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"
24 #include "xfs_trans.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_bmap_btree.h"
31 #include "xfs_alloc_btree.h"
32 #include "xfs_ialloc_btree.h"
33 #include "xfs_btree.h"
34 #include "xfs_dir2_sf.h"
35 #include "xfs_attr_sf.h"
36 #include "xfs_inode.h"
37 #include "xfs_dinode.h"
38 #include "xfs_error.h"
39 #include "xfs_mru_cache.h"
40 #include "xfs_filestream.h"
41 #include "xfs_vnodeops.h"
42 #include "xfs_utils.h"
43 #include "xfs_buf_item.h"
44 #include "xfs_inode_item.h"
46 #include "xfs_quota.h"
47 #include "xfs_trace.h"
49 #include <linux/kthread.h>
50 #include <linux/freezer.h>
56 struct xfs_perag *pag,
57 uint32_t *first_index,
64 * use a gang lookup to find the next inode in the tree
65 * as the tree is sparse and a gang lookup walks to find
66 * the number of objects requested.
68 if (tag == XFS_ICI_NO_TAG) {
69 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
70 (void **)&ip, *first_index, 1);
72 nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
73 (void **)&ip, *first_index, 1, tag);
79 * Update the index for the next lookup. Catch overflows
80 * into the next AG range which can occur if we have inodes
81 * in the last block of the AG and we are currently
82 * pointing to the last inode.
84 *first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
85 if (*first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
93 struct xfs_perag *pag,
94 int (*execute)(struct xfs_inode *ip,
95 struct xfs_perag *pag, int flags),
101 uint32_t first_index;
113 write_lock(&pag->pag_ici_lock);
115 read_lock(&pag->pag_ici_lock);
116 ip = xfs_inode_ag_lookup(mp, pag, &first_index, tag);
119 write_unlock(&pag->pag_ici_lock);
121 read_unlock(&pag->pag_ici_lock);
125 /* execute releases pag->pag_ici_lock */
126 error = execute(ip, pag, flags);
127 if (error == EAGAIN) {
134 /* bail out if the filesystem is corrupted. */
135 if (error == EFSCORRUPTED)
138 } while ((*nr_to_scan)--);
148 xfs_inode_ag_iterator(
149 struct xfs_mount *mp,
150 int (*execute)(struct xfs_inode *ip,
151 struct xfs_perag *pag, int flags),
162 nr = nr_to_scan ? *nr_to_scan : INT_MAX;
163 for (ag = 0; ag < mp->m_sb.sb_agcount; ag++) {
164 struct xfs_perag *pag;
166 pag = xfs_perag_get(mp, ag);
167 if (!pag->pag_ici_init) {
171 error = xfs_inode_ag_walk(mp, pag, execute, flags, tag,
176 if (error == EFSCORRUPTED)
184 return XFS_ERROR(last_error);
187 /* must be called with pag_ici_lock held and releases it */
189 xfs_sync_inode_valid(
190 struct xfs_inode *ip,
191 struct xfs_perag *pag)
193 struct inode *inode = VFS_I(ip);
194 int error = EFSCORRUPTED;
196 /* nothing to sync during shutdown */
197 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
200 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
202 if (xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM))
205 /* If we can't grab the inode, it must on it's way to reclaim. */
209 if (is_bad_inode(inode)) {
217 read_unlock(&pag->pag_ici_lock);
223 struct xfs_inode *ip,
224 struct xfs_perag *pag,
227 struct inode *inode = VFS_I(ip);
228 struct address_space *mapping = inode->i_mapping;
231 error = xfs_sync_inode_valid(ip, pag);
235 if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
238 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED)) {
239 if (flags & SYNC_TRYLOCK)
241 xfs_ilock(ip, XFS_IOLOCK_SHARED);
244 error = xfs_flush_pages(ip, 0, -1, (flags & SYNC_WAIT) ?
245 0 : XBF_ASYNC, FI_NONE);
246 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
249 if (flags & SYNC_WAIT)
257 struct xfs_inode *ip,
258 struct xfs_perag *pag,
263 error = xfs_sync_inode_valid(ip, pag);
267 xfs_ilock(ip, XFS_ILOCK_SHARED);
268 if (xfs_inode_clean(ip))
270 if (!xfs_iflock_nowait(ip)) {
271 if (!(flags & SYNC_WAIT))
276 if (xfs_inode_clean(ip)) {
281 error = xfs_iflush(ip, flags);
284 xfs_iunlock(ip, XFS_ILOCK_SHARED);
290 * Write out pagecache data for the whole filesystem.
294 struct xfs_mount *mp,
299 ASSERT((flags & ~(SYNC_TRYLOCK|SYNC_WAIT)) == 0);
301 error = xfs_inode_ag_iterator(mp, xfs_sync_inode_data, flags,
302 XFS_ICI_NO_TAG, 0, NULL);
304 return XFS_ERROR(error);
306 xfs_log_force(mp, (flags & SYNC_WAIT) ? XFS_LOG_SYNC : 0);
311 * Write out inode metadata (attributes) for the whole filesystem.
315 struct xfs_mount *mp,
318 ASSERT((flags & ~SYNC_WAIT) == 0);
320 return xfs_inode_ag_iterator(mp, xfs_sync_inode_attr, flags,
321 XFS_ICI_NO_TAG, 0, NULL);
325 xfs_commit_dummy_trans(
326 struct xfs_mount *mp,
329 struct xfs_inode *ip = mp->m_rootip;
330 struct xfs_trans *tp;
334 * Put a dummy transaction in the log to tell recovery
335 * that all others are OK.
337 tp = xfs_trans_alloc(mp, XFS_TRANS_DUMMY1);
338 error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES(mp), 0, 0, 0);
340 xfs_trans_cancel(tp, 0);
344 xfs_ilock(ip, XFS_ILOCK_EXCL);
346 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
347 xfs_trans_ihold(tp, ip);
348 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
349 error = xfs_trans_commit(tp, 0);
350 xfs_iunlock(ip, XFS_ILOCK_EXCL);
352 /* the log force ensures this transaction is pushed to disk */
353 xfs_log_force(mp, (flags & SYNC_WAIT) ? XFS_LOG_SYNC : 0);
359 struct xfs_mount *mp,
363 struct xfs_buf_log_item *bip;
367 * If this is xfssyncd() then only sync the superblock if we can
368 * lock it without sleeping and it is not pinned.
370 if (flags & SYNC_TRYLOCK) {
371 ASSERT(!(flags & SYNC_WAIT));
373 bp = xfs_getsb(mp, XBF_TRYLOCK);
377 bip = XFS_BUF_FSPRIVATE(bp, struct xfs_buf_log_item *);
378 if (!bip || !xfs_buf_item_dirty(bip) || XFS_BUF_ISPINNED(bp))
381 bp = xfs_getsb(mp, 0);
384 * If the buffer is pinned then push on the log so we won't
385 * get stuck waiting in the write for someone, maybe
386 * ourselves, to flush the log.
388 * Even though we just pushed the log above, we did not have
389 * the superblock buffer locked at that point so it can
390 * become pinned in between there and here.
392 if (XFS_BUF_ISPINNED(bp))
393 xfs_log_force(mp, 0);
397 if (flags & SYNC_WAIT)
402 error = xfs_bwrite(mp, bp);
407 * If this is a data integrity sync make sure all pending buffers
408 * are flushed out for the log coverage check below.
410 if (flags & SYNC_WAIT)
411 xfs_flush_buftarg(mp->m_ddev_targp, 1);
413 if (xfs_log_need_covered(mp))
414 error = xfs_commit_dummy_trans(mp, flags);
424 * When remounting a filesystem read-only or freezing the filesystem, we have
425 * two phases to execute. This first phase is syncing the data before we
426 * quiesce the filesystem, and the second is flushing all the inodes out after
427 * we've waited for all the transactions created by the first phase to
428 * complete. The second phase ensures that the inodes are written to their
429 * location on disk rather than just existing in transactions in the log. This
430 * means after a quiesce there is no log replay required to write the inodes to
431 * disk (this is the main difference between a sync and a quiesce).
434 * First stage of freeze - no writers will make progress now we are here,
435 * so we flush delwri and delalloc buffers here, then wait for all I/O to
436 * complete. Data is frozen at that point. Metadata is not frozen,
437 * transactions can still occur here so don't bother flushing the buftarg
438 * because it'll just get dirty again.
442 struct xfs_mount *mp)
446 /* push non-blocking */
447 xfs_sync_data(mp, 0);
448 xfs_qm_sync(mp, SYNC_TRYLOCK);
450 /* push and block till complete */
451 xfs_sync_data(mp, SYNC_WAIT);
452 xfs_qm_sync(mp, SYNC_WAIT);
454 /* write superblock and hoover up shutdown errors */
455 error = xfs_sync_fsdata(mp, SYNC_WAIT);
457 /* flush data-only devices */
458 if (mp->m_rtdev_targp)
459 XFS_bflush(mp->m_rtdev_targp);
466 struct xfs_mount *mp)
468 int count = 0, pincount;
470 xfs_reclaim_inodes(mp, 0);
471 xfs_flush_buftarg(mp->m_ddev_targp, 0);
474 * This loop must run at least twice. The first instance of the loop
475 * will flush most meta data but that will generate more meta data
476 * (typically directory updates). Which then must be flushed and
477 * logged before we can write the unmount record. We also so sync
478 * reclaim of inodes to catch any that the above delwri flush skipped.
481 xfs_reclaim_inodes(mp, SYNC_WAIT);
482 xfs_sync_attr(mp, SYNC_WAIT);
483 pincount = xfs_flush_buftarg(mp->m_ddev_targp, 1);
492 * Second stage of a quiesce. The data is already synced, now we have to take
493 * care of the metadata. New transactions are already blocked, so we need to
494 * wait for any remaining transactions to drain out before proceding.
498 struct xfs_mount *mp)
502 /* wait for all modifications to complete */
503 while (atomic_read(&mp->m_active_trans) > 0)
506 /* flush inodes and push all remaining buffers out to disk */
510 * Just warn here till VFS can correctly support
511 * read-only remount without racing.
513 WARN_ON(atomic_read(&mp->m_active_trans) != 0);
515 /* Push the superblock and write an unmount record */
516 error = xfs_log_sbcount(mp, 1);
518 xfs_fs_cmn_err(CE_WARN, mp,
519 "xfs_attr_quiesce: failed to log sb changes. "
520 "Frozen image may not be consistent.");
521 xfs_log_unmount_write(mp);
522 xfs_unmountfs_writesb(mp);
526 * Enqueue a work item to be picked up by the vfs xfssyncd thread.
527 * Doing this has two advantages:
528 * - It saves on stack space, which is tight in certain situations
529 * - It can be used (with care) as a mechanism to avoid deadlocks.
530 * Flushing while allocating in a full filesystem requires both.
533 xfs_syncd_queue_work(
534 struct xfs_mount *mp,
536 void (*syncer)(struct xfs_mount *, void *),
537 struct completion *completion)
539 struct xfs_sync_work *work;
541 work = kmem_alloc(sizeof(struct xfs_sync_work), KM_SLEEP);
542 INIT_LIST_HEAD(&work->w_list);
543 work->w_syncer = syncer;
546 work->w_completion = completion;
547 spin_lock(&mp->m_sync_lock);
548 list_add_tail(&work->w_list, &mp->m_sync_list);
549 spin_unlock(&mp->m_sync_lock);
550 wake_up_process(mp->m_sync_task);
554 * Flush delayed allocate data, attempting to free up reserved space
555 * from existing allocations. At this point a new allocation attempt
556 * has failed with ENOSPC and we are in the process of scratching our
557 * heads, looking about for more room...
560 xfs_flush_inodes_work(
561 struct xfs_mount *mp,
564 struct inode *inode = arg;
565 xfs_sync_data(mp, SYNC_TRYLOCK);
566 xfs_sync_data(mp, SYNC_TRYLOCK | SYNC_WAIT);
574 struct inode *inode = VFS_I(ip);
575 DECLARE_COMPLETION_ONSTACK(completion);
578 xfs_syncd_queue_work(ip->i_mount, inode, xfs_flush_inodes_work, &completion);
579 wait_for_completion(&completion);
580 xfs_log_force(ip->i_mount, XFS_LOG_SYNC);
584 * Every sync period we need to unpin all items, reclaim inodes, sync
585 * quota and write out the superblock. We might need to cover the log
586 * to indicate it is idle.
590 struct xfs_mount *mp,
595 if (!(mp->m_flags & XFS_MOUNT_RDONLY)) {
596 xfs_log_force(mp, 0);
597 xfs_reclaim_inodes(mp, 0);
598 /* dgc: errors ignored here */
599 error = xfs_qm_sync(mp, SYNC_TRYLOCK);
600 error = xfs_sync_fsdata(mp, SYNC_TRYLOCK);
603 wake_up(&mp->m_wait_single_sync_task);
610 struct xfs_mount *mp = arg;
612 xfs_sync_work_t *work, *n;
616 timeleft = xfs_syncd_centisecs * msecs_to_jiffies(10);
618 if (list_empty(&mp->m_sync_list))
619 timeleft = schedule_timeout_interruptible(timeleft);
622 if (kthread_should_stop() && list_empty(&mp->m_sync_list))
625 spin_lock(&mp->m_sync_lock);
627 * We can get woken by laptop mode, to do a sync -
628 * that's the (only!) case where the list would be
629 * empty with time remaining.
631 if (!timeleft || list_empty(&mp->m_sync_list)) {
633 timeleft = xfs_syncd_centisecs *
634 msecs_to_jiffies(10);
635 INIT_LIST_HEAD(&mp->m_sync_work.w_list);
636 list_add_tail(&mp->m_sync_work.w_list,
639 list_splice_init(&mp->m_sync_list, &tmp);
640 spin_unlock(&mp->m_sync_lock);
642 list_for_each_entry_safe(work, n, &tmp, w_list) {
643 (*work->w_syncer)(mp, work->w_data);
644 list_del(&work->w_list);
645 if (work == &mp->m_sync_work)
647 if (work->w_completion)
648 complete(work->w_completion);
658 struct xfs_mount *mp)
660 mp->m_sync_work.w_syncer = xfs_sync_worker;
661 mp->m_sync_work.w_mount = mp;
662 mp->m_sync_work.w_completion = NULL;
663 mp->m_sync_task = kthread_run(xfssyncd, mp, "xfssyncd");
664 if (IS_ERR(mp->m_sync_task))
665 return -PTR_ERR(mp->m_sync_task);
671 struct xfs_mount *mp)
673 kthread_stop(mp->m_sync_task);
677 __xfs_inode_set_reclaim_tag(
678 struct xfs_perag *pag,
679 struct xfs_inode *ip)
681 radix_tree_tag_set(&pag->pag_ici_root,
682 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
683 XFS_ICI_RECLAIM_TAG);
684 pag->pag_ici_reclaimable++;
688 * We set the inode flag atomically with the radix tree tag.
689 * Once we get tag lookups on the radix tree, this inode flag
693 xfs_inode_set_reclaim_tag(
696 struct xfs_mount *mp = ip->i_mount;
697 struct xfs_perag *pag;
699 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
700 write_lock(&pag->pag_ici_lock);
701 spin_lock(&ip->i_flags_lock);
702 __xfs_inode_set_reclaim_tag(pag, ip);
703 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
704 spin_unlock(&ip->i_flags_lock);
705 write_unlock(&pag->pag_ici_lock);
710 __xfs_inode_clear_reclaim_tag(
715 radix_tree_tag_clear(&pag->pag_ici_root,
716 XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
717 pag->pag_ici_reclaimable--;
721 * Inodes in different states need to be treated differently, and the return
722 * value of xfs_iflush is not sufficient to get this right. The following table
723 * lists the inode states and the reclaim actions necessary for non-blocking
727 * inode state iflush ret required action
728 * --------------- ---------- ---------------
730 * shutdown EIO unpin and reclaim
731 * clean, unpinned 0 reclaim
732 * stale, unpinned 0 reclaim
733 * clean, pinned(*) 0 requeue
734 * stale, pinned EAGAIN requeue
735 * dirty, delwri ok 0 requeue
736 * dirty, delwri blocked EAGAIN requeue
737 * dirty, sync flush 0 reclaim
739 * (*) dgc: I don't think the clean, pinned state is possible but it gets
740 * handled anyway given the order of checks implemented.
742 * As can be seen from the table, the return value of xfs_iflush() is not
743 * sufficient to correctly decide the reclaim action here. The checks in
744 * xfs_iflush() might look like duplicates, but they are not.
746 * Also, because we get the flush lock first, we know that any inode that has
747 * been flushed delwri has had the flush completed by the time we check that
748 * the inode is clean. The clean inode check needs to be done before flushing
749 * the inode delwri otherwise we would loop forever requeuing clean inodes as
750 * we cannot tell apart a successful delwri flush and a clean inode from the
751 * return value of xfs_iflush().
753 * Note that because the inode is flushed delayed write by background
754 * writeback, the flush lock may already be held here and waiting on it can
755 * result in very long latencies. Hence for sync reclaims, where we wait on the
756 * flush lock, the caller should push out delayed write inodes first before
757 * trying to reclaim them to minimise the amount of time spent waiting. For
758 * background relaim, we just requeue the inode for the next pass.
760 * Hence the order of actions after gaining the locks should be:
762 * shutdown => unpin and reclaim
763 * pinned, delwri => requeue
764 * pinned, sync => unpin
767 * dirty, delwri => flush and requeue
768 * dirty, sync => flush, wait and reclaim
772 struct xfs_inode *ip,
773 struct xfs_perag *pag,
779 * The radix tree lock here protects a thread in xfs_iget from racing
780 * with us starting reclaim on the inode. Once we have the
781 * XFS_IRECLAIM flag set it will not touch us.
783 spin_lock(&ip->i_flags_lock);
784 ASSERT_ALWAYS(__xfs_iflags_test(ip, XFS_IRECLAIMABLE));
785 if (__xfs_iflags_test(ip, XFS_IRECLAIM)) {
786 /* ignore as it is already under reclaim */
787 spin_unlock(&ip->i_flags_lock);
788 write_unlock(&pag->pag_ici_lock);
791 __xfs_iflags_set(ip, XFS_IRECLAIM);
792 spin_unlock(&ip->i_flags_lock);
793 write_unlock(&pag->pag_ici_lock);
795 xfs_ilock(ip, XFS_ILOCK_EXCL);
796 if (!xfs_iflock_nowait(ip)) {
797 if (!(sync_mode & SYNC_WAIT))
802 if (is_bad_inode(VFS_I(ip)))
804 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
808 if (xfs_ipincount(ip)) {
809 if (!(sync_mode & SYNC_WAIT)) {
815 if (xfs_iflags_test(ip, XFS_ISTALE))
817 if (xfs_inode_clean(ip))
820 /* Now we have an inode that needs flushing */
821 error = xfs_iflush(ip, sync_mode);
822 if (sync_mode & SYNC_WAIT) {
828 * When we have to flush an inode but don't have SYNC_WAIT set, we
829 * flush the inode out using a delwri buffer and wait for the next
830 * call into reclaim to find it in a clean state instead of waiting for
831 * it now. We also don't return errors here - if the error is transient
832 * then the next reclaim pass will flush the inode, and if the error
833 * is permanent then the next sync reclaim will reclaim the inode and
836 if (error && error != EAGAIN && !XFS_FORCED_SHUTDOWN(ip->i_mount)) {
837 xfs_fs_cmn_err(CE_WARN, ip->i_mount,
838 "inode 0x%llx background reclaim flush failed with %d",
839 (long long)ip->i_ino, error);
842 xfs_iflags_clear(ip, XFS_IRECLAIM);
843 xfs_iunlock(ip, XFS_ILOCK_EXCL);
845 * We could return EAGAIN here to make reclaim rescan the inode tree in
846 * a short while. However, this just burns CPU time scanning the tree
847 * waiting for IO to complete and xfssyncd never goes back to the idle
848 * state. Instead, return 0 to let the next scheduled background reclaim
849 * attempt to reclaim the inode again.
855 xfs_iunlock(ip, XFS_ILOCK_EXCL);
866 return xfs_inode_ag_iterator(mp, xfs_reclaim_inode, mode,
867 XFS_ICI_RECLAIM_TAG, 1, NULL);
871 * Shrinker infrastructure.
873 * This is all far more complex than it needs to be. It adds a global list of
874 * mounts because the shrinkers can only call a global context. We need to make
875 * the shrinkers pass a context to avoid the need for global state.
877 static LIST_HEAD(xfs_mount_list);
878 static struct rw_semaphore xfs_mount_list_lock;
881 xfs_reclaim_inode_shrink(
885 struct xfs_mount *mp;
886 struct xfs_perag *pag;
891 if (!(gfp_mask & __GFP_FS))
894 down_read(&xfs_mount_list_lock);
895 list_for_each_entry(mp, &xfs_mount_list, m_mplist) {
896 xfs_inode_ag_iterator(mp, xfs_reclaim_inode, 0,
897 XFS_ICI_RECLAIM_TAG, 1, &nr_to_scan);
901 up_read(&xfs_mount_list_lock);
904 down_read(&xfs_mount_list_lock);
905 list_for_each_entry(mp, &xfs_mount_list, m_mplist) {
906 for (ag = 0; ag < mp->m_sb.sb_agcount; ag++) {
908 pag = xfs_perag_get(mp, ag);
909 if (!pag->pag_ici_init) {
913 reclaimable += pag->pag_ici_reclaimable;
917 up_read(&xfs_mount_list_lock);
921 static struct shrinker xfs_inode_shrinker = {
922 .shrink = xfs_reclaim_inode_shrink,
923 .seeks = DEFAULT_SEEKS,
927 xfs_inode_shrinker_init(void)
929 init_rwsem(&xfs_mount_list_lock);
930 register_shrinker(&xfs_inode_shrinker);
934 xfs_inode_shrinker_destroy(void)
936 ASSERT(list_empty(&xfs_mount_list));
937 unregister_shrinker(&xfs_inode_shrinker);
941 xfs_inode_shrinker_register(
942 struct xfs_mount *mp)
944 down_write(&xfs_mount_list_lock);
945 list_add_tail(&mp->m_mplist, &xfs_mount_list);
946 up_write(&xfs_mount_list_lock);
950 xfs_inode_shrinker_unregister(
951 struct xfs_mount *mp)
953 down_write(&xfs_mount_list_lock);
954 list_del(&mp->m_mplist);
955 up_write(&xfs_mount_list_lock);