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>
41 * Allocate and initialise an xfs_inode.
51 * if this didn't occur in transactions, we could use
52 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
53 * code up to do this anyway.
55 ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
58 if (inode_init_always(mp->m_super, VFS_I(ip))) {
59 kmem_zone_free(xfs_inode_zone, ip);
63 /* VFS doesn't initialise i_mode! */
64 VFS_I(ip)->i_mode = 0;
66 XFS_STATS_INC(mp, 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));
81 ip->i_cformat = XFS_DINODE_FMT_EXTENTS;
82 memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
84 ip->i_delayed_blks = 0;
85 memset(&ip->i_d, 0, sizeof(ip->i_d));
91 xfs_inode_free_callback(
92 struct rcu_head *head)
94 struct inode *inode = container_of(head, struct inode, i_rcu);
95 struct xfs_inode *ip = XFS_I(inode);
97 switch (VFS_I(ip)->i_mode & S_IFMT) {
101 xfs_idestroy_fork(ip, XFS_DATA_FORK);
106 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
108 xfs_idestroy_fork(ip, XFS_COW_FORK);
111 ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
112 xfs_inode_item_destroy(ip);
116 kmem_zone_free(xfs_inode_zone, ip);
121 struct xfs_inode *ip)
123 /* asserts to verify all state is correct here */
124 ASSERT(atomic_read(&ip->i_pincount) == 0);
125 ASSERT(!xfs_isiflocked(ip));
126 XFS_STATS_DEC(ip->i_mount, vn_active);
128 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
133 struct xfs_inode *ip)
136 * Because we use RCU freeing we need to ensure the inode always
137 * appears to be reclaimed with an invalid inode number when in the
138 * free state. The ip->i_flags_lock provides the barrier against lookup
141 spin_lock(&ip->i_flags_lock);
142 ip->i_flags = XFS_IRECLAIM;
144 spin_unlock(&ip->i_flags_lock);
146 __xfs_inode_free(ip);
150 * Queue a new inode reclaim pass if there are reclaimable inodes and there
151 * isn't a reclaim pass already in progress. By default it runs every 5s based
152 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
153 * tunable, but that can be done if this method proves to be ineffective or too
157 xfs_reclaim_work_queue(
158 struct xfs_mount *mp)
162 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
163 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
164 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
170 * This is a fast pass over the inode cache to try to get reclaim moving on as
171 * many inodes as possible in a short period of time. It kicks itself every few
172 * seconds, as well as being kicked by the inode cache shrinker when memory
173 * goes low. It scans as quickly as possible avoiding locked inodes or those
174 * already being flushed, and once done schedules a future pass.
178 struct work_struct *work)
180 struct xfs_mount *mp = container_of(to_delayed_work(work),
181 struct xfs_mount, m_reclaim_work);
183 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
184 xfs_reclaim_work_queue(mp);
188 xfs_perag_set_reclaim_tag(
189 struct xfs_perag *pag)
191 struct xfs_mount *mp = pag->pag_mount;
193 ASSERT(spin_is_locked(&pag->pag_ici_lock));
194 if (pag->pag_ici_reclaimable++)
197 /* propagate the reclaim tag up into the perag radix tree */
198 spin_lock(&mp->m_perag_lock);
199 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno,
200 XFS_ICI_RECLAIM_TAG);
201 spin_unlock(&mp->m_perag_lock);
203 /* schedule periodic background inode reclaim */
204 xfs_reclaim_work_queue(mp);
206 trace_xfs_perag_set_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
210 xfs_perag_clear_reclaim_tag(
211 struct xfs_perag *pag)
213 struct xfs_mount *mp = pag->pag_mount;
215 ASSERT(spin_is_locked(&pag->pag_ici_lock));
216 if (--pag->pag_ici_reclaimable)
219 /* clear the reclaim tag from the perag radix tree */
220 spin_lock(&mp->m_perag_lock);
221 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno,
222 XFS_ICI_RECLAIM_TAG);
223 spin_unlock(&mp->m_perag_lock);
224 trace_xfs_perag_clear_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
229 * We set the inode flag atomically with the radix tree tag.
230 * Once we get tag lookups on the radix tree, this inode flag
234 xfs_inode_set_reclaim_tag(
235 struct xfs_inode *ip)
237 struct xfs_mount *mp = ip->i_mount;
238 struct xfs_perag *pag;
240 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
241 spin_lock(&pag->pag_ici_lock);
242 spin_lock(&ip->i_flags_lock);
244 radix_tree_tag_set(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, ip->i_ino),
245 XFS_ICI_RECLAIM_TAG);
246 xfs_perag_set_reclaim_tag(pag);
247 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
249 spin_unlock(&ip->i_flags_lock);
250 spin_unlock(&pag->pag_ici_lock);
255 xfs_inode_clear_reclaim_tag(
256 struct xfs_perag *pag,
259 radix_tree_tag_clear(&pag->pag_ici_root,
260 XFS_INO_TO_AGINO(pag->pag_mount, ino),
261 XFS_ICI_RECLAIM_TAG);
262 xfs_perag_clear_reclaim_tag(pag);
266 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
267 * part of the structure. This is made more complex by the fact we store
268 * information about the on-disk values in the VFS inode and so we can't just
269 * overwrite the values unconditionally. Hence we save the parameters we
270 * need to retain across reinitialisation, and rewrite them into the VFS inode
271 * after reinitialisation even if it fails.
275 struct xfs_mount *mp,
279 uint32_t nlink = inode->i_nlink;
280 uint32_t generation = inode->i_generation;
281 uint64_t version = inode->i_version;
282 umode_t mode = inode->i_mode;
284 error = inode_init_always(mp->m_super, inode);
286 set_nlink(inode, nlink);
287 inode->i_generation = generation;
288 inode->i_version = version;
289 inode->i_mode = mode;
294 * Check the validity of the inode we just found it the cache
298 struct xfs_perag *pag,
299 struct xfs_inode *ip,
302 int lock_flags) __releases(RCU)
304 struct inode *inode = VFS_I(ip);
305 struct xfs_mount *mp = ip->i_mount;
309 * check for re-use of an inode within an RCU grace period due to the
310 * radix tree nodes not being updated yet. We monitor for this by
311 * setting the inode number to zero before freeing the inode structure.
312 * If the inode has been reallocated and set up, then the inode number
313 * will not match, so check for that, too.
315 spin_lock(&ip->i_flags_lock);
316 if (ip->i_ino != ino) {
317 trace_xfs_iget_skip(ip);
318 XFS_STATS_INC(mp, xs_ig_frecycle);
325 * If we are racing with another cache hit that is currently
326 * instantiating this inode or currently recycling it out of
327 * reclaimabe state, wait for the initialisation to complete
330 * XXX(hch): eventually we should do something equivalent to
331 * wait_on_inode to wait for these flags to be cleared
332 * instead of polling for it.
334 if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
335 trace_xfs_iget_skip(ip);
336 XFS_STATS_INC(mp, xs_ig_frecycle);
342 * If lookup is racing with unlink return an error immediately.
344 if (VFS_I(ip)->i_mode == 0 && !(flags & XFS_IGET_CREATE)) {
350 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
351 * Need to carefully get it back into useable state.
353 if (ip->i_flags & XFS_IRECLAIMABLE) {
354 trace_xfs_iget_reclaim(ip);
357 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
358 * from stomping over us while we recycle the inode. We can't
359 * clear the radix tree reclaimable tag yet as it requires
360 * pag_ici_lock to be held exclusive.
362 ip->i_flags |= XFS_IRECLAIM;
364 spin_unlock(&ip->i_flags_lock);
367 error = xfs_reinit_inode(mp, inode);
370 * Re-initializing the inode failed, and we are in deep
371 * trouble. Try to re-add it to the reclaim list.
374 spin_lock(&ip->i_flags_lock);
376 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
377 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
378 trace_xfs_iget_reclaim_fail(ip);
382 spin_lock(&pag->pag_ici_lock);
383 spin_lock(&ip->i_flags_lock);
386 * Clear the per-lifetime state in the inode as we are now
387 * effectively a new inode and need to return to the initial
388 * state before reuse occurs.
390 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
391 ip->i_flags |= XFS_INEW;
392 xfs_inode_clear_reclaim_tag(pag, ip->i_ino);
393 inode->i_state = I_NEW;
395 ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock));
396 mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
398 spin_unlock(&ip->i_flags_lock);
399 spin_unlock(&pag->pag_ici_lock);
401 /* If the VFS inode is being torn down, pause and try again. */
403 trace_xfs_iget_skip(ip);
408 /* We've got a live one. */
409 spin_unlock(&ip->i_flags_lock);
411 trace_xfs_iget_hit(ip);
415 xfs_ilock(ip, lock_flags);
417 xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
418 XFS_STATS_INC(mp, xs_ig_found);
423 spin_unlock(&ip->i_flags_lock);
431 struct xfs_mount *mp,
432 struct xfs_perag *pag,
435 struct xfs_inode **ipp,
439 struct xfs_inode *ip;
441 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
444 ip = xfs_inode_alloc(mp, ino);
448 error = xfs_iread(mp, tp, ip, flags);
452 trace_xfs_iget_miss(ip);
454 if ((VFS_I(ip)->i_mode == 0) && !(flags & XFS_IGET_CREATE)) {
460 * Preload the radix tree so we can insert safely under the
461 * write spinlock. Note that we cannot sleep inside the preload
462 * region. Since we can be called from transaction context, don't
463 * recurse into the file system.
465 if (radix_tree_preload(GFP_NOFS)) {
471 * Because the inode hasn't been added to the radix-tree yet it can't
472 * be found by another thread, so we can do the non-sleeping lock here.
475 if (!xfs_ilock_nowait(ip, lock_flags))
480 * These values must be set before inserting the inode into the radix
481 * tree as the moment it is inserted a concurrent lookup (allowed by the
482 * RCU locking mechanism) can find it and that lookup must see that this
483 * is an inode currently under construction (i.e. that XFS_INEW is set).
484 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
485 * memory barrier that ensures this detection works correctly at lookup
489 if (flags & XFS_IGET_DONTCACHE)
490 iflags |= XFS_IDONTCACHE;
494 xfs_iflags_set(ip, iflags);
496 /* insert the new inode */
497 spin_lock(&pag->pag_ici_lock);
498 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
499 if (unlikely(error)) {
500 WARN_ON(error != -EEXIST);
501 XFS_STATS_INC(mp, xs_ig_dup);
503 goto out_preload_end;
505 spin_unlock(&pag->pag_ici_lock);
506 radix_tree_preload_end();
512 spin_unlock(&pag->pag_ici_lock);
513 radix_tree_preload_end();
515 xfs_iunlock(ip, lock_flags);
517 __destroy_inode(VFS_I(ip));
523 * Look up an inode by number in the given file system.
524 * The inode is looked up in the cache held in each AG.
525 * If the inode is found in the cache, initialise the vfs inode
528 * If it is not in core, read it in from the file system's device,
529 * add it to the cache and initialise the vfs inode.
531 * The inode is locked according to the value of the lock_flags parameter.
532 * This flag parameter indicates how and if the inode's IO lock and inode lock
535 * mp -- the mount point structure for the current file system. It points
536 * to the inode hash table.
537 * tp -- a pointer to the current transaction if there is one. This is
538 * simply passed through to the xfs_iread() call.
539 * ino -- the number of the inode desired. This is the unique identifier
540 * within the file system for the inode being requested.
541 * lock_flags -- flags indicating how to lock the inode. See the comment
542 * for xfs_ilock() for a list of valid values.
559 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
560 * doesn't get freed while it's being referenced during a
561 * radix tree traversal here. It assumes this function
562 * aqcuires only the ILOCK (and therefore it has no need to
563 * involve the IOLOCK in this synchronization).
565 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
567 /* reject inode numbers outside existing AGs */
568 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
571 XFS_STATS_INC(mp, xs_ig_attempts);
573 /* get the perag structure and ensure that it's inode capable */
574 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
575 agino = XFS_INO_TO_AGINO(mp, ino);
580 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
583 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
585 goto out_error_or_again;
588 XFS_STATS_INC(mp, xs_ig_missed);
590 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
593 goto out_error_or_again;
600 * If we have a real type for an on-disk inode, we can setup the inode
601 * now. If it's a new inode being created, xfs_ialloc will handle it.
603 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
604 xfs_setup_existing_inode(ip);
608 if (error == -EAGAIN) {
617 * The inode lookup is done in batches to keep the amount of lock traffic and
618 * radix tree lookups to a minimum. The batch size is a trade off between
619 * lookup reduction and stack usage. This is in the reclaim path, so we can't
622 #define XFS_LOOKUP_BATCH 32
625 xfs_inode_ag_walk_grab(
626 struct xfs_inode *ip)
628 struct inode *inode = VFS_I(ip);
630 ASSERT(rcu_read_lock_held());
633 * check for stale RCU freed inode
635 * If the inode has been reallocated, it doesn't matter if it's not in
636 * the AG we are walking - we are walking for writeback, so if it
637 * passes all the "valid inode" checks and is dirty, then we'll write
638 * it back anyway. If it has been reallocated and still being
639 * initialised, the XFS_INEW check below will catch it.
641 spin_lock(&ip->i_flags_lock);
643 goto out_unlock_noent;
645 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
646 if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM))
647 goto out_unlock_noent;
648 spin_unlock(&ip->i_flags_lock);
650 /* nothing to sync during shutdown */
651 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
652 return -EFSCORRUPTED;
654 /* If we can't grab the inode, it must on it's way to reclaim. */
662 spin_unlock(&ip->i_flags_lock);
668 struct xfs_mount *mp,
669 struct xfs_perag *pag,
670 int (*execute)(struct xfs_inode *ip, int flags,
676 uint32_t first_index;
688 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
695 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
696 (void **)batch, first_index,
699 nr_found = radix_tree_gang_lookup_tag(
701 (void **) batch, first_index,
702 XFS_LOOKUP_BATCH, tag);
710 * Grab the inodes before we drop the lock. if we found
711 * nothing, nr == 0 and the loop will be skipped.
713 for (i = 0; i < nr_found; i++) {
714 struct xfs_inode *ip = batch[i];
716 if (done || xfs_inode_ag_walk_grab(ip))
720 * Update the index for the next lookup. Catch
721 * overflows into the next AG range which can occur if
722 * we have inodes in the last block of the AG and we
723 * are currently pointing to the last inode.
725 * Because we may see inodes that are from the wrong AG
726 * due to RCU freeing and reallocation, only update the
727 * index if it lies in this AG. It was a race that lead
728 * us to see this inode, so another lookup from the
729 * same index will not find it again.
731 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
733 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
734 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
738 /* unlock now we've grabbed the inodes. */
741 for (i = 0; i < nr_found; i++) {
744 error = execute(batch[i], flags, args);
746 if (error == -EAGAIN) {
750 if (error && last_error != -EFSCORRUPTED)
754 /* bail out if the filesystem is corrupted. */
755 if (error == -EFSCORRUPTED)
760 } while (nr_found && !done);
770 * Background scanning to trim post-EOF preallocated space. This is queued
771 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
775 struct xfs_mount *mp)
778 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG))
779 queue_delayed_work(mp->m_eofblocks_workqueue,
780 &mp->m_eofblocks_work,
781 msecs_to_jiffies(xfs_eofb_secs * 1000));
786 xfs_eofblocks_worker(
787 struct work_struct *work)
789 struct xfs_mount *mp = container_of(to_delayed_work(work),
790 struct xfs_mount, m_eofblocks_work);
791 xfs_icache_free_eofblocks(mp, NULL);
792 xfs_queue_eofblocks(mp);
796 xfs_inode_ag_iterator(
797 struct xfs_mount *mp,
798 int (*execute)(struct xfs_inode *ip, int flags,
803 struct xfs_perag *pag;
809 while ((pag = xfs_perag_get(mp, ag))) {
810 ag = pag->pag_agno + 1;
811 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1);
815 if (error == -EFSCORRUPTED)
823 xfs_inode_ag_iterator_tag(
824 struct xfs_mount *mp,
825 int (*execute)(struct xfs_inode *ip, int flags,
831 struct xfs_perag *pag;
837 while ((pag = xfs_perag_get_tag(mp, ag, tag))) {
838 ag = pag->pag_agno + 1;
839 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag);
843 if (error == -EFSCORRUPTED)
851 * Grab the inode for reclaim exclusively.
852 * Return 0 if we grabbed it, non-zero otherwise.
855 xfs_reclaim_inode_grab(
856 struct xfs_inode *ip,
859 ASSERT(rcu_read_lock_held());
861 /* quick check for stale RCU freed inode */
866 * If we are asked for non-blocking operation, do unlocked checks to
867 * see if the inode already is being flushed or in reclaim to avoid
870 if ((flags & SYNC_TRYLOCK) &&
871 __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
875 * The radix tree lock here protects a thread in xfs_iget from racing
876 * with us starting reclaim on the inode. Once we have the
877 * XFS_IRECLAIM flag set it will not touch us.
879 * Due to RCU lookup, we may find inodes that have been freed and only
880 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
881 * aren't candidates for reclaim at all, so we must check the
882 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
884 spin_lock(&ip->i_flags_lock);
885 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
886 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
887 /* not a reclaim candidate. */
888 spin_unlock(&ip->i_flags_lock);
891 __xfs_iflags_set(ip, XFS_IRECLAIM);
892 spin_unlock(&ip->i_flags_lock);
897 * Inodes in different states need to be treated differently. The following
898 * table lists the inode states and the reclaim actions necessary:
900 * inode state iflush ret required action
901 * --------------- ---------- ---------------
903 * shutdown EIO unpin and reclaim
904 * clean, unpinned 0 reclaim
905 * stale, unpinned 0 reclaim
906 * clean, pinned(*) 0 requeue
907 * stale, pinned EAGAIN requeue
908 * dirty, async - requeue
909 * dirty, sync 0 reclaim
911 * (*) dgc: I don't think the clean, pinned state is possible but it gets
912 * handled anyway given the order of checks implemented.
914 * Also, because we get the flush lock first, we know that any inode that has
915 * been flushed delwri has had the flush completed by the time we check that
916 * the inode is clean.
918 * Note that because the inode is flushed delayed write by AIL pushing, the
919 * flush lock may already be held here and waiting on it can result in very
920 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
921 * the caller should push the AIL first before trying to reclaim inodes to
922 * minimise the amount of time spent waiting. For background relaim, we only
923 * bother to reclaim clean inodes anyway.
925 * Hence the order of actions after gaining the locks should be:
927 * shutdown => unpin and reclaim
928 * pinned, async => requeue
929 * pinned, sync => unpin
932 * dirty, async => requeue
933 * dirty, sync => flush, wait and reclaim
937 struct xfs_inode *ip,
938 struct xfs_perag *pag,
941 struct xfs_buf *bp = NULL;
942 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
947 xfs_ilock(ip, XFS_ILOCK_EXCL);
948 if (!xfs_iflock_nowait(ip)) {
949 if (!(sync_mode & SYNC_WAIT))
954 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
956 xfs_iflush_abort(ip, false);
959 if (xfs_ipincount(ip)) {
960 if (!(sync_mode & SYNC_WAIT))
964 if (xfs_iflags_test(ip, XFS_ISTALE))
966 if (xfs_inode_clean(ip))
970 * Never flush out dirty data during non-blocking reclaim, as it would
971 * just contend with AIL pushing trying to do the same job.
973 if (!(sync_mode & SYNC_WAIT))
977 * Now we have an inode that needs flushing.
979 * Note that xfs_iflush will never block on the inode buffer lock, as
980 * xfs_ifree_cluster() can lock the inode buffer before it locks the
981 * ip->i_lock, and we are doing the exact opposite here. As a result,
982 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
983 * result in an ABBA deadlock with xfs_ifree_cluster().
985 * As xfs_ifree_cluser() must gather all inodes that are active in the
986 * cache to mark them stale, if we hit this case we don't actually want
987 * to do IO here - we want the inode marked stale so we can simply
988 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
989 * inode, back off and try again. Hopefully the next pass through will
990 * see the stale flag set on the inode.
992 error = xfs_iflush(ip, &bp);
993 if (error == -EAGAIN) {
994 xfs_iunlock(ip, XFS_ILOCK_EXCL);
995 /* backoff longer than in xfs_ifree_cluster */
1001 error = xfs_bwrite(bp);
1008 * Because we use RCU freeing we need to ensure the inode always appears
1009 * to be reclaimed with an invalid inode number when in the free state.
1010 * We do this as early as possible under the ILOCK and flush lock so
1011 * that xfs_iflush_cluster() can be guaranteed to detect races with us
1012 * here. By doing this, we guarantee that once xfs_iflush_cluster has
1013 * locked both the XFS_ILOCK and the flush lock that it will see either
1014 * a valid, flushable inode that will serialise correctly against the
1015 * locks below, or it will see a clean (and invalid) inode that it can
1018 spin_lock(&ip->i_flags_lock);
1019 ip->i_flags = XFS_IRECLAIM;
1021 spin_unlock(&ip->i_flags_lock);
1024 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1026 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1028 * Remove the inode from the per-AG radix tree.
1030 * Because radix_tree_delete won't complain even if the item was never
1031 * added to the tree assert that it's been there before to catch
1032 * problems with the inode life time early on.
1034 spin_lock(&pag->pag_ici_lock);
1035 if (!radix_tree_delete(&pag->pag_ici_root,
1036 XFS_INO_TO_AGINO(ip->i_mount, ino)))
1038 xfs_perag_clear_reclaim_tag(pag);
1039 spin_unlock(&pag->pag_ici_lock);
1042 * Here we do an (almost) spurious inode lock in order to coordinate
1043 * with inode cache radix tree lookups. This is because the lookup
1044 * can reference the inodes in the cache without taking references.
1046 * We make that OK here by ensuring that we wait until the inode is
1047 * unlocked after the lookup before we go ahead and free it.
1049 xfs_ilock(ip, XFS_ILOCK_EXCL);
1050 xfs_qm_dqdetach(ip);
1051 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1053 __xfs_inode_free(ip);
1059 xfs_iflags_clear(ip, XFS_IRECLAIM);
1060 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1062 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1063 * a short while. However, this just burns CPU time scanning the tree
1064 * waiting for IO to complete and the reclaim work never goes back to
1065 * the idle state. Instead, return 0 to let the next scheduled
1066 * background reclaim attempt to reclaim the inode again.
1072 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1073 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1074 * then a shut down during filesystem unmount reclaim walk leak all the
1075 * unreclaimed inodes.
1078 xfs_reclaim_inodes_ag(
1079 struct xfs_mount *mp,
1083 struct xfs_perag *pag;
1087 int trylock = flags & SYNC_TRYLOCK;
1093 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1094 unsigned long first_index = 0;
1098 ag = pag->pag_agno + 1;
1101 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
1106 first_index = pag->pag_ici_reclaim_cursor;
1108 mutex_lock(&pag->pag_ici_reclaim_lock);
1111 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1115 nr_found = radix_tree_gang_lookup_tag(
1117 (void **)batch, first_index,
1119 XFS_ICI_RECLAIM_TAG);
1127 * Grab the inodes before we drop the lock. if we found
1128 * nothing, nr == 0 and the loop will be skipped.
1130 for (i = 0; i < nr_found; i++) {
1131 struct xfs_inode *ip = batch[i];
1133 if (done || xfs_reclaim_inode_grab(ip, flags))
1137 * Update the index for the next lookup. Catch
1138 * overflows into the next AG range which can
1139 * occur if we have inodes in the last block of
1140 * the AG and we are currently pointing to the
1143 * Because we may see inodes that are from the
1144 * wrong AG due to RCU freeing and
1145 * reallocation, only update the index if it
1146 * lies in this AG. It was a race that lead us
1147 * to see this inode, so another lookup from
1148 * the same index will not find it again.
1150 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1153 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1154 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1158 /* unlock now we've grabbed the inodes. */
1161 for (i = 0; i < nr_found; i++) {
1164 error = xfs_reclaim_inode(batch[i], pag, flags);
1165 if (error && last_error != -EFSCORRUPTED)
1169 *nr_to_scan -= XFS_LOOKUP_BATCH;
1173 } while (nr_found && !done && *nr_to_scan > 0);
1175 if (trylock && !done)
1176 pag->pag_ici_reclaim_cursor = first_index;
1178 pag->pag_ici_reclaim_cursor = 0;
1179 mutex_unlock(&pag->pag_ici_reclaim_lock);
1184 * if we skipped any AG, and we still have scan count remaining, do
1185 * another pass this time using blocking reclaim semantics (i.e
1186 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1187 * ensure that when we get more reclaimers than AGs we block rather
1188 * than spin trying to execute reclaim.
1190 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1202 int nr_to_scan = INT_MAX;
1204 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1208 * Scan a certain number of inodes for reclaim.
1210 * When called we make sure that there is a background (fast) inode reclaim in
1211 * progress, while we will throttle the speed of reclaim via doing synchronous
1212 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1213 * them to be cleaned, which we hope will not be very long due to the
1214 * background walker having already kicked the IO off on those dirty inodes.
1217 xfs_reclaim_inodes_nr(
1218 struct xfs_mount *mp,
1221 /* kick background reclaimer and push the AIL */
1222 xfs_reclaim_work_queue(mp);
1223 xfs_ail_push_all(mp->m_ail);
1225 return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1229 * Return the number of reclaimable inodes in the filesystem for
1230 * the shrinker to determine how much to reclaim.
1233 xfs_reclaim_inodes_count(
1234 struct xfs_mount *mp)
1236 struct xfs_perag *pag;
1237 xfs_agnumber_t ag = 0;
1238 int reclaimable = 0;
1240 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1241 ag = pag->pag_agno + 1;
1242 reclaimable += pag->pag_ici_reclaimable;
1250 struct xfs_inode *ip,
1251 struct xfs_eofblocks *eofb)
1253 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1254 !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1257 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1258 !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1261 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1262 xfs_get_projid(ip) != eofb->eof_prid)
1269 * A union-based inode filtering algorithm. Process the inode if any of the
1270 * criteria match. This is for global/internal scans only.
1273 xfs_inode_match_id_union(
1274 struct xfs_inode *ip,
1275 struct xfs_eofblocks *eofb)
1277 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1278 uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1281 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1282 gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1285 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1286 xfs_get_projid(ip) == eofb->eof_prid)
1293 xfs_inode_free_eofblocks(
1294 struct xfs_inode *ip,
1299 struct xfs_eofblocks *eofb = args;
1300 bool need_iolock = true;
1303 ASSERT(!eofb || (eofb && eofb->eof_scan_owner != 0));
1305 if (!xfs_can_free_eofblocks(ip, false)) {
1306 /* inode could be preallocated or append-only */
1307 trace_xfs_inode_free_eofblocks_invalid(ip);
1308 xfs_inode_clear_eofblocks_tag(ip);
1313 * If the mapping is dirty the operation can block and wait for some
1314 * time. Unless we are waiting, skip it.
1316 if (!(flags & SYNC_WAIT) &&
1317 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1321 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1322 match = xfs_inode_match_id_union(ip, eofb);
1324 match = xfs_inode_match_id(ip, eofb);
1328 /* skip the inode if the file size is too small */
1329 if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1330 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1334 * A scan owner implies we already hold the iolock. Skip it in
1335 * xfs_free_eofblocks() to avoid deadlock. This also eliminates
1336 * the possibility of EAGAIN being returned.
1338 if (eofb->eof_scan_owner == ip->i_ino)
1339 need_iolock = false;
1342 ret = xfs_free_eofblocks(ip->i_mount, ip, need_iolock);
1344 /* don't revisit the inode if we're not waiting */
1345 if (ret == -EAGAIN && !(flags & SYNC_WAIT))
1352 xfs_icache_free_eofblocks(
1353 struct xfs_mount *mp,
1354 struct xfs_eofblocks *eofb)
1356 int flags = SYNC_TRYLOCK;
1358 if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
1361 return xfs_inode_ag_iterator_tag(mp, xfs_inode_free_eofblocks, flags,
1362 eofb, XFS_ICI_EOFBLOCKS_TAG);
1366 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1367 * multiple quotas, we don't know exactly which quota caused an allocation
1368 * failure. We make a best effort by including each quota under low free space
1369 * conditions (less than 1% free space) in the scan.
1372 xfs_inode_free_quota_eofblocks(
1373 struct xfs_inode *ip)
1376 struct xfs_eofblocks eofb = {0};
1377 struct xfs_dquot *dq;
1379 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1382 * Set the scan owner to avoid a potential livelock. Otherwise, the scan
1383 * can repeatedly trylock on the inode we're currently processing. We
1384 * run a sync scan to increase effectiveness and use the union filter to
1385 * cover all applicable quotas in a single scan.
1387 eofb.eof_scan_owner = ip->i_ino;
1388 eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC;
1390 if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) {
1391 dq = xfs_inode_dquot(ip, XFS_DQ_USER);
1392 if (dq && xfs_dquot_lowsp(dq)) {
1393 eofb.eof_uid = VFS_I(ip)->i_uid;
1394 eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1399 if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) {
1400 dq = xfs_inode_dquot(ip, XFS_DQ_GROUP);
1401 if (dq && xfs_dquot_lowsp(dq)) {
1402 eofb.eof_gid = VFS_I(ip)->i_gid;
1403 eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1409 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
1415 xfs_inode_set_eofblocks_tag(
1418 struct xfs_mount *mp = ip->i_mount;
1419 struct xfs_perag *pag;
1423 * Don't bother locking the AG and looking up in the radix trees
1424 * if we already know that we have the tag set.
1426 if (ip->i_flags & XFS_IEOFBLOCKS)
1428 spin_lock(&ip->i_flags_lock);
1429 ip->i_flags |= XFS_IEOFBLOCKS;
1430 spin_unlock(&ip->i_flags_lock);
1432 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1433 spin_lock(&pag->pag_ici_lock);
1434 trace_xfs_inode_set_eofblocks_tag(ip);
1436 tagged = radix_tree_tagged(&pag->pag_ici_root,
1437 XFS_ICI_EOFBLOCKS_TAG);
1438 radix_tree_tag_set(&pag->pag_ici_root,
1439 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1440 XFS_ICI_EOFBLOCKS_TAG);
1442 /* propagate the eofblocks tag up into the perag radix tree */
1443 spin_lock(&ip->i_mount->m_perag_lock);
1444 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
1445 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1446 XFS_ICI_EOFBLOCKS_TAG);
1447 spin_unlock(&ip->i_mount->m_perag_lock);
1449 /* kick off background trimming */
1450 xfs_queue_eofblocks(ip->i_mount);
1452 trace_xfs_perag_set_eofblocks(ip->i_mount, pag->pag_agno,
1456 spin_unlock(&pag->pag_ici_lock);
1461 xfs_inode_clear_eofblocks_tag(
1464 struct xfs_mount *mp = ip->i_mount;
1465 struct xfs_perag *pag;
1467 spin_lock(&ip->i_flags_lock);
1468 ip->i_flags &= ~XFS_IEOFBLOCKS;
1469 spin_unlock(&ip->i_flags_lock);
1471 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1472 spin_lock(&pag->pag_ici_lock);
1473 trace_xfs_inode_clear_eofblocks_tag(ip);
1475 radix_tree_tag_clear(&pag->pag_ici_root,
1476 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1477 XFS_ICI_EOFBLOCKS_TAG);
1478 if (!radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_EOFBLOCKS_TAG)) {
1479 /* clear the eofblocks tag from the perag radix tree */
1480 spin_lock(&ip->i_mount->m_perag_lock);
1481 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1482 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1483 XFS_ICI_EOFBLOCKS_TAG);
1484 spin_unlock(&ip->i_mount->m_perag_lock);
1485 trace_xfs_perag_clear_eofblocks(ip->i_mount, pag->pag_agno,
1489 spin_unlock(&pag->pag_ici_lock);