2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_format.h"
21 #include "xfs_log_format.h"
22 #include "xfs_trans_resv.h"
24 #include "xfs_mount.h"
25 #include "xfs_inode.h"
26 #include "xfs_error.h"
27 #include "xfs_trans.h"
28 #include "xfs_trans_priv.h"
29 #include "xfs_inode_item.h"
30 #include "xfs_quota.h"
31 #include "xfs_trace.h"
32 #include "xfs_icache.h"
33 #include "xfs_bmap_util.h"
34 #include "xfs_dquot_item.h"
35 #include "xfs_dquot.h"
37 #include <linux/kthread.h>
38 #include <linux/freezer.h>
40 STATIC void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp,
41 struct xfs_perag *pag, struct xfs_inode *ip);
44 * Allocate and initialise an xfs_inode.
54 * if this didn't occur in transactions, we could use
55 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
56 * code up to do this anyway.
58 ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
61 if (inode_init_always(mp->m_super, VFS_I(ip))) {
62 kmem_zone_free(xfs_inode_zone, ip);
66 XFS_STATS_INC(vn_active);
67 ASSERT(atomic_read(&ip->i_pincount) == 0);
68 ASSERT(!spin_is_locked(&ip->i_flags_lock));
69 ASSERT(!xfs_isiflocked(ip));
70 ASSERT(ip->i_ino == 0);
72 mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
74 /* initialise the xfs inode */
77 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
79 memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
81 ip->i_delayed_blks = 0;
82 memset(&ip->i_d, 0, sizeof(xfs_icdinode_t));
88 xfs_inode_free_callback(
89 struct rcu_head *head)
91 struct inode *inode = container_of(head, struct inode, i_rcu);
92 struct xfs_inode *ip = XFS_I(inode);
94 kmem_zone_free(xfs_inode_zone, ip);
101 switch (ip->i_d.di_mode & S_IFMT) {
105 xfs_idestroy_fork(ip, XFS_DATA_FORK);
110 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
113 ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
114 xfs_inode_item_destroy(ip);
119 * Because we use RCU freeing we need to ensure the inode always
120 * appears to be reclaimed with an invalid inode number when in the
121 * free state. The ip->i_flags_lock provides the barrier against lookup
124 spin_lock(&ip->i_flags_lock);
125 ip->i_flags = XFS_IRECLAIM;
127 spin_unlock(&ip->i_flags_lock);
129 /* asserts to verify all state is correct here */
130 ASSERT(atomic_read(&ip->i_pincount) == 0);
131 ASSERT(!xfs_isiflocked(ip));
132 XFS_STATS_DEC(vn_active);
134 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
138 * Check the validity of the inode we just found it the cache
142 struct xfs_perag *pag,
143 struct xfs_inode *ip,
146 int lock_flags) __releases(RCU)
148 struct inode *inode = VFS_I(ip);
149 struct xfs_mount *mp = ip->i_mount;
153 * check for re-use of an inode within an RCU grace period due to the
154 * radix tree nodes not being updated yet. We monitor for this by
155 * setting the inode number to zero before freeing the inode structure.
156 * If the inode has been reallocated and set up, then the inode number
157 * will not match, so check for that, too.
159 spin_lock(&ip->i_flags_lock);
160 if (ip->i_ino != ino) {
161 trace_xfs_iget_skip(ip);
162 XFS_STATS_INC(xs_ig_frecycle);
169 * If we are racing with another cache hit that is currently
170 * instantiating this inode or currently recycling it out of
171 * reclaimabe state, wait for the initialisation to complete
174 * XXX(hch): eventually we should do something equivalent to
175 * wait_on_inode to wait for these flags to be cleared
176 * instead of polling for it.
178 if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
179 trace_xfs_iget_skip(ip);
180 XFS_STATS_INC(xs_ig_frecycle);
186 * If lookup is racing with unlink return an error immediately.
188 if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) {
194 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
195 * Need to carefully get it back into useable state.
197 if (ip->i_flags & XFS_IRECLAIMABLE) {
198 trace_xfs_iget_reclaim(ip);
201 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
202 * from stomping over us while we recycle the inode. We can't
203 * clear the radix tree reclaimable tag yet as it requires
204 * pag_ici_lock to be held exclusive.
206 ip->i_flags |= XFS_IRECLAIM;
208 spin_unlock(&ip->i_flags_lock);
211 error = inode_init_always(mp->m_super, inode);
214 * Re-initializing the inode failed, and we are in deep
215 * trouble. Try to re-add it to the reclaim list.
218 spin_lock(&ip->i_flags_lock);
220 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
221 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
222 trace_xfs_iget_reclaim_fail(ip);
226 spin_lock(&pag->pag_ici_lock);
227 spin_lock(&ip->i_flags_lock);
230 * Clear the per-lifetime state in the inode as we are now
231 * effectively a new inode and need to return to the initial
232 * state before reuse occurs.
234 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
235 ip->i_flags |= XFS_INEW;
236 __xfs_inode_clear_reclaim_tag(mp, pag, ip);
237 inode->i_state = I_NEW;
239 ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock));
240 mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
242 spin_unlock(&ip->i_flags_lock);
243 spin_unlock(&pag->pag_ici_lock);
245 /* If the VFS inode is being torn down, pause and try again. */
247 trace_xfs_iget_skip(ip);
252 /* We've got a live one. */
253 spin_unlock(&ip->i_flags_lock);
255 trace_xfs_iget_hit(ip);
259 xfs_ilock(ip, lock_flags);
261 xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
262 XFS_STATS_INC(xs_ig_found);
267 spin_unlock(&ip->i_flags_lock);
275 struct xfs_mount *mp,
276 struct xfs_perag *pag,
279 struct xfs_inode **ipp,
283 struct xfs_inode *ip;
285 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
288 ip = xfs_inode_alloc(mp, ino);
292 error = xfs_iread(mp, tp, ip, flags);
296 trace_xfs_iget_miss(ip);
298 if ((ip->i_d.di_mode == 0) && !(flags & XFS_IGET_CREATE)) {
304 * Preload the radix tree so we can insert safely under the
305 * write spinlock. Note that we cannot sleep inside the preload
306 * region. Since we can be called from transaction context, don't
307 * recurse into the file system.
309 if (radix_tree_preload(GFP_NOFS)) {
315 * Because the inode hasn't been added to the radix-tree yet it can't
316 * be found by another thread, so we can do the non-sleeping lock here.
319 if (!xfs_ilock_nowait(ip, lock_flags))
324 * These values must be set before inserting the inode into the radix
325 * tree as the moment it is inserted a concurrent lookup (allowed by the
326 * RCU locking mechanism) can find it and that lookup must see that this
327 * is an inode currently under construction (i.e. that XFS_INEW is set).
328 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
329 * memory barrier that ensures this detection works correctly at lookup
333 if (flags & XFS_IGET_DONTCACHE)
334 iflags |= XFS_IDONTCACHE;
338 xfs_iflags_set(ip, iflags);
340 /* insert the new inode */
341 spin_lock(&pag->pag_ici_lock);
342 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
343 if (unlikely(error)) {
344 WARN_ON(error != -EEXIST);
345 XFS_STATS_INC(xs_ig_dup);
347 goto out_preload_end;
349 spin_unlock(&pag->pag_ici_lock);
350 radix_tree_preload_end();
356 spin_unlock(&pag->pag_ici_lock);
357 radix_tree_preload_end();
359 xfs_iunlock(ip, lock_flags);
361 __destroy_inode(VFS_I(ip));
367 * Look up an inode by number in the given file system.
368 * The inode is looked up in the cache held in each AG.
369 * If the inode is found in the cache, initialise the vfs inode
372 * If it is not in core, read it in from the file system's device,
373 * add it to the cache and initialise the vfs inode.
375 * The inode is locked according to the value of the lock_flags parameter.
376 * This flag parameter indicates how and if the inode's IO lock and inode lock
379 * mp -- the mount point structure for the current file system. It points
380 * to the inode hash table.
381 * tp -- a pointer to the current transaction if there is one. This is
382 * simply passed through to the xfs_iread() call.
383 * ino -- the number of the inode desired. This is the unique identifier
384 * within the file system for the inode being requested.
385 * lock_flags -- flags indicating how to lock the inode. See the comment
386 * for xfs_ilock() for a list of valid values.
403 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
404 * doesn't get freed while it's being referenced during a
405 * radix tree traversal here. It assumes this function
406 * aqcuires only the ILOCK (and therefore it has no need to
407 * involve the IOLOCK in this synchronization).
409 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
411 /* reject inode numbers outside existing AGs */
412 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
415 XFS_STATS_INC(xs_ig_attempts);
417 /* get the perag structure and ensure that it's inode capable */
418 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
419 agino = XFS_INO_TO_AGINO(mp, ino);
424 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
427 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
429 goto out_error_or_again;
432 XFS_STATS_INC(xs_ig_missed);
434 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
437 goto out_error_or_again;
444 * If we have a real type for an on-disk inode, we can setup the inode
445 * now. If it's a new inode being created, xfs_ialloc will handle it.
447 if (xfs_iflags_test(ip, XFS_INEW) && ip->i_d.di_mode != 0)
448 xfs_setup_existing_inode(ip);
452 if (error == -EAGAIN) {
461 * The inode lookup is done in batches to keep the amount of lock traffic and
462 * radix tree lookups to a minimum. The batch size is a trade off between
463 * lookup reduction and stack usage. This is in the reclaim path, so we can't
466 #define XFS_LOOKUP_BATCH 32
469 xfs_inode_ag_walk_grab(
470 struct xfs_inode *ip)
472 struct inode *inode = VFS_I(ip);
474 ASSERT(rcu_read_lock_held());
477 * check for stale RCU freed inode
479 * If the inode has been reallocated, it doesn't matter if it's not in
480 * the AG we are walking - we are walking for writeback, so if it
481 * passes all the "valid inode" checks and is dirty, then we'll write
482 * it back anyway. If it has been reallocated and still being
483 * initialised, the XFS_INEW check below will catch it.
485 spin_lock(&ip->i_flags_lock);
487 goto out_unlock_noent;
489 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
490 if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM))
491 goto out_unlock_noent;
492 spin_unlock(&ip->i_flags_lock);
494 /* nothing to sync during shutdown */
495 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
496 return -EFSCORRUPTED;
498 /* If we can't grab the inode, it must on it's way to reclaim. */
506 spin_unlock(&ip->i_flags_lock);
512 struct xfs_mount *mp,
513 struct xfs_perag *pag,
514 int (*execute)(struct xfs_inode *ip, int flags,
520 uint32_t first_index;
532 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
539 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
540 (void **)batch, first_index,
543 nr_found = radix_tree_gang_lookup_tag(
545 (void **) batch, first_index,
546 XFS_LOOKUP_BATCH, tag);
554 * Grab the inodes before we drop the lock. if we found
555 * nothing, nr == 0 and the loop will be skipped.
557 for (i = 0; i < nr_found; i++) {
558 struct xfs_inode *ip = batch[i];
560 if (done || xfs_inode_ag_walk_grab(ip))
564 * Update the index for the next lookup. Catch
565 * overflows into the next AG range which can occur if
566 * we have inodes in the last block of the AG and we
567 * are currently pointing to the last inode.
569 * Because we may see inodes that are from the wrong AG
570 * due to RCU freeing and reallocation, only update the
571 * index if it lies in this AG. It was a race that lead
572 * us to see this inode, so another lookup from the
573 * same index will not find it again.
575 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
577 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
578 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
582 /* unlock now we've grabbed the inodes. */
585 for (i = 0; i < nr_found; i++) {
588 error = execute(batch[i], flags, args);
590 if (error == -EAGAIN) {
594 if (error && last_error != -EFSCORRUPTED)
598 /* bail out if the filesystem is corrupted. */
599 if (error == -EFSCORRUPTED)
604 } while (nr_found && !done);
614 * Background scanning to trim post-EOF preallocated space. This is queued
615 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
619 struct xfs_mount *mp)
622 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG))
623 queue_delayed_work(mp->m_eofblocks_workqueue,
624 &mp->m_eofblocks_work,
625 msecs_to_jiffies(xfs_eofb_secs * 1000));
630 xfs_eofblocks_worker(
631 struct work_struct *work)
633 struct xfs_mount *mp = container_of(to_delayed_work(work),
634 struct xfs_mount, m_eofblocks_work);
635 xfs_icache_free_eofblocks(mp, NULL);
636 xfs_queue_eofblocks(mp);
640 xfs_inode_ag_iterator(
641 struct xfs_mount *mp,
642 int (*execute)(struct xfs_inode *ip, int flags,
647 struct xfs_perag *pag;
653 while ((pag = xfs_perag_get(mp, ag))) {
654 ag = pag->pag_agno + 1;
655 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1);
659 if (error == -EFSCORRUPTED)
667 xfs_inode_ag_iterator_tag(
668 struct xfs_mount *mp,
669 int (*execute)(struct xfs_inode *ip, int flags,
675 struct xfs_perag *pag;
681 while ((pag = xfs_perag_get_tag(mp, ag, tag))) {
682 ag = pag->pag_agno + 1;
683 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag);
687 if (error == -EFSCORRUPTED)
695 * Queue a new inode reclaim pass if there are reclaimable inodes and there
696 * isn't a reclaim pass already in progress. By default it runs every 5s based
697 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
698 * tunable, but that can be done if this method proves to be ineffective or too
702 xfs_reclaim_work_queue(
703 struct xfs_mount *mp)
707 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
708 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
709 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
715 * This is a fast pass over the inode cache to try to get reclaim moving on as
716 * many inodes as possible in a short period of time. It kicks itself every few
717 * seconds, as well as being kicked by the inode cache shrinker when memory
718 * goes low. It scans as quickly as possible avoiding locked inodes or those
719 * already being flushed, and once done schedules a future pass.
723 struct work_struct *work)
725 struct xfs_mount *mp = container_of(to_delayed_work(work),
726 struct xfs_mount, m_reclaim_work);
728 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
729 xfs_reclaim_work_queue(mp);
733 __xfs_inode_set_reclaim_tag(
734 struct xfs_perag *pag,
735 struct xfs_inode *ip)
737 radix_tree_tag_set(&pag->pag_ici_root,
738 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
739 XFS_ICI_RECLAIM_TAG);
741 if (!pag->pag_ici_reclaimable) {
742 /* propagate the reclaim tag up into the perag radix tree */
743 spin_lock(&ip->i_mount->m_perag_lock);
744 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
745 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
746 XFS_ICI_RECLAIM_TAG);
747 spin_unlock(&ip->i_mount->m_perag_lock);
749 /* schedule periodic background inode reclaim */
750 xfs_reclaim_work_queue(ip->i_mount);
752 trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno,
755 pag->pag_ici_reclaimable++;
759 * We set the inode flag atomically with the radix tree tag.
760 * Once we get tag lookups on the radix tree, this inode flag
764 xfs_inode_set_reclaim_tag(
767 struct xfs_mount *mp = ip->i_mount;
768 struct xfs_perag *pag;
770 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
771 spin_lock(&pag->pag_ici_lock);
772 spin_lock(&ip->i_flags_lock);
773 __xfs_inode_set_reclaim_tag(pag, ip);
774 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
775 spin_unlock(&ip->i_flags_lock);
776 spin_unlock(&pag->pag_ici_lock);
781 __xfs_inode_clear_reclaim(
785 pag->pag_ici_reclaimable--;
786 if (!pag->pag_ici_reclaimable) {
787 /* clear the reclaim tag from the perag radix tree */
788 spin_lock(&ip->i_mount->m_perag_lock);
789 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
790 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
791 XFS_ICI_RECLAIM_TAG);
792 spin_unlock(&ip->i_mount->m_perag_lock);
793 trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno,
799 __xfs_inode_clear_reclaim_tag(
804 radix_tree_tag_clear(&pag->pag_ici_root,
805 XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
806 __xfs_inode_clear_reclaim(pag, ip);
810 * Grab the inode for reclaim exclusively.
811 * Return 0 if we grabbed it, non-zero otherwise.
814 xfs_reclaim_inode_grab(
815 struct xfs_inode *ip,
818 ASSERT(rcu_read_lock_held());
820 /* quick check for stale RCU freed inode */
825 * If we are asked for non-blocking operation, do unlocked checks to
826 * see if the inode already is being flushed or in reclaim to avoid
829 if ((flags & SYNC_TRYLOCK) &&
830 __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
834 * The radix tree lock here protects a thread in xfs_iget from racing
835 * with us starting reclaim on the inode. Once we have the
836 * XFS_IRECLAIM flag set it will not touch us.
838 * Due to RCU lookup, we may find inodes that have been freed and only
839 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
840 * aren't candidates for reclaim at all, so we must check the
841 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
843 spin_lock(&ip->i_flags_lock);
844 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
845 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
846 /* not a reclaim candidate. */
847 spin_unlock(&ip->i_flags_lock);
850 __xfs_iflags_set(ip, XFS_IRECLAIM);
851 spin_unlock(&ip->i_flags_lock);
856 * Inodes in different states need to be treated differently. The following
857 * table lists the inode states and the reclaim actions necessary:
859 * inode state iflush ret required action
860 * --------------- ---------- ---------------
862 * shutdown EIO unpin and reclaim
863 * clean, unpinned 0 reclaim
864 * stale, unpinned 0 reclaim
865 * clean, pinned(*) 0 requeue
866 * stale, pinned EAGAIN requeue
867 * dirty, async - requeue
868 * dirty, sync 0 reclaim
870 * (*) dgc: I don't think the clean, pinned state is possible but it gets
871 * handled anyway given the order of checks implemented.
873 * Also, because we get the flush lock first, we know that any inode that has
874 * been flushed delwri has had the flush completed by the time we check that
875 * the inode is clean.
877 * Note that because the inode is flushed delayed write by AIL pushing, the
878 * flush lock may already be held here and waiting on it can result in very
879 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
880 * the caller should push the AIL first before trying to reclaim inodes to
881 * minimise the amount of time spent waiting. For background relaim, we only
882 * bother to reclaim clean inodes anyway.
884 * Hence the order of actions after gaining the locks should be:
886 * shutdown => unpin and reclaim
887 * pinned, async => requeue
888 * pinned, sync => unpin
891 * dirty, async => requeue
892 * dirty, sync => flush, wait and reclaim
896 struct xfs_inode *ip,
897 struct xfs_perag *pag,
900 struct xfs_buf *bp = NULL;
905 xfs_ilock(ip, XFS_ILOCK_EXCL);
906 if (!xfs_iflock_nowait(ip)) {
907 if (!(sync_mode & SYNC_WAIT))
912 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
914 xfs_iflush_abort(ip, false);
917 if (xfs_ipincount(ip)) {
918 if (!(sync_mode & SYNC_WAIT))
922 if (xfs_iflags_test(ip, XFS_ISTALE))
924 if (xfs_inode_clean(ip))
928 * Never flush out dirty data during non-blocking reclaim, as it would
929 * just contend with AIL pushing trying to do the same job.
931 if (!(sync_mode & SYNC_WAIT))
935 * Now we have an inode that needs flushing.
937 * Note that xfs_iflush will never block on the inode buffer lock, as
938 * xfs_ifree_cluster() can lock the inode buffer before it locks the
939 * ip->i_lock, and we are doing the exact opposite here. As a result,
940 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
941 * result in an ABBA deadlock with xfs_ifree_cluster().
943 * As xfs_ifree_cluser() must gather all inodes that are active in the
944 * cache to mark them stale, if we hit this case we don't actually want
945 * to do IO here - we want the inode marked stale so we can simply
946 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
947 * inode, back off and try again. Hopefully the next pass through will
948 * see the stale flag set on the inode.
950 error = xfs_iflush(ip, &bp);
951 if (error == -EAGAIN) {
952 xfs_iunlock(ip, XFS_ILOCK_EXCL);
953 /* backoff longer than in xfs_ifree_cluster */
959 error = xfs_bwrite(bp);
966 xfs_iunlock(ip, XFS_ILOCK_EXCL);
968 XFS_STATS_INC(xs_ig_reclaims);
970 * Remove the inode from the per-AG radix tree.
972 * Because radix_tree_delete won't complain even if the item was never
973 * added to the tree assert that it's been there before to catch
974 * problems with the inode life time early on.
976 spin_lock(&pag->pag_ici_lock);
977 if (!radix_tree_delete(&pag->pag_ici_root,
978 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino)))
980 __xfs_inode_clear_reclaim(pag, ip);
981 spin_unlock(&pag->pag_ici_lock);
984 * Here we do an (almost) spurious inode lock in order to coordinate
985 * with inode cache radix tree lookups. This is because the lookup
986 * can reference the inodes in the cache without taking references.
988 * We make that OK here by ensuring that we wait until the inode is
989 * unlocked after the lookup before we go ahead and free it.
991 xfs_ilock(ip, XFS_ILOCK_EXCL);
993 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1001 xfs_iflags_clear(ip, XFS_IRECLAIM);
1002 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1004 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1005 * a short while. However, this just burns CPU time scanning the tree
1006 * waiting for IO to complete and the reclaim work never goes back to
1007 * the idle state. Instead, return 0 to let the next scheduled
1008 * background reclaim attempt to reclaim the inode again.
1014 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1015 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1016 * then a shut down during filesystem unmount reclaim walk leak all the
1017 * unreclaimed inodes.
1020 xfs_reclaim_inodes_ag(
1021 struct xfs_mount *mp,
1025 struct xfs_perag *pag;
1029 int trylock = flags & SYNC_TRYLOCK;
1035 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1036 unsigned long first_index = 0;
1040 ag = pag->pag_agno + 1;
1043 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
1048 first_index = pag->pag_ici_reclaim_cursor;
1050 mutex_lock(&pag->pag_ici_reclaim_lock);
1053 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1057 nr_found = radix_tree_gang_lookup_tag(
1059 (void **)batch, first_index,
1061 XFS_ICI_RECLAIM_TAG);
1069 * Grab the inodes before we drop the lock. if we found
1070 * nothing, nr == 0 and the loop will be skipped.
1072 for (i = 0; i < nr_found; i++) {
1073 struct xfs_inode *ip = batch[i];
1075 if (done || xfs_reclaim_inode_grab(ip, flags))
1079 * Update the index for the next lookup. Catch
1080 * overflows into the next AG range which can
1081 * occur if we have inodes in the last block of
1082 * the AG and we are currently pointing to the
1085 * Because we may see inodes that are from the
1086 * wrong AG due to RCU freeing and
1087 * reallocation, only update the index if it
1088 * lies in this AG. It was a race that lead us
1089 * to see this inode, so another lookup from
1090 * the same index will not find it again.
1092 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1095 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1096 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1100 /* unlock now we've grabbed the inodes. */
1103 for (i = 0; i < nr_found; i++) {
1106 error = xfs_reclaim_inode(batch[i], pag, flags);
1107 if (error && last_error != -EFSCORRUPTED)
1111 *nr_to_scan -= XFS_LOOKUP_BATCH;
1115 } while (nr_found && !done && *nr_to_scan > 0);
1117 if (trylock && !done)
1118 pag->pag_ici_reclaim_cursor = first_index;
1120 pag->pag_ici_reclaim_cursor = 0;
1121 mutex_unlock(&pag->pag_ici_reclaim_lock);
1126 * if we skipped any AG, and we still have scan count remaining, do
1127 * another pass this time using blocking reclaim semantics (i.e
1128 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1129 * ensure that when we get more reclaimers than AGs we block rather
1130 * than spin trying to execute reclaim.
1132 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1144 int nr_to_scan = INT_MAX;
1146 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1150 * Scan a certain number of inodes for reclaim.
1152 * When called we make sure that there is a background (fast) inode reclaim in
1153 * progress, while we will throttle the speed of reclaim via doing synchronous
1154 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1155 * them to be cleaned, which we hope will not be very long due to the
1156 * background walker having already kicked the IO off on those dirty inodes.
1159 xfs_reclaim_inodes_nr(
1160 struct xfs_mount *mp,
1163 /* kick background reclaimer and push the AIL */
1164 xfs_reclaim_work_queue(mp);
1165 xfs_ail_push_all(mp->m_ail);
1167 return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1171 * Return the number of reclaimable inodes in the filesystem for
1172 * the shrinker to determine how much to reclaim.
1175 xfs_reclaim_inodes_count(
1176 struct xfs_mount *mp)
1178 struct xfs_perag *pag;
1179 xfs_agnumber_t ag = 0;
1180 int reclaimable = 0;
1182 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1183 ag = pag->pag_agno + 1;
1184 reclaimable += pag->pag_ici_reclaimable;
1192 struct xfs_inode *ip,
1193 struct xfs_eofblocks *eofb)
1195 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1196 !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1199 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1200 !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1203 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1204 xfs_get_projid(ip) != eofb->eof_prid)
1211 * A union-based inode filtering algorithm. Process the inode if any of the
1212 * criteria match. This is for global/internal scans only.
1215 xfs_inode_match_id_union(
1216 struct xfs_inode *ip,
1217 struct xfs_eofblocks *eofb)
1219 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1220 uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1223 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1224 gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1227 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1228 xfs_get_projid(ip) == eofb->eof_prid)
1235 xfs_inode_free_eofblocks(
1236 struct xfs_inode *ip,
1241 struct xfs_eofblocks *eofb = args;
1242 bool need_iolock = true;
1245 ASSERT(!eofb || (eofb && eofb->eof_scan_owner != 0));
1247 if (!xfs_can_free_eofblocks(ip, false)) {
1248 /* inode could be preallocated or append-only */
1249 trace_xfs_inode_free_eofblocks_invalid(ip);
1250 xfs_inode_clear_eofblocks_tag(ip);
1255 * If the mapping is dirty the operation can block and wait for some
1256 * time. Unless we are waiting, skip it.
1258 if (!(flags & SYNC_WAIT) &&
1259 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1263 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1264 match = xfs_inode_match_id_union(ip, eofb);
1266 match = xfs_inode_match_id(ip, eofb);
1270 /* skip the inode if the file size is too small */
1271 if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1272 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1276 * A scan owner implies we already hold the iolock. Skip it in
1277 * xfs_free_eofblocks() to avoid deadlock. This also eliminates
1278 * the possibility of EAGAIN being returned.
1280 if (eofb->eof_scan_owner == ip->i_ino)
1281 need_iolock = false;
1284 ret = xfs_free_eofblocks(ip->i_mount, ip, need_iolock);
1286 /* don't revisit the inode if we're not waiting */
1287 if (ret == -EAGAIN && !(flags & SYNC_WAIT))
1294 xfs_icache_free_eofblocks(
1295 struct xfs_mount *mp,
1296 struct xfs_eofblocks *eofb)
1298 int flags = SYNC_TRYLOCK;
1300 if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
1303 return xfs_inode_ag_iterator_tag(mp, xfs_inode_free_eofblocks, flags,
1304 eofb, XFS_ICI_EOFBLOCKS_TAG);
1308 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1309 * multiple quotas, we don't know exactly which quota caused an allocation
1310 * failure. We make a best effort by including each quota under low free space
1311 * conditions (less than 1% free space) in the scan.
1314 xfs_inode_free_quota_eofblocks(
1315 struct xfs_inode *ip)
1318 struct xfs_eofblocks eofb = {0};
1319 struct xfs_dquot *dq;
1321 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1324 * Set the scan owner to avoid a potential livelock. Otherwise, the scan
1325 * can repeatedly trylock on the inode we're currently processing. We
1326 * run a sync scan to increase effectiveness and use the union filter to
1327 * cover all applicable quotas in a single scan.
1329 eofb.eof_scan_owner = ip->i_ino;
1330 eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC;
1332 if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) {
1333 dq = xfs_inode_dquot(ip, XFS_DQ_USER);
1334 if (dq && xfs_dquot_lowsp(dq)) {
1335 eofb.eof_uid = VFS_I(ip)->i_uid;
1336 eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1341 if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) {
1342 dq = xfs_inode_dquot(ip, XFS_DQ_GROUP);
1343 if (dq && xfs_dquot_lowsp(dq)) {
1344 eofb.eof_gid = VFS_I(ip)->i_gid;
1345 eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1351 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
1357 xfs_inode_set_eofblocks_tag(
1360 struct xfs_mount *mp = ip->i_mount;
1361 struct xfs_perag *pag;
1364 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1365 spin_lock(&pag->pag_ici_lock);
1366 trace_xfs_inode_set_eofblocks_tag(ip);
1368 tagged = radix_tree_tagged(&pag->pag_ici_root,
1369 XFS_ICI_EOFBLOCKS_TAG);
1370 radix_tree_tag_set(&pag->pag_ici_root,
1371 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1372 XFS_ICI_EOFBLOCKS_TAG);
1374 /* propagate the eofblocks tag up into the perag radix tree */
1375 spin_lock(&ip->i_mount->m_perag_lock);
1376 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
1377 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1378 XFS_ICI_EOFBLOCKS_TAG);
1379 spin_unlock(&ip->i_mount->m_perag_lock);
1381 /* kick off background trimming */
1382 xfs_queue_eofblocks(ip->i_mount);
1384 trace_xfs_perag_set_eofblocks(ip->i_mount, pag->pag_agno,
1388 spin_unlock(&pag->pag_ici_lock);
1393 xfs_inode_clear_eofblocks_tag(
1396 struct xfs_mount *mp = ip->i_mount;
1397 struct xfs_perag *pag;
1399 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1400 spin_lock(&pag->pag_ici_lock);
1401 trace_xfs_inode_clear_eofblocks_tag(ip);
1403 radix_tree_tag_clear(&pag->pag_ici_root,
1404 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1405 XFS_ICI_EOFBLOCKS_TAG);
1406 if (!radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_EOFBLOCKS_TAG)) {
1407 /* clear the eofblocks tag from the perag radix tree */
1408 spin_lock(&ip->i_mount->m_perag_lock);
1409 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1410 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1411 XFS_ICI_EOFBLOCKS_TAG);
1412 spin_unlock(&ip->i_mount->m_perag_lock);
1413 trace_xfs_perag_clear_eofblocks(ip->i_mount, pag->pag_agno,
1417 spin_unlock(&pag->pag_ici_lock);