1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
13 #include "xfs_mount.h"
14 #include "xfs_inode.h"
15 #include "xfs_trans.h"
16 #include "xfs_trans_priv.h"
17 #include "xfs_inode_item.h"
18 #include "xfs_quota.h"
19 #include "xfs_trace.h"
20 #include "xfs_icache.h"
21 #include "xfs_bmap_util.h"
22 #include "xfs_dquot_item.h"
23 #include "xfs_dquot.h"
24 #include "xfs_reflink.h"
25 #include "xfs_ialloc.h"
27 #include <linux/iversion.h>
30 * Allocate and initialise an xfs_inode.
40 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL
41 * and return NULL here on ENOMEM.
43 ip = kmem_cache_alloc(xfs_inode_zone, GFP_KERNEL | __GFP_NOFAIL);
45 if (inode_init_always(mp->m_super, VFS_I(ip))) {
46 kmem_cache_free(xfs_inode_zone, ip);
50 /* VFS doesn't initialise i_mode! */
51 VFS_I(ip)->i_mode = 0;
53 XFS_STATS_INC(mp, vn_active);
54 ASSERT(atomic_read(&ip->i_pincount) == 0);
55 ASSERT(ip->i_ino == 0);
57 /* initialise the xfs inode */
60 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
63 memset(&ip->i_df, 0, sizeof(ip->i_df));
65 ip->i_delayed_blks = 0;
66 memset(&ip->i_d, 0, sizeof(ip->i_d));
67 ip->i_d.di_flags2 = mp->m_ino_geo.new_diflags2;
70 INIT_WORK(&ip->i_ioend_work, xfs_end_io);
71 INIT_LIST_HEAD(&ip->i_ioend_list);
72 spin_lock_init(&ip->i_ioend_lock);
78 xfs_inode_free_callback(
79 struct rcu_head *head)
81 struct inode *inode = container_of(head, struct inode, i_rcu);
82 struct xfs_inode *ip = XFS_I(inode);
84 switch (VFS_I(ip)->i_mode & S_IFMT) {
88 xfs_idestroy_fork(&ip->i_df);
93 xfs_idestroy_fork(ip->i_afp);
94 kmem_cache_free(xfs_ifork_zone, ip->i_afp);
97 xfs_idestroy_fork(ip->i_cowfp);
98 kmem_cache_free(xfs_ifork_zone, ip->i_cowfp);
101 ASSERT(!test_bit(XFS_LI_IN_AIL,
102 &ip->i_itemp->ili_item.li_flags));
103 xfs_inode_item_destroy(ip);
107 kmem_cache_free(xfs_inode_zone, ip);
112 struct xfs_inode *ip)
114 /* asserts to verify all state is correct here */
115 ASSERT(atomic_read(&ip->i_pincount) == 0);
116 ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
117 XFS_STATS_DEC(ip->i_mount, vn_active);
119 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
124 struct xfs_inode *ip)
126 ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
129 * Because we use RCU freeing we need to ensure the inode always
130 * appears to be reclaimed with an invalid inode number when in the
131 * free state. The ip->i_flags_lock provides the barrier against lookup
134 spin_lock(&ip->i_flags_lock);
135 ip->i_flags = XFS_IRECLAIM;
137 spin_unlock(&ip->i_flags_lock);
139 __xfs_inode_free(ip);
143 * Queue background inode reclaim work if there are reclaimable inodes and there
144 * isn't reclaim work already scheduled or in progress.
147 xfs_reclaim_work_queue(
148 struct xfs_mount *mp)
152 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
153 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
154 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
160 xfs_perag_set_reclaim_tag(
161 struct xfs_perag *pag)
163 struct xfs_mount *mp = pag->pag_mount;
165 lockdep_assert_held(&pag->pag_ici_lock);
166 if (pag->pag_ici_reclaimable++)
169 /* propagate the reclaim tag up into the perag radix tree */
170 spin_lock(&mp->m_perag_lock);
171 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno,
172 XFS_ICI_RECLAIM_TAG);
173 spin_unlock(&mp->m_perag_lock);
175 /* schedule periodic background inode reclaim */
176 xfs_reclaim_work_queue(mp);
178 trace_xfs_perag_set_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
182 xfs_perag_clear_reclaim_tag(
183 struct xfs_perag *pag)
185 struct xfs_mount *mp = pag->pag_mount;
187 lockdep_assert_held(&pag->pag_ici_lock);
188 if (--pag->pag_ici_reclaimable)
191 /* clear the reclaim tag from the perag radix tree */
192 spin_lock(&mp->m_perag_lock);
193 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno,
194 XFS_ICI_RECLAIM_TAG);
195 spin_unlock(&mp->m_perag_lock);
196 trace_xfs_perag_clear_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
201 * We set the inode flag atomically with the radix tree tag.
202 * Once we get tag lookups on the radix tree, this inode flag
206 xfs_inode_set_reclaim_tag(
207 struct xfs_inode *ip)
209 struct xfs_mount *mp = ip->i_mount;
210 struct xfs_perag *pag;
212 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
213 spin_lock(&pag->pag_ici_lock);
214 spin_lock(&ip->i_flags_lock);
216 radix_tree_tag_set(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, ip->i_ino),
217 XFS_ICI_RECLAIM_TAG);
218 xfs_perag_set_reclaim_tag(pag);
219 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
221 spin_unlock(&ip->i_flags_lock);
222 spin_unlock(&pag->pag_ici_lock);
227 xfs_inode_clear_reclaim_tag(
228 struct xfs_perag *pag,
231 radix_tree_tag_clear(&pag->pag_ici_root,
232 XFS_INO_TO_AGINO(pag->pag_mount, ino),
233 XFS_ICI_RECLAIM_TAG);
234 xfs_perag_clear_reclaim_tag(pag);
239 struct xfs_inode *ip)
241 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_INEW_BIT);
242 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_INEW_BIT);
245 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
246 if (!xfs_iflags_test(ip, XFS_INEW))
250 finish_wait(wq, &wait.wq_entry);
254 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
255 * part of the structure. This is made more complex by the fact we store
256 * information about the on-disk values in the VFS inode and so we can't just
257 * overwrite the values unconditionally. Hence we save the parameters we
258 * need to retain across reinitialisation, and rewrite them into the VFS inode
259 * after reinitialisation even if it fails.
263 struct xfs_mount *mp,
267 uint32_t nlink = inode->i_nlink;
268 uint32_t generation = inode->i_generation;
269 uint64_t version = inode_peek_iversion(inode);
270 umode_t mode = inode->i_mode;
271 dev_t dev = inode->i_rdev;
272 kuid_t uid = inode->i_uid;
273 kgid_t gid = inode->i_gid;
275 error = inode_init_always(mp->m_super, inode);
277 set_nlink(inode, nlink);
278 inode->i_generation = generation;
279 inode_set_iversion_queried(inode, version);
280 inode->i_mode = mode;
288 * If we are allocating a new inode, then check what was returned is
289 * actually a free, empty inode. If we are not allocating an inode,
290 * then check we didn't find a free inode.
293 * 0 if the inode free state matches the lookup context
294 * -ENOENT if the inode is free and we are not allocating
295 * -EFSCORRUPTED if there is any state mismatch at all
298 xfs_iget_check_free_state(
299 struct xfs_inode *ip,
302 if (flags & XFS_IGET_CREATE) {
303 /* should be a free inode */
304 if (VFS_I(ip)->i_mode != 0) {
305 xfs_warn(ip->i_mount,
306 "Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
307 ip->i_ino, VFS_I(ip)->i_mode);
308 return -EFSCORRUPTED;
311 if (ip->i_d.di_nblocks != 0) {
312 xfs_warn(ip->i_mount,
313 "Corruption detected! Free inode 0x%llx has blocks allocated!",
315 return -EFSCORRUPTED;
320 /* should be an allocated inode */
321 if (VFS_I(ip)->i_mode == 0)
328 * Check the validity of the inode we just found it the cache
332 struct xfs_perag *pag,
333 struct xfs_inode *ip,
336 int lock_flags) __releases(RCU)
338 struct inode *inode = VFS_I(ip);
339 struct xfs_mount *mp = ip->i_mount;
343 * check for re-use of an inode within an RCU grace period due to the
344 * radix tree nodes not being updated yet. We monitor for this by
345 * setting the inode number to zero before freeing the inode structure.
346 * If the inode has been reallocated and set up, then the inode number
347 * will not match, so check for that, too.
349 spin_lock(&ip->i_flags_lock);
350 if (ip->i_ino != ino) {
351 trace_xfs_iget_skip(ip);
352 XFS_STATS_INC(mp, xs_ig_frecycle);
359 * If we are racing with another cache hit that is currently
360 * instantiating this inode or currently recycling it out of
361 * reclaimabe state, wait for the initialisation to complete
364 * XXX(hch): eventually we should do something equivalent to
365 * wait_on_inode to wait for these flags to be cleared
366 * instead of polling for it.
368 if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
369 trace_xfs_iget_skip(ip);
370 XFS_STATS_INC(mp, xs_ig_frecycle);
376 * Check the inode free state is valid. This also detects lookup
377 * racing with unlinks.
379 error = xfs_iget_check_free_state(ip, flags);
384 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
385 * Need to carefully get it back into useable state.
387 if (ip->i_flags & XFS_IRECLAIMABLE) {
388 trace_xfs_iget_reclaim(ip);
390 if (flags & XFS_IGET_INCORE) {
396 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
397 * from stomping over us while we recycle the inode. We can't
398 * clear the radix tree reclaimable tag yet as it requires
399 * pag_ici_lock to be held exclusive.
401 ip->i_flags |= XFS_IRECLAIM;
403 spin_unlock(&ip->i_flags_lock);
406 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
407 error = xfs_reinit_inode(mp, inode);
411 * Re-initializing the inode failed, and we are in deep
412 * trouble. Try to re-add it to the reclaim list.
415 spin_lock(&ip->i_flags_lock);
416 wake = !!__xfs_iflags_test(ip, XFS_INEW);
417 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
419 wake_up_bit(&ip->i_flags, __XFS_INEW_BIT);
420 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
421 trace_xfs_iget_reclaim_fail(ip);
425 spin_lock(&pag->pag_ici_lock);
426 spin_lock(&ip->i_flags_lock);
429 * Clear the per-lifetime state in the inode as we are now
430 * effectively a new inode and need to return to the initial
431 * state before reuse occurs.
433 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
434 ip->i_flags |= XFS_INEW;
435 xfs_inode_clear_reclaim_tag(pag, ip->i_ino);
436 inode->i_state = I_NEW;
440 spin_unlock(&ip->i_flags_lock);
441 spin_unlock(&pag->pag_ici_lock);
443 /* If the VFS inode is being torn down, pause and try again. */
445 trace_xfs_iget_skip(ip);
450 /* We've got a live one. */
451 spin_unlock(&ip->i_flags_lock);
453 trace_xfs_iget_hit(ip);
457 xfs_ilock(ip, lock_flags);
459 if (!(flags & XFS_IGET_INCORE))
460 xfs_iflags_clear(ip, XFS_ISTALE);
461 XFS_STATS_INC(mp, xs_ig_found);
466 spin_unlock(&ip->i_flags_lock);
474 struct xfs_mount *mp,
475 struct xfs_perag *pag,
478 struct xfs_inode **ipp,
482 struct xfs_inode *ip;
484 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
487 ip = xfs_inode_alloc(mp, ino);
491 error = xfs_imap(mp, tp, ip->i_ino, &ip->i_imap, flags);
496 * For version 5 superblocks, if we are initialising a new inode and we
497 * are not utilising the XFS_MOUNT_IKEEP inode cluster mode, we can
498 * simply build the new inode core with a random generation number.
500 * For version 4 (and older) superblocks, log recovery is dependent on
501 * the di_flushiter field being initialised from the current on-disk
502 * value and hence we must also read the inode off disk even when
503 * initializing new inodes.
505 if (xfs_sb_version_has_v3inode(&mp->m_sb) &&
506 (flags & XFS_IGET_CREATE) && !(mp->m_flags & XFS_MOUNT_IKEEP)) {
507 VFS_I(ip)->i_generation = prandom_u32();
511 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
515 error = xfs_inode_from_disk(ip,
516 xfs_buf_offset(bp, ip->i_imap.im_boffset));
518 xfs_buf_set_ref(bp, XFS_INO_REF);
519 xfs_trans_brelse(tp, bp);
525 trace_xfs_iget_miss(ip);
528 * Check the inode free state is valid. This also detects lookup
529 * racing with unlinks.
531 error = xfs_iget_check_free_state(ip, flags);
536 * Preload the radix tree so we can insert safely under the
537 * write spinlock. Note that we cannot sleep inside the preload
538 * region. Since we can be called from transaction context, don't
539 * recurse into the file system.
541 if (radix_tree_preload(GFP_NOFS)) {
547 * Because the inode hasn't been added to the radix-tree yet it can't
548 * be found by another thread, so we can do the non-sleeping lock here.
551 if (!xfs_ilock_nowait(ip, lock_flags))
556 * These values must be set before inserting the inode into the radix
557 * tree as the moment it is inserted a concurrent lookup (allowed by the
558 * RCU locking mechanism) can find it and that lookup must see that this
559 * is an inode currently under construction (i.e. that XFS_INEW is set).
560 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
561 * memory barrier that ensures this detection works correctly at lookup
565 if (flags & XFS_IGET_DONTCACHE)
566 d_mark_dontcache(VFS_I(ip));
570 xfs_iflags_set(ip, iflags);
572 /* insert the new inode */
573 spin_lock(&pag->pag_ici_lock);
574 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
575 if (unlikely(error)) {
576 WARN_ON(error != -EEXIST);
577 XFS_STATS_INC(mp, xs_ig_dup);
579 goto out_preload_end;
581 spin_unlock(&pag->pag_ici_lock);
582 radix_tree_preload_end();
588 spin_unlock(&pag->pag_ici_lock);
589 radix_tree_preload_end();
591 xfs_iunlock(ip, lock_flags);
593 __destroy_inode(VFS_I(ip));
599 * Look up an inode by number in the given file system. The inode is looked up
600 * in the cache held in each AG. If the inode is found in the cache, initialise
601 * the vfs inode if necessary.
603 * If it is not in core, read it in from the file system's device, add it to the
604 * cache and initialise the vfs inode.
606 * The inode is locked according to the value of the lock_flags parameter.
607 * Inode lookup is only done during metadata operations and not as part of the
608 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
612 struct xfs_mount *mp,
613 struct xfs_trans *tp,
617 struct xfs_inode **ipp)
619 struct xfs_inode *ip;
620 struct xfs_perag *pag;
624 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
626 /* reject inode numbers outside existing AGs */
627 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
630 XFS_STATS_INC(mp, xs_ig_attempts);
632 /* get the perag structure and ensure that it's inode capable */
633 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
634 agino = XFS_INO_TO_AGINO(mp, ino);
639 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
642 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
644 goto out_error_or_again;
647 if (flags & XFS_IGET_INCORE) {
649 goto out_error_or_again;
651 XFS_STATS_INC(mp, xs_ig_missed);
653 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
656 goto out_error_or_again;
663 * If we have a real type for an on-disk inode, we can setup the inode
664 * now. If it's a new inode being created, xfs_ialloc will handle it.
666 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
667 xfs_setup_existing_inode(ip);
671 if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
680 * "Is this a cached inode that's also allocated?"
682 * Look up an inode by number in the given file system. If the inode is
683 * in cache and isn't in purgatory, return 1 if the inode is allocated
684 * and 0 if it is not. For all other cases (not in cache, being torn
685 * down, etc.), return a negative error code.
687 * The caller has to prevent inode allocation and freeing activity,
688 * presumably by locking the AGI buffer. This is to ensure that an
689 * inode cannot transition from allocated to freed until the caller is
690 * ready to allow that. If the inode is in an intermediate state (new,
691 * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
692 * inode is not in the cache, -ENOENT will be returned. The caller must
693 * deal with these scenarios appropriately.
695 * This is a specialized use case for the online scrubber; if you're
696 * reading this, you probably want xfs_iget.
699 xfs_icache_inode_is_allocated(
700 struct xfs_mount *mp,
701 struct xfs_trans *tp,
705 struct xfs_inode *ip;
708 error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
712 *inuse = !!(VFS_I(ip)->i_mode);
718 * The inode lookup is done in batches to keep the amount of lock traffic and
719 * radix tree lookups to a minimum. The batch size is a trade off between
720 * lookup reduction and stack usage. This is in the reclaim path, so we can't
723 #define XFS_LOOKUP_BATCH 32
726 * Decide if the given @ip is eligible to be a part of the inode walk, and
727 * grab it if so. Returns true if it's ready to go or false if we should just
731 xfs_inode_walk_ag_grab(
732 struct xfs_inode *ip,
735 struct inode *inode = VFS_I(ip);
736 bool newinos = !!(flags & XFS_INODE_WALK_INEW_WAIT);
738 ASSERT(rcu_read_lock_held());
740 /* Check for stale RCU freed inode */
741 spin_lock(&ip->i_flags_lock);
743 goto out_unlock_noent;
745 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
746 if ((!newinos && __xfs_iflags_test(ip, XFS_INEW)) ||
747 __xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM))
748 goto out_unlock_noent;
749 spin_unlock(&ip->i_flags_lock);
751 /* nothing to sync during shutdown */
752 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
755 /* If we can't grab the inode, it must on it's way to reclaim. */
763 spin_unlock(&ip->i_flags_lock);
768 * For a given per-AG structure @pag, grab, @execute, and rele all incore
769 * inodes with the given radix tree @tag.
773 struct xfs_perag *pag,
775 int (*execute)(struct xfs_inode *ip, void *args),
779 struct xfs_mount *mp = pag->pag_mount;
780 uint32_t first_index;
792 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
798 if (tag == XFS_ICI_NO_TAG)
799 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
800 (void **)batch, first_index,
803 nr_found = radix_tree_gang_lookup_tag(
805 (void **) batch, first_index,
806 XFS_LOOKUP_BATCH, tag);
814 * Grab the inodes before we drop the lock. if we found
815 * nothing, nr == 0 and the loop will be skipped.
817 for (i = 0; i < nr_found; i++) {
818 struct xfs_inode *ip = batch[i];
820 if (done || !xfs_inode_walk_ag_grab(ip, iter_flags))
824 * Update the index for the next lookup. Catch
825 * overflows into the next AG range which can occur if
826 * we have inodes in the last block of the AG and we
827 * are currently pointing to the last inode.
829 * Because we may see inodes that are from the wrong AG
830 * due to RCU freeing and reallocation, only update the
831 * index if it lies in this AG. It was a race that lead
832 * us to see this inode, so another lookup from the
833 * same index will not find it again.
835 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
837 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
838 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
842 /* unlock now we've grabbed the inodes. */
845 for (i = 0; i < nr_found; i++) {
848 if ((iter_flags & XFS_INODE_WALK_INEW_WAIT) &&
849 xfs_iflags_test(batch[i], XFS_INEW))
850 xfs_inew_wait(batch[i]);
851 error = execute(batch[i], args);
853 if (error == -EAGAIN) {
857 if (error && last_error != -EFSCORRUPTED)
861 /* bail out if the filesystem is corrupted. */
862 if (error == -EFSCORRUPTED)
867 } while (nr_found && !done);
876 /* Fetch the next (possibly tagged) per-AG structure. */
877 static inline struct xfs_perag *
878 xfs_inode_walk_get_perag(
879 struct xfs_mount *mp,
883 if (tag == XFS_ICI_NO_TAG)
884 return xfs_perag_get(mp, agno);
885 return xfs_perag_get_tag(mp, agno, tag);
889 * Call the @execute function on all incore inodes matching the radix tree
894 struct xfs_mount *mp,
896 int (*execute)(struct xfs_inode *ip, void *args),
900 struct xfs_perag *pag;
906 while ((pag = xfs_inode_walk_get_perag(mp, ag, tag))) {
907 ag = pag->pag_agno + 1;
908 error = xfs_inode_walk_ag(pag, iter_flags, execute, args, tag);
912 if (error == -EFSCORRUPTED)
920 * Grab the inode for reclaim exclusively.
922 * We have found this inode via a lookup under RCU, so the inode may have
923 * already been freed, or it may be in the process of being recycled by
924 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
925 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
926 * will not be set. Hence we need to check for both these flag conditions to
927 * avoid inodes that are no longer reclaim candidates.
929 * Note: checking for other state flags here, under the i_flags_lock or not, is
930 * racy and should be avoided. Those races should be resolved only after we have
931 * ensured that we are able to reclaim this inode and the world can see that we
932 * are going to reclaim it.
934 * Return true if we grabbed it, false otherwise.
937 xfs_reclaim_inode_grab(
938 struct xfs_inode *ip)
940 ASSERT(rcu_read_lock_held());
942 spin_lock(&ip->i_flags_lock);
943 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
944 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
945 /* not a reclaim candidate. */
946 spin_unlock(&ip->i_flags_lock);
949 __xfs_iflags_set(ip, XFS_IRECLAIM);
950 spin_unlock(&ip->i_flags_lock);
955 * Inode reclaim is non-blocking, so the default action if progress cannot be
956 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
957 * XFS_IRECLAIM flag. If we are in a shutdown state, we don't care about
958 * blocking anymore and hence we can wait for the inode to be able to reclaim
961 * We do no IO here - if callers require inodes to be cleaned they must push the
962 * AIL first to trigger writeback of dirty inodes. This enables writeback to be
963 * done in the background in a non-blocking manner, and enables memory reclaim
964 * to make progress without blocking.
968 struct xfs_inode *ip,
969 struct xfs_perag *pag)
971 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
973 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
975 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
978 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
980 xfs_iflush_abort(ip);
983 if (xfs_ipincount(ip))
984 goto out_clear_flush;
985 if (!xfs_inode_clean(ip))
986 goto out_clear_flush;
988 xfs_iflags_clear(ip, XFS_IFLUSHING);
992 * Because we use RCU freeing we need to ensure the inode always appears
993 * to be reclaimed with an invalid inode number when in the free state.
994 * We do this as early as possible under the ILOCK so that
995 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
996 * detect races with us here. By doing this, we guarantee that once
997 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
998 * it will see either a valid inode that will serialise correctly, or it
999 * will see an invalid inode that it can skip.
1001 spin_lock(&ip->i_flags_lock);
1002 ip->i_flags = XFS_IRECLAIM;
1004 spin_unlock(&ip->i_flags_lock);
1006 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1008 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1010 * Remove the inode from the per-AG radix tree.
1012 * Because radix_tree_delete won't complain even if the item was never
1013 * added to the tree assert that it's been there before to catch
1014 * problems with the inode life time early on.
1016 spin_lock(&pag->pag_ici_lock);
1017 if (!radix_tree_delete(&pag->pag_ici_root,
1018 XFS_INO_TO_AGINO(ip->i_mount, ino)))
1020 xfs_perag_clear_reclaim_tag(pag);
1021 spin_unlock(&pag->pag_ici_lock);
1024 * Here we do an (almost) spurious inode lock in order to coordinate
1025 * with inode cache radix tree lookups. This is because the lookup
1026 * can reference the inodes in the cache without taking references.
1028 * We make that OK here by ensuring that we wait until the inode is
1029 * unlocked after the lookup before we go ahead and free it.
1031 xfs_ilock(ip, XFS_ILOCK_EXCL);
1032 xfs_qm_dqdetach(ip);
1033 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1034 ASSERT(xfs_inode_clean(ip));
1036 __xfs_inode_free(ip);
1040 xfs_iflags_clear(ip, XFS_IFLUSHING);
1042 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1044 xfs_iflags_clear(ip, XFS_IRECLAIM);
1048 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1049 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1050 * then a shut down during filesystem unmount reclaim walk leak all the
1051 * unreclaimed inodes.
1053 * Returns non-zero if any AGs or inodes were skipped in the reclaim pass
1054 * so that callers that want to block until all dirty inodes are written back
1055 * and reclaimed can sanely loop.
1058 xfs_reclaim_inodes_ag(
1059 struct xfs_mount *mp,
1062 struct xfs_perag *pag;
1063 xfs_agnumber_t ag = 0;
1065 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1066 unsigned long first_index = 0;
1070 ag = pag->pag_agno + 1;
1072 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1074 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1078 nr_found = radix_tree_gang_lookup_tag(
1080 (void **)batch, first_index,
1082 XFS_ICI_RECLAIM_TAG);
1090 * Grab the inodes before we drop the lock. if we found
1091 * nothing, nr == 0 and the loop will be skipped.
1093 for (i = 0; i < nr_found; i++) {
1094 struct xfs_inode *ip = batch[i];
1096 if (done || !xfs_reclaim_inode_grab(ip))
1100 * Update the index for the next lookup. Catch
1101 * overflows into the next AG range which can
1102 * occur if we have inodes in the last block of
1103 * the AG and we are currently pointing to the
1106 * Because we may see inodes that are from the
1107 * wrong AG due to RCU freeing and
1108 * reallocation, only update the index if it
1109 * lies in this AG. It was a race that lead us
1110 * to see this inode, so another lookup from
1111 * the same index will not find it again.
1113 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1116 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1117 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1121 /* unlock now we've grabbed the inodes. */
1124 for (i = 0; i < nr_found; i++) {
1126 xfs_reclaim_inode(batch[i], pag);
1129 *nr_to_scan -= XFS_LOOKUP_BATCH;
1131 } while (nr_found && !done && *nr_to_scan > 0);
1135 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1142 struct xfs_mount *mp)
1144 int nr_to_scan = INT_MAX;
1146 while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
1147 xfs_ail_push_all_sync(mp->m_ail);
1148 xfs_reclaim_inodes_ag(mp, &nr_to_scan);
1153 * The shrinker infrastructure determines how many inodes we should scan for
1154 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
1155 * push the AIL here. We also want to proactively free up memory if we can to
1156 * minimise the amount of work memory reclaim has to do so we kick the
1157 * background reclaim if it isn't already scheduled.
1160 xfs_reclaim_inodes_nr(
1161 struct xfs_mount *mp,
1164 /* kick background reclaimer and push the AIL */
1165 xfs_reclaim_work_queue(mp);
1166 xfs_ail_push_all(mp->m_ail);
1168 xfs_reclaim_inodes_ag(mp, &nr_to_scan);
1173 * Return the number of reclaimable inodes in the filesystem for
1174 * the shrinker to determine how much to reclaim.
1177 xfs_reclaim_inodes_count(
1178 struct xfs_mount *mp)
1180 struct xfs_perag *pag;
1181 xfs_agnumber_t ag = 0;
1182 int reclaimable = 0;
1184 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1185 ag = pag->pag_agno + 1;
1186 reclaimable += pag->pag_ici_reclaimable;
1194 struct xfs_inode *ip,
1195 struct xfs_eofblocks *eofb)
1197 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1198 !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1201 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1202 !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1205 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1206 ip->i_d.di_projid != eofb->eof_prid)
1213 * A union-based inode filtering algorithm. Process the inode if any of the
1214 * criteria match. This is for global/internal scans only.
1217 xfs_inode_match_id_union(
1218 struct xfs_inode *ip,
1219 struct xfs_eofblocks *eofb)
1221 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1222 uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1225 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1226 gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1229 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1230 ip->i_d.di_projid == eofb->eof_prid)
1237 * Is this inode @ip eligible for eof/cow block reclamation, given some
1238 * filtering parameters @eofb? The inode is eligible if @eofb is null or
1239 * if the predicate functions match.
1242 xfs_inode_matches_eofb(
1243 struct xfs_inode *ip,
1244 struct xfs_eofblocks *eofb)
1251 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1252 match = xfs_inode_match_id_union(ip, eofb);
1254 match = xfs_inode_match_id(ip, eofb);
1258 /* skip the inode if the file size is too small */
1259 if ((eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE) &&
1260 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1267 * This is a fast pass over the inode cache to try to get reclaim moving on as
1268 * many inodes as possible in a short period of time. It kicks itself every few
1269 * seconds, as well as being kicked by the inode cache shrinker when memory
1274 struct work_struct *work)
1276 struct xfs_mount *mp = container_of(to_delayed_work(work),
1277 struct xfs_mount, m_reclaim_work);
1278 int nr_to_scan = INT_MAX;
1280 xfs_reclaim_inodes_ag(mp, &nr_to_scan);
1281 xfs_reclaim_work_queue(mp);
1285 xfs_inode_free_eofblocks(
1286 struct xfs_inode *ip,
1288 unsigned int *lockflags)
1290 struct xfs_eofblocks *eofb = args;
1293 wait = eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC);
1295 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1298 if (!xfs_can_free_eofblocks(ip, false)) {
1299 /* inode could be preallocated or append-only */
1300 trace_xfs_inode_free_eofblocks_invalid(ip);
1301 xfs_inode_clear_eofblocks_tag(ip);
1306 * If the mapping is dirty the operation can block and wait for some
1307 * time. Unless we are waiting, skip it.
1309 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1312 if (!xfs_inode_matches_eofb(ip, eofb))
1316 * If the caller is waiting, return -EAGAIN to keep the background
1317 * scanner moving and revisit the inode in a subsequent pass.
1319 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1324 *lockflags |= XFS_IOLOCK_EXCL;
1326 return xfs_free_eofblocks(ip);
1330 * Background scanning to trim preallocated space. This is queued based on the
1331 * 'speculative_prealloc_lifetime' tunable (5m by default).
1335 struct xfs_perag *pag)
1338 if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
1339 queue_delayed_work(pag->pag_mount->m_gc_workqueue,
1340 &pag->pag_blockgc_work,
1341 msecs_to_jiffies(xfs_blockgc_secs * 1000));
1346 xfs_blockgc_set_iflag(
1347 struct xfs_inode *ip,
1348 unsigned long iflag)
1350 struct xfs_mount *mp = ip->i_mount;
1351 struct xfs_perag *pag;
1354 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1357 * Don't bother locking the AG and looking up in the radix trees
1358 * if we already know that we have the tag set.
1360 if (ip->i_flags & iflag)
1362 spin_lock(&ip->i_flags_lock);
1363 ip->i_flags |= iflag;
1364 spin_unlock(&ip->i_flags_lock);
1366 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1367 spin_lock(&pag->pag_ici_lock);
1369 tagged = radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG);
1370 radix_tree_tag_set(&pag->pag_ici_root,
1371 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1372 XFS_ICI_BLOCKGC_TAG);
1374 /* propagate the blockgc 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_BLOCKGC_TAG);
1379 spin_unlock(&ip->i_mount->m_perag_lock);
1381 /* kick off background trimming */
1382 xfs_blockgc_queue(pag);
1384 trace_xfs_perag_set_blockgc(ip->i_mount, pag->pag_agno, -1,
1388 spin_unlock(&pag->pag_ici_lock);
1393 xfs_inode_set_eofblocks_tag(
1396 trace_xfs_inode_set_eofblocks_tag(ip);
1397 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1401 xfs_blockgc_clear_iflag(
1402 struct xfs_inode *ip,
1403 unsigned long iflag)
1405 struct xfs_mount *mp = ip->i_mount;
1406 struct xfs_perag *pag;
1409 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1411 spin_lock(&ip->i_flags_lock);
1412 ip->i_flags &= ~iflag;
1413 clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1414 spin_unlock(&ip->i_flags_lock);
1419 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1420 spin_lock(&pag->pag_ici_lock);
1422 radix_tree_tag_clear(&pag->pag_ici_root,
1423 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1424 XFS_ICI_BLOCKGC_TAG);
1425 if (!radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG)) {
1426 /* clear the blockgc tag from the perag radix tree */
1427 spin_lock(&ip->i_mount->m_perag_lock);
1428 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1429 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1430 XFS_ICI_BLOCKGC_TAG);
1431 spin_unlock(&ip->i_mount->m_perag_lock);
1432 trace_xfs_perag_clear_blockgc(ip->i_mount, pag->pag_agno, -1,
1436 spin_unlock(&pag->pag_ici_lock);
1441 xfs_inode_clear_eofblocks_tag(
1444 trace_xfs_inode_clear_eofblocks_tag(ip);
1445 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1449 * Set ourselves up to free CoW blocks from this file. If it's already clean
1450 * then we can bail out quickly, but otherwise we must back off if the file
1451 * is undergoing some kind of write.
1454 xfs_prep_free_cowblocks(
1455 struct xfs_inode *ip)
1458 * Just clear the tag if we have an empty cow fork or none at all. It's
1459 * possible the inode was fully unshared since it was originally tagged.
1461 if (!xfs_inode_has_cow_data(ip)) {
1462 trace_xfs_inode_free_cowblocks_invalid(ip);
1463 xfs_inode_clear_cowblocks_tag(ip);
1468 * If the mapping is dirty or under writeback we cannot touch the
1469 * CoW fork. Leave it alone if we're in the midst of a directio.
1471 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1472 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1473 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1474 atomic_read(&VFS_I(ip)->i_dio_count))
1481 * Automatic CoW Reservation Freeing
1483 * These functions automatically garbage collect leftover CoW reservations
1484 * that were made on behalf of a cowextsize hint when we start to run out
1485 * of quota or when the reservations sit around for too long. If the file
1486 * has dirty pages or is undergoing writeback, its CoW reservations will
1489 * The actual garbage collection piggybacks off the same code that runs
1490 * the speculative EOF preallocation garbage collector.
1493 xfs_inode_free_cowblocks(
1494 struct xfs_inode *ip,
1496 unsigned int *lockflags)
1498 struct xfs_eofblocks *eofb = args;
1502 wait = eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC);
1504 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1507 if (!xfs_prep_free_cowblocks(ip))
1510 if (!xfs_inode_matches_eofb(ip, eofb))
1514 * If the caller is waiting, return -EAGAIN to keep the background
1515 * scanner moving and revisit the inode in a subsequent pass.
1517 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1518 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1523 *lockflags |= XFS_IOLOCK_EXCL;
1525 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1530 *lockflags |= XFS_MMAPLOCK_EXCL;
1533 * Check again, nobody else should be able to dirty blocks or change
1534 * the reflink iflag now that we have the first two locks held.
1536 if (xfs_prep_free_cowblocks(ip))
1537 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1542 xfs_inode_set_cowblocks_tag(
1545 trace_xfs_inode_set_cowblocks_tag(ip);
1546 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1550 xfs_inode_clear_cowblocks_tag(
1553 trace_xfs_inode_clear_cowblocks_tag(ip);
1554 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1557 #define for_each_perag_tag(mp, next_agno, pag, tag) \
1558 for ((next_agno) = 0, (pag) = xfs_perag_get_tag((mp), 0, (tag)); \
1560 (next_agno) = (pag)->pag_agno + 1, \
1561 xfs_perag_put(pag), \
1562 (pag) = xfs_perag_get_tag((mp), (next_agno), (tag)))
1565 /* Disable post-EOF and CoW block auto-reclamation. */
1568 struct xfs_mount *mp)
1570 struct xfs_perag *pag;
1571 xfs_agnumber_t agno;
1573 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1574 cancel_delayed_work_sync(&pag->pag_blockgc_work);
1577 /* Enable post-EOF and CoW block auto-reclamation. */
1580 struct xfs_mount *mp)
1582 struct xfs_perag *pag;
1583 xfs_agnumber_t agno;
1585 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1586 xfs_blockgc_queue(pag);
1589 /* Scan one incore inode for block preallocations that we can remove. */
1591 xfs_blockgc_scan_inode(
1592 struct xfs_inode *ip,
1595 unsigned int lockflags = 0;
1598 error = xfs_inode_free_eofblocks(ip, args, &lockflags);
1602 error = xfs_inode_free_cowblocks(ip, args, &lockflags);
1605 xfs_iunlock(ip, lockflags);
1609 /* Background worker that trims preallocated space. */
1612 struct work_struct *work)
1614 struct xfs_perag *pag = container_of(to_delayed_work(work),
1615 struct xfs_perag, pag_blockgc_work);
1616 struct xfs_mount *mp = pag->pag_mount;
1619 if (!sb_start_write_trylock(mp->m_super))
1621 error = xfs_inode_walk_ag(pag, 0, xfs_blockgc_scan_inode, NULL,
1622 XFS_ICI_BLOCKGC_TAG);
1624 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1625 pag->pag_agno, error);
1626 sb_end_write(mp->m_super);
1627 xfs_blockgc_queue(pag);
1631 * Try to free space in the filesystem by purging eofblocks and cowblocks.
1634 xfs_blockgc_free_space(
1635 struct xfs_mount *mp,
1636 struct xfs_eofblocks *eofb)
1638 trace_xfs_blockgc_free_space(mp, eofb, _RET_IP_);
1640 return xfs_inode_walk(mp, 0, xfs_blockgc_scan_inode, eofb,
1641 XFS_ICI_BLOCKGC_TAG);
1645 * Run cow/eofblocks scans on the supplied dquots. We don't know exactly which
1646 * quota caused an allocation failure, so we make a best effort by including
1647 * each quota under low free space conditions (less than 1% free space) in the
1650 * Callers must not hold any inode's ILOCK. If requesting a synchronous scan
1651 * (XFS_EOF_FLAGS_SYNC), the caller also must not hold any inode's IOLOCK or
1655 xfs_blockgc_free_dquots(
1656 struct xfs_mount *mp,
1657 struct xfs_dquot *udqp,
1658 struct xfs_dquot *gdqp,
1659 struct xfs_dquot *pdqp,
1660 unsigned int eof_flags)
1662 struct xfs_eofblocks eofb = {0};
1663 bool do_work = false;
1665 if (!udqp && !gdqp && !pdqp)
1669 * Run a scan to free blocks using the union filter to cover all
1670 * applicable quotas in a single scan.
1672 eofb.eof_flags = XFS_EOF_FLAGS_UNION | eof_flags;
1674 if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1675 eofb.eof_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1676 eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1680 if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1681 eofb.eof_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1682 eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1686 if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1687 eofb.eof_prid = pdqp->q_id;
1688 eofb.eof_flags |= XFS_EOF_FLAGS_PRID;
1695 return xfs_blockgc_free_space(mp, &eofb);
1698 /* Run cow/eofblocks scans on the quotas attached to the inode. */
1700 xfs_blockgc_free_quota(
1701 struct xfs_inode *ip,
1702 unsigned int eof_flags)
1704 return xfs_blockgc_free_dquots(ip->i_mount,
1705 xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1706 xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1707 xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), eof_flags);