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));
103 memset(&ip->i_df, 0, sizeof(ip->i_df));
105 ip->i_delayed_blks = 0;
106 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
111 INIT_WORK(&ip->i_ioend_work, xfs_end_io);
112 INIT_LIST_HEAD(&ip->i_ioend_list);
113 spin_lock_init(&ip->i_ioend_lock);
119 xfs_inode_free_callback(
120 struct rcu_head *head)
122 struct inode *inode = container_of(head, struct inode, i_rcu);
123 struct xfs_inode *ip = XFS_I(inode);
125 switch (VFS_I(ip)->i_mode & S_IFMT) {
129 xfs_idestroy_fork(&ip->i_df);
134 xfs_idestroy_fork(ip->i_afp);
135 kmem_cache_free(xfs_ifork_cache, ip->i_afp);
138 xfs_idestroy_fork(ip->i_cowfp);
139 kmem_cache_free(xfs_ifork_cache, ip->i_cowfp);
142 ASSERT(!test_bit(XFS_LI_IN_AIL,
143 &ip->i_itemp->ili_item.li_flags));
144 xfs_inode_item_destroy(ip);
148 kmem_cache_free(xfs_inode_cache, ip);
153 struct xfs_inode *ip)
155 /* asserts to verify all state is correct here */
156 ASSERT(atomic_read(&ip->i_pincount) == 0);
157 ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
158 XFS_STATS_DEC(ip->i_mount, vn_active);
160 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
165 struct xfs_inode *ip)
167 ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
170 * Because we use RCU freeing we need to ensure the inode always
171 * appears to be reclaimed with an invalid inode number when in the
172 * free state. The ip->i_flags_lock provides the barrier against lookup
175 spin_lock(&ip->i_flags_lock);
176 ip->i_flags = XFS_IRECLAIM;
178 spin_unlock(&ip->i_flags_lock);
180 __xfs_inode_free(ip);
184 * Queue background inode reclaim work if there are reclaimable inodes and there
185 * isn't reclaim work already scheduled or in progress.
188 xfs_reclaim_work_queue(
189 struct xfs_mount *mp)
193 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
194 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
195 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
201 * Background scanning to trim preallocated space. This is queued based on the
202 * 'speculative_prealloc_lifetime' tunable (5m by default).
206 struct xfs_perag *pag)
208 struct xfs_mount *mp = pag->pag_mount;
210 if (!xfs_is_blockgc_enabled(mp))
214 if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
215 queue_delayed_work(pag->pag_mount->m_blockgc_wq,
216 &pag->pag_blockgc_work,
217 msecs_to_jiffies(xfs_blockgc_secs * 1000));
221 /* Set a tag on both the AG incore inode tree and the AG radix tree. */
223 xfs_perag_set_inode_tag(
224 struct xfs_perag *pag,
228 struct xfs_mount *mp = pag->pag_mount;
231 lockdep_assert_held(&pag->pag_ici_lock);
233 was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
234 radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
236 if (tag == XFS_ICI_RECLAIM_TAG)
237 pag->pag_ici_reclaimable++;
242 /* propagate the tag up into the perag radix tree */
243 spin_lock(&mp->m_perag_lock);
244 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno, tag);
245 spin_unlock(&mp->m_perag_lock);
247 /* start background work */
249 case XFS_ICI_RECLAIM_TAG:
250 xfs_reclaim_work_queue(mp);
252 case XFS_ICI_BLOCKGC_TAG:
253 xfs_blockgc_queue(pag);
257 trace_xfs_perag_set_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
260 /* Clear a tag on both the AG incore inode tree and the AG radix tree. */
262 xfs_perag_clear_inode_tag(
263 struct xfs_perag *pag,
267 struct xfs_mount *mp = pag->pag_mount;
269 lockdep_assert_held(&pag->pag_ici_lock);
272 * Reclaim can signal (with a null agino) that it cleared its own tag
273 * by removing the inode from the radix tree.
275 if (agino != NULLAGINO)
276 radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
278 ASSERT(tag == XFS_ICI_RECLAIM_TAG);
280 if (tag == XFS_ICI_RECLAIM_TAG)
281 pag->pag_ici_reclaimable--;
283 if (radix_tree_tagged(&pag->pag_ici_root, tag))
286 /* clear the tag from the perag radix tree */
287 spin_lock(&mp->m_perag_lock);
288 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno, tag);
289 spin_unlock(&mp->m_perag_lock);
291 trace_xfs_perag_clear_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
295 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
296 * part of the structure. This is made more complex by the fact we store
297 * information about the on-disk values in the VFS inode and so we can't just
298 * overwrite the values unconditionally. Hence we save the parameters we
299 * need to retain across reinitialisation, and rewrite them into the VFS inode
300 * after reinitialisation even if it fails.
304 struct xfs_mount *mp,
308 uint32_t nlink = inode->i_nlink;
309 uint32_t generation = inode->i_generation;
310 uint64_t version = inode_peek_iversion(inode);
311 umode_t mode = inode->i_mode;
312 dev_t dev = inode->i_rdev;
313 kuid_t uid = inode->i_uid;
314 kgid_t gid = inode->i_gid;
316 error = inode_init_always(mp->m_super, inode);
318 set_nlink(inode, nlink);
319 inode->i_generation = generation;
320 inode_set_iversion_queried(inode, version);
321 inode->i_mode = mode;
325 mapping_set_large_folios(inode->i_mapping);
330 * Carefully nudge an inode whose VFS state has been torn down back into a
331 * usable state. Drops the i_flags_lock and the rcu read lock.
335 struct xfs_perag *pag,
336 struct xfs_inode *ip) __releases(&ip->i_flags_lock)
338 struct xfs_mount *mp = ip->i_mount;
339 struct inode *inode = VFS_I(ip);
342 trace_xfs_iget_recycle(ip);
345 * We need to make it look like the inode is being reclaimed to prevent
346 * the actual reclaim workers from stomping over us while we recycle
347 * the inode. We can't clear the radix tree tag yet as it requires
348 * pag_ici_lock to be held exclusive.
350 ip->i_flags |= XFS_IRECLAIM;
352 spin_unlock(&ip->i_flags_lock);
355 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
356 error = xfs_reinit_inode(mp, inode);
359 * Re-initializing the inode failed, and we are in deep
360 * trouble. Try to re-add it to the reclaim list.
363 spin_lock(&ip->i_flags_lock);
364 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
365 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
366 spin_unlock(&ip->i_flags_lock);
369 trace_xfs_iget_recycle_fail(ip);
373 spin_lock(&pag->pag_ici_lock);
374 spin_lock(&ip->i_flags_lock);
377 * Clear the per-lifetime state in the inode as we are now effectively
378 * a new inode and need to return to the initial state before reuse
381 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
382 ip->i_flags |= XFS_INEW;
383 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
384 XFS_ICI_RECLAIM_TAG);
385 inode->i_state = I_NEW;
386 spin_unlock(&ip->i_flags_lock);
387 spin_unlock(&pag->pag_ici_lock);
393 * If we are allocating a new inode, then check what was returned is
394 * actually a free, empty inode. If we are not allocating an inode,
395 * then check we didn't find a free inode.
398 * 0 if the inode free state matches the lookup context
399 * -ENOENT if the inode is free and we are not allocating
400 * -EFSCORRUPTED if there is any state mismatch at all
403 xfs_iget_check_free_state(
404 struct xfs_inode *ip,
407 if (flags & XFS_IGET_CREATE) {
408 /* should be a free inode */
409 if (VFS_I(ip)->i_mode != 0) {
410 xfs_warn(ip->i_mount,
411 "Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
412 ip->i_ino, VFS_I(ip)->i_mode);
413 return -EFSCORRUPTED;
416 if (ip->i_nblocks != 0) {
417 xfs_warn(ip->i_mount,
418 "Corruption detected! Free inode 0x%llx has blocks allocated!",
420 return -EFSCORRUPTED;
425 /* should be an allocated inode */
426 if (VFS_I(ip)->i_mode == 0)
432 /* Make all pending inactivation work start immediately. */
434 xfs_inodegc_queue_all(
435 struct xfs_mount *mp)
437 struct xfs_inodegc *gc;
440 for_each_online_cpu(cpu) {
441 gc = per_cpu_ptr(mp->m_inodegc, cpu);
442 if (!llist_empty(&gc->list))
443 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
448 * Check the validity of the inode we just found it the cache
452 struct xfs_perag *pag,
453 struct xfs_inode *ip,
456 int lock_flags) __releases(RCU)
458 struct inode *inode = VFS_I(ip);
459 struct xfs_mount *mp = ip->i_mount;
463 * check for re-use of an inode within an RCU grace period due to the
464 * radix tree nodes not being updated yet. We monitor for this by
465 * setting the inode number to zero before freeing the inode structure.
466 * If the inode has been reallocated and set up, then the inode number
467 * will not match, so check for that, too.
469 spin_lock(&ip->i_flags_lock);
470 if (ip->i_ino != ino)
474 * If we are racing with another cache hit that is currently
475 * instantiating this inode or currently recycling it out of
476 * reclaimable state, wait for the initialisation to complete
479 * If we're racing with the inactivation worker we also want to wait.
480 * If we're creating a new file, it's possible that the worker
481 * previously marked the inode as free on disk but hasn't finished
482 * updating the incore state yet. The AGI buffer will be dirty and
483 * locked to the icreate transaction, so a synchronous push of the
484 * inodegc workers would result in deadlock. For a regular iget, the
485 * worker is running already, so we might as well wait.
487 * XXX(hch): eventually we should do something equivalent to
488 * wait_on_inode to wait for these flags to be cleared
489 * instead of polling for it.
491 if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
494 if (ip->i_flags & XFS_NEED_INACTIVE) {
495 /* Unlinked inodes cannot be re-grabbed. */
496 if (VFS_I(ip)->i_nlink == 0) {
500 goto out_inodegc_flush;
504 * Check the inode free state is valid. This also detects lookup
505 * racing with unlinks.
507 error = xfs_iget_check_free_state(ip, flags);
511 /* Skip inodes that have no vfs state. */
512 if ((flags & XFS_IGET_INCORE) &&
513 (ip->i_flags & XFS_IRECLAIMABLE))
516 /* The inode fits the selection criteria; process it. */
517 if (ip->i_flags & XFS_IRECLAIMABLE) {
518 /* Drops i_flags_lock and RCU read lock. */
519 error = xfs_iget_recycle(pag, ip);
523 /* If the VFS inode is being torn down, pause and try again. */
527 /* We've got a live one. */
528 spin_unlock(&ip->i_flags_lock);
530 trace_xfs_iget_hit(ip);
534 xfs_ilock(ip, lock_flags);
536 if (!(flags & XFS_IGET_INCORE))
537 xfs_iflags_clear(ip, XFS_ISTALE);
538 XFS_STATS_INC(mp, xs_ig_found);
543 trace_xfs_iget_skip(ip);
544 XFS_STATS_INC(mp, xs_ig_frecycle);
547 spin_unlock(&ip->i_flags_lock);
552 spin_unlock(&ip->i_flags_lock);
555 * Do not wait for the workers, because the caller could hold an AGI
556 * buffer lock. We're just going to sleep in a loop anyway.
558 if (xfs_is_inodegc_enabled(mp))
559 xfs_inodegc_queue_all(mp);
565 struct xfs_mount *mp,
566 struct xfs_perag *pag,
569 struct xfs_inode **ipp,
573 struct xfs_inode *ip;
575 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
578 ip = xfs_inode_alloc(mp, ino);
582 error = xfs_imap(mp, tp, ip->i_ino, &ip->i_imap, flags);
587 * For version 5 superblocks, if we are initialising a new inode and we
588 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can
589 * simply build the new inode core with a random generation number.
591 * For version 4 (and older) superblocks, log recovery is dependent on
592 * the i_flushiter field being initialised from the current on-disk
593 * value and hence we must also read the inode off disk even when
594 * initializing new inodes.
596 if (xfs_has_v3inodes(mp) &&
597 (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
598 VFS_I(ip)->i_generation = prandom_u32();
602 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
606 error = xfs_inode_from_disk(ip,
607 xfs_buf_offset(bp, ip->i_imap.im_boffset));
609 xfs_buf_set_ref(bp, XFS_INO_REF);
610 xfs_trans_brelse(tp, bp);
616 trace_xfs_iget_miss(ip);
619 * Check the inode free state is valid. This also detects lookup
620 * racing with unlinks.
622 error = xfs_iget_check_free_state(ip, flags);
627 * Preload the radix tree so we can insert safely under the
628 * write spinlock. Note that we cannot sleep inside the preload
629 * region. Since we can be called from transaction context, don't
630 * recurse into the file system.
632 if (radix_tree_preload(GFP_NOFS)) {
638 * Because the inode hasn't been added to the radix-tree yet it can't
639 * be found by another thread, so we can do the non-sleeping lock here.
642 if (!xfs_ilock_nowait(ip, lock_flags))
647 * These values must be set before inserting the inode into the radix
648 * tree as the moment it is inserted a concurrent lookup (allowed by the
649 * RCU locking mechanism) can find it and that lookup must see that this
650 * is an inode currently under construction (i.e. that XFS_INEW is set).
651 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
652 * memory barrier that ensures this detection works correctly at lookup
656 if (flags & XFS_IGET_DONTCACHE)
657 d_mark_dontcache(VFS_I(ip));
661 xfs_iflags_set(ip, iflags);
663 /* insert the new inode */
664 spin_lock(&pag->pag_ici_lock);
665 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
666 if (unlikely(error)) {
667 WARN_ON(error != -EEXIST);
668 XFS_STATS_INC(mp, xs_ig_dup);
670 goto out_preload_end;
672 spin_unlock(&pag->pag_ici_lock);
673 radix_tree_preload_end();
679 spin_unlock(&pag->pag_ici_lock);
680 radix_tree_preload_end();
682 xfs_iunlock(ip, lock_flags);
684 __destroy_inode(VFS_I(ip));
690 * Look up an inode by number in the given file system. The inode is looked up
691 * in the cache held in each AG. If the inode is found in the cache, initialise
692 * the vfs inode if necessary.
694 * If it is not in core, read it in from the file system's device, add it to the
695 * cache and initialise the vfs inode.
697 * The inode is locked according to the value of the lock_flags parameter.
698 * Inode lookup is only done during metadata operations and not as part of the
699 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
703 struct xfs_mount *mp,
704 struct xfs_trans *tp,
708 struct xfs_inode **ipp)
710 struct xfs_inode *ip;
711 struct xfs_perag *pag;
715 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
717 /* reject inode numbers outside existing AGs */
718 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
721 XFS_STATS_INC(mp, xs_ig_attempts);
723 /* get the perag structure and ensure that it's inode capable */
724 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
725 agino = XFS_INO_TO_AGINO(mp, ino);
730 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
733 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
735 goto out_error_or_again;
738 if (flags & XFS_IGET_INCORE) {
740 goto out_error_or_again;
742 XFS_STATS_INC(mp, xs_ig_missed);
744 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
747 goto out_error_or_again;
754 * If we have a real type for an on-disk inode, we can setup the inode
755 * now. If it's a new inode being created, xfs_init_new_inode will
758 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
759 xfs_setup_existing_inode(ip);
763 if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
772 * "Is this a cached inode that's also allocated?"
774 * Look up an inode by number in the given file system. If the inode is
775 * in cache and isn't in purgatory, return 1 if the inode is allocated
776 * and 0 if it is not. For all other cases (not in cache, being torn
777 * down, etc.), return a negative error code.
779 * The caller has to prevent inode allocation and freeing activity,
780 * presumably by locking the AGI buffer. This is to ensure that an
781 * inode cannot transition from allocated to freed until the caller is
782 * ready to allow that. If the inode is in an intermediate state (new,
783 * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
784 * inode is not in the cache, -ENOENT will be returned. The caller must
785 * deal with these scenarios appropriately.
787 * This is a specialized use case for the online scrubber; if you're
788 * reading this, you probably want xfs_iget.
791 xfs_icache_inode_is_allocated(
792 struct xfs_mount *mp,
793 struct xfs_trans *tp,
797 struct xfs_inode *ip;
800 error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
804 *inuse = !!(VFS_I(ip)->i_mode);
810 * Grab the inode for reclaim exclusively.
812 * We have found this inode via a lookup under RCU, so the inode may have
813 * already been freed, or it may be in the process of being recycled by
814 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
815 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
816 * will not be set. Hence we need to check for both these flag conditions to
817 * avoid inodes that are no longer reclaim candidates.
819 * Note: checking for other state flags here, under the i_flags_lock or not, is
820 * racy and should be avoided. Those races should be resolved only after we have
821 * ensured that we are able to reclaim this inode and the world can see that we
822 * are going to reclaim it.
824 * Return true if we grabbed it, false otherwise.
828 struct xfs_inode *ip,
829 struct xfs_icwalk *icw)
831 ASSERT(rcu_read_lock_held());
833 spin_lock(&ip->i_flags_lock);
834 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
835 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
836 /* not a reclaim candidate. */
837 spin_unlock(&ip->i_flags_lock);
841 /* Don't reclaim a sick inode unless the caller asked for it. */
843 (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
844 spin_unlock(&ip->i_flags_lock);
848 __xfs_iflags_set(ip, XFS_IRECLAIM);
849 spin_unlock(&ip->i_flags_lock);
854 * Inode reclaim is non-blocking, so the default action if progress cannot be
855 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
856 * XFS_IRECLAIM flag. If we are in a shutdown state, we don't care about
857 * blocking anymore and hence we can wait for the inode to be able to reclaim
860 * We do no IO here - if callers require inodes to be cleaned they must push the
861 * AIL first to trigger writeback of dirty inodes. This enables writeback to be
862 * done in the background in a non-blocking manner, and enables memory reclaim
863 * to make progress without blocking.
867 struct xfs_inode *ip,
868 struct xfs_perag *pag)
870 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
872 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
874 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
878 * Check for log shutdown because aborting the inode can move the log
879 * tail and corrupt in memory state. This is fine if the log is shut
880 * down, but if the log is still active and only the mount is shut down
881 * then the in-memory log tail movement caused by the abort can be
882 * incorrectly propagated to disk.
884 if (xlog_is_shutdown(ip->i_mount->m_log)) {
886 xfs_iflush_shutdown_abort(ip);
889 if (xfs_ipincount(ip))
890 goto out_clear_flush;
891 if (!xfs_inode_clean(ip))
892 goto out_clear_flush;
894 xfs_iflags_clear(ip, XFS_IFLUSHING);
896 trace_xfs_inode_reclaiming(ip);
899 * Because we use RCU freeing we need to ensure the inode always appears
900 * to be reclaimed with an invalid inode number when in the free state.
901 * We do this as early as possible under the ILOCK so that
902 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
903 * detect races with us here. By doing this, we guarantee that once
904 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
905 * it will see either a valid inode that will serialise correctly, or it
906 * will see an invalid inode that it can skip.
908 spin_lock(&ip->i_flags_lock);
909 ip->i_flags = XFS_IRECLAIM;
913 spin_unlock(&ip->i_flags_lock);
915 xfs_iunlock(ip, XFS_ILOCK_EXCL);
917 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
919 * Remove the inode from the per-AG radix tree.
921 * Because radix_tree_delete won't complain even if the item was never
922 * added to the tree assert that it's been there before to catch
923 * problems with the inode life time early on.
925 spin_lock(&pag->pag_ici_lock);
926 if (!radix_tree_delete(&pag->pag_ici_root,
927 XFS_INO_TO_AGINO(ip->i_mount, ino)))
929 xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
930 spin_unlock(&pag->pag_ici_lock);
933 * Here we do an (almost) spurious inode lock in order to coordinate
934 * with inode cache radix tree lookups. This is because the lookup
935 * can reference the inodes in the cache without taking references.
937 * We make that OK here by ensuring that we wait until the inode is
938 * unlocked after the lookup before we go ahead and free it.
940 xfs_ilock(ip, XFS_ILOCK_EXCL);
941 ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
942 xfs_iunlock(ip, XFS_ILOCK_EXCL);
943 ASSERT(xfs_inode_clean(ip));
945 __xfs_inode_free(ip);
949 xfs_iflags_clear(ip, XFS_IFLUSHING);
951 xfs_iunlock(ip, XFS_ILOCK_EXCL);
953 xfs_iflags_clear(ip, XFS_IRECLAIM);
956 /* Reclaim sick inodes if we're unmounting or the fs went down. */
958 xfs_want_reclaim_sick(
959 struct xfs_mount *mp)
961 return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
967 struct xfs_mount *mp)
969 struct xfs_icwalk icw = {
973 if (xfs_want_reclaim_sick(mp))
974 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
976 while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
977 xfs_ail_push_all_sync(mp->m_ail);
978 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
983 * The shrinker infrastructure determines how many inodes we should scan for
984 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
985 * push the AIL here. We also want to proactively free up memory if we can to
986 * minimise the amount of work memory reclaim has to do so we kick the
987 * background reclaim if it isn't already scheduled.
990 xfs_reclaim_inodes_nr(
991 struct xfs_mount *mp,
992 unsigned long nr_to_scan)
994 struct xfs_icwalk icw = {
995 .icw_flags = XFS_ICWALK_FLAG_SCAN_LIMIT,
996 .icw_scan_limit = min_t(unsigned long, LONG_MAX, nr_to_scan),
999 if (xfs_want_reclaim_sick(mp))
1000 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1002 /* kick background reclaimer and push the AIL */
1003 xfs_reclaim_work_queue(mp);
1004 xfs_ail_push_all(mp->m_ail);
1006 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1011 * Return the number of reclaimable inodes in the filesystem for
1012 * the shrinker to determine how much to reclaim.
1015 xfs_reclaim_inodes_count(
1016 struct xfs_mount *mp)
1018 struct xfs_perag *pag;
1019 xfs_agnumber_t ag = 0;
1020 long reclaimable = 0;
1022 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1023 ag = pag->pag_agno + 1;
1024 reclaimable += pag->pag_ici_reclaimable;
1031 xfs_icwalk_match_id(
1032 struct xfs_inode *ip,
1033 struct xfs_icwalk *icw)
1035 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1036 !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1039 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1040 !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1043 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1044 ip->i_projid != icw->icw_prid)
1051 * A union-based inode filtering algorithm. Process the inode if any of the
1052 * criteria match. This is for global/internal scans only.
1055 xfs_icwalk_match_id_union(
1056 struct xfs_inode *ip,
1057 struct xfs_icwalk *icw)
1059 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1060 uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1063 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1064 gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1067 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1068 ip->i_projid == icw->icw_prid)
1075 * Is this inode @ip eligible for eof/cow block reclamation, given some
1076 * filtering parameters @icw? The inode is eligible if @icw is null or
1077 * if the predicate functions match.
1081 struct xfs_inode *ip,
1082 struct xfs_icwalk *icw)
1089 if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1090 match = xfs_icwalk_match_id_union(ip, icw);
1092 match = xfs_icwalk_match_id(ip, icw);
1096 /* skip the inode if the file size is too small */
1097 if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1098 XFS_ISIZE(ip) < icw->icw_min_file_size)
1105 * This is a fast pass over the inode cache to try to get reclaim moving on as
1106 * many inodes as possible in a short period of time. It kicks itself every few
1107 * seconds, as well as being kicked by the inode cache shrinker when memory
1112 struct work_struct *work)
1114 struct xfs_mount *mp = container_of(to_delayed_work(work),
1115 struct xfs_mount, m_reclaim_work);
1117 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1118 xfs_reclaim_work_queue(mp);
1122 xfs_inode_free_eofblocks(
1123 struct xfs_inode *ip,
1124 struct xfs_icwalk *icw,
1125 unsigned int *lockflags)
1129 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1131 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1135 * If the mapping is dirty the operation can block and wait for some
1136 * time. Unless we are waiting, skip it.
1138 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1141 if (!xfs_icwalk_match(ip, icw))
1145 * If the caller is waiting, return -EAGAIN to keep the background
1146 * scanner moving and revisit the inode in a subsequent pass.
1148 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1153 *lockflags |= XFS_IOLOCK_EXCL;
1155 if (xfs_can_free_eofblocks(ip, false))
1156 return xfs_free_eofblocks(ip);
1158 /* inode could be preallocated or append-only */
1159 trace_xfs_inode_free_eofblocks_invalid(ip);
1160 xfs_inode_clear_eofblocks_tag(ip);
1165 xfs_blockgc_set_iflag(
1166 struct xfs_inode *ip,
1167 unsigned long iflag)
1169 struct xfs_mount *mp = ip->i_mount;
1170 struct xfs_perag *pag;
1172 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1175 * Don't bother locking the AG and looking up in the radix trees
1176 * if we already know that we have the tag set.
1178 if (ip->i_flags & iflag)
1180 spin_lock(&ip->i_flags_lock);
1181 ip->i_flags |= iflag;
1182 spin_unlock(&ip->i_flags_lock);
1184 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1185 spin_lock(&pag->pag_ici_lock);
1187 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1188 XFS_ICI_BLOCKGC_TAG);
1190 spin_unlock(&pag->pag_ici_lock);
1195 xfs_inode_set_eofblocks_tag(
1198 trace_xfs_inode_set_eofblocks_tag(ip);
1199 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1203 xfs_blockgc_clear_iflag(
1204 struct xfs_inode *ip,
1205 unsigned long iflag)
1207 struct xfs_mount *mp = ip->i_mount;
1208 struct xfs_perag *pag;
1211 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1213 spin_lock(&ip->i_flags_lock);
1214 ip->i_flags &= ~iflag;
1215 clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1216 spin_unlock(&ip->i_flags_lock);
1221 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1222 spin_lock(&pag->pag_ici_lock);
1224 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1225 XFS_ICI_BLOCKGC_TAG);
1227 spin_unlock(&pag->pag_ici_lock);
1232 xfs_inode_clear_eofblocks_tag(
1235 trace_xfs_inode_clear_eofblocks_tag(ip);
1236 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1240 * Set ourselves up to free CoW blocks from this file. If it's already clean
1241 * then we can bail out quickly, but otherwise we must back off if the file
1242 * is undergoing some kind of write.
1245 xfs_prep_free_cowblocks(
1246 struct xfs_inode *ip)
1249 * Just clear the tag if we have an empty cow fork or none at all. It's
1250 * possible the inode was fully unshared since it was originally tagged.
1252 if (!xfs_inode_has_cow_data(ip)) {
1253 trace_xfs_inode_free_cowblocks_invalid(ip);
1254 xfs_inode_clear_cowblocks_tag(ip);
1259 * If the mapping is dirty or under writeback we cannot touch the
1260 * CoW fork. Leave it alone if we're in the midst of a directio.
1262 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1263 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1264 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1265 atomic_read(&VFS_I(ip)->i_dio_count))
1272 * Automatic CoW Reservation Freeing
1274 * These functions automatically garbage collect leftover CoW reservations
1275 * that were made on behalf of a cowextsize hint when we start to run out
1276 * of quota or when the reservations sit around for too long. If the file
1277 * has dirty pages or is undergoing writeback, its CoW reservations will
1280 * The actual garbage collection piggybacks off the same code that runs
1281 * the speculative EOF preallocation garbage collector.
1284 xfs_inode_free_cowblocks(
1285 struct xfs_inode *ip,
1286 struct xfs_icwalk *icw,
1287 unsigned int *lockflags)
1292 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1294 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1297 if (!xfs_prep_free_cowblocks(ip))
1300 if (!xfs_icwalk_match(ip, icw))
1304 * If the caller is waiting, return -EAGAIN to keep the background
1305 * scanner moving and revisit the inode in a subsequent pass.
1307 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1308 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1313 *lockflags |= XFS_IOLOCK_EXCL;
1315 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1320 *lockflags |= XFS_MMAPLOCK_EXCL;
1323 * Check again, nobody else should be able to dirty blocks or change
1324 * the reflink iflag now that we have the first two locks held.
1326 if (xfs_prep_free_cowblocks(ip))
1327 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1332 xfs_inode_set_cowblocks_tag(
1335 trace_xfs_inode_set_cowblocks_tag(ip);
1336 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1340 xfs_inode_clear_cowblocks_tag(
1343 trace_xfs_inode_clear_cowblocks_tag(ip);
1344 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1347 /* Disable post-EOF and CoW block auto-reclamation. */
1350 struct xfs_mount *mp)
1352 struct xfs_perag *pag;
1353 xfs_agnumber_t agno;
1355 if (!xfs_clear_blockgc_enabled(mp))
1358 for_each_perag(mp, agno, pag)
1359 cancel_delayed_work_sync(&pag->pag_blockgc_work);
1360 trace_xfs_blockgc_stop(mp, __return_address);
1363 /* Enable post-EOF and CoW block auto-reclamation. */
1366 struct xfs_mount *mp)
1368 struct xfs_perag *pag;
1369 xfs_agnumber_t agno;
1371 if (xfs_set_blockgc_enabled(mp))
1374 trace_xfs_blockgc_start(mp, __return_address);
1375 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1376 xfs_blockgc_queue(pag);
1379 /* Don't try to run block gc on an inode that's in any of these states. */
1380 #define XFS_BLOCKGC_NOGRAB_IFLAGS (XFS_INEW | \
1381 XFS_NEED_INACTIVE | \
1382 XFS_INACTIVATING | \
1383 XFS_IRECLAIMABLE | \
1386 * Decide if the given @ip is eligible for garbage collection of speculative
1387 * preallocations, and grab it if so. Returns true if it's ready to go or
1388 * false if we should just ignore it.
1392 struct xfs_inode *ip)
1394 struct inode *inode = VFS_I(ip);
1396 ASSERT(rcu_read_lock_held());
1398 /* Check for stale RCU freed inode */
1399 spin_lock(&ip->i_flags_lock);
1401 goto out_unlock_noent;
1403 if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1404 goto out_unlock_noent;
1405 spin_unlock(&ip->i_flags_lock);
1407 /* nothing to sync during shutdown */
1408 if (xfs_is_shutdown(ip->i_mount))
1411 /* If we can't grab the inode, it must on it's way to reclaim. */
1415 /* inode is valid */
1419 spin_unlock(&ip->i_flags_lock);
1423 /* Scan one incore inode for block preallocations that we can remove. */
1425 xfs_blockgc_scan_inode(
1426 struct xfs_inode *ip,
1427 struct xfs_icwalk *icw)
1429 unsigned int lockflags = 0;
1432 error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1436 error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1439 xfs_iunlock(ip, lockflags);
1444 /* Background worker that trims preallocated space. */
1447 struct work_struct *work)
1449 struct xfs_perag *pag = container_of(to_delayed_work(work),
1450 struct xfs_perag, pag_blockgc_work);
1451 struct xfs_mount *mp = pag->pag_mount;
1454 trace_xfs_blockgc_worker(mp, __return_address);
1456 error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1458 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1459 pag->pag_agno, error);
1460 xfs_blockgc_queue(pag);
1464 * Try to free space in the filesystem by purging inactive inodes, eofblocks
1468 xfs_blockgc_free_space(
1469 struct xfs_mount *mp,
1470 struct xfs_icwalk *icw)
1474 trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1476 error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1480 xfs_inodegc_flush(mp);
1485 * Reclaim all the free space that we can by scheduling the background blockgc
1486 * and inodegc workers immediately and waiting for them all to clear.
1489 xfs_blockgc_flush_all(
1490 struct xfs_mount *mp)
1492 struct xfs_perag *pag;
1493 xfs_agnumber_t agno;
1495 trace_xfs_blockgc_flush_all(mp, __return_address);
1498 * For each blockgc worker, move its queue time up to now. If it
1499 * wasn't queued, it will not be requeued. Then flush whatever's
1502 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1503 mod_delayed_work(pag->pag_mount->m_blockgc_wq,
1504 &pag->pag_blockgc_work, 0);
1506 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1507 flush_delayed_work(&pag->pag_blockgc_work);
1509 xfs_inodegc_flush(mp);
1513 * Run cow/eofblocks scans on the supplied dquots. We don't know exactly which
1514 * quota caused an allocation failure, so we make a best effort by including
1515 * each quota under low free space conditions (less than 1% free space) in the
1518 * Callers must not hold any inode's ILOCK. If requesting a synchronous scan
1519 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1523 xfs_blockgc_free_dquots(
1524 struct xfs_mount *mp,
1525 struct xfs_dquot *udqp,
1526 struct xfs_dquot *gdqp,
1527 struct xfs_dquot *pdqp,
1528 unsigned int iwalk_flags)
1530 struct xfs_icwalk icw = {0};
1531 bool do_work = false;
1533 if (!udqp && !gdqp && !pdqp)
1537 * Run a scan to free blocks using the union filter to cover all
1538 * applicable quotas in a single scan.
1540 icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1542 if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1543 icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1544 icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1548 if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1549 icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1550 icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1554 if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1555 icw.icw_prid = pdqp->q_id;
1556 icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1563 return xfs_blockgc_free_space(mp, &icw);
1566 /* Run cow/eofblocks scans on the quotas attached to the inode. */
1568 xfs_blockgc_free_quota(
1569 struct xfs_inode *ip,
1570 unsigned int iwalk_flags)
1572 return xfs_blockgc_free_dquots(ip->i_mount,
1573 xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1574 xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1575 xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1578 /* XFS Inode Cache Walking Code */
1581 * The inode lookup is done in batches to keep the amount of lock traffic and
1582 * radix tree lookups to a minimum. The batch size is a trade off between
1583 * lookup reduction and stack usage. This is in the reclaim path, so we can't
1586 #define XFS_LOOKUP_BATCH 32
1590 * Decide if we want to grab this inode in anticipation of doing work towards
1595 enum xfs_icwalk_goal goal,
1596 struct xfs_inode *ip,
1597 struct xfs_icwalk *icw)
1600 case XFS_ICWALK_BLOCKGC:
1601 return xfs_blockgc_igrab(ip);
1602 case XFS_ICWALK_RECLAIM:
1603 return xfs_reclaim_igrab(ip, icw);
1610 * Process an inode. Each processing function must handle any state changes
1611 * made by the icwalk igrab function. Return -EAGAIN to skip an inode.
1614 xfs_icwalk_process_inode(
1615 enum xfs_icwalk_goal goal,
1616 struct xfs_inode *ip,
1617 struct xfs_perag *pag,
1618 struct xfs_icwalk *icw)
1623 case XFS_ICWALK_BLOCKGC:
1624 error = xfs_blockgc_scan_inode(ip, icw);
1626 case XFS_ICWALK_RECLAIM:
1627 xfs_reclaim_inode(ip, pag);
1634 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1635 * process them in some manner.
1639 struct xfs_perag *pag,
1640 enum xfs_icwalk_goal goal,
1641 struct xfs_icwalk *icw)
1643 struct xfs_mount *mp = pag->pag_mount;
1644 uint32_t first_index;
1653 if (goal == XFS_ICWALK_RECLAIM)
1654 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1659 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1665 nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1666 (void **) batch, first_index,
1667 XFS_LOOKUP_BATCH, goal);
1675 * Grab the inodes before we drop the lock. if we found
1676 * nothing, nr == 0 and the loop will be skipped.
1678 for (i = 0; i < nr_found; i++) {
1679 struct xfs_inode *ip = batch[i];
1681 if (done || !xfs_icwalk_igrab(goal, ip, icw))
1685 * Update the index for the next lookup. Catch
1686 * overflows into the next AG range which can occur if
1687 * we have inodes in the last block of the AG and we
1688 * are currently pointing to the last inode.
1690 * Because we may see inodes that are from the wrong AG
1691 * due to RCU freeing and reallocation, only update the
1692 * index if it lies in this AG. It was a race that lead
1693 * us to see this inode, so another lookup from the
1694 * same index will not find it again.
1696 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
1698 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1699 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1703 /* unlock now we've grabbed the inodes. */
1706 for (i = 0; i < nr_found; i++) {
1709 error = xfs_icwalk_process_inode(goal, batch[i], pag,
1711 if (error == -EAGAIN) {
1715 if (error && last_error != -EFSCORRUPTED)
1719 /* bail out if the filesystem is corrupted. */
1720 if (error == -EFSCORRUPTED)
1725 if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1726 icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1727 if (icw->icw_scan_limit <= 0)
1730 } while (nr_found && !done);
1732 if (goal == XFS_ICWALK_RECLAIM) {
1735 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1745 /* Walk all incore inodes to achieve a given goal. */
1748 struct xfs_mount *mp,
1749 enum xfs_icwalk_goal goal,
1750 struct xfs_icwalk *icw)
1752 struct xfs_perag *pag;
1755 xfs_agnumber_t agno;
1757 for_each_perag_tag(mp, agno, pag, goal) {
1758 error = xfs_icwalk_ag(pag, goal, icw);
1761 if (error == -EFSCORRUPTED) {
1768 BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1774 struct xfs_inode *ip,
1777 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, whichfork);
1778 struct xfs_bmbt_irec got;
1779 struct xfs_iext_cursor icur;
1781 if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1784 if (isnullstartblock(got.br_startblock)) {
1785 xfs_warn(ip->i_mount,
1786 "ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",
1788 whichfork == XFS_DATA_FORK ? "data" : "cow",
1789 got.br_startoff, got.br_blockcount);
1791 } while (xfs_iext_next_extent(ifp, &icur, &got));
1794 #define xfs_check_delalloc(ip, whichfork) do { } while (0)
1797 /* Schedule the inode for reclaim. */
1799 xfs_inodegc_set_reclaimable(
1800 struct xfs_inode *ip)
1802 struct xfs_mount *mp = ip->i_mount;
1803 struct xfs_perag *pag;
1805 if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {
1806 xfs_check_delalloc(ip, XFS_DATA_FORK);
1807 xfs_check_delalloc(ip, XFS_COW_FORK);
1811 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1812 spin_lock(&pag->pag_ici_lock);
1813 spin_lock(&ip->i_flags_lock);
1815 trace_xfs_inode_set_reclaimable(ip);
1816 ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);
1817 ip->i_flags |= XFS_IRECLAIMABLE;
1818 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1819 XFS_ICI_RECLAIM_TAG);
1821 spin_unlock(&ip->i_flags_lock);
1822 spin_unlock(&pag->pag_ici_lock);
1827 * Free all speculative preallocations and possibly even the inode itself.
1828 * This is the last chance to make changes to an otherwise unreferenced file
1829 * before incore reclamation happens.
1832 xfs_inodegc_inactivate(
1833 struct xfs_inode *ip)
1835 trace_xfs_inode_inactivating(ip);
1837 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;
1849 WRITE_ONCE(gc->items, 0);
1854 ip = llist_entry(node, struct xfs_inode, i_gclist);
1855 trace_xfs_inodegc_worker(ip->i_mount, READ_ONCE(gc->shrinker_hits));
1857 WRITE_ONCE(gc->shrinker_hits, 0);
1858 llist_for_each_entry_safe(ip, n, node, i_gclist) {
1859 xfs_iflags_set(ip, XFS_INACTIVATING);
1860 xfs_inodegc_inactivate(ip);
1865 * Expedite all pending inodegc work to run immediately. This does not wait for
1866 * completion of the work.
1870 struct xfs_mount *mp)
1872 if (!xfs_is_inodegc_enabled(mp))
1874 trace_xfs_inodegc_push(mp, __return_address);
1875 xfs_inodegc_queue_all(mp);
1879 * Force all currently queued inode inactivation work to run immediately and
1880 * wait for the work to finish.
1884 struct xfs_mount *mp)
1886 xfs_inodegc_push(mp);
1887 trace_xfs_inodegc_flush(mp, __return_address);
1888 flush_workqueue(mp->m_inodegc_wq);
1892 * Flush all the pending work and then disable the inode inactivation background
1893 * workers and wait for them to stop.
1897 struct xfs_mount *mp)
1899 if (!xfs_clear_inodegc_enabled(mp))
1902 xfs_inodegc_queue_all(mp);
1903 drain_workqueue(mp->m_inodegc_wq);
1905 trace_xfs_inodegc_stop(mp, __return_address);
1909 * Enable the inode inactivation background workers and schedule deferred inode
1910 * inactivation work if there is any.
1914 struct xfs_mount *mp)
1916 if (xfs_set_inodegc_enabled(mp))
1919 trace_xfs_inodegc_start(mp, __return_address);
1920 xfs_inodegc_queue_all(mp);
1923 #ifdef CONFIG_XFS_RT
1925 xfs_inodegc_want_queue_rt_file(
1926 struct xfs_inode *ip)
1928 struct xfs_mount *mp = ip->i_mount;
1930 if (!XFS_IS_REALTIME_INODE(ip))
1933 if (__percpu_counter_compare(&mp->m_frextents,
1934 mp->m_low_rtexts[XFS_LOWSP_5_PCNT],
1935 XFS_FDBLOCKS_BATCH) < 0)
1941 # define xfs_inodegc_want_queue_rt_file(ip) (false)
1942 #endif /* CONFIG_XFS_RT */
1945 * Schedule the inactivation worker when:
1947 * - We've accumulated more than one inode cluster buffer's worth of inodes.
1948 * - There is less than 5% free space left.
1949 * - Any of the quotas for this inode are near an enforcement limit.
1952 xfs_inodegc_want_queue_work(
1953 struct xfs_inode *ip,
1956 struct xfs_mount *mp = ip->i_mount;
1958 if (items > mp->m_ino_geo.inodes_per_cluster)
1961 if (__percpu_counter_compare(&mp->m_fdblocks,
1962 mp->m_low_space[XFS_LOWSP_5_PCNT],
1963 XFS_FDBLOCKS_BATCH) < 0)
1966 if (xfs_inodegc_want_queue_rt_file(ip))
1969 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
1972 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
1975 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
1982 * Upper bound on the number of inodes in each AG that can be queued for
1983 * inactivation at any given time, to avoid monopolizing the workqueue.
1985 #define XFS_INODEGC_MAX_BACKLOG (4 * XFS_INODES_PER_CHUNK)
1988 * Make the frontend wait for inactivations when:
1990 * - Memory shrinkers queued the inactivation worker and it hasn't finished.
1991 * - The queue depth exceeds the maximum allowable percpu backlog.
1993 * Note: If the current thread is running a transaction, we don't ever want to
1994 * wait for other transactions because that could introduce a deadlock.
1997 xfs_inodegc_want_flush_work(
1998 struct xfs_inode *ip,
2000 unsigned int shrinker_hits)
2002 if (current->journal_info)
2005 if (shrinker_hits > 0)
2008 if (items > XFS_INODEGC_MAX_BACKLOG)
2015 * Queue a background inactivation worker if there are inodes that need to be
2016 * inactivated and higher level xfs code hasn't disabled the background
2021 struct xfs_inode *ip)
2023 struct xfs_mount *mp = ip->i_mount;
2024 struct xfs_inodegc *gc;
2026 unsigned int shrinker_hits;
2027 unsigned long queue_delay = 1;
2029 trace_xfs_inode_set_need_inactive(ip);
2030 spin_lock(&ip->i_flags_lock);
2031 ip->i_flags |= XFS_NEED_INACTIVE;
2032 spin_unlock(&ip->i_flags_lock);
2034 gc = get_cpu_ptr(mp->m_inodegc);
2035 llist_add(&ip->i_gclist, &gc->list);
2036 items = READ_ONCE(gc->items);
2037 WRITE_ONCE(gc->items, items + 1);
2038 shrinker_hits = READ_ONCE(gc->shrinker_hits);
2041 * We queue the work while holding the current CPU so that the work
2042 * is scheduled to run on this CPU.
2044 if (!xfs_is_inodegc_enabled(mp)) {
2049 if (xfs_inodegc_want_queue_work(ip, items))
2052 trace_xfs_inodegc_queue(mp, __return_address);
2053 mod_delayed_work(mp->m_inodegc_wq, &gc->work, queue_delay);
2056 if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
2057 trace_xfs_inodegc_throttle(mp, __return_address);
2058 flush_delayed_work(&gc->work);
2063 * Fold the dead CPU inodegc queue into the current CPUs queue.
2066 xfs_inodegc_cpu_dead(
2067 struct xfs_mount *mp,
2068 unsigned int dead_cpu)
2070 struct xfs_inodegc *dead_gc, *gc;
2071 struct llist_node *first, *last;
2072 unsigned int count = 0;
2074 dead_gc = per_cpu_ptr(mp->m_inodegc, dead_cpu);
2075 cancel_delayed_work_sync(&dead_gc->work);
2077 if (llist_empty(&dead_gc->list))
2080 first = dead_gc->list.first;
2082 while (last->next) {
2086 dead_gc->list.first = NULL;
2089 /* Add pending work to current CPU */
2090 gc = get_cpu_ptr(mp->m_inodegc);
2091 llist_add_batch(first, last, &gc->list);
2092 count += READ_ONCE(gc->items);
2093 WRITE_ONCE(gc->items, count);
2095 if (xfs_is_inodegc_enabled(mp)) {
2096 trace_xfs_inodegc_queue(mp, __return_address);
2097 mod_delayed_work(mp->m_inodegc_wq, &gc->work, 0);
2103 * We set the inode flag atomically with the radix tree tag. Once we get tag
2104 * lookups on the radix tree, this inode flag can go away.
2106 * We always use background reclaim here because even if the inode is clean, it
2107 * still may be under IO and hence we have wait for IO completion to occur
2108 * before we can reclaim the inode. The background reclaim path handles this
2109 * more efficiently than we can here, so simply let background reclaim tear down
2113 xfs_inode_mark_reclaimable(
2114 struct xfs_inode *ip)
2116 struct xfs_mount *mp = ip->i_mount;
2119 XFS_STATS_INC(mp, vn_reclaim);
2122 * We should never get here with any of the reclaim flags already set.
2124 ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2126 need_inactive = xfs_inode_needs_inactive(ip);
2127 if (need_inactive) {
2128 xfs_inodegc_queue(ip);
2132 /* Going straight to reclaim, so drop the dquots. */
2133 xfs_qm_dqdetach(ip);
2134 xfs_inodegc_set_reclaimable(ip);
2138 * Register a phony shrinker so that we can run background inodegc sooner when
2139 * there's memory pressure. Inactivation does not itself free any memory but
2140 * it does make inodes reclaimable, which eventually frees memory.
2142 * The count function, seek value, and batch value are crafted to trigger the
2143 * scan function during the second round of scanning. Hopefully this means
2144 * that we reclaimed enough memory that initiating metadata transactions won't
2145 * make things worse.
2147 #define XFS_INODEGC_SHRINKER_COUNT (1UL << DEF_PRIORITY)
2148 #define XFS_INODEGC_SHRINKER_BATCH ((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2150 static unsigned long
2151 xfs_inodegc_shrinker_count(
2152 struct shrinker *shrink,
2153 struct shrink_control *sc)
2155 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2156 m_inodegc_shrinker);
2157 struct xfs_inodegc *gc;
2160 if (!xfs_is_inodegc_enabled(mp))
2163 for_each_online_cpu(cpu) {
2164 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2165 if (!llist_empty(&gc->list))
2166 return XFS_INODEGC_SHRINKER_COUNT;
2172 static unsigned long
2173 xfs_inodegc_shrinker_scan(
2174 struct shrinker *shrink,
2175 struct shrink_control *sc)
2177 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2178 m_inodegc_shrinker);
2179 struct xfs_inodegc *gc;
2181 bool no_items = true;
2183 if (!xfs_is_inodegc_enabled(mp))
2186 trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2188 for_each_online_cpu(cpu) {
2189 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2190 if (!llist_empty(&gc->list)) {
2191 unsigned int h = READ_ONCE(gc->shrinker_hits);
2193 WRITE_ONCE(gc->shrinker_hits, h + 1);
2194 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
2200 * If there are no inodes to inactivate, we don't want the shrinker
2201 * to think there's deferred work to call us back about.
2209 /* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2211 xfs_inodegc_register_shrinker(
2212 struct xfs_mount *mp)
2214 struct shrinker *shrink = &mp->m_inodegc_shrinker;
2216 shrink->count_objects = xfs_inodegc_shrinker_count;
2217 shrink->scan_objects = xfs_inodegc_shrinker_scan;
2219 shrink->flags = SHRINKER_NONSLAB;
2220 shrink->batch = XFS_INODEGC_SHRINKER_BATCH;
2222 return register_shrinker(shrink);