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"
12 #include "xfs_mount.h"
13 #include "xfs_inode.h"
14 #include "xfs_trans.h"
15 #include "xfs_trans_priv.h"
16 #include "xfs_inode_item.h"
17 #include "xfs_quota.h"
18 #include "xfs_trace.h"
19 #include "xfs_icache.h"
20 #include "xfs_bmap_util.h"
21 #include "xfs_dquot_item.h"
22 #include "xfs_dquot.h"
23 #include "xfs_reflink.h"
24 #include "xfs_ialloc.h"
26 #include "xfs_log_priv.h"
28 #include <linux/iversion.h>
30 /* Radix tree tags for incore inode tree. */
32 /* inode is to be reclaimed */
33 #define XFS_ICI_RECLAIM_TAG 0
34 /* Inode has speculative preallocations (posteof or cow) to clean. */
35 #define XFS_ICI_BLOCKGC_TAG 1
38 * The goal for walking incore inodes. These can correspond with incore inode
39 * radix tree tags when convenient. Avoid existing XFS_IWALK namespace.
41 enum xfs_icwalk_goal {
42 /* Goals directly associated with tagged inodes. */
43 XFS_ICWALK_BLOCKGC = XFS_ICI_BLOCKGC_TAG,
44 XFS_ICWALK_RECLAIM = XFS_ICI_RECLAIM_TAG,
47 static int xfs_icwalk(struct xfs_mount *mp,
48 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
49 static int xfs_icwalk_ag(struct xfs_perag *pag,
50 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
53 * Private inode cache walk flags for struct xfs_icwalk. Must not
54 * coincide with XFS_ICWALK_FLAGS_VALID.
57 /* Stop scanning after icw_scan_limit inodes. */
58 #define XFS_ICWALK_FLAG_SCAN_LIMIT (1U << 28)
60 #define XFS_ICWALK_FLAG_RECLAIM_SICK (1U << 27)
61 #define XFS_ICWALK_FLAG_UNION (1U << 26) /* union filter algorithm */
63 #define XFS_ICWALK_PRIVATE_FLAGS (XFS_ICWALK_FLAG_SCAN_LIMIT | \
64 XFS_ICWALK_FLAG_RECLAIM_SICK | \
65 XFS_ICWALK_FLAG_UNION)
68 * Allocate and initialise an xfs_inode.
78 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL
79 * and return NULL here on ENOMEM.
81 ip = alloc_inode_sb(mp->m_super, xfs_inode_cache, GFP_KERNEL | __GFP_NOFAIL);
83 if (inode_init_always(mp->m_super, VFS_I(ip))) {
84 kmem_cache_free(xfs_inode_cache, ip);
88 /* VFS doesn't initialise i_mode or i_state! */
89 VFS_I(ip)->i_mode = 0;
90 VFS_I(ip)->i_state = 0;
91 mapping_set_large_folios(VFS_I(ip)->i_mapping);
93 XFS_STATS_INC(mp, vn_active);
94 ASSERT(atomic_read(&ip->i_pincount) == 0);
95 ASSERT(ip->i_ino == 0);
97 /* initialise the xfs inode */
100 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
102 memset(&ip->i_af, 0, sizeof(ip->i_af));
103 ip->i_af.if_format = XFS_DINODE_FMT_EXTENTS;
104 memset(&ip->i_df, 0, sizeof(ip->i_df));
106 ip->i_delayed_blks = 0;
107 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
112 INIT_WORK(&ip->i_ioend_work, xfs_end_io);
113 INIT_LIST_HEAD(&ip->i_ioend_list);
114 spin_lock_init(&ip->i_ioend_lock);
115 ip->i_next_unlinked = NULLAGINO;
116 ip->i_prev_unlinked = 0;
122 xfs_inode_free_callback(
123 struct rcu_head *head)
125 struct inode *inode = container_of(head, struct inode, i_rcu);
126 struct xfs_inode *ip = XFS_I(inode);
128 switch (VFS_I(ip)->i_mode & S_IFMT) {
132 xfs_idestroy_fork(&ip->i_df);
136 xfs_ifork_zap_attr(ip);
139 xfs_idestroy_fork(ip->i_cowfp);
140 kmem_cache_free(xfs_ifork_cache, ip->i_cowfp);
143 ASSERT(!test_bit(XFS_LI_IN_AIL,
144 &ip->i_itemp->ili_item.li_flags));
145 xfs_inode_item_destroy(ip);
149 kmem_cache_free(xfs_inode_cache, ip);
154 struct xfs_inode *ip)
156 /* asserts to verify all state is correct here */
157 ASSERT(atomic_read(&ip->i_pincount) == 0);
158 ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
159 XFS_STATS_DEC(ip->i_mount, vn_active);
161 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
166 struct xfs_inode *ip)
168 ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
171 * Because we use RCU freeing we need to ensure the inode always
172 * appears to be reclaimed with an invalid inode number when in the
173 * free state. The ip->i_flags_lock provides the barrier against lookup
176 spin_lock(&ip->i_flags_lock);
177 ip->i_flags = XFS_IRECLAIM;
179 spin_unlock(&ip->i_flags_lock);
181 __xfs_inode_free(ip);
185 * Queue background inode reclaim work if there are reclaimable inodes and there
186 * isn't reclaim work already scheduled or in progress.
189 xfs_reclaim_work_queue(
190 struct xfs_mount *mp)
194 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
195 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
196 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
202 * Background scanning to trim preallocated space. This is queued based on the
203 * 'speculative_prealloc_lifetime' tunable (5m by default).
207 struct xfs_perag *pag)
209 struct xfs_mount *mp = pag->pag_mount;
211 if (!xfs_is_blockgc_enabled(mp))
215 if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
216 queue_delayed_work(pag->pag_mount->m_blockgc_wq,
217 &pag->pag_blockgc_work,
218 msecs_to_jiffies(xfs_blockgc_secs * 1000));
222 /* Set a tag on both the AG incore inode tree and the AG radix tree. */
224 xfs_perag_set_inode_tag(
225 struct xfs_perag *pag,
229 struct xfs_mount *mp = pag->pag_mount;
232 lockdep_assert_held(&pag->pag_ici_lock);
234 was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
235 radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
237 if (tag == XFS_ICI_RECLAIM_TAG)
238 pag->pag_ici_reclaimable++;
243 /* propagate the tag up into the perag radix tree */
244 spin_lock(&mp->m_perag_lock);
245 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno, tag);
246 spin_unlock(&mp->m_perag_lock);
248 /* start background work */
250 case XFS_ICI_RECLAIM_TAG:
251 xfs_reclaim_work_queue(mp);
253 case XFS_ICI_BLOCKGC_TAG:
254 xfs_blockgc_queue(pag);
258 trace_xfs_perag_set_inode_tag(pag, _RET_IP_);
261 /* Clear a tag on both the AG incore inode tree and the AG radix tree. */
263 xfs_perag_clear_inode_tag(
264 struct xfs_perag *pag,
268 struct xfs_mount *mp = pag->pag_mount;
270 lockdep_assert_held(&pag->pag_ici_lock);
273 * Reclaim can signal (with a null agino) that it cleared its own tag
274 * by removing the inode from the radix tree.
276 if (agino != NULLAGINO)
277 radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
279 ASSERT(tag == XFS_ICI_RECLAIM_TAG);
281 if (tag == XFS_ICI_RECLAIM_TAG)
282 pag->pag_ici_reclaimable--;
284 if (radix_tree_tagged(&pag->pag_ici_root, tag))
287 /* clear the tag from the perag radix tree */
288 spin_lock(&mp->m_perag_lock);
289 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno, tag);
290 spin_unlock(&mp->m_perag_lock);
292 trace_xfs_perag_clear_inode_tag(pag, _RET_IP_);
296 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
297 * part of the structure. This is made more complex by the fact we store
298 * information about the on-disk values in the VFS inode and so we can't just
299 * overwrite the values unconditionally. Hence we save the parameters we
300 * need to retain across reinitialisation, and rewrite them into the VFS inode
301 * after reinitialisation even if it fails.
305 struct xfs_mount *mp,
309 uint32_t nlink = inode->i_nlink;
310 uint32_t generation = inode->i_generation;
311 uint64_t version = inode_peek_iversion(inode);
312 umode_t mode = inode->i_mode;
313 dev_t dev = inode->i_rdev;
314 kuid_t uid = inode->i_uid;
315 kgid_t gid = inode->i_gid;
317 error = inode_init_always(mp->m_super, inode);
319 set_nlink(inode, nlink);
320 inode->i_generation = generation;
321 inode_set_iversion_queried(inode, version);
322 inode->i_mode = mode;
326 mapping_set_large_folios(inode->i_mapping);
331 * Carefully nudge an inode whose VFS state has been torn down back into a
332 * usable state. Drops the i_flags_lock and the rcu read lock.
336 struct xfs_perag *pag,
337 struct xfs_inode *ip) __releases(&ip->i_flags_lock)
339 struct xfs_mount *mp = ip->i_mount;
340 struct inode *inode = VFS_I(ip);
343 trace_xfs_iget_recycle(ip);
345 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
349 * We need to make it look like the inode is being reclaimed to prevent
350 * the actual reclaim workers from stomping over us while we recycle
351 * the inode. We can't clear the radix tree tag yet as it requires
352 * pag_ici_lock to be held exclusive.
354 ip->i_flags |= XFS_IRECLAIM;
356 spin_unlock(&ip->i_flags_lock);
359 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
360 error = xfs_reinit_inode(mp, inode);
361 xfs_iunlock(ip, XFS_ILOCK_EXCL);
364 * Re-initializing the inode failed, and we are in deep
365 * trouble. Try to re-add it to the reclaim list.
368 spin_lock(&ip->i_flags_lock);
369 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
370 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
371 spin_unlock(&ip->i_flags_lock);
374 trace_xfs_iget_recycle_fail(ip);
378 spin_lock(&pag->pag_ici_lock);
379 spin_lock(&ip->i_flags_lock);
382 * Clear the per-lifetime state in the inode as we are now effectively
383 * a new inode and need to return to the initial state before reuse
386 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
387 ip->i_flags |= XFS_INEW;
388 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
389 XFS_ICI_RECLAIM_TAG);
390 inode->i_state = I_NEW;
391 spin_unlock(&ip->i_flags_lock);
392 spin_unlock(&pag->pag_ici_lock);
398 * If we are allocating a new inode, then check what was returned is
399 * actually a free, empty inode. If we are not allocating an inode,
400 * then check we didn't find a free inode.
403 * 0 if the inode free state matches the lookup context
404 * -ENOENT if the inode is free and we are not allocating
405 * -EFSCORRUPTED if there is any state mismatch at all
408 xfs_iget_check_free_state(
409 struct xfs_inode *ip,
412 if (flags & XFS_IGET_CREATE) {
413 /* should be a free inode */
414 if (VFS_I(ip)->i_mode != 0) {
415 xfs_warn(ip->i_mount,
416 "Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
417 ip->i_ino, VFS_I(ip)->i_mode);
418 return -EFSCORRUPTED;
421 if (ip->i_nblocks != 0) {
422 xfs_warn(ip->i_mount,
423 "Corruption detected! Free inode 0x%llx has blocks allocated!",
425 return -EFSCORRUPTED;
430 /* should be an allocated inode */
431 if (VFS_I(ip)->i_mode == 0)
437 /* Make all pending inactivation work start immediately. */
439 xfs_inodegc_queue_all(
440 struct xfs_mount *mp)
442 struct xfs_inodegc *gc;
446 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
447 gc = per_cpu_ptr(mp->m_inodegc, cpu);
448 if (!llist_empty(&gc->list)) {
449 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
457 /* Wait for all queued work and collect errors */
459 xfs_inodegc_wait_all(
460 struct xfs_mount *mp)
465 flush_workqueue(mp->m_inodegc_wq);
466 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
467 struct xfs_inodegc *gc;
469 gc = per_cpu_ptr(mp->m_inodegc, cpu);
470 if (gc->error && !error)
479 * Check the validity of the inode we just found it the cache
483 struct xfs_perag *pag,
484 struct xfs_inode *ip,
487 int lock_flags) __releases(RCU)
489 struct inode *inode = VFS_I(ip);
490 struct xfs_mount *mp = ip->i_mount;
494 * check for re-use of an inode within an RCU grace period due to the
495 * radix tree nodes not being updated yet. We monitor for this by
496 * setting the inode number to zero before freeing the inode structure.
497 * If the inode has been reallocated and set up, then the inode number
498 * will not match, so check for that, too.
500 spin_lock(&ip->i_flags_lock);
501 if (ip->i_ino != ino)
505 * If we are racing with another cache hit that is currently
506 * instantiating this inode or currently recycling it out of
507 * reclaimable state, wait for the initialisation to complete
510 * If we're racing with the inactivation worker we also want to wait.
511 * If we're creating a new file, it's possible that the worker
512 * previously marked the inode as free on disk but hasn't finished
513 * updating the incore state yet. The AGI buffer will be dirty and
514 * locked to the icreate transaction, so a synchronous push of the
515 * inodegc workers would result in deadlock. For a regular iget, the
516 * worker is running already, so we might as well wait.
518 * XXX(hch): eventually we should do something equivalent to
519 * wait_on_inode to wait for these flags to be cleared
520 * instead of polling for it.
522 if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
525 if (ip->i_flags & XFS_NEED_INACTIVE) {
526 /* Unlinked inodes cannot be re-grabbed. */
527 if (VFS_I(ip)->i_nlink == 0) {
531 goto out_inodegc_flush;
535 * Check the inode free state is valid. This also detects lookup
536 * racing with unlinks.
538 error = xfs_iget_check_free_state(ip, flags);
542 /* Skip inodes that have no vfs state. */
543 if ((flags & XFS_IGET_INCORE) &&
544 (ip->i_flags & XFS_IRECLAIMABLE))
547 /* The inode fits the selection criteria; process it. */
548 if (ip->i_flags & XFS_IRECLAIMABLE) {
549 /* Drops i_flags_lock and RCU read lock. */
550 error = xfs_iget_recycle(pag, ip);
551 if (error == -EAGAIN)
556 /* If the VFS inode is being torn down, pause and try again. */
560 /* We've got a live one. */
561 spin_unlock(&ip->i_flags_lock);
563 trace_xfs_iget_hit(ip);
567 xfs_ilock(ip, lock_flags);
569 if (!(flags & XFS_IGET_INCORE))
570 xfs_iflags_clear(ip, XFS_ISTALE);
571 XFS_STATS_INC(mp, xs_ig_found);
576 trace_xfs_iget_skip(ip);
577 XFS_STATS_INC(mp, xs_ig_frecycle);
580 spin_unlock(&ip->i_flags_lock);
585 spin_unlock(&ip->i_flags_lock);
588 * Do not wait for the workers, because the caller could hold an AGI
589 * buffer lock. We're just going to sleep in a loop anyway.
591 if (xfs_is_inodegc_enabled(mp))
592 xfs_inodegc_queue_all(mp);
598 struct xfs_mount *mp,
599 struct xfs_perag *pag,
602 struct xfs_inode **ipp,
606 struct xfs_inode *ip;
608 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
611 ip = xfs_inode_alloc(mp, ino);
615 error = xfs_imap(pag, tp, ip->i_ino, &ip->i_imap, flags);
620 * For version 5 superblocks, if we are initialising a new inode and we
621 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can
622 * simply build the new inode core with a random generation number.
624 * For version 4 (and older) superblocks, log recovery is dependent on
625 * the i_flushiter field being initialised from the current on-disk
626 * value and hence we must also read the inode off disk even when
627 * initializing new inodes.
629 if (xfs_has_v3inodes(mp) &&
630 (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
631 VFS_I(ip)->i_generation = get_random_u32();
635 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
639 error = xfs_inode_from_disk(ip,
640 xfs_buf_offset(bp, ip->i_imap.im_boffset));
642 xfs_buf_set_ref(bp, XFS_INO_REF);
643 xfs_trans_brelse(tp, bp);
649 trace_xfs_iget_miss(ip);
652 * Check the inode free state is valid. This also detects lookup
653 * racing with unlinks.
655 error = xfs_iget_check_free_state(ip, flags);
660 * Preload the radix tree so we can insert safely under the
661 * write spinlock. Note that we cannot sleep inside the preload
662 * region. Since we can be called from transaction context, don't
663 * recurse into the file system.
665 if (radix_tree_preload(GFP_NOFS)) {
671 * Because the inode hasn't been added to the radix-tree yet it can't
672 * be found by another thread, so we can do the non-sleeping lock here.
675 if (!xfs_ilock_nowait(ip, lock_flags))
680 * These values must be set before inserting the inode into the radix
681 * tree as the moment it is inserted a concurrent lookup (allowed by the
682 * RCU locking mechanism) can find it and that lookup must see that this
683 * is an inode currently under construction (i.e. that XFS_INEW is set).
684 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
685 * memory barrier that ensures this detection works correctly at lookup
689 if (flags & XFS_IGET_DONTCACHE)
690 d_mark_dontcache(VFS_I(ip));
694 xfs_iflags_set(ip, iflags);
696 /* insert the new inode */
697 spin_lock(&pag->pag_ici_lock);
698 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
699 if (unlikely(error)) {
700 WARN_ON(error != -EEXIST);
701 XFS_STATS_INC(mp, xs_ig_dup);
703 goto out_preload_end;
705 spin_unlock(&pag->pag_ici_lock);
706 radix_tree_preload_end();
712 spin_unlock(&pag->pag_ici_lock);
713 radix_tree_preload_end();
715 xfs_iunlock(ip, lock_flags);
717 __destroy_inode(VFS_I(ip));
723 * Look up an inode by number in the given file system. The inode is looked up
724 * in the cache held in each AG. If the inode is found in the cache, initialise
725 * the vfs inode if necessary.
727 * If it is not in core, read it in from the file system's device, add it to the
728 * cache and initialise the vfs inode.
730 * The inode is locked according to the value of the lock_flags parameter.
731 * Inode lookup is only done during metadata operations and not as part of the
732 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
736 struct xfs_mount *mp,
737 struct xfs_trans *tp,
741 struct xfs_inode **ipp)
743 struct xfs_inode *ip;
744 struct xfs_perag *pag;
748 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
750 /* reject inode numbers outside existing AGs */
751 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
754 XFS_STATS_INC(mp, xs_ig_attempts);
756 /* get the perag structure and ensure that it's inode capable */
757 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
758 agino = XFS_INO_TO_AGINO(mp, ino);
763 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
766 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
768 goto out_error_or_again;
771 if (flags & XFS_IGET_INCORE) {
773 goto out_error_or_again;
775 XFS_STATS_INC(mp, xs_ig_missed);
777 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
780 goto out_error_or_again;
787 * If we have a real type for an on-disk inode, we can setup the inode
788 * now. If it's a new inode being created, xfs_init_new_inode will
791 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
792 xfs_setup_existing_inode(ip);
796 if (!(flags & (XFS_IGET_INCORE | XFS_IGET_NORETRY)) &&
806 * Grab the inode for reclaim exclusively.
808 * We have found this inode via a lookup under RCU, so the inode may have
809 * already been freed, or it may be in the process of being recycled by
810 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
811 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
812 * will not be set. Hence we need to check for both these flag conditions to
813 * avoid inodes that are no longer reclaim candidates.
815 * Note: checking for other state flags here, under the i_flags_lock or not, is
816 * racy and should be avoided. Those races should be resolved only after we have
817 * ensured that we are able to reclaim this inode and the world can see that we
818 * are going to reclaim it.
820 * Return true if we grabbed it, false otherwise.
824 struct xfs_inode *ip,
825 struct xfs_icwalk *icw)
827 ASSERT(rcu_read_lock_held());
829 spin_lock(&ip->i_flags_lock);
830 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
831 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
832 /* not a reclaim candidate. */
833 spin_unlock(&ip->i_flags_lock);
837 /* Don't reclaim a sick inode unless the caller asked for it. */
839 (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
840 spin_unlock(&ip->i_flags_lock);
844 __xfs_iflags_set(ip, XFS_IRECLAIM);
845 spin_unlock(&ip->i_flags_lock);
850 * Inode reclaim is non-blocking, so the default action if progress cannot be
851 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
852 * XFS_IRECLAIM flag. If we are in a shutdown state, we don't care about
853 * blocking anymore and hence we can wait for the inode to be able to reclaim
856 * We do no IO here - if callers require inodes to be cleaned they must push the
857 * AIL first to trigger writeback of dirty inodes. This enables writeback to be
858 * done in the background in a non-blocking manner, and enables memory reclaim
859 * to make progress without blocking.
863 struct xfs_inode *ip,
864 struct xfs_perag *pag)
866 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
868 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
870 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
874 * Check for log shutdown because aborting the inode can move the log
875 * tail and corrupt in memory state. This is fine if the log is shut
876 * down, but if the log is still active and only the mount is shut down
877 * then the in-memory log tail movement caused by the abort can be
878 * incorrectly propagated to disk.
880 if (xlog_is_shutdown(ip->i_mount->m_log)) {
882 xfs_iflush_shutdown_abort(ip);
885 if (xfs_ipincount(ip))
886 goto out_clear_flush;
887 if (!xfs_inode_clean(ip))
888 goto out_clear_flush;
890 xfs_iflags_clear(ip, XFS_IFLUSHING);
892 trace_xfs_inode_reclaiming(ip);
895 * Because we use RCU freeing we need to ensure the inode always appears
896 * to be reclaimed with an invalid inode number when in the free state.
897 * We do this as early as possible under the ILOCK so that
898 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
899 * detect races with us here. By doing this, we guarantee that once
900 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
901 * it will see either a valid inode that will serialise correctly, or it
902 * will see an invalid inode that it can skip.
904 spin_lock(&ip->i_flags_lock);
905 ip->i_flags = XFS_IRECLAIM;
909 spin_unlock(&ip->i_flags_lock);
911 ASSERT(!ip->i_itemp || ip->i_itemp->ili_item.li_buf == NULL);
912 xfs_iunlock(ip, XFS_ILOCK_EXCL);
914 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
916 * Remove the inode from the per-AG radix tree.
918 * Because radix_tree_delete won't complain even if the item was never
919 * added to the tree assert that it's been there before to catch
920 * problems with the inode life time early on.
922 spin_lock(&pag->pag_ici_lock);
923 if (!radix_tree_delete(&pag->pag_ici_root,
924 XFS_INO_TO_AGINO(ip->i_mount, ino)))
926 xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
927 spin_unlock(&pag->pag_ici_lock);
930 * Here we do an (almost) spurious inode lock in order to coordinate
931 * with inode cache radix tree lookups. This is because the lookup
932 * can reference the inodes in the cache without taking references.
934 * We make that OK here by ensuring that we wait until the inode is
935 * unlocked after the lookup before we go ahead and free it.
937 xfs_ilock(ip, XFS_ILOCK_EXCL);
938 ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
939 xfs_iunlock(ip, XFS_ILOCK_EXCL);
940 ASSERT(xfs_inode_clean(ip));
942 __xfs_inode_free(ip);
946 xfs_iflags_clear(ip, XFS_IFLUSHING);
948 xfs_iunlock(ip, XFS_ILOCK_EXCL);
950 xfs_iflags_clear(ip, XFS_IRECLAIM);
953 /* Reclaim sick inodes if we're unmounting or the fs went down. */
955 xfs_want_reclaim_sick(
956 struct xfs_mount *mp)
958 return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
964 struct xfs_mount *mp)
966 struct xfs_icwalk icw = {
970 if (xfs_want_reclaim_sick(mp))
971 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
973 while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
974 xfs_ail_push_all_sync(mp->m_ail);
975 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
980 * The shrinker infrastructure determines how many inodes we should scan for
981 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
982 * push the AIL here. We also want to proactively free up memory if we can to
983 * minimise the amount of work memory reclaim has to do so we kick the
984 * background reclaim if it isn't already scheduled.
987 xfs_reclaim_inodes_nr(
988 struct xfs_mount *mp,
989 unsigned long nr_to_scan)
991 struct xfs_icwalk icw = {
992 .icw_flags = XFS_ICWALK_FLAG_SCAN_LIMIT,
993 .icw_scan_limit = min_t(unsigned long, LONG_MAX, nr_to_scan),
996 if (xfs_want_reclaim_sick(mp))
997 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
999 /* kick background reclaimer and push the AIL */
1000 xfs_reclaim_work_queue(mp);
1001 xfs_ail_push_all(mp->m_ail);
1003 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1008 * Return the number of reclaimable inodes in the filesystem for
1009 * the shrinker to determine how much to reclaim.
1012 xfs_reclaim_inodes_count(
1013 struct xfs_mount *mp)
1015 struct xfs_perag *pag;
1016 xfs_agnumber_t ag = 0;
1017 long reclaimable = 0;
1019 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1020 ag = pag->pag_agno + 1;
1021 reclaimable += pag->pag_ici_reclaimable;
1028 xfs_icwalk_match_id(
1029 struct xfs_inode *ip,
1030 struct xfs_icwalk *icw)
1032 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1033 !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1036 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1037 !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1040 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1041 ip->i_projid != icw->icw_prid)
1048 * A union-based inode filtering algorithm. Process the inode if any of the
1049 * criteria match. This is for global/internal scans only.
1052 xfs_icwalk_match_id_union(
1053 struct xfs_inode *ip,
1054 struct xfs_icwalk *icw)
1056 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1057 uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1060 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1061 gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1064 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1065 ip->i_projid == icw->icw_prid)
1072 * Is this inode @ip eligible for eof/cow block reclamation, given some
1073 * filtering parameters @icw? The inode is eligible if @icw is null or
1074 * if the predicate functions match.
1078 struct xfs_inode *ip,
1079 struct xfs_icwalk *icw)
1086 if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1087 match = xfs_icwalk_match_id_union(ip, icw);
1089 match = xfs_icwalk_match_id(ip, icw);
1093 /* skip the inode if the file size is too small */
1094 if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1095 XFS_ISIZE(ip) < icw->icw_min_file_size)
1102 * This is a fast pass over the inode cache to try to get reclaim moving on as
1103 * many inodes as possible in a short period of time. It kicks itself every few
1104 * seconds, as well as being kicked by the inode cache shrinker when memory
1109 struct work_struct *work)
1111 struct xfs_mount *mp = container_of(to_delayed_work(work),
1112 struct xfs_mount, m_reclaim_work);
1114 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1115 xfs_reclaim_work_queue(mp);
1119 xfs_inode_free_eofblocks(
1120 struct xfs_inode *ip,
1121 struct xfs_icwalk *icw,
1122 unsigned int *lockflags)
1126 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1128 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1132 * If the mapping is dirty the operation can block and wait for some
1133 * time. Unless we are waiting, skip it.
1135 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1138 if (!xfs_icwalk_match(ip, icw))
1142 * If the caller is waiting, return -EAGAIN to keep the background
1143 * scanner moving and revisit the inode in a subsequent pass.
1145 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1150 *lockflags |= XFS_IOLOCK_EXCL;
1152 if (xfs_can_free_eofblocks(ip, false))
1153 return xfs_free_eofblocks(ip);
1155 /* inode could be preallocated or append-only */
1156 trace_xfs_inode_free_eofblocks_invalid(ip);
1157 xfs_inode_clear_eofblocks_tag(ip);
1162 xfs_blockgc_set_iflag(
1163 struct xfs_inode *ip,
1164 unsigned long iflag)
1166 struct xfs_mount *mp = ip->i_mount;
1167 struct xfs_perag *pag;
1169 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1172 * Don't bother locking the AG and looking up in the radix trees
1173 * if we already know that we have the tag set.
1175 if (ip->i_flags & iflag)
1177 spin_lock(&ip->i_flags_lock);
1178 ip->i_flags |= iflag;
1179 spin_unlock(&ip->i_flags_lock);
1181 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1182 spin_lock(&pag->pag_ici_lock);
1184 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1185 XFS_ICI_BLOCKGC_TAG);
1187 spin_unlock(&pag->pag_ici_lock);
1192 xfs_inode_set_eofblocks_tag(
1195 trace_xfs_inode_set_eofblocks_tag(ip);
1196 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1200 xfs_blockgc_clear_iflag(
1201 struct xfs_inode *ip,
1202 unsigned long iflag)
1204 struct xfs_mount *mp = ip->i_mount;
1205 struct xfs_perag *pag;
1208 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1210 spin_lock(&ip->i_flags_lock);
1211 ip->i_flags &= ~iflag;
1212 clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1213 spin_unlock(&ip->i_flags_lock);
1218 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1219 spin_lock(&pag->pag_ici_lock);
1221 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1222 XFS_ICI_BLOCKGC_TAG);
1224 spin_unlock(&pag->pag_ici_lock);
1229 xfs_inode_clear_eofblocks_tag(
1232 trace_xfs_inode_clear_eofblocks_tag(ip);
1233 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1237 * Set ourselves up to free CoW blocks from this file. If it's already clean
1238 * then we can bail out quickly, but otherwise we must back off if the file
1239 * is undergoing some kind of write.
1242 xfs_prep_free_cowblocks(
1243 struct xfs_inode *ip)
1246 * Just clear the tag if we have an empty cow fork or none at all. It's
1247 * possible the inode was fully unshared since it was originally tagged.
1249 if (!xfs_inode_has_cow_data(ip)) {
1250 trace_xfs_inode_free_cowblocks_invalid(ip);
1251 xfs_inode_clear_cowblocks_tag(ip);
1256 * If the mapping is dirty or under writeback we cannot touch the
1257 * CoW fork. Leave it alone if we're in the midst of a directio.
1259 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1260 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1261 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1262 atomic_read(&VFS_I(ip)->i_dio_count))
1269 * Automatic CoW Reservation Freeing
1271 * These functions automatically garbage collect leftover CoW reservations
1272 * that were made on behalf of a cowextsize hint when we start to run out
1273 * of quota or when the reservations sit around for too long. If the file
1274 * has dirty pages or is undergoing writeback, its CoW reservations will
1277 * The actual garbage collection piggybacks off the same code that runs
1278 * the speculative EOF preallocation garbage collector.
1281 xfs_inode_free_cowblocks(
1282 struct xfs_inode *ip,
1283 struct xfs_icwalk *icw,
1284 unsigned int *lockflags)
1289 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1291 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1294 if (!xfs_prep_free_cowblocks(ip))
1297 if (!xfs_icwalk_match(ip, icw))
1301 * If the caller is waiting, return -EAGAIN to keep the background
1302 * scanner moving and revisit the inode in a subsequent pass.
1304 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1305 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1310 *lockflags |= XFS_IOLOCK_EXCL;
1312 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1317 *lockflags |= XFS_MMAPLOCK_EXCL;
1320 * Check again, nobody else should be able to dirty blocks or change
1321 * the reflink iflag now that we have the first two locks held.
1323 if (xfs_prep_free_cowblocks(ip))
1324 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1329 xfs_inode_set_cowblocks_tag(
1332 trace_xfs_inode_set_cowblocks_tag(ip);
1333 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1337 xfs_inode_clear_cowblocks_tag(
1340 trace_xfs_inode_clear_cowblocks_tag(ip);
1341 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1344 /* Disable post-EOF and CoW block auto-reclamation. */
1347 struct xfs_mount *mp)
1349 struct xfs_perag *pag;
1350 xfs_agnumber_t agno;
1352 if (!xfs_clear_blockgc_enabled(mp))
1355 for_each_perag(mp, agno, pag)
1356 cancel_delayed_work_sync(&pag->pag_blockgc_work);
1357 trace_xfs_blockgc_stop(mp, __return_address);
1360 /* Enable post-EOF and CoW block auto-reclamation. */
1363 struct xfs_mount *mp)
1365 struct xfs_perag *pag;
1366 xfs_agnumber_t agno;
1368 if (xfs_set_blockgc_enabled(mp))
1371 trace_xfs_blockgc_start(mp, __return_address);
1372 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1373 xfs_blockgc_queue(pag);
1376 /* Don't try to run block gc on an inode that's in any of these states. */
1377 #define XFS_BLOCKGC_NOGRAB_IFLAGS (XFS_INEW | \
1378 XFS_NEED_INACTIVE | \
1379 XFS_INACTIVATING | \
1380 XFS_IRECLAIMABLE | \
1383 * Decide if the given @ip is eligible for garbage collection of speculative
1384 * preallocations, and grab it if so. Returns true if it's ready to go or
1385 * false if we should just ignore it.
1389 struct xfs_inode *ip)
1391 struct inode *inode = VFS_I(ip);
1393 ASSERT(rcu_read_lock_held());
1395 /* Check for stale RCU freed inode */
1396 spin_lock(&ip->i_flags_lock);
1398 goto out_unlock_noent;
1400 if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1401 goto out_unlock_noent;
1402 spin_unlock(&ip->i_flags_lock);
1404 /* nothing to sync during shutdown */
1405 if (xfs_is_shutdown(ip->i_mount))
1408 /* If we can't grab the inode, it must on it's way to reclaim. */
1412 /* inode is valid */
1416 spin_unlock(&ip->i_flags_lock);
1420 /* Scan one incore inode for block preallocations that we can remove. */
1422 xfs_blockgc_scan_inode(
1423 struct xfs_inode *ip,
1424 struct xfs_icwalk *icw)
1426 unsigned int lockflags = 0;
1429 error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1433 error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1436 xfs_iunlock(ip, lockflags);
1441 /* Background worker that trims preallocated space. */
1444 struct work_struct *work)
1446 struct xfs_perag *pag = container_of(to_delayed_work(work),
1447 struct xfs_perag, pag_blockgc_work);
1448 struct xfs_mount *mp = pag->pag_mount;
1451 trace_xfs_blockgc_worker(mp, __return_address);
1453 error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1455 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1456 pag->pag_agno, error);
1457 xfs_blockgc_queue(pag);
1461 * Try to free space in the filesystem by purging inactive inodes, eofblocks
1465 xfs_blockgc_free_space(
1466 struct xfs_mount *mp,
1467 struct xfs_icwalk *icw)
1471 trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1473 error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1477 return xfs_inodegc_flush(mp);
1481 * Reclaim all the free space that we can by scheduling the background blockgc
1482 * and inodegc workers immediately and waiting for them all to clear.
1485 xfs_blockgc_flush_all(
1486 struct xfs_mount *mp)
1488 struct xfs_perag *pag;
1489 xfs_agnumber_t agno;
1491 trace_xfs_blockgc_flush_all(mp, __return_address);
1494 * For each blockgc worker, move its queue time up to now. If it
1495 * wasn't queued, it will not be requeued. Then flush whatever's
1498 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1499 mod_delayed_work(pag->pag_mount->m_blockgc_wq,
1500 &pag->pag_blockgc_work, 0);
1502 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1503 flush_delayed_work(&pag->pag_blockgc_work);
1505 return xfs_inodegc_flush(mp);
1509 * Run cow/eofblocks scans on the supplied dquots. We don't know exactly which
1510 * quota caused an allocation failure, so we make a best effort by including
1511 * each quota under low free space conditions (less than 1% free space) in the
1514 * Callers must not hold any inode's ILOCK. If requesting a synchronous scan
1515 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1519 xfs_blockgc_free_dquots(
1520 struct xfs_mount *mp,
1521 struct xfs_dquot *udqp,
1522 struct xfs_dquot *gdqp,
1523 struct xfs_dquot *pdqp,
1524 unsigned int iwalk_flags)
1526 struct xfs_icwalk icw = {0};
1527 bool do_work = false;
1529 if (!udqp && !gdqp && !pdqp)
1533 * Run a scan to free blocks using the union filter to cover all
1534 * applicable quotas in a single scan.
1536 icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1538 if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1539 icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1540 icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1544 if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1545 icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1546 icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1550 if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1551 icw.icw_prid = pdqp->q_id;
1552 icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1559 return xfs_blockgc_free_space(mp, &icw);
1562 /* Run cow/eofblocks scans on the quotas attached to the inode. */
1564 xfs_blockgc_free_quota(
1565 struct xfs_inode *ip,
1566 unsigned int iwalk_flags)
1568 return xfs_blockgc_free_dquots(ip->i_mount,
1569 xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1570 xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1571 xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1574 /* XFS Inode Cache Walking Code */
1577 * The inode lookup is done in batches to keep the amount of lock traffic and
1578 * radix tree lookups to a minimum. The batch size is a trade off between
1579 * lookup reduction and stack usage. This is in the reclaim path, so we can't
1582 #define XFS_LOOKUP_BATCH 32
1586 * Decide if we want to grab this inode in anticipation of doing work towards
1591 enum xfs_icwalk_goal goal,
1592 struct xfs_inode *ip,
1593 struct xfs_icwalk *icw)
1596 case XFS_ICWALK_BLOCKGC:
1597 return xfs_blockgc_igrab(ip);
1598 case XFS_ICWALK_RECLAIM:
1599 return xfs_reclaim_igrab(ip, icw);
1606 * Process an inode. Each processing function must handle any state changes
1607 * made by the icwalk igrab function. Return -EAGAIN to skip an inode.
1610 xfs_icwalk_process_inode(
1611 enum xfs_icwalk_goal goal,
1612 struct xfs_inode *ip,
1613 struct xfs_perag *pag,
1614 struct xfs_icwalk *icw)
1619 case XFS_ICWALK_BLOCKGC:
1620 error = xfs_blockgc_scan_inode(ip, icw);
1622 case XFS_ICWALK_RECLAIM:
1623 xfs_reclaim_inode(ip, pag);
1630 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1631 * process them in some manner.
1635 struct xfs_perag *pag,
1636 enum xfs_icwalk_goal goal,
1637 struct xfs_icwalk *icw)
1639 struct xfs_mount *mp = pag->pag_mount;
1640 uint32_t first_index;
1649 if (goal == XFS_ICWALK_RECLAIM)
1650 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1655 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1661 nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1662 (void **) batch, first_index,
1663 XFS_LOOKUP_BATCH, goal);
1671 * Grab the inodes before we drop the lock. if we found
1672 * nothing, nr == 0 and the loop will be skipped.
1674 for (i = 0; i < nr_found; i++) {
1675 struct xfs_inode *ip = batch[i];
1677 if (done || !xfs_icwalk_igrab(goal, ip, icw))
1681 * Update the index for the next lookup. Catch
1682 * overflows into the next AG range which can occur if
1683 * we have inodes in the last block of the AG and we
1684 * are currently pointing to the last inode.
1686 * Because we may see inodes that are from the wrong AG
1687 * due to RCU freeing and reallocation, only update the
1688 * index if it lies in this AG. It was a race that lead
1689 * us to see this inode, so another lookup from the
1690 * same index will not find it again.
1692 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
1694 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1695 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1699 /* unlock now we've grabbed the inodes. */
1702 for (i = 0; i < nr_found; i++) {
1705 error = xfs_icwalk_process_inode(goal, batch[i], pag,
1707 if (error == -EAGAIN) {
1711 if (error && last_error != -EFSCORRUPTED)
1715 /* bail out if the filesystem is corrupted. */
1716 if (error == -EFSCORRUPTED)
1721 if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1722 icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1723 if (icw->icw_scan_limit <= 0)
1726 } while (nr_found && !done);
1728 if (goal == XFS_ICWALK_RECLAIM) {
1731 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1741 /* Walk all incore inodes to achieve a given goal. */
1744 struct xfs_mount *mp,
1745 enum xfs_icwalk_goal goal,
1746 struct xfs_icwalk *icw)
1748 struct xfs_perag *pag;
1751 xfs_agnumber_t agno;
1753 for_each_perag_tag(mp, agno, pag, goal) {
1754 error = xfs_icwalk_ag(pag, goal, icw);
1757 if (error == -EFSCORRUPTED) {
1758 xfs_perag_rele(pag);
1764 BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1770 struct xfs_inode *ip,
1773 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, whichfork);
1774 struct xfs_bmbt_irec got;
1775 struct xfs_iext_cursor icur;
1777 if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1780 if (isnullstartblock(got.br_startblock)) {
1781 xfs_warn(ip->i_mount,
1782 "ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",
1784 whichfork == XFS_DATA_FORK ? "data" : "cow",
1785 got.br_startoff, got.br_blockcount);
1787 } while (xfs_iext_next_extent(ifp, &icur, &got));
1790 #define xfs_check_delalloc(ip, whichfork) do { } while (0)
1793 /* Schedule the inode for reclaim. */
1795 xfs_inodegc_set_reclaimable(
1796 struct xfs_inode *ip)
1798 struct xfs_mount *mp = ip->i_mount;
1799 struct xfs_perag *pag;
1801 if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {
1802 xfs_check_delalloc(ip, XFS_DATA_FORK);
1803 xfs_check_delalloc(ip, XFS_COW_FORK);
1807 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1808 spin_lock(&pag->pag_ici_lock);
1809 spin_lock(&ip->i_flags_lock);
1811 trace_xfs_inode_set_reclaimable(ip);
1812 ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);
1813 ip->i_flags |= XFS_IRECLAIMABLE;
1814 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1815 XFS_ICI_RECLAIM_TAG);
1817 spin_unlock(&ip->i_flags_lock);
1818 spin_unlock(&pag->pag_ici_lock);
1823 * Free all speculative preallocations and possibly even the inode itself.
1824 * This is the last chance to make changes to an otherwise unreferenced file
1825 * before incore reclamation happens.
1828 xfs_inodegc_inactivate(
1829 struct xfs_inode *ip)
1833 trace_xfs_inode_inactivating(ip);
1834 error = xfs_inactive(ip);
1835 xfs_inodegc_set_reclaimable(ip);
1842 struct work_struct *work)
1844 struct xfs_inodegc *gc = container_of(to_delayed_work(work),
1845 struct xfs_inodegc, work);
1846 struct llist_node *node = llist_del_all(&gc->list);
1847 struct xfs_inode *ip, *n;
1848 struct xfs_mount *mp = gc->mp;
1849 unsigned int nofs_flag;
1852 * Clear the cpu mask bit and ensure that we have seen the latest
1853 * update of the gc structure associated with this CPU. This matches
1854 * with the release semantics used when setting the cpumask bit in
1855 * xfs_inodegc_queue.
1857 cpumask_clear_cpu(gc->cpu, &mp->m_inodegc_cpumask);
1858 smp_mb__after_atomic();
1860 WRITE_ONCE(gc->items, 0);
1866 * We can allocate memory here while doing writeback on behalf of
1867 * memory reclaim. To avoid memory allocation deadlocks set the
1868 * task-wide nofs context for the following operations.
1870 nofs_flag = memalloc_nofs_save();
1872 ip = llist_entry(node, struct xfs_inode, i_gclist);
1873 trace_xfs_inodegc_worker(mp, READ_ONCE(gc->shrinker_hits));
1875 WRITE_ONCE(gc->shrinker_hits, 0);
1876 llist_for_each_entry_safe(ip, n, node, i_gclist) {
1879 xfs_iflags_set(ip, XFS_INACTIVATING);
1880 error = xfs_inodegc_inactivate(ip);
1881 if (error && !gc->error)
1885 memalloc_nofs_restore(nofs_flag);
1889 * Expedite all pending inodegc work to run immediately. This does not wait for
1890 * completion of the work.
1894 struct xfs_mount *mp)
1896 if (!xfs_is_inodegc_enabled(mp))
1898 trace_xfs_inodegc_push(mp, __return_address);
1899 xfs_inodegc_queue_all(mp);
1903 * Force all currently queued inode inactivation work to run immediately and
1904 * wait for the work to finish.
1908 struct xfs_mount *mp)
1910 xfs_inodegc_push(mp);
1911 trace_xfs_inodegc_flush(mp, __return_address);
1912 return xfs_inodegc_wait_all(mp);
1916 * Flush all the pending work and then disable the inode inactivation background
1917 * workers and wait for them to stop. Caller must hold sb->s_umount to
1918 * coordinate changes in the inodegc_enabled state.
1922 struct xfs_mount *mp)
1926 if (!xfs_clear_inodegc_enabled(mp))
1930 * Drain all pending inodegc work, including inodes that could be
1931 * queued by racing xfs_inodegc_queue or xfs_inodegc_shrinker_scan
1932 * threads that sample the inodegc state just prior to us clearing it.
1933 * The inodegc flag state prevents new threads from queuing more
1934 * inodes, so we queue pending work items and flush the workqueue until
1935 * all inodegc lists are empty. IOWs, we cannot use drain_workqueue
1936 * here because it does not allow other unserialized mechanisms to
1937 * reschedule inodegc work while this draining is in progress.
1939 xfs_inodegc_queue_all(mp);
1941 flush_workqueue(mp->m_inodegc_wq);
1942 rerun = xfs_inodegc_queue_all(mp);
1945 trace_xfs_inodegc_stop(mp, __return_address);
1949 * Enable the inode inactivation background workers and schedule deferred inode
1950 * inactivation work if there is any. Caller must hold sb->s_umount to
1951 * coordinate changes in the inodegc_enabled state.
1955 struct xfs_mount *mp)
1957 if (xfs_set_inodegc_enabled(mp))
1960 trace_xfs_inodegc_start(mp, __return_address);
1961 xfs_inodegc_queue_all(mp);
1964 #ifdef CONFIG_XFS_RT
1966 xfs_inodegc_want_queue_rt_file(
1967 struct xfs_inode *ip)
1969 struct xfs_mount *mp = ip->i_mount;
1971 if (!XFS_IS_REALTIME_INODE(ip))
1974 if (__percpu_counter_compare(&mp->m_frextents,
1975 mp->m_low_rtexts[XFS_LOWSP_5_PCNT],
1976 XFS_FDBLOCKS_BATCH) < 0)
1982 # define xfs_inodegc_want_queue_rt_file(ip) (false)
1983 #endif /* CONFIG_XFS_RT */
1986 * Schedule the inactivation worker when:
1988 * - We've accumulated more than one inode cluster buffer's worth of inodes.
1989 * - There is less than 5% free space left.
1990 * - Any of the quotas for this inode are near an enforcement limit.
1993 xfs_inodegc_want_queue_work(
1994 struct xfs_inode *ip,
1997 struct xfs_mount *mp = ip->i_mount;
1999 if (items > mp->m_ino_geo.inodes_per_cluster)
2002 if (__percpu_counter_compare(&mp->m_fdblocks,
2003 mp->m_low_space[XFS_LOWSP_5_PCNT],
2004 XFS_FDBLOCKS_BATCH) < 0)
2007 if (xfs_inodegc_want_queue_rt_file(ip))
2010 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
2013 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
2016 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
2023 * Upper bound on the number of inodes in each AG that can be queued for
2024 * inactivation at any given time, to avoid monopolizing the workqueue.
2026 #define XFS_INODEGC_MAX_BACKLOG (4 * XFS_INODES_PER_CHUNK)
2029 * Make the frontend wait for inactivations when:
2031 * - Memory shrinkers queued the inactivation worker and it hasn't finished.
2032 * - The queue depth exceeds the maximum allowable percpu backlog.
2034 * Note: If the current thread is running a transaction, we don't ever want to
2035 * wait for other transactions because that could introduce a deadlock.
2038 xfs_inodegc_want_flush_work(
2039 struct xfs_inode *ip,
2041 unsigned int shrinker_hits)
2043 if (current->journal_info)
2046 if (shrinker_hits > 0)
2049 if (items > XFS_INODEGC_MAX_BACKLOG)
2056 * Queue a background inactivation worker if there are inodes that need to be
2057 * inactivated and higher level xfs code hasn't disabled the background
2062 struct xfs_inode *ip)
2064 struct xfs_mount *mp = ip->i_mount;
2065 struct xfs_inodegc *gc;
2067 unsigned int shrinker_hits;
2068 unsigned int cpu_nr;
2069 unsigned long queue_delay = 1;
2071 trace_xfs_inode_set_need_inactive(ip);
2072 spin_lock(&ip->i_flags_lock);
2073 ip->i_flags |= XFS_NEED_INACTIVE;
2074 spin_unlock(&ip->i_flags_lock);
2077 gc = this_cpu_ptr(mp->m_inodegc);
2078 llist_add(&ip->i_gclist, &gc->list);
2079 items = READ_ONCE(gc->items);
2080 WRITE_ONCE(gc->items, items + 1);
2081 shrinker_hits = READ_ONCE(gc->shrinker_hits);
2084 * Ensure the list add is always seen by anyone who finds the cpumask
2085 * bit set. This effectively gives the cpumask bit set operation
2086 * release ordering semantics.
2088 smp_mb__before_atomic();
2089 if (!cpumask_test_cpu(cpu_nr, &mp->m_inodegc_cpumask))
2090 cpumask_test_and_set_cpu(cpu_nr, &mp->m_inodegc_cpumask);
2093 * We queue the work while holding the current CPU so that the work
2094 * is scheduled to run on this CPU.
2096 if (!xfs_is_inodegc_enabled(mp)) {
2101 if (xfs_inodegc_want_queue_work(ip, items))
2104 trace_xfs_inodegc_queue(mp, __return_address);
2105 mod_delayed_work_on(current_cpu(), mp->m_inodegc_wq, &gc->work,
2109 if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
2110 trace_xfs_inodegc_throttle(mp, __return_address);
2111 flush_delayed_work(&gc->work);
2116 * We set the inode flag atomically with the radix tree tag. Once we get tag
2117 * lookups on the radix tree, this inode flag can go away.
2119 * We always use background reclaim here because even if the inode is clean, it
2120 * still may be under IO and hence we have wait for IO completion to occur
2121 * before we can reclaim the inode. The background reclaim path handles this
2122 * more efficiently than we can here, so simply let background reclaim tear down
2126 xfs_inode_mark_reclaimable(
2127 struct xfs_inode *ip)
2129 struct xfs_mount *mp = ip->i_mount;
2132 XFS_STATS_INC(mp, vn_reclaim);
2135 * We should never get here with any of the reclaim flags already set.
2137 ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2139 need_inactive = xfs_inode_needs_inactive(ip);
2140 if (need_inactive) {
2141 xfs_inodegc_queue(ip);
2145 /* Going straight to reclaim, so drop the dquots. */
2146 xfs_qm_dqdetach(ip);
2147 xfs_inodegc_set_reclaimable(ip);
2151 * Register a phony shrinker so that we can run background inodegc sooner when
2152 * there's memory pressure. Inactivation does not itself free any memory but
2153 * it does make inodes reclaimable, which eventually frees memory.
2155 * The count function, seek value, and batch value are crafted to trigger the
2156 * scan function during the second round of scanning. Hopefully this means
2157 * that we reclaimed enough memory that initiating metadata transactions won't
2158 * make things worse.
2160 #define XFS_INODEGC_SHRINKER_COUNT (1UL << DEF_PRIORITY)
2161 #define XFS_INODEGC_SHRINKER_BATCH ((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2163 static unsigned long
2164 xfs_inodegc_shrinker_count(
2165 struct shrinker *shrink,
2166 struct shrink_control *sc)
2168 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2169 m_inodegc_shrinker);
2170 struct xfs_inodegc *gc;
2173 if (!xfs_is_inodegc_enabled(mp))
2176 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
2177 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2178 if (!llist_empty(&gc->list))
2179 return XFS_INODEGC_SHRINKER_COUNT;
2185 static unsigned long
2186 xfs_inodegc_shrinker_scan(
2187 struct shrinker *shrink,
2188 struct shrink_control *sc)
2190 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2191 m_inodegc_shrinker);
2192 struct xfs_inodegc *gc;
2194 bool no_items = true;
2196 if (!xfs_is_inodegc_enabled(mp))
2199 trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2201 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
2202 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2203 if (!llist_empty(&gc->list)) {
2204 unsigned int h = READ_ONCE(gc->shrinker_hits);
2206 WRITE_ONCE(gc->shrinker_hits, h + 1);
2207 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
2213 * If there are no inodes to inactivate, we don't want the shrinker
2214 * to think there's deferred work to call us back about.
2222 /* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2224 xfs_inodegc_register_shrinker(
2225 struct xfs_mount *mp)
2227 struct shrinker *shrink = &mp->m_inodegc_shrinker;
2229 shrink->count_objects = xfs_inodegc_shrinker_count;
2230 shrink->scan_objects = xfs_inodegc_shrinker_scan;
2232 shrink->flags = SHRINKER_NONSLAB;
2233 shrink->batch = XFS_INODEGC_SHRINKER_BATCH;
2235 return register_shrinker(shrink, "xfs-inodegc:%s", mp->m_super->s_id);