Loading...
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
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.
8 *
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.
13 *
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
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_types.h"
21#include "xfs_bit.h"
22#include "xfs_log.h"
23#include "xfs_inum.h"
24#include "xfs_trans.h"
25#include "xfs_trans_priv.h"
26#include "xfs_sb.h"
27#include "xfs_ag.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"
39
40#include <linux/kthread.h>
41#include <linux/freezer.h>
42
43struct workqueue_struct *xfs_syncd_wq; /* sync workqueue */
44
45/*
46 * The inode lookup is done in batches to keep the amount of lock traffic and
47 * radix tree lookups to a minimum. The batch size is a trade off between
48 * lookup reduction and stack usage. This is in the reclaim path, so we can't
49 * be too greedy.
50 */
51#define XFS_LOOKUP_BATCH 32
52
53STATIC int
54xfs_inode_ag_walk_grab(
55 struct xfs_inode *ip)
56{
57 struct inode *inode = VFS_I(ip);
58
59 ASSERT(rcu_read_lock_held());
60
61 /*
62 * check for stale RCU freed inode
63 *
64 * If the inode has been reallocated, it doesn't matter if it's not in
65 * the AG we are walking - we are walking for writeback, so if it
66 * passes all the "valid inode" checks and is dirty, then we'll write
67 * it back anyway. If it has been reallocated and still being
68 * initialised, the XFS_INEW check below will catch it.
69 */
70 spin_lock(&ip->i_flags_lock);
71 if (!ip->i_ino)
72 goto out_unlock_noent;
73
74 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
75 if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM))
76 goto out_unlock_noent;
77 spin_unlock(&ip->i_flags_lock);
78
79 /* nothing to sync during shutdown */
80 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
81 return EFSCORRUPTED;
82
83 /* If we can't grab the inode, it must on it's way to reclaim. */
84 if (!igrab(inode))
85 return ENOENT;
86
87 if (is_bad_inode(inode)) {
88 IRELE(ip);
89 return ENOENT;
90 }
91
92 /* inode is valid */
93 return 0;
94
95out_unlock_noent:
96 spin_unlock(&ip->i_flags_lock);
97 return ENOENT;
98}
99
100STATIC int
101xfs_inode_ag_walk(
102 struct xfs_mount *mp,
103 struct xfs_perag *pag,
104 int (*execute)(struct xfs_inode *ip,
105 struct xfs_perag *pag, int flags),
106 int flags)
107{
108 uint32_t first_index;
109 int last_error = 0;
110 int skipped;
111 int done;
112 int nr_found;
113
114restart:
115 done = 0;
116 skipped = 0;
117 first_index = 0;
118 nr_found = 0;
119 do {
120 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
121 int error = 0;
122 int i;
123
124 rcu_read_lock();
125 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
126 (void **)batch, first_index,
127 XFS_LOOKUP_BATCH);
128 if (!nr_found) {
129 rcu_read_unlock();
130 break;
131 }
132
133 /*
134 * Grab the inodes before we drop the lock. if we found
135 * nothing, nr == 0 and the loop will be skipped.
136 */
137 for (i = 0; i < nr_found; i++) {
138 struct xfs_inode *ip = batch[i];
139
140 if (done || xfs_inode_ag_walk_grab(ip))
141 batch[i] = NULL;
142
143 /*
144 * Update the index for the next lookup. Catch
145 * overflows into the next AG range which can occur if
146 * we have inodes in the last block of the AG and we
147 * are currently pointing to the last inode.
148 *
149 * Because we may see inodes that are from the wrong AG
150 * due to RCU freeing and reallocation, only update the
151 * index if it lies in this AG. It was a race that lead
152 * us to see this inode, so another lookup from the
153 * same index will not find it again.
154 */
155 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
156 continue;
157 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
158 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
159 done = 1;
160 }
161
162 /* unlock now we've grabbed the inodes. */
163 rcu_read_unlock();
164
165 for (i = 0; i < nr_found; i++) {
166 if (!batch[i])
167 continue;
168 error = execute(batch[i], pag, flags);
169 IRELE(batch[i]);
170 if (error == EAGAIN) {
171 skipped++;
172 continue;
173 }
174 if (error && last_error != EFSCORRUPTED)
175 last_error = error;
176 }
177
178 /* bail out if the filesystem is corrupted. */
179 if (error == EFSCORRUPTED)
180 break;
181
182 cond_resched();
183
184 } while (nr_found && !done);
185
186 if (skipped) {
187 delay(1);
188 goto restart;
189 }
190 return last_error;
191}
192
193int
194xfs_inode_ag_iterator(
195 struct xfs_mount *mp,
196 int (*execute)(struct xfs_inode *ip,
197 struct xfs_perag *pag, int flags),
198 int flags)
199{
200 struct xfs_perag *pag;
201 int error = 0;
202 int last_error = 0;
203 xfs_agnumber_t ag;
204
205 ag = 0;
206 while ((pag = xfs_perag_get(mp, ag))) {
207 ag = pag->pag_agno + 1;
208 error = xfs_inode_ag_walk(mp, pag, execute, flags);
209 xfs_perag_put(pag);
210 if (error) {
211 last_error = error;
212 if (error == EFSCORRUPTED)
213 break;
214 }
215 }
216 return XFS_ERROR(last_error);
217}
218
219STATIC int
220xfs_sync_inode_data(
221 struct xfs_inode *ip,
222 struct xfs_perag *pag,
223 int flags)
224{
225 struct inode *inode = VFS_I(ip);
226 struct address_space *mapping = inode->i_mapping;
227 int error = 0;
228
229 if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
230 goto out_wait;
231
232 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED)) {
233 if (flags & SYNC_TRYLOCK)
234 goto out_wait;
235 xfs_ilock(ip, XFS_IOLOCK_SHARED);
236 }
237
238 error = xfs_flush_pages(ip, 0, -1, (flags & SYNC_WAIT) ?
239 0 : XBF_ASYNC, FI_NONE);
240 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
241
242 out_wait:
243 if (flags & SYNC_WAIT)
244 xfs_ioend_wait(ip);
245 return error;
246}
247
248STATIC int
249xfs_sync_inode_attr(
250 struct xfs_inode *ip,
251 struct xfs_perag *pag,
252 int flags)
253{
254 int error = 0;
255
256 xfs_ilock(ip, XFS_ILOCK_SHARED);
257 if (xfs_inode_clean(ip))
258 goto out_unlock;
259 if (!xfs_iflock_nowait(ip)) {
260 if (!(flags & SYNC_WAIT))
261 goto out_unlock;
262 xfs_iflock(ip);
263 }
264
265 if (xfs_inode_clean(ip)) {
266 xfs_ifunlock(ip);
267 goto out_unlock;
268 }
269
270 error = xfs_iflush(ip, flags);
271
272 /*
273 * We don't want to try again on non-blocking flushes that can't run
274 * again immediately. If an inode really must be written, then that's
275 * what the SYNC_WAIT flag is for.
276 */
277 if (error == EAGAIN) {
278 ASSERT(!(flags & SYNC_WAIT));
279 error = 0;
280 }
281
282 out_unlock:
283 xfs_iunlock(ip, XFS_ILOCK_SHARED);
284 return error;
285}
286
287/*
288 * Write out pagecache data for the whole filesystem.
289 */
290STATIC int
291xfs_sync_data(
292 struct xfs_mount *mp,
293 int flags)
294{
295 int error;
296
297 ASSERT((flags & ~(SYNC_TRYLOCK|SYNC_WAIT)) == 0);
298
299 error = xfs_inode_ag_iterator(mp, xfs_sync_inode_data, flags);
300 if (error)
301 return XFS_ERROR(error);
302
303 xfs_log_force(mp, (flags & SYNC_WAIT) ? XFS_LOG_SYNC : 0);
304 return 0;
305}
306
307/*
308 * Write out inode metadata (attributes) for the whole filesystem.
309 */
310STATIC int
311xfs_sync_attr(
312 struct xfs_mount *mp,
313 int flags)
314{
315 ASSERT((flags & ~SYNC_WAIT) == 0);
316
317 return xfs_inode_ag_iterator(mp, xfs_sync_inode_attr, flags);
318}
319
320STATIC int
321xfs_sync_fsdata(
322 struct xfs_mount *mp)
323{
324 struct xfs_buf *bp;
325
326 /*
327 * If the buffer is pinned then push on the log so we won't get stuck
328 * waiting in the write for someone, maybe ourselves, to flush the log.
329 *
330 * Even though we just pushed the log above, we did not have the
331 * superblock buffer locked at that point so it can become pinned in
332 * between there and here.
333 */
334 bp = xfs_getsb(mp, 0);
335 if (xfs_buf_ispinned(bp))
336 xfs_log_force(mp, 0);
337
338 return xfs_bwrite(mp, bp);
339}
340
341/*
342 * When remounting a filesystem read-only or freezing the filesystem, we have
343 * two phases to execute. This first phase is syncing the data before we
344 * quiesce the filesystem, and the second is flushing all the inodes out after
345 * we've waited for all the transactions created by the first phase to
346 * complete. The second phase ensures that the inodes are written to their
347 * location on disk rather than just existing in transactions in the log. This
348 * means after a quiesce there is no log replay required to write the inodes to
349 * disk (this is the main difference between a sync and a quiesce).
350 */
351/*
352 * First stage of freeze - no writers will make progress now we are here,
353 * so we flush delwri and delalloc buffers here, then wait for all I/O to
354 * complete. Data is frozen at that point. Metadata is not frozen,
355 * transactions can still occur here so don't bother flushing the buftarg
356 * because it'll just get dirty again.
357 */
358int
359xfs_quiesce_data(
360 struct xfs_mount *mp)
361{
362 int error, error2 = 0;
363
364 xfs_qm_sync(mp, SYNC_TRYLOCK);
365 xfs_qm_sync(mp, SYNC_WAIT);
366
367 /* force out the newly dirtied log buffers */
368 xfs_log_force(mp, XFS_LOG_SYNC);
369
370 /* write superblock and hoover up shutdown errors */
371 error = xfs_sync_fsdata(mp);
372
373 /* make sure all delwri buffers are written out */
374 xfs_flush_buftarg(mp->m_ddev_targp, 1);
375
376 /* mark the log as covered if needed */
377 if (xfs_log_need_covered(mp))
378 error2 = xfs_fs_log_dummy(mp);
379
380 /* flush data-only devices */
381 if (mp->m_rtdev_targp)
382 XFS_bflush(mp->m_rtdev_targp);
383
384 return error ? error : error2;
385}
386
387STATIC void
388xfs_quiesce_fs(
389 struct xfs_mount *mp)
390{
391 int count = 0, pincount;
392
393 xfs_reclaim_inodes(mp, 0);
394 xfs_flush_buftarg(mp->m_ddev_targp, 0);
395
396 /*
397 * This loop must run at least twice. The first instance of the loop
398 * will flush most meta data but that will generate more meta data
399 * (typically directory updates). Which then must be flushed and
400 * logged before we can write the unmount record. We also so sync
401 * reclaim of inodes to catch any that the above delwri flush skipped.
402 */
403 do {
404 xfs_reclaim_inodes(mp, SYNC_WAIT);
405 xfs_sync_attr(mp, SYNC_WAIT);
406 pincount = xfs_flush_buftarg(mp->m_ddev_targp, 1);
407 if (!pincount) {
408 delay(50);
409 count++;
410 }
411 } while (count < 2);
412}
413
414/*
415 * Second stage of a quiesce. The data is already synced, now we have to take
416 * care of the metadata. New transactions are already blocked, so we need to
417 * wait for any remaining transactions to drain out before proceeding.
418 */
419void
420xfs_quiesce_attr(
421 struct xfs_mount *mp)
422{
423 int error = 0;
424
425 /* wait for all modifications to complete */
426 while (atomic_read(&mp->m_active_trans) > 0)
427 delay(100);
428
429 /* flush inodes and push all remaining buffers out to disk */
430 xfs_quiesce_fs(mp);
431
432 /*
433 * Just warn here till VFS can correctly support
434 * read-only remount without racing.
435 */
436 WARN_ON(atomic_read(&mp->m_active_trans) != 0);
437
438 /* Push the superblock and write an unmount record */
439 error = xfs_log_sbcount(mp);
440 if (error)
441 xfs_warn(mp, "xfs_attr_quiesce: failed to log sb changes. "
442 "Frozen image may not be consistent.");
443 xfs_log_unmount_write(mp);
444 xfs_unmountfs_writesb(mp);
445}
446
447static void
448xfs_syncd_queue_sync(
449 struct xfs_mount *mp)
450{
451 queue_delayed_work(xfs_syncd_wq, &mp->m_sync_work,
452 msecs_to_jiffies(xfs_syncd_centisecs * 10));
453}
454
455/*
456 * Every sync period we need to unpin all items, reclaim inodes and sync
457 * disk quotas. We might need to cover the log to indicate that the
458 * filesystem is idle and not frozen.
459 */
460STATIC void
461xfs_sync_worker(
462 struct work_struct *work)
463{
464 struct xfs_mount *mp = container_of(to_delayed_work(work),
465 struct xfs_mount, m_sync_work);
466 int error;
467
468 if (!(mp->m_flags & XFS_MOUNT_RDONLY)) {
469 /* dgc: errors ignored here */
470 if (mp->m_super->s_frozen == SB_UNFROZEN &&
471 xfs_log_need_covered(mp))
472 error = xfs_fs_log_dummy(mp);
473 else
474 xfs_log_force(mp, 0);
475 error = xfs_qm_sync(mp, SYNC_TRYLOCK);
476
477 /* start pushing all the metadata that is currently dirty */
478 xfs_ail_push_all(mp->m_ail);
479 }
480
481 /* queue us up again */
482 xfs_syncd_queue_sync(mp);
483}
484
485/*
486 * Queue a new inode reclaim pass if there are reclaimable inodes and there
487 * isn't a reclaim pass already in progress. By default it runs every 5s based
488 * on the xfs syncd work default of 30s. Perhaps this should have it's own
489 * tunable, but that can be done if this method proves to be ineffective or too
490 * aggressive.
491 */
492static void
493xfs_syncd_queue_reclaim(
494 struct xfs_mount *mp)
495{
496
497 /*
498 * We can have inodes enter reclaim after we've shut down the syncd
499 * workqueue during unmount, so don't allow reclaim work to be queued
500 * during unmount.
501 */
502 if (!(mp->m_super->s_flags & MS_ACTIVE))
503 return;
504
505 rcu_read_lock();
506 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
507 queue_delayed_work(xfs_syncd_wq, &mp->m_reclaim_work,
508 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
509 }
510 rcu_read_unlock();
511}
512
513/*
514 * This is a fast pass over the inode cache to try to get reclaim moving on as
515 * many inodes as possible in a short period of time. It kicks itself every few
516 * seconds, as well as being kicked by the inode cache shrinker when memory
517 * goes low. It scans as quickly as possible avoiding locked inodes or those
518 * already being flushed, and once done schedules a future pass.
519 */
520STATIC void
521xfs_reclaim_worker(
522 struct work_struct *work)
523{
524 struct xfs_mount *mp = container_of(to_delayed_work(work),
525 struct xfs_mount, m_reclaim_work);
526
527 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
528 xfs_syncd_queue_reclaim(mp);
529}
530
531/*
532 * Flush delayed allocate data, attempting to free up reserved space
533 * from existing allocations. At this point a new allocation attempt
534 * has failed with ENOSPC and we are in the process of scratching our
535 * heads, looking about for more room.
536 *
537 * Queue a new data flush if there isn't one already in progress and
538 * wait for completion of the flush. This means that we only ever have one
539 * inode flush in progress no matter how many ENOSPC events are occurring and
540 * so will prevent the system from bogging down due to every concurrent
541 * ENOSPC event scanning all the active inodes in the system for writeback.
542 */
543void
544xfs_flush_inodes(
545 struct xfs_inode *ip)
546{
547 struct xfs_mount *mp = ip->i_mount;
548
549 queue_work(xfs_syncd_wq, &mp->m_flush_work);
550 flush_work_sync(&mp->m_flush_work);
551}
552
553STATIC void
554xfs_flush_worker(
555 struct work_struct *work)
556{
557 struct xfs_mount *mp = container_of(work,
558 struct xfs_mount, m_flush_work);
559
560 xfs_sync_data(mp, SYNC_TRYLOCK);
561 xfs_sync_data(mp, SYNC_TRYLOCK | SYNC_WAIT);
562}
563
564int
565xfs_syncd_init(
566 struct xfs_mount *mp)
567{
568 INIT_WORK(&mp->m_flush_work, xfs_flush_worker);
569 INIT_DELAYED_WORK(&mp->m_sync_work, xfs_sync_worker);
570 INIT_DELAYED_WORK(&mp->m_reclaim_work, xfs_reclaim_worker);
571
572 xfs_syncd_queue_sync(mp);
573 xfs_syncd_queue_reclaim(mp);
574
575 return 0;
576}
577
578void
579xfs_syncd_stop(
580 struct xfs_mount *mp)
581{
582 cancel_delayed_work_sync(&mp->m_sync_work);
583 cancel_delayed_work_sync(&mp->m_reclaim_work);
584 cancel_work_sync(&mp->m_flush_work);
585}
586
587void
588__xfs_inode_set_reclaim_tag(
589 struct xfs_perag *pag,
590 struct xfs_inode *ip)
591{
592 radix_tree_tag_set(&pag->pag_ici_root,
593 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
594 XFS_ICI_RECLAIM_TAG);
595
596 if (!pag->pag_ici_reclaimable) {
597 /* propagate the reclaim tag up into the perag radix tree */
598 spin_lock(&ip->i_mount->m_perag_lock);
599 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
600 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
601 XFS_ICI_RECLAIM_TAG);
602 spin_unlock(&ip->i_mount->m_perag_lock);
603
604 /* schedule periodic background inode reclaim */
605 xfs_syncd_queue_reclaim(ip->i_mount);
606
607 trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno,
608 -1, _RET_IP_);
609 }
610 pag->pag_ici_reclaimable++;
611}
612
613/*
614 * We set the inode flag atomically with the radix tree tag.
615 * Once we get tag lookups on the radix tree, this inode flag
616 * can go away.
617 */
618void
619xfs_inode_set_reclaim_tag(
620 xfs_inode_t *ip)
621{
622 struct xfs_mount *mp = ip->i_mount;
623 struct xfs_perag *pag;
624
625 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
626 spin_lock(&pag->pag_ici_lock);
627 spin_lock(&ip->i_flags_lock);
628 __xfs_inode_set_reclaim_tag(pag, ip);
629 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
630 spin_unlock(&ip->i_flags_lock);
631 spin_unlock(&pag->pag_ici_lock);
632 xfs_perag_put(pag);
633}
634
635STATIC void
636__xfs_inode_clear_reclaim(
637 xfs_perag_t *pag,
638 xfs_inode_t *ip)
639{
640 pag->pag_ici_reclaimable--;
641 if (!pag->pag_ici_reclaimable) {
642 /* clear the reclaim tag from the perag radix tree */
643 spin_lock(&ip->i_mount->m_perag_lock);
644 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
645 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
646 XFS_ICI_RECLAIM_TAG);
647 spin_unlock(&ip->i_mount->m_perag_lock);
648 trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno,
649 -1, _RET_IP_);
650 }
651}
652
653void
654__xfs_inode_clear_reclaim_tag(
655 xfs_mount_t *mp,
656 xfs_perag_t *pag,
657 xfs_inode_t *ip)
658{
659 radix_tree_tag_clear(&pag->pag_ici_root,
660 XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
661 __xfs_inode_clear_reclaim(pag, ip);
662}
663
664/*
665 * Grab the inode for reclaim exclusively.
666 * Return 0 if we grabbed it, non-zero otherwise.
667 */
668STATIC int
669xfs_reclaim_inode_grab(
670 struct xfs_inode *ip,
671 int flags)
672{
673 ASSERT(rcu_read_lock_held());
674
675 /* quick check for stale RCU freed inode */
676 if (!ip->i_ino)
677 return 1;
678
679 /*
680 * do some unlocked checks first to avoid unnecessary lock traffic.
681 * The first is a flush lock check, the second is a already in reclaim
682 * check. Only do these checks if we are not going to block on locks.
683 */
684 if ((flags & SYNC_TRYLOCK) &&
685 (!ip->i_flush.done || __xfs_iflags_test(ip, XFS_IRECLAIM))) {
686 return 1;
687 }
688
689 /*
690 * The radix tree lock here protects a thread in xfs_iget from racing
691 * with us starting reclaim on the inode. Once we have the
692 * XFS_IRECLAIM flag set it will not touch us.
693 *
694 * Due to RCU lookup, we may find inodes that have been freed and only
695 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
696 * aren't candidates for reclaim at all, so we must check the
697 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
698 */
699 spin_lock(&ip->i_flags_lock);
700 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
701 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
702 /* not a reclaim candidate. */
703 spin_unlock(&ip->i_flags_lock);
704 return 1;
705 }
706 __xfs_iflags_set(ip, XFS_IRECLAIM);
707 spin_unlock(&ip->i_flags_lock);
708 return 0;
709}
710
711/*
712 * Inodes in different states need to be treated differently, and the return
713 * value of xfs_iflush is not sufficient to get this right. The following table
714 * lists the inode states and the reclaim actions necessary for non-blocking
715 * reclaim:
716 *
717 *
718 * inode state iflush ret required action
719 * --------------- ---------- ---------------
720 * bad - reclaim
721 * shutdown EIO unpin and reclaim
722 * clean, unpinned 0 reclaim
723 * stale, unpinned 0 reclaim
724 * clean, pinned(*) 0 requeue
725 * stale, pinned EAGAIN requeue
726 * dirty, delwri ok 0 requeue
727 * dirty, delwri blocked EAGAIN requeue
728 * dirty, sync flush 0 reclaim
729 *
730 * (*) dgc: I don't think the clean, pinned state is possible but it gets
731 * handled anyway given the order of checks implemented.
732 *
733 * As can be seen from the table, the return value of xfs_iflush() is not
734 * sufficient to correctly decide the reclaim action here. The checks in
735 * xfs_iflush() might look like duplicates, but they are not.
736 *
737 * Also, because we get the flush lock first, we know that any inode that has
738 * been flushed delwri has had the flush completed by the time we check that
739 * the inode is clean. The clean inode check needs to be done before flushing
740 * the inode delwri otherwise we would loop forever requeuing clean inodes as
741 * we cannot tell apart a successful delwri flush and a clean inode from the
742 * return value of xfs_iflush().
743 *
744 * Note that because the inode is flushed delayed write by background
745 * writeback, the flush lock may already be held here and waiting on it can
746 * result in very long latencies. Hence for sync reclaims, where we wait on the
747 * flush lock, the caller should push out delayed write inodes first before
748 * trying to reclaim them to minimise the amount of time spent waiting. For
749 * background relaim, we just requeue the inode for the next pass.
750 *
751 * Hence the order of actions after gaining the locks should be:
752 * bad => reclaim
753 * shutdown => unpin and reclaim
754 * pinned, delwri => requeue
755 * pinned, sync => unpin
756 * stale => reclaim
757 * clean => reclaim
758 * dirty, delwri => flush and requeue
759 * dirty, sync => flush, wait and reclaim
760 */
761STATIC int
762xfs_reclaim_inode(
763 struct xfs_inode *ip,
764 struct xfs_perag *pag,
765 int sync_mode)
766{
767 int error;
768
769restart:
770 error = 0;
771 xfs_ilock(ip, XFS_ILOCK_EXCL);
772 if (!xfs_iflock_nowait(ip)) {
773 if (!(sync_mode & SYNC_WAIT))
774 goto out;
775 xfs_iflock(ip);
776 }
777
778 if (is_bad_inode(VFS_I(ip)))
779 goto reclaim;
780 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
781 xfs_iunpin_wait(ip);
782 goto reclaim;
783 }
784 if (xfs_ipincount(ip)) {
785 if (!(sync_mode & SYNC_WAIT)) {
786 xfs_ifunlock(ip);
787 goto out;
788 }
789 xfs_iunpin_wait(ip);
790 }
791 if (xfs_iflags_test(ip, XFS_ISTALE))
792 goto reclaim;
793 if (xfs_inode_clean(ip))
794 goto reclaim;
795
796 /*
797 * Now we have an inode that needs flushing.
798 *
799 * We do a nonblocking flush here even if we are doing a SYNC_WAIT
800 * reclaim as we can deadlock with inode cluster removal.
801 * xfs_ifree_cluster() can lock the inode buffer before it locks the
802 * ip->i_lock, and we are doing the exact opposite here. As a result,
803 * doing a blocking xfs_itobp() to get the cluster buffer will result
804 * in an ABBA deadlock with xfs_ifree_cluster().
805 *
806 * As xfs_ifree_cluser() must gather all inodes that are active in the
807 * cache to mark them stale, if we hit this case we don't actually want
808 * to do IO here - we want the inode marked stale so we can simply
809 * reclaim it. Hence if we get an EAGAIN error on a SYNC_WAIT flush,
810 * just unlock the inode, back off and try again. Hopefully the next
811 * pass through will see the stale flag set on the inode.
812 */
813 error = xfs_iflush(ip, SYNC_TRYLOCK | sync_mode);
814 if (sync_mode & SYNC_WAIT) {
815 if (error == EAGAIN) {
816 xfs_iunlock(ip, XFS_ILOCK_EXCL);
817 /* backoff longer than in xfs_ifree_cluster */
818 delay(2);
819 goto restart;
820 }
821 xfs_iflock(ip);
822 goto reclaim;
823 }
824
825 /*
826 * When we have to flush an inode but don't have SYNC_WAIT set, we
827 * flush the inode out using a delwri buffer and wait for the next
828 * call into reclaim to find it in a clean state instead of waiting for
829 * it now. We also don't return errors here - if the error is transient
830 * then the next reclaim pass will flush the inode, and if the error
831 * is permanent then the next sync reclaim will reclaim the inode and
832 * pass on the error.
833 */
834 if (error && error != EAGAIN && !XFS_FORCED_SHUTDOWN(ip->i_mount)) {
835 xfs_warn(ip->i_mount,
836 "inode 0x%llx background reclaim flush failed with %d",
837 (long long)ip->i_ino, error);
838 }
839out:
840 xfs_iflags_clear(ip, XFS_IRECLAIM);
841 xfs_iunlock(ip, XFS_ILOCK_EXCL);
842 /*
843 * We could return EAGAIN here to make reclaim rescan the inode tree in
844 * a short while. However, this just burns CPU time scanning the tree
845 * waiting for IO to complete and xfssyncd never goes back to the idle
846 * state. Instead, return 0 to let the next scheduled background reclaim
847 * attempt to reclaim the inode again.
848 */
849 return 0;
850
851reclaim:
852 xfs_ifunlock(ip);
853 xfs_iunlock(ip, XFS_ILOCK_EXCL);
854
855 XFS_STATS_INC(xs_ig_reclaims);
856 /*
857 * Remove the inode from the per-AG radix tree.
858 *
859 * Because radix_tree_delete won't complain even if the item was never
860 * added to the tree assert that it's been there before to catch
861 * problems with the inode life time early on.
862 */
863 spin_lock(&pag->pag_ici_lock);
864 if (!radix_tree_delete(&pag->pag_ici_root,
865 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino)))
866 ASSERT(0);
867 __xfs_inode_clear_reclaim(pag, ip);
868 spin_unlock(&pag->pag_ici_lock);
869
870 /*
871 * Here we do an (almost) spurious inode lock in order to coordinate
872 * with inode cache radix tree lookups. This is because the lookup
873 * can reference the inodes in the cache without taking references.
874 *
875 * We make that OK here by ensuring that we wait until the inode is
876 * unlocked after the lookup before we go ahead and free it. We get
877 * both the ilock and the iolock because the code may need to drop the
878 * ilock one but will still hold the iolock.
879 */
880 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
881 xfs_qm_dqdetach(ip);
882 xfs_iunlock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
883
884 xfs_inode_free(ip);
885 return error;
886
887}
888
889/*
890 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
891 * corrupted, we still want to try to reclaim all the inodes. If we don't,
892 * then a shut down during filesystem unmount reclaim walk leak all the
893 * unreclaimed inodes.
894 */
895int
896xfs_reclaim_inodes_ag(
897 struct xfs_mount *mp,
898 int flags,
899 int *nr_to_scan)
900{
901 struct xfs_perag *pag;
902 int error = 0;
903 int last_error = 0;
904 xfs_agnumber_t ag;
905 int trylock = flags & SYNC_TRYLOCK;
906 int skipped;
907
908restart:
909 ag = 0;
910 skipped = 0;
911 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
912 unsigned long first_index = 0;
913 int done = 0;
914 int nr_found = 0;
915
916 ag = pag->pag_agno + 1;
917
918 if (trylock) {
919 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
920 skipped++;
921 xfs_perag_put(pag);
922 continue;
923 }
924 first_index = pag->pag_ici_reclaim_cursor;
925 } else
926 mutex_lock(&pag->pag_ici_reclaim_lock);
927
928 do {
929 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
930 int i;
931
932 rcu_read_lock();
933 nr_found = radix_tree_gang_lookup_tag(
934 &pag->pag_ici_root,
935 (void **)batch, first_index,
936 XFS_LOOKUP_BATCH,
937 XFS_ICI_RECLAIM_TAG);
938 if (!nr_found) {
939 done = 1;
940 rcu_read_unlock();
941 break;
942 }
943
944 /*
945 * Grab the inodes before we drop the lock. if we found
946 * nothing, nr == 0 and the loop will be skipped.
947 */
948 for (i = 0; i < nr_found; i++) {
949 struct xfs_inode *ip = batch[i];
950
951 if (done || xfs_reclaim_inode_grab(ip, flags))
952 batch[i] = NULL;
953
954 /*
955 * Update the index for the next lookup. Catch
956 * overflows into the next AG range which can
957 * occur if we have inodes in the last block of
958 * the AG and we are currently pointing to the
959 * last inode.
960 *
961 * Because we may see inodes that are from the
962 * wrong AG due to RCU freeing and
963 * reallocation, only update the index if it
964 * lies in this AG. It was a race that lead us
965 * to see this inode, so another lookup from
966 * the same index will not find it again.
967 */
968 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
969 pag->pag_agno)
970 continue;
971 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
972 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
973 done = 1;
974 }
975
976 /* unlock now we've grabbed the inodes. */
977 rcu_read_unlock();
978
979 for (i = 0; i < nr_found; i++) {
980 if (!batch[i])
981 continue;
982 error = xfs_reclaim_inode(batch[i], pag, flags);
983 if (error && last_error != EFSCORRUPTED)
984 last_error = error;
985 }
986
987 *nr_to_scan -= XFS_LOOKUP_BATCH;
988
989 cond_resched();
990
991 } while (nr_found && !done && *nr_to_scan > 0);
992
993 if (trylock && !done)
994 pag->pag_ici_reclaim_cursor = first_index;
995 else
996 pag->pag_ici_reclaim_cursor = 0;
997 mutex_unlock(&pag->pag_ici_reclaim_lock);
998 xfs_perag_put(pag);
999 }
1000
1001 /*
1002 * if we skipped any AG, and we still have scan count remaining, do
1003 * another pass this time using blocking reclaim semantics (i.e
1004 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1005 * ensure that when we get more reclaimers than AGs we block rather
1006 * than spin trying to execute reclaim.
1007 */
1008 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1009 trylock = 0;
1010 goto restart;
1011 }
1012 return XFS_ERROR(last_error);
1013}
1014
1015int
1016xfs_reclaim_inodes(
1017 xfs_mount_t *mp,
1018 int mode)
1019{
1020 int nr_to_scan = INT_MAX;
1021
1022 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1023}
1024
1025/*
1026 * Scan a certain number of inodes for reclaim.
1027 *
1028 * When called we make sure that there is a background (fast) inode reclaim in
1029 * progress, while we will throttle the speed of reclaim via doing synchronous
1030 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1031 * them to be cleaned, which we hope will not be very long due to the
1032 * background walker having already kicked the IO off on those dirty inodes.
1033 */
1034void
1035xfs_reclaim_inodes_nr(
1036 struct xfs_mount *mp,
1037 int nr_to_scan)
1038{
1039 /* kick background reclaimer and push the AIL */
1040 xfs_syncd_queue_reclaim(mp);
1041 xfs_ail_push_all(mp->m_ail);
1042
1043 xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1044}
1045
1046/*
1047 * Return the number of reclaimable inodes in the filesystem for
1048 * the shrinker to determine how much to reclaim.
1049 */
1050int
1051xfs_reclaim_inodes_count(
1052 struct xfs_mount *mp)
1053{
1054 struct xfs_perag *pag;
1055 xfs_agnumber_t ag = 0;
1056 int reclaimable = 0;
1057
1058 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1059 ag = pag->pag_agno + 1;
1060 reclaimable += pag->pag_ici_reclaimable;
1061 xfs_perag_put(pag);
1062 }
1063 return reclaimable;
1064}
1065
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
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.
8 *
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.
13 *
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
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_types.h"
21#include "xfs_log.h"
22#include "xfs_inum.h"
23#include "xfs_trans.h"
24#include "xfs_trans_priv.h"
25#include "xfs_sb.h"
26#include "xfs_ag.h"
27#include "xfs_mount.h"
28#include "xfs_bmap_btree.h"
29#include "xfs_inode.h"
30#include "xfs_dinode.h"
31#include "xfs_error.h"
32#include "xfs_filestream.h"
33#include "xfs_vnodeops.h"
34#include "xfs_inode_item.h"
35#include "xfs_quota.h"
36#include "xfs_trace.h"
37#include "xfs_fsops.h"
38
39#include <linux/kthread.h>
40#include <linux/freezer.h>
41
42struct workqueue_struct *xfs_syncd_wq; /* sync workqueue */
43
44/*
45 * The inode lookup is done in batches to keep the amount of lock traffic and
46 * radix tree lookups to a minimum. The batch size is a trade off between
47 * lookup reduction and stack usage. This is in the reclaim path, so we can't
48 * be too greedy.
49 */
50#define XFS_LOOKUP_BATCH 32
51
52STATIC int
53xfs_inode_ag_walk_grab(
54 struct xfs_inode *ip)
55{
56 struct inode *inode = VFS_I(ip);
57
58 ASSERT(rcu_read_lock_held());
59
60 /*
61 * check for stale RCU freed inode
62 *
63 * If the inode has been reallocated, it doesn't matter if it's not in
64 * the AG we are walking - we are walking for writeback, so if it
65 * passes all the "valid inode" checks and is dirty, then we'll write
66 * it back anyway. If it has been reallocated and still being
67 * initialised, the XFS_INEW check below will catch it.
68 */
69 spin_lock(&ip->i_flags_lock);
70 if (!ip->i_ino)
71 goto out_unlock_noent;
72
73 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
74 if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM))
75 goto out_unlock_noent;
76 spin_unlock(&ip->i_flags_lock);
77
78 /* nothing to sync during shutdown */
79 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
80 return EFSCORRUPTED;
81
82 /* If we can't grab the inode, it must on it's way to reclaim. */
83 if (!igrab(inode))
84 return ENOENT;
85
86 if (is_bad_inode(inode)) {
87 IRELE(ip);
88 return ENOENT;
89 }
90
91 /* inode is valid */
92 return 0;
93
94out_unlock_noent:
95 spin_unlock(&ip->i_flags_lock);
96 return ENOENT;
97}
98
99STATIC int
100xfs_inode_ag_walk(
101 struct xfs_mount *mp,
102 struct xfs_perag *pag,
103 int (*execute)(struct xfs_inode *ip,
104 struct xfs_perag *pag, int flags),
105 int flags)
106{
107 uint32_t first_index;
108 int last_error = 0;
109 int skipped;
110 int done;
111 int nr_found;
112
113restart:
114 done = 0;
115 skipped = 0;
116 first_index = 0;
117 nr_found = 0;
118 do {
119 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
120 int error = 0;
121 int i;
122
123 rcu_read_lock();
124 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
125 (void **)batch, first_index,
126 XFS_LOOKUP_BATCH);
127 if (!nr_found) {
128 rcu_read_unlock();
129 break;
130 }
131
132 /*
133 * Grab the inodes before we drop the lock. if we found
134 * nothing, nr == 0 and the loop will be skipped.
135 */
136 for (i = 0; i < nr_found; i++) {
137 struct xfs_inode *ip = batch[i];
138
139 if (done || xfs_inode_ag_walk_grab(ip))
140 batch[i] = NULL;
141
142 /*
143 * Update the index for the next lookup. Catch
144 * overflows into the next AG range which can occur if
145 * we have inodes in the last block of the AG and we
146 * are currently pointing to the last inode.
147 *
148 * Because we may see inodes that are from the wrong AG
149 * due to RCU freeing and reallocation, only update the
150 * index if it lies in this AG. It was a race that lead
151 * us to see this inode, so another lookup from the
152 * same index will not find it again.
153 */
154 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
155 continue;
156 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
157 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
158 done = 1;
159 }
160
161 /* unlock now we've grabbed the inodes. */
162 rcu_read_unlock();
163
164 for (i = 0; i < nr_found; i++) {
165 if (!batch[i])
166 continue;
167 error = execute(batch[i], pag, flags);
168 IRELE(batch[i]);
169 if (error == EAGAIN) {
170 skipped++;
171 continue;
172 }
173 if (error && last_error != EFSCORRUPTED)
174 last_error = error;
175 }
176
177 /* bail out if the filesystem is corrupted. */
178 if (error == EFSCORRUPTED)
179 break;
180
181 cond_resched();
182
183 } while (nr_found && !done);
184
185 if (skipped) {
186 delay(1);
187 goto restart;
188 }
189 return last_error;
190}
191
192int
193xfs_inode_ag_iterator(
194 struct xfs_mount *mp,
195 int (*execute)(struct xfs_inode *ip,
196 struct xfs_perag *pag, int flags),
197 int flags)
198{
199 struct xfs_perag *pag;
200 int error = 0;
201 int last_error = 0;
202 xfs_agnumber_t ag;
203
204 ag = 0;
205 while ((pag = xfs_perag_get(mp, ag))) {
206 ag = pag->pag_agno + 1;
207 error = xfs_inode_ag_walk(mp, pag, execute, flags);
208 xfs_perag_put(pag);
209 if (error) {
210 last_error = error;
211 if (error == EFSCORRUPTED)
212 break;
213 }
214 }
215 return XFS_ERROR(last_error);
216}
217
218STATIC int
219xfs_sync_inode_data(
220 struct xfs_inode *ip,
221 struct xfs_perag *pag,
222 int flags)
223{
224 struct inode *inode = VFS_I(ip);
225 struct address_space *mapping = inode->i_mapping;
226 int error = 0;
227
228 if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
229 return 0;
230
231 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED)) {
232 if (flags & SYNC_TRYLOCK)
233 return 0;
234 xfs_ilock(ip, XFS_IOLOCK_SHARED);
235 }
236
237 error = xfs_flush_pages(ip, 0, -1, (flags & SYNC_WAIT) ?
238 0 : XBF_ASYNC, FI_NONE);
239 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
240 return error;
241}
242
243/*
244 * Write out pagecache data for the whole filesystem.
245 */
246STATIC int
247xfs_sync_data(
248 struct xfs_mount *mp,
249 int flags)
250{
251 int error;
252
253 ASSERT((flags & ~(SYNC_TRYLOCK|SYNC_WAIT)) == 0);
254
255 error = xfs_inode_ag_iterator(mp, xfs_sync_inode_data, flags);
256 if (error)
257 return XFS_ERROR(error);
258
259 xfs_log_force(mp, (flags & SYNC_WAIT) ? XFS_LOG_SYNC : 0);
260 return 0;
261}
262
263STATIC int
264xfs_sync_fsdata(
265 struct xfs_mount *mp)
266{
267 struct xfs_buf *bp;
268 int error;
269
270 /*
271 * If the buffer is pinned then push on the log so we won't get stuck
272 * waiting in the write for someone, maybe ourselves, to flush the log.
273 *
274 * Even though we just pushed the log above, we did not have the
275 * superblock buffer locked at that point so it can become pinned in
276 * between there and here.
277 */
278 bp = xfs_getsb(mp, 0);
279 if (xfs_buf_ispinned(bp))
280 xfs_log_force(mp, 0);
281 error = xfs_bwrite(bp);
282 xfs_buf_relse(bp);
283 return error;
284}
285
286/*
287 * When remounting a filesystem read-only or freezing the filesystem, we have
288 * two phases to execute. This first phase is syncing the data before we
289 * quiesce the filesystem, and the second is flushing all the inodes out after
290 * we've waited for all the transactions created by the first phase to
291 * complete. The second phase ensures that the inodes are written to their
292 * location on disk rather than just existing in transactions in the log. This
293 * means after a quiesce there is no log replay required to write the inodes to
294 * disk (this is the main difference between a sync and a quiesce).
295 */
296/*
297 * First stage of freeze - no writers will make progress now we are here,
298 * so we flush delwri and delalloc buffers here, then wait for all I/O to
299 * complete. Data is frozen at that point. Metadata is not frozen,
300 * transactions can still occur here so don't bother emptying the AIL
301 * because it'll just get dirty again.
302 */
303int
304xfs_quiesce_data(
305 struct xfs_mount *mp)
306{
307 int error, error2 = 0;
308
309 /* force out the log */
310 xfs_log_force(mp, XFS_LOG_SYNC);
311
312 /* write superblock and hoover up shutdown errors */
313 error = xfs_sync_fsdata(mp);
314
315 /* mark the log as covered if needed */
316 if (xfs_log_need_covered(mp))
317 error2 = xfs_fs_log_dummy(mp);
318
319 return error ? error : error2;
320}
321
322/*
323 * Second stage of a quiesce. The data is already synced, now we have to take
324 * care of the metadata. New transactions are already blocked, so we need to
325 * wait for any remaining transactions to drain out before proceeding.
326 */
327void
328xfs_quiesce_attr(
329 struct xfs_mount *mp)
330{
331 int error = 0;
332
333 /* wait for all modifications to complete */
334 while (atomic_read(&mp->m_active_trans) > 0)
335 delay(100);
336
337 /* reclaim inodes to do any IO before the freeze completes */
338 xfs_reclaim_inodes(mp, 0);
339 xfs_reclaim_inodes(mp, SYNC_WAIT);
340
341 /* flush all pending changes from the AIL */
342 xfs_ail_push_all_sync(mp->m_ail);
343
344 /*
345 * Just warn here till VFS can correctly support
346 * read-only remount without racing.
347 */
348 WARN_ON(atomic_read(&mp->m_active_trans) != 0);
349
350 /* Push the superblock and write an unmount record */
351 error = xfs_log_sbcount(mp);
352 if (error)
353 xfs_warn(mp, "xfs_attr_quiesce: failed to log sb changes. "
354 "Frozen image may not be consistent.");
355 xfs_log_unmount_write(mp);
356
357 /*
358 * At this point we might have modified the superblock again and thus
359 * added an item to the AIL, thus flush it again.
360 */
361 xfs_ail_push_all_sync(mp->m_ail);
362}
363
364static void
365xfs_syncd_queue_sync(
366 struct xfs_mount *mp)
367{
368 queue_delayed_work(xfs_syncd_wq, &mp->m_sync_work,
369 msecs_to_jiffies(xfs_syncd_centisecs * 10));
370}
371
372/*
373 * Every sync period we need to unpin all items, reclaim inodes and sync
374 * disk quotas. We might need to cover the log to indicate that the
375 * filesystem is idle and not frozen.
376 */
377STATIC void
378xfs_sync_worker(
379 struct work_struct *work)
380{
381 struct xfs_mount *mp = container_of(to_delayed_work(work),
382 struct xfs_mount, m_sync_work);
383 int error;
384
385 /*
386 * We shouldn't write/force the log if we are in the mount/unmount
387 * process or on a read only filesystem. The workqueue still needs to be
388 * active in both cases, however, because it is used for inode reclaim
389 * during these times. Use the MS_ACTIVE flag to avoid doing anything
390 * during mount. Doing work during unmount is avoided by calling
391 * cancel_delayed_work_sync on this work queue before tearing down
392 * the ail and the log in xfs_log_unmount.
393 */
394 if (!(mp->m_super->s_flags & MS_ACTIVE) &&
395 !(mp->m_flags & XFS_MOUNT_RDONLY)) {
396 /* dgc: errors ignored here */
397 if (mp->m_super->s_frozen == SB_UNFROZEN &&
398 xfs_log_need_covered(mp))
399 error = xfs_fs_log_dummy(mp);
400 else
401 xfs_log_force(mp, 0);
402
403 /* start pushing all the metadata that is currently
404 * dirty */
405 xfs_ail_push_all(mp->m_ail);
406 }
407
408 /* queue us up again */
409 xfs_syncd_queue_sync(mp);
410}
411
412/*
413 * Queue a new inode reclaim pass if there are reclaimable inodes and there
414 * isn't a reclaim pass already in progress. By default it runs every 5s based
415 * on the xfs syncd work default of 30s. Perhaps this should have it's own
416 * tunable, but that can be done if this method proves to be ineffective or too
417 * aggressive.
418 */
419static void
420xfs_syncd_queue_reclaim(
421 struct xfs_mount *mp)
422{
423
424 rcu_read_lock();
425 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
426 queue_delayed_work(xfs_syncd_wq, &mp->m_reclaim_work,
427 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
428 }
429 rcu_read_unlock();
430}
431
432/*
433 * This is a fast pass over the inode cache to try to get reclaim moving on as
434 * many inodes as possible in a short period of time. It kicks itself every few
435 * seconds, as well as being kicked by the inode cache shrinker when memory
436 * goes low. It scans as quickly as possible avoiding locked inodes or those
437 * already being flushed, and once done schedules a future pass.
438 */
439STATIC void
440xfs_reclaim_worker(
441 struct work_struct *work)
442{
443 struct xfs_mount *mp = container_of(to_delayed_work(work),
444 struct xfs_mount, m_reclaim_work);
445
446 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
447 xfs_syncd_queue_reclaim(mp);
448}
449
450/*
451 * Flush delayed allocate data, attempting to free up reserved space
452 * from existing allocations. At this point a new allocation attempt
453 * has failed with ENOSPC and we are in the process of scratching our
454 * heads, looking about for more room.
455 *
456 * Queue a new data flush if there isn't one already in progress and
457 * wait for completion of the flush. This means that we only ever have one
458 * inode flush in progress no matter how many ENOSPC events are occurring and
459 * so will prevent the system from bogging down due to every concurrent
460 * ENOSPC event scanning all the active inodes in the system for writeback.
461 */
462void
463xfs_flush_inodes(
464 struct xfs_inode *ip)
465{
466 struct xfs_mount *mp = ip->i_mount;
467
468 queue_work(xfs_syncd_wq, &mp->m_flush_work);
469 flush_work_sync(&mp->m_flush_work);
470}
471
472STATIC void
473xfs_flush_worker(
474 struct work_struct *work)
475{
476 struct xfs_mount *mp = container_of(work,
477 struct xfs_mount, m_flush_work);
478
479 xfs_sync_data(mp, SYNC_TRYLOCK);
480 xfs_sync_data(mp, SYNC_TRYLOCK | SYNC_WAIT);
481}
482
483int
484xfs_syncd_init(
485 struct xfs_mount *mp)
486{
487 INIT_WORK(&mp->m_flush_work, xfs_flush_worker);
488 INIT_DELAYED_WORK(&mp->m_sync_work, xfs_sync_worker);
489 INIT_DELAYED_WORK(&mp->m_reclaim_work, xfs_reclaim_worker);
490
491 xfs_syncd_queue_sync(mp);
492
493 return 0;
494}
495
496void
497xfs_syncd_stop(
498 struct xfs_mount *mp)
499{
500 cancel_delayed_work_sync(&mp->m_sync_work);
501 cancel_delayed_work_sync(&mp->m_reclaim_work);
502 cancel_work_sync(&mp->m_flush_work);
503}
504
505void
506__xfs_inode_set_reclaim_tag(
507 struct xfs_perag *pag,
508 struct xfs_inode *ip)
509{
510 radix_tree_tag_set(&pag->pag_ici_root,
511 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
512 XFS_ICI_RECLAIM_TAG);
513
514 if (!pag->pag_ici_reclaimable) {
515 /* propagate the reclaim tag up into the perag radix tree */
516 spin_lock(&ip->i_mount->m_perag_lock);
517 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
518 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
519 XFS_ICI_RECLAIM_TAG);
520 spin_unlock(&ip->i_mount->m_perag_lock);
521
522 /* schedule periodic background inode reclaim */
523 xfs_syncd_queue_reclaim(ip->i_mount);
524
525 trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno,
526 -1, _RET_IP_);
527 }
528 pag->pag_ici_reclaimable++;
529}
530
531/*
532 * We set the inode flag atomically with the radix tree tag.
533 * Once we get tag lookups on the radix tree, this inode flag
534 * can go away.
535 */
536void
537xfs_inode_set_reclaim_tag(
538 xfs_inode_t *ip)
539{
540 struct xfs_mount *mp = ip->i_mount;
541 struct xfs_perag *pag;
542
543 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
544 spin_lock(&pag->pag_ici_lock);
545 spin_lock(&ip->i_flags_lock);
546 __xfs_inode_set_reclaim_tag(pag, ip);
547 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
548 spin_unlock(&ip->i_flags_lock);
549 spin_unlock(&pag->pag_ici_lock);
550 xfs_perag_put(pag);
551}
552
553STATIC void
554__xfs_inode_clear_reclaim(
555 xfs_perag_t *pag,
556 xfs_inode_t *ip)
557{
558 pag->pag_ici_reclaimable--;
559 if (!pag->pag_ici_reclaimable) {
560 /* clear the reclaim tag from the perag radix tree */
561 spin_lock(&ip->i_mount->m_perag_lock);
562 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
563 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
564 XFS_ICI_RECLAIM_TAG);
565 spin_unlock(&ip->i_mount->m_perag_lock);
566 trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno,
567 -1, _RET_IP_);
568 }
569}
570
571void
572__xfs_inode_clear_reclaim_tag(
573 xfs_mount_t *mp,
574 xfs_perag_t *pag,
575 xfs_inode_t *ip)
576{
577 radix_tree_tag_clear(&pag->pag_ici_root,
578 XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
579 __xfs_inode_clear_reclaim(pag, ip);
580}
581
582/*
583 * Grab the inode for reclaim exclusively.
584 * Return 0 if we grabbed it, non-zero otherwise.
585 */
586STATIC int
587xfs_reclaim_inode_grab(
588 struct xfs_inode *ip,
589 int flags)
590{
591 ASSERT(rcu_read_lock_held());
592
593 /* quick check for stale RCU freed inode */
594 if (!ip->i_ino)
595 return 1;
596
597 /*
598 * If we are asked for non-blocking operation, do unlocked checks to
599 * see if the inode already is being flushed or in reclaim to avoid
600 * lock traffic.
601 */
602 if ((flags & SYNC_TRYLOCK) &&
603 __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
604 return 1;
605
606 /*
607 * The radix tree lock here protects a thread in xfs_iget from racing
608 * with us starting reclaim on the inode. Once we have the
609 * XFS_IRECLAIM flag set it will not touch us.
610 *
611 * Due to RCU lookup, we may find inodes that have been freed and only
612 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
613 * aren't candidates for reclaim at all, so we must check the
614 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
615 */
616 spin_lock(&ip->i_flags_lock);
617 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
618 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
619 /* not a reclaim candidate. */
620 spin_unlock(&ip->i_flags_lock);
621 return 1;
622 }
623 __xfs_iflags_set(ip, XFS_IRECLAIM);
624 spin_unlock(&ip->i_flags_lock);
625 return 0;
626}
627
628/*
629 * Inodes in different states need to be treated differently. The following
630 * table lists the inode states and the reclaim actions necessary:
631 *
632 * inode state iflush ret required action
633 * --------------- ---------- ---------------
634 * bad - reclaim
635 * shutdown EIO unpin and reclaim
636 * clean, unpinned 0 reclaim
637 * stale, unpinned 0 reclaim
638 * clean, pinned(*) 0 requeue
639 * stale, pinned EAGAIN requeue
640 * dirty, async - requeue
641 * dirty, sync 0 reclaim
642 *
643 * (*) dgc: I don't think the clean, pinned state is possible but it gets
644 * handled anyway given the order of checks implemented.
645 *
646 * Also, because we get the flush lock first, we know that any inode that has
647 * been flushed delwri has had the flush completed by the time we check that
648 * the inode is clean.
649 *
650 * Note that because the inode is flushed delayed write by AIL pushing, the
651 * flush lock may already be held here and waiting on it can result in very
652 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
653 * the caller should push the AIL first before trying to reclaim inodes to
654 * minimise the amount of time spent waiting. For background relaim, we only
655 * bother to reclaim clean inodes anyway.
656 *
657 * Hence the order of actions after gaining the locks should be:
658 * bad => reclaim
659 * shutdown => unpin and reclaim
660 * pinned, async => requeue
661 * pinned, sync => unpin
662 * stale => reclaim
663 * clean => reclaim
664 * dirty, async => requeue
665 * dirty, sync => flush, wait and reclaim
666 */
667STATIC int
668xfs_reclaim_inode(
669 struct xfs_inode *ip,
670 struct xfs_perag *pag,
671 int sync_mode)
672{
673 struct xfs_buf *bp = NULL;
674 int error;
675
676restart:
677 error = 0;
678 xfs_ilock(ip, XFS_ILOCK_EXCL);
679 if (!xfs_iflock_nowait(ip)) {
680 if (!(sync_mode & SYNC_WAIT))
681 goto out;
682 xfs_iflock(ip);
683 }
684
685 if (is_bad_inode(VFS_I(ip)))
686 goto reclaim;
687 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
688 xfs_iunpin_wait(ip);
689 xfs_iflush_abort(ip, false);
690 goto reclaim;
691 }
692 if (xfs_ipincount(ip)) {
693 if (!(sync_mode & SYNC_WAIT))
694 goto out_ifunlock;
695 xfs_iunpin_wait(ip);
696 }
697 if (xfs_iflags_test(ip, XFS_ISTALE))
698 goto reclaim;
699 if (xfs_inode_clean(ip))
700 goto reclaim;
701
702 /*
703 * Never flush out dirty data during non-blocking reclaim, as it would
704 * just contend with AIL pushing trying to do the same job.
705 */
706 if (!(sync_mode & SYNC_WAIT))
707 goto out_ifunlock;
708
709 /*
710 * Now we have an inode that needs flushing.
711 *
712 * Note that xfs_iflush will never block on the inode buffer lock, as
713 * xfs_ifree_cluster() can lock the inode buffer before it locks the
714 * ip->i_lock, and we are doing the exact opposite here. As a result,
715 * doing a blocking xfs_itobp() to get the cluster buffer would result
716 * in an ABBA deadlock with xfs_ifree_cluster().
717 *
718 * As xfs_ifree_cluser() must gather all inodes that are active in the
719 * cache to mark them stale, if we hit this case we don't actually want
720 * to do IO here - we want the inode marked stale so we can simply
721 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
722 * inode, back off and try again. Hopefully the next pass through will
723 * see the stale flag set on the inode.
724 */
725 error = xfs_iflush(ip, &bp);
726 if (error == EAGAIN) {
727 xfs_iunlock(ip, XFS_ILOCK_EXCL);
728 /* backoff longer than in xfs_ifree_cluster */
729 delay(2);
730 goto restart;
731 }
732
733 if (!error) {
734 error = xfs_bwrite(bp);
735 xfs_buf_relse(bp);
736 }
737
738 xfs_iflock(ip);
739reclaim:
740 xfs_ifunlock(ip);
741 xfs_iunlock(ip, XFS_ILOCK_EXCL);
742
743 XFS_STATS_INC(xs_ig_reclaims);
744 /*
745 * Remove the inode from the per-AG radix tree.
746 *
747 * Because radix_tree_delete won't complain even if the item was never
748 * added to the tree assert that it's been there before to catch
749 * problems with the inode life time early on.
750 */
751 spin_lock(&pag->pag_ici_lock);
752 if (!radix_tree_delete(&pag->pag_ici_root,
753 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino)))
754 ASSERT(0);
755 __xfs_inode_clear_reclaim(pag, ip);
756 spin_unlock(&pag->pag_ici_lock);
757
758 /*
759 * Here we do an (almost) spurious inode lock in order to coordinate
760 * with inode cache radix tree lookups. This is because the lookup
761 * can reference the inodes in the cache without taking references.
762 *
763 * We make that OK here by ensuring that we wait until the inode is
764 * unlocked after the lookup before we go ahead and free it.
765 */
766 xfs_ilock(ip, XFS_ILOCK_EXCL);
767 xfs_qm_dqdetach(ip);
768 xfs_iunlock(ip, XFS_ILOCK_EXCL);
769
770 xfs_inode_free(ip);
771 return error;
772
773out_ifunlock:
774 xfs_ifunlock(ip);
775out:
776 xfs_iflags_clear(ip, XFS_IRECLAIM);
777 xfs_iunlock(ip, XFS_ILOCK_EXCL);
778 /*
779 * We could return EAGAIN here to make reclaim rescan the inode tree in
780 * a short while. However, this just burns CPU time scanning the tree
781 * waiting for IO to complete and xfssyncd never goes back to the idle
782 * state. Instead, return 0 to let the next scheduled background reclaim
783 * attempt to reclaim the inode again.
784 */
785 return 0;
786}
787
788/*
789 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
790 * corrupted, we still want to try to reclaim all the inodes. If we don't,
791 * then a shut down during filesystem unmount reclaim walk leak all the
792 * unreclaimed inodes.
793 */
794int
795xfs_reclaim_inodes_ag(
796 struct xfs_mount *mp,
797 int flags,
798 int *nr_to_scan)
799{
800 struct xfs_perag *pag;
801 int error = 0;
802 int last_error = 0;
803 xfs_agnumber_t ag;
804 int trylock = flags & SYNC_TRYLOCK;
805 int skipped;
806
807restart:
808 ag = 0;
809 skipped = 0;
810 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
811 unsigned long first_index = 0;
812 int done = 0;
813 int nr_found = 0;
814
815 ag = pag->pag_agno + 1;
816
817 if (trylock) {
818 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
819 skipped++;
820 xfs_perag_put(pag);
821 continue;
822 }
823 first_index = pag->pag_ici_reclaim_cursor;
824 } else
825 mutex_lock(&pag->pag_ici_reclaim_lock);
826
827 do {
828 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
829 int i;
830
831 rcu_read_lock();
832 nr_found = radix_tree_gang_lookup_tag(
833 &pag->pag_ici_root,
834 (void **)batch, first_index,
835 XFS_LOOKUP_BATCH,
836 XFS_ICI_RECLAIM_TAG);
837 if (!nr_found) {
838 done = 1;
839 rcu_read_unlock();
840 break;
841 }
842
843 /*
844 * Grab the inodes before we drop the lock. if we found
845 * nothing, nr == 0 and the loop will be skipped.
846 */
847 for (i = 0; i < nr_found; i++) {
848 struct xfs_inode *ip = batch[i];
849
850 if (done || xfs_reclaim_inode_grab(ip, flags))
851 batch[i] = NULL;
852
853 /*
854 * Update the index for the next lookup. Catch
855 * overflows into the next AG range which can
856 * occur if we have inodes in the last block of
857 * the AG and we are currently pointing to the
858 * last inode.
859 *
860 * Because we may see inodes that are from the
861 * wrong AG due to RCU freeing and
862 * reallocation, only update the index if it
863 * lies in this AG. It was a race that lead us
864 * to see this inode, so another lookup from
865 * the same index will not find it again.
866 */
867 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
868 pag->pag_agno)
869 continue;
870 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
871 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
872 done = 1;
873 }
874
875 /* unlock now we've grabbed the inodes. */
876 rcu_read_unlock();
877
878 for (i = 0; i < nr_found; i++) {
879 if (!batch[i])
880 continue;
881 error = xfs_reclaim_inode(batch[i], pag, flags);
882 if (error && last_error != EFSCORRUPTED)
883 last_error = error;
884 }
885
886 *nr_to_scan -= XFS_LOOKUP_BATCH;
887
888 cond_resched();
889
890 } while (nr_found && !done && *nr_to_scan > 0);
891
892 if (trylock && !done)
893 pag->pag_ici_reclaim_cursor = first_index;
894 else
895 pag->pag_ici_reclaim_cursor = 0;
896 mutex_unlock(&pag->pag_ici_reclaim_lock);
897 xfs_perag_put(pag);
898 }
899
900 /*
901 * if we skipped any AG, and we still have scan count remaining, do
902 * another pass this time using blocking reclaim semantics (i.e
903 * waiting on the reclaim locks and ignoring the reclaim cursors). This
904 * ensure that when we get more reclaimers than AGs we block rather
905 * than spin trying to execute reclaim.
906 */
907 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
908 trylock = 0;
909 goto restart;
910 }
911 return XFS_ERROR(last_error);
912}
913
914int
915xfs_reclaim_inodes(
916 xfs_mount_t *mp,
917 int mode)
918{
919 int nr_to_scan = INT_MAX;
920
921 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
922}
923
924/*
925 * Scan a certain number of inodes for reclaim.
926 *
927 * When called we make sure that there is a background (fast) inode reclaim in
928 * progress, while we will throttle the speed of reclaim via doing synchronous
929 * reclaim of inodes. That means if we come across dirty inodes, we wait for
930 * them to be cleaned, which we hope will not be very long due to the
931 * background walker having already kicked the IO off on those dirty inodes.
932 */
933void
934xfs_reclaim_inodes_nr(
935 struct xfs_mount *mp,
936 int nr_to_scan)
937{
938 /* kick background reclaimer and push the AIL */
939 xfs_syncd_queue_reclaim(mp);
940 xfs_ail_push_all(mp->m_ail);
941
942 xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
943}
944
945/*
946 * Return the number of reclaimable inodes in the filesystem for
947 * the shrinker to determine how much to reclaim.
948 */
949int
950xfs_reclaim_inodes_count(
951 struct xfs_mount *mp)
952{
953 struct xfs_perag *pag;
954 xfs_agnumber_t ag = 0;
955 int reclaimable = 0;
956
957 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
958 ag = pag->pag_agno + 1;
959 reclaimable += pag->pag_ici_reclaimable;
960 xfs_perag_put(pag);
961 }
962 return reclaimable;
963}
964