici_root
,
735 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
),
736 XFS_ICI_RECLAIM_TAG
);
738 if (!pag
->pag_ici_reclaimable
) {
739 /* propagate the reclaim tag up into the perag radix tree */
740 spin_lock(&ip
->i_mount
->m_perag_lock
);
741 radix_tree_tag_set(&ip
->i_mount
->m_perag_tree
,
742 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
743 XFS_ICI_RECLAIM_TAG
);
744 spin_unlock(&ip
->i_mount
->m_perag_lock
);
746 /* schedule periodic background inode reclaim */
747 xfs_reclaim_work_queue(ip
->i_mount
);
749 trace_xfs_perag_set_reclaim(ip
->i_mount
, pag
->pag_agno
,
752 pag
->pag_ici_reclaimable
++;
756 * We set the inode flag atomically with the radix tree tag.
757 * Once we get tag lookups on the radix tree, this inode flag
761 xfs_inode_set_reclaim_tag(
764 struct xfs_mount
*mp
= ip
->i_mount
;
765 struct xfs_perag
*pag
;
767 pag
= xfs_perag_get(mp
, XFS_INO_TO_AGNO(mp
, ip
->i_ino
));
768 spin_lock(&pag
->pag_ici_lock
);
769 spin_lock(&ip
->i_flags_lock
);
770 __xfs_inode_set_reclaim_tag(pag
, ip
);
771 __xfs_iflags_set(ip
, XFS_IRECLAIMABLE
);
772 spin_unlock(&ip
->i_flags_lock
);
773 spin_unlock(&pag
->pag_ici_lock
);
778 __xfs_inode_clear_reclaim(
782 pag
->pag_ici_reclaimable
--;
783 if (!pag
->pag_ici_reclaimable
) {
784 /* clear the reclaim tag from the perag radix tree */
785 spin_lock(&ip
->i_mount
->m_perag_lock
);
786 radix_tree_tag_clear(&ip
->i_mount
->m_perag_tree
,
787 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
788 XFS_ICI_RECLAIM_TAG
);
789 spin_unlock(&ip
->i_mount
->m_perag_lock
);
790 trace_xfs_perag_clear_reclaim(ip
->i_mount
, pag
->pag_agno
,
796 __xfs_inode_clear_reclaim_tag(
801 radix_tree_tag_clear(&pag
->pag_ici_root
,
802 XFS_INO_TO_AGINO(mp
, ip
->i_ino
), XFS_ICI_RECLAIM_TAG
);
803 __xfs_inode_clear_reclaim(pag
, ip
);
807 * Grab the inode for reclaim exclusively.
808 * Return 0 if we grabbed it, non-zero otherwise.
811 xfs_reclaim_inode_grab(
812 struct xfs_inode
*ip
,
815 ASSERT(rcu_read_lock_held());
817 /* quick check for stale RCU freed inode */
822 * If we are asked for non-blocking operation, do unlocked checks to
823 * see if the inode already is being flushed or in reclaim to avoid
826 if ((flags
& SYNC_TRYLOCK
) &&
827 __xfs_iflags_test(ip
, XFS_IFLOCK
| XFS_IRECLAIM
))
831 * The radix tree lock here protects a thread in xfs_iget from racing
832 * with us starting reclaim on the inode. Once we have the
833 * XFS_IRECLAIM flag set it will not touch us.
835 * Due to RCU lookup, we may find inodes that have been freed and only
836 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
837 * aren't candidates for reclaim at all, so we must check the
838 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
840 spin_lock(&ip
->i_flags_lock
);
841 if (!__xfs_iflags_test(ip
, XFS_IRECLAIMABLE
) ||
842 __xfs_iflags_test(ip
, XFS_IRECLAIM
)) {
843 /* not a reclaim candidate. */
844 spin_unlock(&ip
->i_flags_lock
);
847 __xfs_iflags_set(ip
, XFS_IRECLAIM
);
848 spin_unlock(&ip
->i_flags_lock
);
853 * Inodes in different states need to be treated differently. The following
854 * table lists the inode states and the reclaim actions necessary:
856 * inode state iflush ret required action
857 * --------------- ---------- ---------------
859 * shutdown EIO unpin and reclaim
860 * clean, unpinned 0 reclaim
861 * stale, unpinned 0 reclaim
862 * clean, pinned(*) 0 requeue
863 * stale, pinned EAGAIN requeue
864 * dirty, async - requeue
865 * dirty, sync 0 reclaim
867 * (*) dgc: I don't think the clean, pinned state is possible but it gets
868 * handled anyway given the order of checks implemented.
870 * Also, because we get the flush lock first, we know that any inode that has
871 * been flushed delwri has had the flush completed by the time we check that
872 * the inode is clean.
874 * Note that because the inode is flushed delayed write by AIL pushing, the
875 * flush lock may already be held here and waiting on it can result in very
876 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
877 * the caller should push the AIL first before trying to reclaim inodes to
878 * minimise the amount of time spent waiting. For background relaim, we only
879 * bother to reclaim clean inodes anyway.
881 * Hence the order of actions after gaining the locks should be:
883 * shutdown => unpin and reclaim
884 * pinned, async => requeue
885 * pinned, sync => unpin
888 * dirty, async => requeue
889 * dirty, sync => flush, wait and reclaim
893 struct xfs_inode
*ip
,
894 struct xfs_perag
*pag
,
897 struct xfs_buf
*bp
= NULL
;
902 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
903 if (!xfs_iflock_nowait(ip
)) {
904 if (!(sync_mode
& SYNC_WAIT
))
909 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
)) {
911 xfs_iflush_abort(ip
, false);
914 if (xfs_ipincount(ip
)) {
915 if (!(sync_mode
& SYNC_WAIT
))
919 if (xfs_iflags_test(ip
, XFS_ISTALE
))
921 if (xfs_inode_clean(ip
))
925 * Never flush out dirty data during non-blocking reclaim, as it would
926 * just contend with AIL pushing trying to do the same job.
928 if (!(sync_mode
& SYNC_WAIT
))
932 * Now we have an inode that needs flushing.
934 * Note that xfs_iflush will never block on the inode buffer lock, as
935 * xfs_ifree_cluster() can lock the inode buffer before it locks the
936 * ip->i_lock, and we are doing the exact opposite here. As a result,
937 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
938 * result in an ABBA deadlock with xfs_ifree_cluster().
940 * As xfs_ifree_cluser() must gather all inodes that are active in the
941 * cache to mark them stale, if we hit this case we don't actually want
942 * to do IO here - we want the inode marked stale so we can simply
943 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
944 * inode, back off and try again. Hopefully the next pass through will
945 * see the stale flag set on the inode.
947 error
= xfs_iflush(ip
, &bp
);
948 if (error
== -EAGAIN
) {
949 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
950 /* backoff longer than in xfs_ifree_cluster */
956 error
= xfs_bwrite(bp
);
963 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
965 XFS_STATS_INC(xs_ig_reclaims
);
967 * Remove the inode from the per-AG radix tree.
969 * Because radix_tree_delete won't complain even if the item was never
970 * added to the tree assert that it's been there before to catch
971 * problems with the inode life time early on.
973 spin_lock(&pag
->pag_ici_lock
);
974 if (!radix_tree_delete(&pag
->pag_ici_root
,
975 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
)))
977 __xfs_inode_clear_reclaim(pag
, ip
);
978 spin_unlock(&pag
->pag_ici_lock
);
981 * Here we do an (almost) spurious inode lock in order to coordinate
982 * with inode cache radix tree lookups. This is because the lookup
983 * can reference the inodes in the cache without taking references.
985 * We make that OK here by ensuring that we wait until the inode is
986 * unlocked after the lookup before we go ahead and free it.
988 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
990 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
998 xfs_iflags_clear(ip
, XFS_IRECLAIM
);
999 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
1001 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1002 * a short while. However, this just burns CPU time scanning the tree
1003 * waiting for IO to complete and the reclaim work never goes back to
1004 * the idle state. Instead, return 0 to let the next scheduled
1005 * background reclaim attempt to reclaim the inode again.
1011 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1012 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1013 * then a shut down during filesystem unmount reclaim walk leak all the
1014 * unreclaimed inodes.
1017 xfs_reclaim_inodes_ag(
1018 struct xfs_mount
*mp
,
1022 struct xfs_perag
*pag
;
1026 int trylock
= flags
& SYNC_TRYLOCK
;
1032 while ((pag
= xfs_perag_get_tag(mp
, ag
, XFS_ICI_RECLAIM_TAG
))) {
1033 unsigned long first_index
= 0;
1037 ag
= pag
->pag_agno
+ 1;
1040 if (!mutex_trylock(&pag
->pag_ici_reclaim_lock
)) {
1045 first_index
= pag
->pag_ici_reclaim_cursor
;
1047 mutex_lock(&pag
->pag_ici_reclaim_lock
);
1050 struct xfs_inode
*batch
[XFS_LOOKUP_BATCH
];
1054 nr_found
= radix_tree_gang_lookup_tag(
1056 (void **)batch
, first_index
,
1058 XFS_ICI_RECLAIM_TAG
);
1066 * Grab the inodes before we drop the lock. if we found
1067 * nothing, nr == 0 and the loop will be skipped.
1069 for (i
= 0; i
< nr_found
; i
++) {
1070 struct xfs_inode
*ip
= batch
[i
];
1072 if (done
|| xfs_reclaim_inode_grab(ip
, flags
))
1076 * Update the index for the next lookup. Catch
1077 * overflows into the next AG range which can
1078 * occur if we have inodes in the last block of
1079 * the AG and we are currently pointing to the
1082 * Because we may see inodes that are from the
1083 * wrong AG due to RCU freeing and
1084 * reallocation, only update the index if it
1085 * lies in this AG. It was a race that lead us
1086 * to see this inode, so another lookup from
1087 * the same index will not find it again.
1089 if (XFS_INO_TO_AGNO(mp
, ip
->i_ino
) !=
1092 first_index
= XFS_INO_TO_AGINO(mp
, ip
->i_ino
+ 1);
1093 if (first_index
< XFS_INO_TO_AGINO(mp
, ip
->i_ino
))
1097 /* unlock now we've grabbed the inodes. */
1100 for (i
= 0; i
< nr_found
; i
++) {
1103 error
= xfs_reclaim_inode(batch
[i
], pag
, flags
);
1104 if (error
&& last_error
!= -EFSCORRUPTED
)
1108 *nr_to_scan
-= XFS_LOOKUP_BATCH
;
1112 } while (nr_found
&& !done
&& *nr_to_scan
> 0);
1114 if (trylock
&& !done
)
1115 pag
->pag_ici_reclaim_cursor
= first_index
;
1117 pag
->pag_ici_reclaim_cursor
= 0;
1118 mutex_unlock(&pag
->pag_ici_reclaim_lock
);
1123 * if we skipped any AG, and we still have scan count remaining, do
1124 * another pass this time using blocking reclaim semantics (i.e
1125 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1126 * ensure that when we get more reclaimers than AGs we block rather
1127 * than spin trying to execute reclaim.
1129 if (skipped
&& (flags
& SYNC_WAIT
) && *nr_to_scan
> 0) {
1141 int nr_to_scan
= INT_MAX
;
1143 return xfs_reclaim_inodes_ag(mp
, mode
, &nr_to_scan
);
1147 * Scan a certain number of inodes for reclaim.
1149 * When called we make sure that there is a background (fast) inode reclaim in
1150 * progress, while we will throttle the speed of reclaim via doing synchronous
1151 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1152 * them to be cleaned, which we hope will not be very long due to the
1153 * background walker having already kicked the IO off on those dirty inodes.
1156 xfs_reclaim_inodes_nr(
1157 struct xfs_mount
*mp
,
1160 /* kick background reclaimer and push the AIL */
1161 xfs_reclaim_work_queue(mp
);
1162 xfs_ail_push_all(mp
->m_ail
);
1164 return xfs_reclaim_inodes_ag(mp
, SYNC_TRYLOCK
| SYNC_WAIT
, &nr_to_scan
);
1168 * Return the number of reclaimable inodes in the filesystem for
1169 * the shrinker to determine how much to reclaim.
1172 xfs_reclaim_inodes_count(
1173 struct xfs_mount
*mp
)
1175 struct xfs_perag
*pag
;
1176 xfs_agnumber_t ag
= 0;
1177 int reclaimable
= 0;
1179 while ((pag
= xfs_perag_get_tag(mp
, ag
, XFS_ICI_RECLAIM_TAG
))) {
1180 ag
= pag
->pag_agno
+ 1;
1181 reclaimable
+= pag
->pag_ici_reclaimable
;
1189 struct xfs_inode
*ip
,
1190 struct xfs_eofblocks
*eofb
)
1192 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_UID
) &&
1193 !uid_eq(VFS_I(ip
)->i_uid
, eofb
->eof_uid
))
1196 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_GID
) &&
1197 !gid_eq(VFS_I(ip
)->i_gid
, eofb
->eof_gid
))
1200 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_PRID
) &&
1201 xfs_get_projid(ip
) != eofb
->eof_prid
)
1208 * A union-based inode filtering algorithm. Process the inode if any of the
1209 * criteria match. This is for global/internal scans only.
1212 xfs_inode_match_id_union(
1213 struct xfs_inode
*ip
,
1214 struct xfs_eofblocks
*eofb
)
1216 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_UID
) &&
1217 uid_eq(VFS_I(ip
)->i_uid
, eofb
->eof_uid
))
1220 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_GID
) &&
1221 gid_eq(VFS_I(ip
)->i_gid
, eofb
->eof_gid
))
1224 if ((eofb
->eof_flags
& XFS_EOF_FLAGS_PRID
) &&
1225 xfs_get_projid(ip
) == eofb
->eof_prid
)
1232 xfs_inode_free_eofblocks(
1233 struct xfs_inode
*ip
,
1238 struct xfs_eofblocks
*eofb
= args
;
1239 bool need_iolock
= true;
1242 ASSERT(!eofb
|| (eofb
&& eofb
->eof_scan_owner
!= 0));
1244 if (!xfs_can_free_eofblocks(ip
, false)) {
1245 /* inode could be preallocated or append-only */
1246 trace_xfs_inode_free_eofblocks_invalid(ip
);
1247 xfs_inode_clear_eofblocks_tag(ip
);
1252 * If the mapping is dirty the operation can block and wait for some
1253 * time. Unless we are waiting, skip it.
1255 if (!(flags
& SYNC_WAIT
) &&
1256 mapping_tagged(VFS_I(ip
)->i_mapping
, PAGECACHE_TAG_DIRTY
))
1260 if (eofb
->eof_flags
& XFS_EOF_FLAGS_UNION
)
1261 match
= xfs_inode_match_id_union(ip
, eofb
);
1263 match
= xfs_inode_match_id(ip
, eofb
);
1267 /* skip the inode if the file size is too small */
1268 if (eofb
->eof_flags
& XFS_EOF_FLAGS_MINFILESIZE
&&
1269 XFS_ISIZE(ip
) < eofb
->eof_min_file_size
)
1273 * A scan owner implies we already hold the iolock. Skip it in
1274 * xfs_free_eofblocks() to avoid deadlock. This also eliminates
1275 * the possibility of EAGAIN being returned.
1277 if (eofb
->eof_scan_owner
== ip
->i_ino
)
1278 need_iolock
= false;
1281 ret
= xfs_free_eofblocks(ip
->i_mount
, ip
, need_iolock
);
1283 /* don't revisit the inode if we're not waiting */
1284 if (ret
== -EAGAIN
&& !(flags
& SYNC_WAIT
))
1291 xfs_icache_free_eofblocks(
1292 struct xfs_mount
*mp
,
1293 struct xfs_eofblocks
*eofb
)
1295 int flags
= SYNC_TRYLOCK
;
1297 if (eofb
&& (eofb
->eof_flags
& XFS_EOF_FLAGS_SYNC
))
1300 return xfs_inode_ag_iterator_tag(mp
, xfs_inode_free_eofblocks
, flags
,
1301 eofb
, XFS_ICI_EOFBLOCKS_TAG
);
1305 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1306 * multiple quotas, we don't know exactly which quota caused an allocation
1307 * failure. We make a best effort by including each quota under low free space
1308 * conditions (less than 1% free space) in the scan.
1311 xfs_inode_free_quota_eofblocks(
1312 struct xfs_inode
*ip
)
1315 struct xfs_eofblocks eofb
= {0};
1316 struct xfs_dquot
*dq
;
1318 ASSERT(xfs_isilocked(ip
, XFS_IOLOCK_EXCL
));
1321 * Set the scan owner to avoid a potential livelock. Otherwise, the scan
1322 * can repeatedly trylock on the inode we're currently processing. We
1323 * run a sync scan to increase effectiveness and use the union filter to
1324 * cover all applicable quotas in a single scan.
1326 eofb
.eof_scan_owner
= ip
->i_ino
;
1327 eofb
.eof_flags
= XFS_EOF_FLAGS_UNION
|XFS_EOF_FLAGS_SYNC
;
1329 if (XFS_IS_UQUOTA_ENFORCED(ip
->i_mount
)) {
1330 dq
= xfs_inode_dquot(ip
, XFS_DQ_USER
);
1331 if (dq
&& xfs_dquot_lowsp(dq
)) {
1332 eofb
.eof_uid
= VFS_I(ip
)->i_uid
;
1333 eofb
.eof_flags
|= XFS_EOF_FLAGS_UID
;
1338 if (XFS_IS_GQUOTA_ENFORCED(ip
->i_mount
)) {
1339 dq
= xfs_inode_dquot(ip
, XFS_DQ_GROUP
);
1340 if (dq
&& xfs_dquot_lowsp(dq
)) {
1341 eofb
.eof_gid
= VFS_I(ip
)->i_gid
;
1342 eofb
.eof_flags
|= XFS_EOF_FLAGS_GID
;
1348 xfs_icache_free_eofblocks(ip
->i_mount
, &eofb
);
1354 xfs_inode_set_eofblocks_tag(
1357 struct xfs_mount
*mp
= ip
->i_mount
;
1358 struct xfs_perag
*pag
;
1361 pag
= xfs_perag_get(mp
, XFS_INO_TO_AGNO(mp
, ip
->i_ino
));
1362 spin_lock(&pag
->pag_ici_lock
);
1363 trace_xfs_inode_set_eofblocks_tag(ip
);
1365 tagged
= radix_tree_tagged(&pag
->pag_ici_root
,
1366 XFS_ICI_EOFBLOCKS_TAG
);
1367 radix_tree_tag_set(&pag
->pag_ici_root
,
1368 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
),
1369 XFS_ICI_EOFBLOCKS_TAG
);
1371 /* propagate the eofblocks tag up into the perag radix tree */
1372 spin_lock(&ip
->i_mount
->m_perag_lock
);
1373 radix_tree_tag_set(&ip
->i_mount
->m_perag_tree
,
1374 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
1375 XFS_ICI_EOFBLOCKS_TAG
);
1376 spin_unlock(&ip
->i_mount
->m_perag_lock
);
1378 /* kick off background trimming */
1379 xfs_queue_eofblocks(ip
->i_mount
);
1381 trace_xfs_perag_set_eofblocks(ip
->i_mount
, pag
->pag_agno
,
1385 spin_unlock(&pag
->pag_ici_lock
);
1390 xfs_inode_clear_eofblocks_tag(
1393 struct xfs_mount
*mp
= ip
->i_mount
;
1394 struct xfs_perag
*pag
;
1396 pag
= xfs_perag_get(mp
, XFS_INO_TO_AGNO(mp
, ip
->i_ino
));
1397 spin_lock(&pag
->pag_ici_lock
);
1398 trace_xfs_inode_clear_eofblocks_tag(ip
);
1400 radix_tree_tag_clear(&pag
->pag_ici_root
,
1401 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
),
1402 XFS_ICI_EOFBLOCKS_TAG
);
1403 if (!radix_tree_tagged(&pag
->pag_ici_root
, XFS_ICI_EOFBLOCKS_TAG
)) {
1404 /* clear the eofblocks tag from the perag radix tree */
1405 spin_lock(&ip
->i_mount
->m_perag_lock
);
1406 radix_tree_tag_clear(&ip
->i_mount
->m_perag_tree
,
1407 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
1408 XFS_ICI_EOFBLOCKS_TAG
);
1409 spin_unlock(&ip
->i_mount
->m_perag_lock
);
1410 trace_xfs_perag_clear_eofblocks(ip
->i_mount
, pag
->pag_agno
,
1414 spin_unlock(&pag
->pag_ici_lock
);