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_types.h"
22 #include "xfs_log_priv.h"
24 #include "xfs_trans.h"
25 #include "xfs_trans_priv.h"
28 #include "xfs_mount.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_inode.h"
31 #include "xfs_dinode.h"
32 #include "xfs_error.h"
33 #include "xfs_filestream.h"
34 #include "xfs_vnodeops.h"
35 #include "xfs_inode_item.h"
36 #include "xfs_quota.h"
37 #include "xfs_trace.h"
38 #include "xfs_fsops.h"
40 #include <linux/kthread.h>
41 #include <linux/freezer.h>
44 * The inode lookup is done in batches to keep the amount of lock traffic and
45 * radix tree lookups to a minimum. The batch size is a trade off between
46 * lookup reduction and stack usage. This is in the reclaim path, so we can't
49 #define XFS_LOOKUP_BATCH 32
52 xfs_inode_ag_walk_grab(
55 struct inode
*inode
= VFS_I(ip
);
57 ASSERT(rcu_read_lock_held());
60 * check for stale RCU freed inode
62 * If the inode has been reallocated, it doesn't matter if it's not in
63 * the AG we are walking - we are walking for writeback, so if it
64 * passes all the "valid inode" checks and is dirty, then we'll write
65 * it back anyway. If it has been reallocated and still being
66 * initialised, the XFS_INEW check below will catch it.
68 spin_lock(&ip
->i_flags_lock
);
70 goto out_unlock_noent
;
72 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
73 if (__xfs_iflags_test(ip
, XFS_INEW
| XFS_IRECLAIMABLE
| XFS_IRECLAIM
))
74 goto out_unlock_noent
;
75 spin_unlock(&ip
->i_flags_lock
);
77 /* nothing to sync during shutdown */
78 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
))
81 /* If we can't grab the inode, it must on it's way to reclaim. */
85 if (is_bad_inode(inode
)) {
94 spin_unlock(&ip
->i_flags_lock
);
100 struct xfs_mount
*mp
,
101 struct xfs_perag
*pag
,
102 int (*execute
)(struct xfs_inode
*ip
,
103 struct xfs_perag
*pag
, int flags
),
106 uint32_t first_index
;
118 struct xfs_inode
*batch
[XFS_LOOKUP_BATCH
];
123 nr_found
= radix_tree_gang_lookup(&pag
->pag_ici_root
,
124 (void **)batch
, first_index
,
132 * Grab the inodes before we drop the lock. if we found
133 * nothing, nr == 0 and the loop will be skipped.
135 for (i
= 0; i
< nr_found
; i
++) {
136 struct xfs_inode
*ip
= batch
[i
];
138 if (done
|| xfs_inode_ag_walk_grab(ip
))
142 * Update the index for the next lookup. Catch
143 * overflows into the next AG range which can occur if
144 * we have inodes in the last block of the AG and we
145 * are currently pointing to the last inode.
147 * Because we may see inodes that are from the wrong AG
148 * due to RCU freeing and reallocation, only update the
149 * index if it lies in this AG. It was a race that lead
150 * us to see this inode, so another lookup from the
151 * same index will not find it again.
153 if (XFS_INO_TO_AGNO(mp
, ip
->i_ino
) != pag
->pag_agno
)
155 first_index
= XFS_INO_TO_AGINO(mp
, ip
->i_ino
+ 1);
156 if (first_index
< XFS_INO_TO_AGINO(mp
, ip
->i_ino
))
160 /* unlock now we've grabbed the inodes. */
163 for (i
= 0; i
< nr_found
; i
++) {
166 error
= execute(batch
[i
], pag
, flags
);
168 if (error
== EAGAIN
) {
172 if (error
&& last_error
!= EFSCORRUPTED
)
176 /* bail out if the filesystem is corrupted. */
177 if (error
== EFSCORRUPTED
)
182 } while (nr_found
&& !done
);
192 xfs_inode_ag_iterator(
193 struct xfs_mount
*mp
,
194 int (*execute
)(struct xfs_inode
*ip
,
195 struct xfs_perag
*pag
, int flags
),
198 struct xfs_perag
*pag
;
204 while ((pag
= xfs_perag_get(mp
, ag
))) {
205 ag
= pag
->pag_agno
+ 1;
206 error
= xfs_inode_ag_walk(mp
, pag
, execute
, flags
);
210 if (error
== EFSCORRUPTED
)
214 return XFS_ERROR(last_error
);
219 struct xfs_mount
*mp
)
225 * If the buffer is pinned then push on the log so we won't get stuck
226 * waiting in the write for someone, maybe ourselves, to flush the log.
228 * Even though we just pushed the log above, we did not have the
229 * superblock buffer locked at that point so it can become pinned in
230 * between there and here.
232 bp
= xfs_getsb(mp
, 0);
233 if (xfs_buf_ispinned(bp
))
234 xfs_log_force(mp
, 0);
235 error
= xfs_bwrite(bp
);
241 * When remounting a filesystem read-only or freezing the filesystem, we have
242 * two phases to execute. This first phase is syncing the data before we
243 * quiesce the filesystem, and the second is flushing all the inodes out after
244 * we've waited for all the transactions created by the first phase to
245 * complete. The second phase ensures that the inodes are written to their
246 * location on disk rather than just existing in transactions in the log. This
247 * means after a quiesce there is no log replay required to write the inodes to
248 * disk (this is the main difference between a sync and a quiesce).
251 * First stage of freeze - no writers will make progress now we are here,
252 * so we flush delwri and delalloc buffers here, then wait for all I/O to
253 * complete. Data is frozen at that point. Metadata is not frozen,
254 * transactions can still occur here so don't bother emptying the AIL
255 * because it'll just get dirty again.
259 struct xfs_mount
*mp
)
261 int error
, error2
= 0;
263 /* force out the log */
264 xfs_log_force(mp
, XFS_LOG_SYNC
);
266 /* write superblock and hoover up shutdown errors */
267 error
= xfs_sync_fsdata(mp
);
269 /* mark the log as covered if needed */
270 if (xfs_log_need_covered(mp
))
271 error2
= xfs_fs_log_dummy(mp
);
273 return error
? error
: error2
;
277 * Second stage of a quiesce. The data is already synced, now we have to take
278 * care of the metadata. New transactions are already blocked, so we need to
279 * wait for any remaining transactions to drain out before proceeding.
281 * Note: this stops background sync work - the callers must ensure it is started
282 * again when appropriate.
286 struct xfs_mount
*mp
)
290 /* wait for all modifications to complete */
291 while (atomic_read(&mp
->m_active_trans
) > 0)
294 /* reclaim inodes to do any IO before the freeze completes */
295 xfs_reclaim_inodes(mp
, 0);
296 xfs_reclaim_inodes(mp
, SYNC_WAIT
);
298 /* flush all pending changes from the AIL */
299 xfs_ail_push_all_sync(mp
->m_ail
);
301 /* stop background log work */
302 cancel_delayed_work_sync(&mp
->m_log
->l_work
);
305 * Just warn here till VFS can correctly support
306 * read-only remount without racing.
308 WARN_ON(atomic_read(&mp
->m_active_trans
) != 0);
310 /* Push the superblock and write an unmount record */
311 error
= xfs_log_sbcount(mp
);
313 xfs_warn(mp
, "xfs_attr_quiesce: failed to log sb changes. "
314 "Frozen image may not be consistent.");
315 xfs_log_unmount_write(mp
);
318 * At this point we might have modified the superblock again and thus
319 * added an item to the AIL, thus flush it again.
321 xfs_ail_push_all_sync(mp
->m_ail
);
324 * The superblock buffer is uncached and xfsaild_push() will lock and
325 * set the XBF_ASYNC flag on the buffer. We cannot do xfs_buf_iowait()
326 * here but a lock on the superblock buffer will block until iodone()
329 xfs_buf_lock(mp
->m_sb_bp
);
330 xfs_buf_unlock(mp
->m_sb_bp
);
334 * Queue a new inode reclaim pass if there are reclaimable inodes and there
335 * isn't a reclaim pass already in progress. By default it runs every 5s based
336 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
337 * tunable, but that can be done if this method proves to be ineffective or too
341 xfs_reclaim_work_queue(
342 struct xfs_mount
*mp
)
346 if (radix_tree_tagged(&mp
->m_perag_tree
, XFS_ICI_RECLAIM_TAG
)) {
347 queue_delayed_work(mp
->m_reclaim_workqueue
, &mp
->m_reclaim_work
,
348 msecs_to_jiffies(xfs_syncd_centisecs
/ 6 * 10));
354 * This is a fast pass over the inode cache to try to get reclaim moving on as
355 * many inodes as possible in a short period of time. It kicks itself every few
356 * seconds, as well as being kicked by the inode cache shrinker when memory
357 * goes low. It scans as quickly as possible avoiding locked inodes or those
358 * already being flushed, and once done schedules a future pass.
362 struct work_struct
*work
)
364 struct xfs_mount
*mp
= container_of(to_delayed_work(work
),
365 struct xfs_mount
, m_reclaim_work
);
367 xfs_reclaim_inodes(mp
, SYNC_TRYLOCK
);
368 xfs_reclaim_work_queue(mp
);
372 __xfs_inode_set_reclaim_tag(
373 struct xfs_perag
*pag
,
374 struct xfs_inode
*ip
)
376 radix_tree_tag_set(&pag
->pag_ici_root
,
377 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
),
378 XFS_ICI_RECLAIM_TAG
);
380 if (!pag
->pag_ici_reclaimable
) {
381 /* propagate the reclaim tag up into the perag radix tree */
382 spin_lock(&ip
->i_mount
->m_perag_lock
);
383 radix_tree_tag_set(&ip
->i_mount
->m_perag_tree
,
384 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
385 XFS_ICI_RECLAIM_TAG
);
386 spin_unlock(&ip
->i_mount
->m_perag_lock
);
388 /* schedule periodic background inode reclaim */
389 xfs_reclaim_work_queue(ip
->i_mount
);
391 trace_xfs_perag_set_reclaim(ip
->i_mount
, pag
->pag_agno
,
394 pag
->pag_ici_reclaimable
++;
398 * We set the inode flag atomically with the radix tree tag.
399 * Once we get tag lookups on the radix tree, this inode flag
403 xfs_inode_set_reclaim_tag(
406 struct xfs_mount
*mp
= ip
->i_mount
;
407 struct xfs_perag
*pag
;
409 pag
= xfs_perag_get(mp
, XFS_INO_TO_AGNO(mp
, ip
->i_ino
));
410 spin_lock(&pag
->pag_ici_lock
);
411 spin_lock(&ip
->i_flags_lock
);
412 __xfs_inode_set_reclaim_tag(pag
, ip
);
413 __xfs_iflags_set(ip
, XFS_IRECLAIMABLE
);
414 spin_unlock(&ip
->i_flags_lock
);
415 spin_unlock(&pag
->pag_ici_lock
);
420 __xfs_inode_clear_reclaim(
424 pag
->pag_ici_reclaimable
--;
425 if (!pag
->pag_ici_reclaimable
) {
426 /* clear the reclaim tag from the perag radix tree */
427 spin_lock(&ip
->i_mount
->m_perag_lock
);
428 radix_tree_tag_clear(&ip
->i_mount
->m_perag_tree
,
429 XFS_INO_TO_AGNO(ip
->i_mount
, ip
->i_ino
),
430 XFS_ICI_RECLAIM_TAG
);
431 spin_unlock(&ip
->i_mount
->m_perag_lock
);
432 trace_xfs_perag_clear_reclaim(ip
->i_mount
, pag
->pag_agno
,
438 __xfs_inode_clear_reclaim_tag(
443 radix_tree_tag_clear(&pag
->pag_ici_root
,
444 XFS_INO_TO_AGINO(mp
, ip
->i_ino
), XFS_ICI_RECLAIM_TAG
);
445 __xfs_inode_clear_reclaim(pag
, ip
);
449 * Grab the inode for reclaim exclusively.
450 * Return 0 if we grabbed it, non-zero otherwise.
453 xfs_reclaim_inode_grab(
454 struct xfs_inode
*ip
,
457 ASSERT(rcu_read_lock_held());
459 /* quick check for stale RCU freed inode */
464 * If we are asked for non-blocking operation, do unlocked checks to
465 * see if the inode already is being flushed or in reclaim to avoid
468 if ((flags
& SYNC_TRYLOCK
) &&
469 __xfs_iflags_test(ip
, XFS_IFLOCK
| XFS_IRECLAIM
))
473 * The radix tree lock here protects a thread in xfs_iget from racing
474 * with us starting reclaim on the inode. Once we have the
475 * XFS_IRECLAIM flag set it will not touch us.
477 * Due to RCU lookup, we may find inodes that have been freed and only
478 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
479 * aren't candidates for reclaim at all, so we must check the
480 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
482 spin_lock(&ip
->i_flags_lock
);
483 if (!__xfs_iflags_test(ip
, XFS_IRECLAIMABLE
) ||
484 __xfs_iflags_test(ip
, XFS_IRECLAIM
)) {
485 /* not a reclaim candidate. */
486 spin_unlock(&ip
->i_flags_lock
);
489 __xfs_iflags_set(ip
, XFS_IRECLAIM
);
490 spin_unlock(&ip
->i_flags_lock
);
495 * Inodes in different states need to be treated differently. The following
496 * table lists the inode states and the reclaim actions necessary:
498 * inode state iflush ret required action
499 * --------------- ---------- ---------------
501 * shutdown EIO unpin and reclaim
502 * clean, unpinned 0 reclaim
503 * stale, unpinned 0 reclaim
504 * clean, pinned(*) 0 requeue
505 * stale, pinned EAGAIN requeue
506 * dirty, async - requeue
507 * dirty, sync 0 reclaim
509 * (*) dgc: I don't think the clean, pinned state is possible but it gets
510 * handled anyway given the order of checks implemented.
512 * Also, because we get the flush lock first, we know that any inode that has
513 * been flushed delwri has had the flush completed by the time we check that
514 * the inode is clean.
516 * Note that because the inode is flushed delayed write by AIL pushing, the
517 * flush lock may already be held here and waiting on it can result in very
518 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
519 * the caller should push the AIL first before trying to reclaim inodes to
520 * minimise the amount of time spent waiting. For background relaim, we only
521 * bother to reclaim clean inodes anyway.
523 * Hence the order of actions after gaining the locks should be:
525 * shutdown => unpin and reclaim
526 * pinned, async => requeue
527 * pinned, sync => unpin
530 * dirty, async => requeue
531 * dirty, sync => flush, wait and reclaim
535 struct xfs_inode
*ip
,
536 struct xfs_perag
*pag
,
539 struct xfs_buf
*bp
= NULL
;
544 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
545 if (!xfs_iflock_nowait(ip
)) {
546 if (!(sync_mode
& SYNC_WAIT
))
551 if (is_bad_inode(VFS_I(ip
)))
553 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
)) {
555 xfs_iflush_abort(ip
, false);
558 if (xfs_ipincount(ip
)) {
559 if (!(sync_mode
& SYNC_WAIT
))
563 if (xfs_iflags_test(ip
, XFS_ISTALE
))
565 if (xfs_inode_clean(ip
))
569 * Never flush out dirty data during non-blocking reclaim, as it would
570 * just contend with AIL pushing trying to do the same job.
572 if (!(sync_mode
& SYNC_WAIT
))
576 * Now we have an inode that needs flushing.
578 * Note that xfs_iflush will never block on the inode buffer lock, as
579 * xfs_ifree_cluster() can lock the inode buffer before it locks the
580 * ip->i_lock, and we are doing the exact opposite here. As a result,
581 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
582 * result in an ABBA deadlock with xfs_ifree_cluster().
584 * As xfs_ifree_cluser() must gather all inodes that are active in the
585 * cache to mark them stale, if we hit this case we don't actually want
586 * to do IO here - we want the inode marked stale so we can simply
587 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
588 * inode, back off and try again. Hopefully the next pass through will
589 * see the stale flag set on the inode.
591 error
= xfs_iflush(ip
, &bp
);
592 if (error
== EAGAIN
) {
593 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
594 /* backoff longer than in xfs_ifree_cluster */
600 error
= xfs_bwrite(bp
);
607 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
609 XFS_STATS_INC(xs_ig_reclaims
);
611 * Remove the inode from the per-AG radix tree.
613 * Because radix_tree_delete won't complain even if the item was never
614 * added to the tree assert that it's been there before to catch
615 * problems with the inode life time early on.
617 spin_lock(&pag
->pag_ici_lock
);
618 if (!radix_tree_delete(&pag
->pag_ici_root
,
619 XFS_INO_TO_AGINO(ip
->i_mount
, ip
->i_ino
)))
621 __xfs_inode_clear_reclaim(pag
, ip
);
622 spin_unlock(&pag
->pag_ici_lock
);
625 * Here we do an (almost) spurious inode lock in order to coordinate
626 * with inode cache radix tree lookups. This is because the lookup
627 * can reference the inodes in the cache without taking references.
629 * We make that OK here by ensuring that we wait until the inode is
630 * unlocked after the lookup before we go ahead and free it.
632 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
634 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
642 xfs_iflags_clear(ip
, XFS_IRECLAIM
);
643 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
645 * We could return EAGAIN here to make reclaim rescan the inode tree in
646 * a short while. However, this just burns CPU time scanning the tree
647 * waiting for IO to complete and the reclaim work never goes back to
648 * the idle state. Instead, return 0 to let the next scheduled
649 * background reclaim attempt to reclaim the inode again.
655 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
656 * corrupted, we still want to try to reclaim all the inodes. If we don't,
657 * then a shut down during filesystem unmount reclaim walk leak all the
658 * unreclaimed inodes.
661 xfs_reclaim_inodes_ag(
662 struct xfs_mount
*mp
,
666 struct xfs_perag
*pag
;
670 int trylock
= flags
& SYNC_TRYLOCK
;
676 while ((pag
= xfs_perag_get_tag(mp
, ag
, XFS_ICI_RECLAIM_TAG
))) {
677 unsigned long first_index
= 0;
681 ag
= pag
->pag_agno
+ 1;
684 if (!mutex_trylock(&pag
->pag_ici_reclaim_lock
)) {
689 first_index
= pag
->pag_ici_reclaim_cursor
;
691 mutex_lock(&pag
->pag_ici_reclaim_lock
);
694 struct xfs_inode
*batch
[XFS_LOOKUP_BATCH
];
698 nr_found
= radix_tree_gang_lookup_tag(
700 (void **)batch
, first_index
,
702 XFS_ICI_RECLAIM_TAG
);
710 * Grab the inodes before we drop the lock. if we found
711 * nothing, nr == 0 and the loop will be skipped.
713 for (i
= 0; i
< nr_found
; i
++) {
714 struct xfs_inode
*ip
= batch
[i
];
716 if (done
|| xfs_reclaim_inode_grab(ip
, flags
))
720 * Update the index for the next lookup. Catch
721 * overflows into the next AG range which can
722 * occur if we have inodes in the last block of
723 * the AG and we are currently pointing to the
726 * Because we may see inodes that are from the
727 * wrong AG due to RCU freeing and
728 * reallocation, only update the index if it
729 * lies in this AG. It was a race that lead us
730 * to see this inode, so another lookup from
731 * the same index will not find it again.
733 if (XFS_INO_TO_AGNO(mp
, ip
->i_ino
) !=
736 first_index
= XFS_INO_TO_AGINO(mp
, ip
->i_ino
+ 1);
737 if (first_index
< XFS_INO_TO_AGINO(mp
, ip
->i_ino
))
741 /* unlock now we've grabbed the inodes. */
744 for (i
= 0; i
< nr_found
; i
++) {
747 error
= xfs_reclaim_inode(batch
[i
], pag
, flags
);
748 if (error
&& last_error
!= EFSCORRUPTED
)
752 *nr_to_scan
-= XFS_LOOKUP_BATCH
;
756 } while (nr_found
&& !done
&& *nr_to_scan
> 0);
758 if (trylock
&& !done
)
759 pag
->pag_ici_reclaim_cursor
= first_index
;
761 pag
->pag_ici_reclaim_cursor
= 0;
762 mutex_unlock(&pag
->pag_ici_reclaim_lock
);
767 * if we skipped any AG, and we still have scan count remaining, do
768 * another pass this time using blocking reclaim semantics (i.e
769 * waiting on the reclaim locks and ignoring the reclaim cursors). This
770 * ensure that when we get more reclaimers than AGs we block rather
771 * than spin trying to execute reclaim.
773 if (skipped
&& (flags
& SYNC_WAIT
) && *nr_to_scan
> 0) {
777 return XFS_ERROR(last_error
);
785 int nr_to_scan
= INT_MAX
;
787 return xfs_reclaim_inodes_ag(mp
, mode
, &nr_to_scan
);
791 * Scan a certain number of inodes for reclaim.
793 * When called we make sure that there is a background (fast) inode reclaim in
794 * progress, while we will throttle the speed of reclaim via doing synchronous
795 * reclaim of inodes. That means if we come across dirty inodes, we wait for
796 * them to be cleaned, which we hope will not be very long due to the
797 * background walker having already kicked the IO off on those dirty inodes.
800 xfs_reclaim_inodes_nr(
801 struct xfs_mount
*mp
,
804 /* kick background reclaimer and push the AIL */
805 xfs_reclaim_work_queue(mp
);
806 xfs_ail_push_all(mp
->m_ail
);
808 xfs_reclaim_inodes_ag(mp
, SYNC_TRYLOCK
| SYNC_WAIT
, &nr_to_scan
);
812 * Return the number of reclaimable inodes in the filesystem for
813 * the shrinker to determine how much to reclaim.
816 xfs_reclaim_inodes_count(
817 struct xfs_mount
*mp
)
819 struct xfs_perag
*pag
;
820 xfs_agnumber_t ag
= 0;
823 while ((pag
= xfs_perag_get_tag(mp
, ag
, XFS_ICI_RECLAIM_TAG
))) {
824 ag
= pag
->pag_agno
+ 1;
825 reclaimable
+= pag
->pag_ici_reclaimable
;
This page took 0.047851 seconds and 5 git commands to generate.