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"
27 #include <linux/iversion.h>
29 /* Radix tree tags for incore inode tree. */
31 /* inode is to be reclaimed */
32 #define XFS_ICI_RECLAIM_TAG 0
33 /* Inode has speculative preallocations (posteof or cow) to clean. */
34 #define XFS_ICI_BLOCKGC_TAG 1
37 * The goal for walking incore inodes. These can correspond with incore inode
38 * radix tree tags when convenient. Avoid existing XFS_IWALK namespace.
40 enum xfs_icwalk_goal {
41 /* Goals directly associated with tagged inodes. */
42 XFS_ICWALK_BLOCKGC = XFS_ICI_BLOCKGC_TAG,
43 XFS_ICWALK_RECLAIM = XFS_ICI_RECLAIM_TAG,
46 static int xfs_icwalk(struct xfs_mount *mp,
47 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
48 static int xfs_icwalk_ag(struct xfs_perag *pag,
49 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
52 * Private inode cache walk flags for struct xfs_icwalk. Must not
53 * coincide with XFS_ICWALK_FLAGS_VALID.
56 /* Stop scanning after icw_scan_limit inodes. */
57 #define XFS_ICWALK_FLAG_SCAN_LIMIT (1U << 28)
59 #define XFS_ICWALK_FLAG_RECLAIM_SICK (1U << 27)
60 #define XFS_ICWALK_FLAG_UNION (1U << 26) /* union filter algorithm */
62 #define XFS_ICWALK_PRIVATE_FLAGS (XFS_ICWALK_FLAG_SCAN_LIMIT | \
63 XFS_ICWALK_FLAG_RECLAIM_SICK | \
64 XFS_ICWALK_FLAG_UNION)
67 * Allocate and initialise an xfs_inode.
77 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL
78 * and return NULL here on ENOMEM.
80 ip = kmem_cache_alloc(xfs_inode_cache, GFP_KERNEL | __GFP_NOFAIL);
82 if (inode_init_always(mp->m_super, VFS_I(ip))) {
83 kmem_cache_free(xfs_inode_cache, ip);
87 /* VFS doesn't initialise i_mode or i_state! */
88 VFS_I(ip)->i_mode = 0;
89 VFS_I(ip)->i_state = 0;
90 mapping_set_large_folios(VFS_I(ip)->i_mapping);
92 XFS_STATS_INC(mp, vn_active);
93 ASSERT(atomic_read(&ip->i_pincount) == 0);
94 ASSERT(ip->i_ino == 0);
96 /* initialise the xfs inode */
99 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
102 memset(&ip->i_df, 0, sizeof(ip->i_df));
104 ip->i_delayed_blks = 0;
105 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
110 INIT_WORK(&ip->i_ioend_work, xfs_end_io);
111 INIT_LIST_HEAD(&ip->i_ioend_list);
112 spin_lock_init(&ip->i_ioend_lock);
118 xfs_inode_free_callback(
119 struct rcu_head *head)
121 struct inode *inode = container_of(head, struct inode, i_rcu);
122 struct xfs_inode *ip = XFS_I(inode);
124 switch (VFS_I(ip)->i_mode & S_IFMT) {
128 xfs_idestroy_fork(&ip->i_df);
133 xfs_idestroy_fork(ip->i_afp);
134 kmem_cache_free(xfs_ifork_cache, ip->i_afp);
137 xfs_idestroy_fork(ip->i_cowfp);
138 kmem_cache_free(xfs_ifork_cache, ip->i_cowfp);
141 ASSERT(!test_bit(XFS_LI_IN_AIL,
142 &ip->i_itemp->ili_item.li_flags));
143 xfs_inode_item_destroy(ip);
147 kmem_cache_free(xfs_inode_cache, ip);
152 struct xfs_inode *ip)
154 /* asserts to verify all state is correct here */
155 ASSERT(atomic_read(&ip->i_pincount) == 0);
156 ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
157 XFS_STATS_DEC(ip->i_mount, vn_active);
159 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
164 struct xfs_inode *ip)
166 ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
169 * Because we use RCU freeing we need to ensure the inode always
170 * appears to be reclaimed with an invalid inode number when in the
171 * free state. The ip->i_flags_lock provides the barrier against lookup
174 spin_lock(&ip->i_flags_lock);
175 ip->i_flags = XFS_IRECLAIM;
177 spin_unlock(&ip->i_flags_lock);
179 __xfs_inode_free(ip);
183 * Queue background inode reclaim work if there are reclaimable inodes and there
184 * isn't reclaim work already scheduled or in progress.
187 xfs_reclaim_work_queue(
188 struct xfs_mount *mp)
192 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
193 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
194 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
200 * Background scanning to trim preallocated space. This is queued based on the
201 * 'speculative_prealloc_lifetime' tunable (5m by default).
205 struct xfs_perag *pag)
207 struct xfs_mount *mp = pag->pag_mount;
209 if (!xfs_is_blockgc_enabled(mp))
213 if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
214 queue_delayed_work(pag->pag_mount->m_blockgc_wq,
215 &pag->pag_blockgc_work,
216 msecs_to_jiffies(xfs_blockgc_secs * 1000));
220 /* Set a tag on both the AG incore inode tree and the AG radix tree. */
222 xfs_perag_set_inode_tag(
223 struct xfs_perag *pag,
227 struct xfs_mount *mp = pag->pag_mount;
230 lockdep_assert_held(&pag->pag_ici_lock);
232 was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
233 radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
235 if (tag == XFS_ICI_RECLAIM_TAG)
236 pag->pag_ici_reclaimable++;
241 /* propagate the tag up into the perag radix tree */
242 spin_lock(&mp->m_perag_lock);
243 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno, tag);
244 spin_unlock(&mp->m_perag_lock);
246 /* start background work */
248 case XFS_ICI_RECLAIM_TAG:
249 xfs_reclaim_work_queue(mp);
251 case XFS_ICI_BLOCKGC_TAG:
252 xfs_blockgc_queue(pag);
256 trace_xfs_perag_set_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
259 /* Clear a tag on both the AG incore inode tree and the AG radix tree. */
261 xfs_perag_clear_inode_tag(
262 struct xfs_perag *pag,
266 struct xfs_mount *mp = pag->pag_mount;
268 lockdep_assert_held(&pag->pag_ici_lock);
271 * Reclaim can signal (with a null agino) that it cleared its own tag
272 * by removing the inode from the radix tree.
274 if (agino != NULLAGINO)
275 radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
277 ASSERT(tag == XFS_ICI_RECLAIM_TAG);
279 if (tag == XFS_ICI_RECLAIM_TAG)
280 pag->pag_ici_reclaimable--;
282 if (radix_tree_tagged(&pag->pag_ici_root, tag))
285 /* clear the tag from the perag radix tree */
286 spin_lock(&mp->m_perag_lock);
287 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno, tag);
288 spin_unlock(&mp->m_perag_lock);
290 trace_xfs_perag_clear_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
294 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
295 * part of the structure. This is made more complex by the fact we store
296 * information about the on-disk values in the VFS inode and so we can't just
297 * overwrite the values unconditionally. Hence we save the parameters we
298 * need to retain across reinitialisation, and rewrite them into the VFS inode
299 * after reinitialisation even if it fails.
303 struct xfs_mount *mp,
307 uint32_t nlink = inode->i_nlink;
308 uint32_t generation = inode->i_generation;
309 uint64_t version = inode_peek_iversion(inode);
310 umode_t mode = inode->i_mode;
311 dev_t dev = inode->i_rdev;
312 kuid_t uid = inode->i_uid;
313 kgid_t gid = inode->i_gid;
315 error = inode_init_always(mp->m_super, inode);
317 set_nlink(inode, nlink);
318 inode->i_generation = generation;
319 inode_set_iversion_queried(inode, version);
320 inode->i_mode = mode;
324 mapping_set_large_folios(inode->i_mapping);
329 * Carefully nudge an inode whose VFS state has been torn down back into a
330 * usable state. Drops the i_flags_lock and the rcu read lock.
334 struct xfs_perag *pag,
335 struct xfs_inode *ip) __releases(&ip->i_flags_lock)
337 struct xfs_mount *mp = ip->i_mount;
338 struct inode *inode = VFS_I(ip);
341 trace_xfs_iget_recycle(ip);
344 * We need to make it look like the inode is being reclaimed to prevent
345 * the actual reclaim workers from stomping over us while we recycle
346 * the inode. We can't clear the radix tree tag yet as it requires
347 * pag_ici_lock to be held exclusive.
349 ip->i_flags |= XFS_IRECLAIM;
351 spin_unlock(&ip->i_flags_lock);
354 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
355 error = xfs_reinit_inode(mp, inode);
358 * Re-initializing the inode failed, and we are in deep
359 * trouble. Try to re-add it to the reclaim list.
362 spin_lock(&ip->i_flags_lock);
363 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
364 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
365 spin_unlock(&ip->i_flags_lock);
368 trace_xfs_iget_recycle_fail(ip);
372 spin_lock(&pag->pag_ici_lock);
373 spin_lock(&ip->i_flags_lock);
376 * Clear the per-lifetime state in the inode as we are now effectively
377 * a new inode and need to return to the initial state before reuse
380 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
381 ip->i_flags |= XFS_INEW;
382 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
383 XFS_ICI_RECLAIM_TAG);
384 inode->i_state = I_NEW;
385 spin_unlock(&ip->i_flags_lock);
386 spin_unlock(&pag->pag_ici_lock);
392 * If we are allocating a new inode, then check what was returned is
393 * actually a free, empty inode. If we are not allocating an inode,
394 * then check we didn't find a free inode.
397 * 0 if the inode free state matches the lookup context
398 * -ENOENT if the inode is free and we are not allocating
399 * -EFSCORRUPTED if there is any state mismatch at all
402 xfs_iget_check_free_state(
403 struct xfs_inode *ip,
406 if (flags & XFS_IGET_CREATE) {
407 /* should be a free inode */
408 if (VFS_I(ip)->i_mode != 0) {
409 xfs_warn(ip->i_mount,
410 "Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
411 ip->i_ino, VFS_I(ip)->i_mode);
412 return -EFSCORRUPTED;
415 if (ip->i_nblocks != 0) {
416 xfs_warn(ip->i_mount,
417 "Corruption detected! Free inode 0x%llx has blocks allocated!",
419 return -EFSCORRUPTED;
424 /* should be an allocated inode */
425 if (VFS_I(ip)->i_mode == 0)
431 /* Make all pending inactivation work start immediately. */
433 xfs_inodegc_queue_all(
434 struct xfs_mount *mp)
436 struct xfs_inodegc *gc;
439 for_each_online_cpu(cpu) {
440 gc = per_cpu_ptr(mp->m_inodegc, cpu);
441 if (!llist_empty(&gc->list))
442 queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
447 * Check the validity of the inode we just found it the cache
451 struct xfs_perag *pag,
452 struct xfs_inode *ip,
455 int lock_flags) __releases(RCU)
457 struct inode *inode = VFS_I(ip);
458 struct xfs_mount *mp = ip->i_mount;
462 * check for re-use of an inode within an RCU grace period due to the
463 * radix tree nodes not being updated yet. We monitor for this by
464 * setting the inode number to zero before freeing the inode structure.
465 * If the inode has been reallocated and set up, then the inode number
466 * will not match, so check for that, too.
468 spin_lock(&ip->i_flags_lock);
469 if (ip->i_ino != ino)
473 * If we are racing with another cache hit that is currently
474 * instantiating this inode or currently recycling it out of
475 * reclaimable state, wait for the initialisation to complete
478 * If we're racing with the inactivation worker we also want to wait.
479 * If we're creating a new file, it's possible that the worker
480 * previously marked the inode as free on disk but hasn't finished
481 * updating the incore state yet. The AGI buffer will be dirty and
482 * locked to the icreate transaction, so a synchronous push of the
483 * inodegc workers would result in deadlock. For a regular iget, the
484 * worker is running already, so we might as well wait.
486 * XXX(hch): eventually we should do something equivalent to
487 * wait_on_inode to wait for these flags to be cleared
488 * instead of polling for it.
490 if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
493 if (ip->i_flags & XFS_NEED_INACTIVE) {
494 /* Unlinked inodes cannot be re-grabbed. */
495 if (VFS_I(ip)->i_nlink == 0) {
499 goto out_inodegc_flush;
503 * Check the inode free state is valid. This also detects lookup
504 * racing with unlinks.
506 error = xfs_iget_check_free_state(ip, flags);
510 /* Skip inodes that have no vfs state. */
511 if ((flags & XFS_IGET_INCORE) &&
512 (ip->i_flags & XFS_IRECLAIMABLE))
515 /* The inode fits the selection criteria; process it. */
516 if (ip->i_flags & XFS_IRECLAIMABLE) {
517 /* Drops i_flags_lock and RCU read lock. */
518 error = xfs_iget_recycle(pag, ip);
522 /* If the VFS inode is being torn down, pause and try again. */
526 /* We've got a live one. */
527 spin_unlock(&ip->i_flags_lock);
529 trace_xfs_iget_hit(ip);
533 xfs_ilock(ip, lock_flags);
535 if (!(flags & XFS_IGET_INCORE))
536 xfs_iflags_clear(ip, XFS_ISTALE);
537 XFS_STATS_INC(mp, xs_ig_found);
542 trace_xfs_iget_skip(ip);
543 XFS_STATS_INC(mp, xs_ig_frecycle);
546 spin_unlock(&ip->i_flags_lock);
551 spin_unlock(&ip->i_flags_lock);
554 * Do not wait for the workers, because the caller could hold an AGI
555 * buffer lock. We're just going to sleep in a loop anyway.
557 if (xfs_is_inodegc_enabled(mp))
558 xfs_inodegc_queue_all(mp);
564 struct xfs_mount *mp,
565 struct xfs_perag *pag,
568 struct xfs_inode **ipp,
572 struct xfs_inode *ip;
574 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
577 ip = xfs_inode_alloc(mp, ino);
581 error = xfs_imap(mp, tp, ip->i_ino, &ip->i_imap, flags);
586 * For version 5 superblocks, if we are initialising a new inode and we
587 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can
588 * simply build the new inode core with a random generation number.
590 * For version 4 (and older) superblocks, log recovery is dependent on
591 * the i_flushiter field being initialised from the current on-disk
592 * value and hence we must also read the inode off disk even when
593 * initializing new inodes.
595 if (xfs_has_v3inodes(mp) &&
596 (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
597 VFS_I(ip)->i_generation = prandom_u32();
601 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
605 error = xfs_inode_from_disk(ip,
606 xfs_buf_offset(bp, ip->i_imap.im_boffset));
608 xfs_buf_set_ref(bp, XFS_INO_REF);
609 xfs_trans_brelse(tp, bp);
615 trace_xfs_iget_miss(ip);
618 * Check the inode free state is valid. This also detects lookup
619 * racing with unlinks.
621 error = xfs_iget_check_free_state(ip, flags);
626 * Preload the radix tree so we can insert safely under the
627 * write spinlock. Note that we cannot sleep inside the preload
628 * region. Since we can be called from transaction context, don't
629 * recurse into the file system.
631 if (radix_tree_preload(GFP_NOFS)) {
637 * Because the inode hasn't been added to the radix-tree yet it can't
638 * be found by another thread, so we can do the non-sleeping lock here.
641 if (!xfs_ilock_nowait(ip, lock_flags))
646 * These values must be set before inserting the inode into the radix
647 * tree as the moment it is inserted a concurrent lookup (allowed by the
648 * RCU locking mechanism) can find it and that lookup must see that this
649 * is an inode currently under construction (i.e. that XFS_INEW is set).
650 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
651 * memory barrier that ensures this detection works correctly at lookup
655 if (flags & XFS_IGET_DONTCACHE)
656 d_mark_dontcache(VFS_I(ip));
660 xfs_iflags_set(ip, iflags);
662 /* insert the new inode */
663 spin_lock(&pag->pag_ici_lock);
664 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
665 if (unlikely(error)) {
666 WARN_ON(error != -EEXIST);
667 XFS_STATS_INC(mp, xs_ig_dup);
669 goto out_preload_end;
671 spin_unlock(&pag->pag_ici_lock);
672 radix_tree_preload_end();
678 spin_unlock(&pag->pag_ici_lock);
679 radix_tree_preload_end();
681 xfs_iunlock(ip, lock_flags);
683 __destroy_inode(VFS_I(ip));
689 * Look up an inode by number in the given file system. The inode is looked up
690 * in the cache held in each AG. If the inode is found in the cache, initialise
691 * the vfs inode if necessary.
693 * If it is not in core, read it in from the file system's device, add it to the
694 * cache and initialise the vfs inode.
696 * The inode is locked according to the value of the lock_flags parameter.
697 * Inode lookup is only done during metadata operations and not as part of the
698 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
702 struct xfs_mount *mp,
703 struct xfs_trans *tp,
707 struct xfs_inode **ipp)
709 struct xfs_inode *ip;
710 struct xfs_perag *pag;
714 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
716 /* reject inode numbers outside existing AGs */
717 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
720 XFS_STATS_INC(mp, xs_ig_attempts);
722 /* get the perag structure and ensure that it's inode capable */
723 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
724 agino = XFS_INO_TO_AGINO(mp, ino);
729 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
732 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
734 goto out_error_or_again;
737 if (flags & XFS_IGET_INCORE) {
739 goto out_error_or_again;
741 XFS_STATS_INC(mp, xs_ig_missed);
743 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
746 goto out_error_or_again;
753 * If we have a real type for an on-disk inode, we can setup the inode
754 * now. If it's a new inode being created, xfs_init_new_inode will
757 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
758 xfs_setup_existing_inode(ip);
762 if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
771 * "Is this a cached inode that's also allocated?"
773 * Look up an inode by number in the given file system. If the inode is
774 * in cache and isn't in purgatory, return 1 if the inode is allocated
775 * and 0 if it is not. For all other cases (not in cache, being torn
776 * down, etc.), return a negative error code.
778 * The caller has to prevent inode allocation and freeing activity,
779 * presumably by locking the AGI buffer. This is to ensure that an
780 * inode cannot transition from allocated to freed until the caller is
781 * ready to allow that. If the inode is in an intermediate state (new,
782 * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
783 * inode is not in the cache, -ENOENT will be returned. The caller must
784 * deal with these scenarios appropriately.
786 * This is a specialized use case for the online scrubber; if you're
787 * reading this, you probably want xfs_iget.
790 xfs_icache_inode_is_allocated(
791 struct xfs_mount *mp,
792 struct xfs_trans *tp,
796 struct xfs_inode *ip;
799 error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
803 *inuse = !!(VFS_I(ip)->i_mode);
809 * Grab the inode for reclaim exclusively.
811 * We have found this inode via a lookup under RCU, so the inode may have
812 * already been freed, or it may be in the process of being recycled by
813 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
814 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
815 * will not be set. Hence we need to check for both these flag conditions to
816 * avoid inodes that are no longer reclaim candidates.
818 * Note: checking for other state flags here, under the i_flags_lock or not, is
819 * racy and should be avoided. Those races should be resolved only after we have
820 * ensured that we are able to reclaim this inode and the world can see that we
821 * are going to reclaim it.
823 * Return true if we grabbed it, false otherwise.
827 struct xfs_inode *ip,
828 struct xfs_icwalk *icw)
830 ASSERT(rcu_read_lock_held());
832 spin_lock(&ip->i_flags_lock);
833 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
834 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
835 /* not a reclaim candidate. */
836 spin_unlock(&ip->i_flags_lock);
840 /* Don't reclaim a sick inode unless the caller asked for it. */
842 (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
843 spin_unlock(&ip->i_flags_lock);
847 __xfs_iflags_set(ip, XFS_IRECLAIM);
848 spin_unlock(&ip->i_flags_lock);
853 * Inode reclaim is non-blocking, so the default action if progress cannot be
854 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
855 * XFS_IRECLAIM flag. If we are in a shutdown state, we don't care about
856 * blocking anymore and hence we can wait for the inode to be able to reclaim
859 * We do no IO here - if callers require inodes to be cleaned they must push the
860 * AIL first to trigger writeback of dirty inodes. This enables writeback to be
861 * done in the background in a non-blocking manner, and enables memory reclaim
862 * to make progress without blocking.
866 struct xfs_inode *ip,
867 struct xfs_perag *pag)
869 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
871 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
873 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
876 if (xfs_is_shutdown(ip->i_mount)) {
878 xfs_iflush_abort(ip);
881 if (xfs_ipincount(ip))
882 goto out_clear_flush;
883 if (!xfs_inode_clean(ip))
884 goto out_clear_flush;
886 xfs_iflags_clear(ip, XFS_IFLUSHING);
888 trace_xfs_inode_reclaiming(ip);
891 * Because we use RCU freeing we need to ensure the inode always appears
892 * to be reclaimed with an invalid inode number when in the free state.
893 * We do this as early as possible under the ILOCK so that
894 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
895 * detect races with us here. By doing this, we guarantee that once
896 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
897 * it will see either a valid inode that will serialise correctly, or it
898 * will see an invalid inode that it can skip.
900 spin_lock(&ip->i_flags_lock);
901 ip->i_flags = XFS_IRECLAIM;
905 spin_unlock(&ip->i_flags_lock);
907 xfs_iunlock(ip, XFS_ILOCK_EXCL);
909 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
911 * Remove the inode from the per-AG radix tree.
913 * Because radix_tree_delete won't complain even if the item was never
914 * added to the tree assert that it's been there before to catch
915 * problems with the inode life time early on.
917 spin_lock(&pag->pag_ici_lock);
918 if (!radix_tree_delete(&pag->pag_ici_root,
919 XFS_INO_TO_AGINO(ip->i_mount, ino)))
921 xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
922 spin_unlock(&pag->pag_ici_lock);
925 * Here we do an (almost) spurious inode lock in order to coordinate
926 * with inode cache radix tree lookups. This is because the lookup
927 * can reference the inodes in the cache without taking references.
929 * We make that OK here by ensuring that we wait until the inode is
930 * unlocked after the lookup before we go ahead and free it.
932 xfs_ilock(ip, XFS_ILOCK_EXCL);
933 ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
934 xfs_iunlock(ip, XFS_ILOCK_EXCL);
935 ASSERT(xfs_inode_clean(ip));
937 __xfs_inode_free(ip);
941 xfs_iflags_clear(ip, XFS_IFLUSHING);
943 xfs_iunlock(ip, XFS_ILOCK_EXCL);
945 xfs_iflags_clear(ip, XFS_IRECLAIM);
948 /* Reclaim sick inodes if we're unmounting or the fs went down. */
950 xfs_want_reclaim_sick(
951 struct xfs_mount *mp)
953 return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
959 struct xfs_mount *mp)
961 struct xfs_icwalk icw = {
965 if (xfs_want_reclaim_sick(mp))
966 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
968 while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
969 xfs_ail_push_all_sync(mp->m_ail);
970 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
975 * The shrinker infrastructure determines how many inodes we should scan for
976 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
977 * push the AIL here. We also want to proactively free up memory if we can to
978 * minimise the amount of work memory reclaim has to do so we kick the
979 * background reclaim if it isn't already scheduled.
982 xfs_reclaim_inodes_nr(
983 struct xfs_mount *mp,
984 unsigned long nr_to_scan)
986 struct xfs_icwalk icw = {
987 .icw_flags = XFS_ICWALK_FLAG_SCAN_LIMIT,
988 .icw_scan_limit = min_t(unsigned long, LONG_MAX, nr_to_scan),
991 if (xfs_want_reclaim_sick(mp))
992 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
994 /* kick background reclaimer and push the AIL */
995 xfs_reclaim_work_queue(mp);
996 xfs_ail_push_all(mp->m_ail);
998 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1003 * Return the number of reclaimable inodes in the filesystem for
1004 * the shrinker to determine how much to reclaim.
1007 xfs_reclaim_inodes_count(
1008 struct xfs_mount *mp)
1010 struct xfs_perag *pag;
1011 xfs_agnumber_t ag = 0;
1012 long reclaimable = 0;
1014 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1015 ag = pag->pag_agno + 1;
1016 reclaimable += pag->pag_ici_reclaimable;
1023 xfs_icwalk_match_id(
1024 struct xfs_inode *ip,
1025 struct xfs_icwalk *icw)
1027 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1028 !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1031 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1032 !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1035 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1036 ip->i_projid != icw->icw_prid)
1043 * A union-based inode filtering algorithm. Process the inode if any of the
1044 * criteria match. This is for global/internal scans only.
1047 xfs_icwalk_match_id_union(
1048 struct xfs_inode *ip,
1049 struct xfs_icwalk *icw)
1051 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1052 uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1055 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1056 gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1059 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1060 ip->i_projid == icw->icw_prid)
1067 * Is this inode @ip eligible for eof/cow block reclamation, given some
1068 * filtering parameters @icw? The inode is eligible if @icw is null or
1069 * if the predicate functions match.
1073 struct xfs_inode *ip,
1074 struct xfs_icwalk *icw)
1081 if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1082 match = xfs_icwalk_match_id_union(ip, icw);
1084 match = xfs_icwalk_match_id(ip, icw);
1088 /* skip the inode if the file size is too small */
1089 if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1090 XFS_ISIZE(ip) < icw->icw_min_file_size)
1097 * This is a fast pass over the inode cache to try to get reclaim moving on as
1098 * many inodes as possible in a short period of time. It kicks itself every few
1099 * seconds, as well as being kicked by the inode cache shrinker when memory
1104 struct work_struct *work)
1106 struct xfs_mount *mp = container_of(to_delayed_work(work),
1107 struct xfs_mount, m_reclaim_work);
1109 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1110 xfs_reclaim_work_queue(mp);
1114 xfs_inode_free_eofblocks(
1115 struct xfs_inode *ip,
1116 struct xfs_icwalk *icw,
1117 unsigned int *lockflags)
1121 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1123 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1127 * If the mapping is dirty the operation can block and wait for some
1128 * time. Unless we are waiting, skip it.
1130 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1133 if (!xfs_icwalk_match(ip, icw))
1137 * If the caller is waiting, return -EAGAIN to keep the background
1138 * scanner moving and revisit the inode in a subsequent pass.
1140 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1145 *lockflags |= XFS_IOLOCK_EXCL;
1147 if (xfs_can_free_eofblocks(ip, false))
1148 return xfs_free_eofblocks(ip);
1150 /* inode could be preallocated or append-only */
1151 trace_xfs_inode_free_eofblocks_invalid(ip);
1152 xfs_inode_clear_eofblocks_tag(ip);
1157 xfs_blockgc_set_iflag(
1158 struct xfs_inode *ip,
1159 unsigned long iflag)
1161 struct xfs_mount *mp = ip->i_mount;
1162 struct xfs_perag *pag;
1164 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1167 * Don't bother locking the AG and looking up in the radix trees
1168 * if we already know that we have the tag set.
1170 if (ip->i_flags & iflag)
1172 spin_lock(&ip->i_flags_lock);
1173 ip->i_flags |= iflag;
1174 spin_unlock(&ip->i_flags_lock);
1176 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1177 spin_lock(&pag->pag_ici_lock);
1179 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1180 XFS_ICI_BLOCKGC_TAG);
1182 spin_unlock(&pag->pag_ici_lock);
1187 xfs_inode_set_eofblocks_tag(
1190 trace_xfs_inode_set_eofblocks_tag(ip);
1191 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1195 xfs_blockgc_clear_iflag(
1196 struct xfs_inode *ip,
1197 unsigned long iflag)
1199 struct xfs_mount *mp = ip->i_mount;
1200 struct xfs_perag *pag;
1203 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1205 spin_lock(&ip->i_flags_lock);
1206 ip->i_flags &= ~iflag;
1207 clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1208 spin_unlock(&ip->i_flags_lock);
1213 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1214 spin_lock(&pag->pag_ici_lock);
1216 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1217 XFS_ICI_BLOCKGC_TAG);
1219 spin_unlock(&pag->pag_ici_lock);
1224 xfs_inode_clear_eofblocks_tag(
1227 trace_xfs_inode_clear_eofblocks_tag(ip);
1228 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1232 * Set ourselves up to free CoW blocks from this file. If it's already clean
1233 * then we can bail out quickly, but otherwise we must back off if the file
1234 * is undergoing some kind of write.
1237 xfs_prep_free_cowblocks(
1238 struct xfs_inode *ip)
1241 * Just clear the tag if we have an empty cow fork or none at all. It's
1242 * possible the inode was fully unshared since it was originally tagged.
1244 if (!xfs_inode_has_cow_data(ip)) {
1245 trace_xfs_inode_free_cowblocks_invalid(ip);
1246 xfs_inode_clear_cowblocks_tag(ip);
1251 * If the mapping is dirty or under writeback we cannot touch the
1252 * CoW fork. Leave it alone if we're in the midst of a directio.
1254 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1255 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1256 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1257 atomic_read(&VFS_I(ip)->i_dio_count))
1264 * Automatic CoW Reservation Freeing
1266 * These functions automatically garbage collect leftover CoW reservations
1267 * that were made on behalf of a cowextsize hint when we start to run out
1268 * of quota or when the reservations sit around for too long. If the file
1269 * has dirty pages or is undergoing writeback, its CoW reservations will
1272 * The actual garbage collection piggybacks off the same code that runs
1273 * the speculative EOF preallocation garbage collector.
1276 xfs_inode_free_cowblocks(
1277 struct xfs_inode *ip,
1278 struct xfs_icwalk *icw,
1279 unsigned int *lockflags)
1284 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1286 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1289 if (!xfs_prep_free_cowblocks(ip))
1292 if (!xfs_icwalk_match(ip, icw))
1296 * If the caller is waiting, return -EAGAIN to keep the background
1297 * scanner moving and revisit the inode in a subsequent pass.
1299 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1300 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1305 *lockflags |= XFS_IOLOCK_EXCL;
1307 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1312 *lockflags |= XFS_MMAPLOCK_EXCL;
1315 * Check again, nobody else should be able to dirty blocks or change
1316 * the reflink iflag now that we have the first two locks held.
1318 if (xfs_prep_free_cowblocks(ip))
1319 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1324 xfs_inode_set_cowblocks_tag(
1327 trace_xfs_inode_set_cowblocks_tag(ip);
1328 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1332 xfs_inode_clear_cowblocks_tag(
1335 trace_xfs_inode_clear_cowblocks_tag(ip);
1336 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1339 /* Disable post-EOF and CoW block auto-reclamation. */
1342 struct xfs_mount *mp)
1344 struct xfs_perag *pag;
1345 xfs_agnumber_t agno;
1347 if (!xfs_clear_blockgc_enabled(mp))
1350 for_each_perag(mp, agno, pag)
1351 cancel_delayed_work_sync(&pag->pag_blockgc_work);
1352 trace_xfs_blockgc_stop(mp, __return_address);
1355 /* Enable post-EOF and CoW block auto-reclamation. */
1358 struct xfs_mount *mp)
1360 struct xfs_perag *pag;
1361 xfs_agnumber_t agno;
1363 if (xfs_set_blockgc_enabled(mp))
1366 trace_xfs_blockgc_start(mp, __return_address);
1367 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1368 xfs_blockgc_queue(pag);
1371 /* Don't try to run block gc on an inode that's in any of these states. */
1372 #define XFS_BLOCKGC_NOGRAB_IFLAGS (XFS_INEW | \
1373 XFS_NEED_INACTIVE | \
1374 XFS_INACTIVATING | \
1375 XFS_IRECLAIMABLE | \
1378 * Decide if the given @ip is eligible for garbage collection of speculative
1379 * preallocations, and grab it if so. Returns true if it's ready to go or
1380 * false if we should just ignore it.
1384 struct xfs_inode *ip)
1386 struct inode *inode = VFS_I(ip);
1388 ASSERT(rcu_read_lock_held());
1390 /* Check for stale RCU freed inode */
1391 spin_lock(&ip->i_flags_lock);
1393 goto out_unlock_noent;
1395 if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1396 goto out_unlock_noent;
1397 spin_unlock(&ip->i_flags_lock);
1399 /* nothing to sync during shutdown */
1400 if (xfs_is_shutdown(ip->i_mount))
1403 /* If we can't grab the inode, it must on it's way to reclaim. */
1407 /* inode is valid */
1411 spin_unlock(&ip->i_flags_lock);
1415 /* Scan one incore inode for block preallocations that we can remove. */
1417 xfs_blockgc_scan_inode(
1418 struct xfs_inode *ip,
1419 struct xfs_icwalk *icw)
1421 unsigned int lockflags = 0;
1424 error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1428 error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1431 xfs_iunlock(ip, lockflags);
1436 /* Background worker that trims preallocated space. */
1439 struct work_struct *work)
1441 struct xfs_perag *pag = container_of(to_delayed_work(work),
1442 struct xfs_perag, pag_blockgc_work);
1443 struct xfs_mount *mp = pag->pag_mount;
1446 trace_xfs_blockgc_worker(mp, __return_address);
1448 error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1450 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1451 pag->pag_agno, error);
1452 xfs_blockgc_queue(pag);
1456 * Try to free space in the filesystem by purging inactive inodes, eofblocks
1460 xfs_blockgc_free_space(
1461 struct xfs_mount *mp,
1462 struct xfs_icwalk *icw)
1466 trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1468 error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1472 xfs_inodegc_flush(mp);
1477 * Reclaim all the free space that we can by scheduling the background blockgc
1478 * and inodegc workers immediately and waiting for them all to clear.
1481 xfs_blockgc_flush_all(
1482 struct xfs_mount *mp)
1484 struct xfs_perag *pag;
1485 xfs_agnumber_t agno;
1487 trace_xfs_blockgc_flush_all(mp, __return_address);
1490 * For each blockgc worker, move its queue time up to now. If it
1491 * wasn't queued, it will not be requeued. Then flush whatever's
1494 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1495 mod_delayed_work(pag->pag_mount->m_blockgc_wq,
1496 &pag->pag_blockgc_work, 0);
1498 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1499 flush_delayed_work(&pag->pag_blockgc_work);
1501 xfs_inodegc_flush(mp);
1505 * Run cow/eofblocks scans on the supplied dquots. We don't know exactly which
1506 * quota caused an allocation failure, so we make a best effort by including
1507 * each quota under low free space conditions (less than 1% free space) in the
1510 * Callers must not hold any inode's ILOCK. If requesting a synchronous scan
1511 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1515 xfs_blockgc_free_dquots(
1516 struct xfs_mount *mp,
1517 struct xfs_dquot *udqp,
1518 struct xfs_dquot *gdqp,
1519 struct xfs_dquot *pdqp,
1520 unsigned int iwalk_flags)
1522 struct xfs_icwalk icw = {0};
1523 bool do_work = false;
1525 if (!udqp && !gdqp && !pdqp)
1529 * Run a scan to free blocks using the union filter to cover all
1530 * applicable quotas in a single scan.
1532 icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1534 if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1535 icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1536 icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1540 if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1541 icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1542 icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1546 if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1547 icw.icw_prid = pdqp->q_id;
1548 icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1555 return xfs_blockgc_free_space(mp, &icw);
1558 /* Run cow/eofblocks scans on the quotas attached to the inode. */
1560 xfs_blockgc_free_quota(
1561 struct xfs_inode *ip,
1562 unsigned int iwalk_flags)
1564 return xfs_blockgc_free_dquots(ip->i_mount,
1565 xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1566 xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1567 xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1570 /* XFS Inode Cache Walking Code */
1573 * The inode lookup is done in batches to keep the amount of lock traffic and
1574 * radix tree lookups to a minimum. The batch size is a trade off between
1575 * lookup reduction and stack usage. This is in the reclaim path, so we can't
1578 #define XFS_LOOKUP_BATCH 32
1582 * Decide if we want to grab this inode in anticipation of doing work towards
1587 enum xfs_icwalk_goal goal,
1588 struct xfs_inode *ip,
1589 struct xfs_icwalk *icw)
1592 case XFS_ICWALK_BLOCKGC:
1593 return xfs_blockgc_igrab(ip);
1594 case XFS_ICWALK_RECLAIM:
1595 return xfs_reclaim_igrab(ip, icw);
1602 * Process an inode. Each processing function must handle any state changes
1603 * made by the icwalk igrab function. Return -EAGAIN to skip an inode.
1606 xfs_icwalk_process_inode(
1607 enum xfs_icwalk_goal goal,
1608 struct xfs_inode *ip,
1609 struct xfs_perag *pag,
1610 struct xfs_icwalk *icw)
1615 case XFS_ICWALK_BLOCKGC:
1616 error = xfs_blockgc_scan_inode(ip, icw);
1618 case XFS_ICWALK_RECLAIM:
1619 xfs_reclaim_inode(ip, pag);
1626 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1627 * process them in some manner.
1631 struct xfs_perag *pag,
1632 enum xfs_icwalk_goal goal,
1633 struct xfs_icwalk *icw)
1635 struct xfs_mount *mp = pag->pag_mount;
1636 uint32_t first_index;
1645 if (goal == XFS_ICWALK_RECLAIM)
1646 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1651 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1657 nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1658 (void **) batch, first_index,
1659 XFS_LOOKUP_BATCH, goal);
1667 * Grab the inodes before we drop the lock. if we found
1668 * nothing, nr == 0 and the loop will be skipped.
1670 for (i = 0; i < nr_found; i++) {
1671 struct xfs_inode *ip = batch[i];
1673 if (done || !xfs_icwalk_igrab(goal, ip, icw))
1677 * Update the index for the next lookup. Catch
1678 * overflows into the next AG range which can occur if
1679 * we have inodes in the last block of the AG and we
1680 * are currently pointing to the last inode.
1682 * Because we may see inodes that are from the wrong AG
1683 * due to RCU freeing and reallocation, only update the
1684 * index if it lies in this AG. It was a race that lead
1685 * us to see this inode, so another lookup from the
1686 * same index will not find it again.
1688 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
1690 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1691 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1695 /* unlock now we've grabbed the inodes. */
1698 for (i = 0; i < nr_found; i++) {
1701 error = xfs_icwalk_process_inode(goal, batch[i], pag,
1703 if (error == -EAGAIN) {
1707 if (error && last_error != -EFSCORRUPTED)
1711 /* bail out if the filesystem is corrupted. */
1712 if (error == -EFSCORRUPTED)
1717 if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1718 icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1719 if (icw->icw_scan_limit <= 0)
1722 } while (nr_found && !done);
1724 if (goal == XFS_ICWALK_RECLAIM) {
1727 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1737 /* Walk all incore inodes to achieve a given goal. */
1740 struct xfs_mount *mp,
1741 enum xfs_icwalk_goal goal,
1742 struct xfs_icwalk *icw)
1744 struct xfs_perag *pag;
1747 xfs_agnumber_t agno;
1749 for_each_perag_tag(mp, agno, pag, goal) {
1750 error = xfs_icwalk_ag(pag, goal, icw);
1753 if (error == -EFSCORRUPTED) {
1760 BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1766 struct xfs_inode *ip,
1769 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, whichfork);
1770 struct xfs_bmbt_irec got;
1771 struct xfs_iext_cursor icur;
1773 if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1776 if (isnullstartblock(got.br_startblock)) {
1777 xfs_warn(ip->i_mount,
1778 "ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",
1780 whichfork == XFS_DATA_FORK ? "data" : "cow",
1781 got.br_startoff, got.br_blockcount);
1783 } while (xfs_iext_next_extent(ifp, &icur, &got));
1786 #define xfs_check_delalloc(ip, whichfork) do { } while (0)
1789 /* Schedule the inode for reclaim. */
1791 xfs_inodegc_set_reclaimable(
1792 struct xfs_inode *ip)
1794 struct xfs_mount *mp = ip->i_mount;
1795 struct xfs_perag *pag;
1797 if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {
1798 xfs_check_delalloc(ip, XFS_DATA_FORK);
1799 xfs_check_delalloc(ip, XFS_COW_FORK);
1803 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1804 spin_lock(&pag->pag_ici_lock);
1805 spin_lock(&ip->i_flags_lock);
1807 trace_xfs_inode_set_reclaimable(ip);
1808 ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);
1809 ip->i_flags |= XFS_IRECLAIMABLE;
1810 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1811 XFS_ICI_RECLAIM_TAG);
1813 spin_unlock(&ip->i_flags_lock);
1814 spin_unlock(&pag->pag_ici_lock);
1819 * Free all speculative preallocations and possibly even the inode itself.
1820 * This is the last chance to make changes to an otherwise unreferenced file
1821 * before incore reclamation happens.
1824 xfs_inodegc_inactivate(
1825 struct xfs_inode *ip)
1827 trace_xfs_inode_inactivating(ip);
1829 xfs_inodegc_set_reclaimable(ip);
1834 struct work_struct *work)
1836 struct xfs_inodegc *gc = container_of(work, struct xfs_inodegc,
1838 struct llist_node *node = llist_del_all(&gc->list);
1839 struct xfs_inode *ip, *n;
1841 WRITE_ONCE(gc->items, 0);
1846 ip = llist_entry(node, struct xfs_inode, i_gclist);
1847 trace_xfs_inodegc_worker(ip->i_mount, READ_ONCE(gc->shrinker_hits));
1849 WRITE_ONCE(gc->shrinker_hits, 0);
1850 llist_for_each_entry_safe(ip, n, node, i_gclist) {
1851 xfs_iflags_set(ip, XFS_INACTIVATING);
1852 xfs_inodegc_inactivate(ip);
1857 * Force all currently queued inode inactivation work to run immediately, and
1858 * wait for the work to finish. Two pass - queue all the work first pass, wait
1859 * for it in a second pass.
1863 struct xfs_mount *mp)
1865 struct xfs_inodegc *gc;
1868 if (!xfs_is_inodegc_enabled(mp))
1871 trace_xfs_inodegc_flush(mp, __return_address);
1873 xfs_inodegc_queue_all(mp);
1875 for_each_online_cpu(cpu) {
1876 gc = per_cpu_ptr(mp->m_inodegc, cpu);
1877 flush_work(&gc->work);
1882 * Flush all the pending work and then disable the inode inactivation background
1883 * workers and wait for them to stop.
1887 struct xfs_mount *mp)
1889 struct xfs_inodegc *gc;
1892 if (!xfs_clear_inodegc_enabled(mp))
1895 xfs_inodegc_queue_all(mp);
1897 for_each_online_cpu(cpu) {
1898 gc = per_cpu_ptr(mp->m_inodegc, cpu);
1899 cancel_work_sync(&gc->work);
1901 trace_xfs_inodegc_stop(mp, __return_address);
1905 * Enable the inode inactivation background workers and schedule deferred inode
1906 * inactivation work if there is any.
1910 struct xfs_mount *mp)
1912 if (xfs_set_inodegc_enabled(mp))
1915 trace_xfs_inodegc_start(mp, __return_address);
1916 xfs_inodegc_queue_all(mp);
1919 #ifdef CONFIG_XFS_RT
1921 xfs_inodegc_want_queue_rt_file(
1922 struct xfs_inode *ip)
1924 struct xfs_mount *mp = ip->i_mount;
1927 if (!XFS_IS_REALTIME_INODE(ip))
1930 freertx = READ_ONCE(mp->m_sb.sb_frextents);
1931 return freertx < mp->m_low_rtexts[XFS_LOWSP_5_PCNT];
1934 # define xfs_inodegc_want_queue_rt_file(ip) (false)
1935 #endif /* CONFIG_XFS_RT */
1938 * Schedule the inactivation worker when:
1940 * - We've accumulated more than one inode cluster buffer's worth of inodes.
1941 * - There is less than 5% free space left.
1942 * - Any of the quotas for this inode are near an enforcement limit.
1945 xfs_inodegc_want_queue_work(
1946 struct xfs_inode *ip,
1949 struct xfs_mount *mp = ip->i_mount;
1951 if (items > mp->m_ino_geo.inodes_per_cluster)
1954 if (__percpu_counter_compare(&mp->m_fdblocks,
1955 mp->m_low_space[XFS_LOWSP_5_PCNT],
1956 XFS_FDBLOCKS_BATCH) < 0)
1959 if (xfs_inodegc_want_queue_rt_file(ip))
1962 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
1965 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
1968 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
1975 * Upper bound on the number of inodes in each AG that can be queued for
1976 * inactivation at any given time, to avoid monopolizing the workqueue.
1978 #define XFS_INODEGC_MAX_BACKLOG (4 * XFS_INODES_PER_CHUNK)
1981 * Make the frontend wait for inactivations when:
1983 * - Memory shrinkers queued the inactivation worker and it hasn't finished.
1984 * - The queue depth exceeds the maximum allowable percpu backlog.
1986 * Note: If the current thread is running a transaction, we don't ever want to
1987 * wait for other transactions because that could introduce a deadlock.
1990 xfs_inodegc_want_flush_work(
1991 struct xfs_inode *ip,
1993 unsigned int shrinker_hits)
1995 if (current->journal_info)
1998 if (shrinker_hits > 0)
2001 if (items > XFS_INODEGC_MAX_BACKLOG)
2008 * Queue a background inactivation worker if there are inodes that need to be
2009 * inactivated and higher level xfs code hasn't disabled the background
2014 struct xfs_inode *ip)
2016 struct xfs_mount *mp = ip->i_mount;
2017 struct xfs_inodegc *gc;
2019 unsigned int shrinker_hits;
2021 trace_xfs_inode_set_need_inactive(ip);
2022 spin_lock(&ip->i_flags_lock);
2023 ip->i_flags |= XFS_NEED_INACTIVE;
2024 spin_unlock(&ip->i_flags_lock);
2026 gc = get_cpu_ptr(mp->m_inodegc);
2027 llist_add(&ip->i_gclist, &gc->list);
2028 items = READ_ONCE(gc->items);
2029 WRITE_ONCE(gc->items, items + 1);
2030 shrinker_hits = READ_ONCE(gc->shrinker_hits);
2033 if (!xfs_is_inodegc_enabled(mp))
2036 if (xfs_inodegc_want_queue_work(ip, items)) {
2037 trace_xfs_inodegc_queue(mp, __return_address);
2038 queue_work(mp->m_inodegc_wq, &gc->work);
2041 if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
2042 trace_xfs_inodegc_throttle(mp, __return_address);
2043 flush_work(&gc->work);
2048 * Fold the dead CPU inodegc queue into the current CPUs queue.
2051 xfs_inodegc_cpu_dead(
2052 struct xfs_mount *mp,
2053 unsigned int dead_cpu)
2055 struct xfs_inodegc *dead_gc, *gc;
2056 struct llist_node *first, *last;
2057 unsigned int count = 0;
2059 dead_gc = per_cpu_ptr(mp->m_inodegc, dead_cpu);
2060 cancel_work_sync(&dead_gc->work);
2062 if (llist_empty(&dead_gc->list))
2065 first = dead_gc->list.first;
2067 while (last->next) {
2071 dead_gc->list.first = NULL;
2074 /* Add pending work to current CPU */
2075 gc = get_cpu_ptr(mp->m_inodegc);
2076 llist_add_batch(first, last, &gc->list);
2077 count += READ_ONCE(gc->items);
2078 WRITE_ONCE(gc->items, count);
2081 if (xfs_is_inodegc_enabled(mp)) {
2082 trace_xfs_inodegc_queue(mp, __return_address);
2083 queue_work(mp->m_inodegc_wq, &gc->work);
2088 * We set the inode flag atomically with the radix tree tag. Once we get tag
2089 * lookups on the radix tree, this inode flag can go away.
2091 * We always use background reclaim here because even if the inode is clean, it
2092 * still may be under IO and hence we have wait for IO completion to occur
2093 * before we can reclaim the inode. The background reclaim path handles this
2094 * more efficiently than we can here, so simply let background reclaim tear down
2098 xfs_inode_mark_reclaimable(
2099 struct xfs_inode *ip)
2101 struct xfs_mount *mp = ip->i_mount;
2104 XFS_STATS_INC(mp, vn_reclaim);
2107 * We should never get here with any of the reclaim flags already set.
2109 ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2111 need_inactive = xfs_inode_needs_inactive(ip);
2112 if (need_inactive) {
2113 xfs_inodegc_queue(ip);
2117 /* Going straight to reclaim, so drop the dquots. */
2118 xfs_qm_dqdetach(ip);
2119 xfs_inodegc_set_reclaimable(ip);
2123 * Register a phony shrinker so that we can run background inodegc sooner when
2124 * there's memory pressure. Inactivation does not itself free any memory but
2125 * it does make inodes reclaimable, which eventually frees memory.
2127 * The count function, seek value, and batch value are crafted to trigger the
2128 * scan function during the second round of scanning. Hopefully this means
2129 * that we reclaimed enough memory that initiating metadata transactions won't
2130 * make things worse.
2132 #define XFS_INODEGC_SHRINKER_COUNT (1UL << DEF_PRIORITY)
2133 #define XFS_INODEGC_SHRINKER_BATCH ((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2135 static unsigned long
2136 xfs_inodegc_shrinker_count(
2137 struct shrinker *shrink,
2138 struct shrink_control *sc)
2140 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2141 m_inodegc_shrinker);
2142 struct xfs_inodegc *gc;
2145 if (!xfs_is_inodegc_enabled(mp))
2148 for_each_online_cpu(cpu) {
2149 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2150 if (!llist_empty(&gc->list))
2151 return XFS_INODEGC_SHRINKER_COUNT;
2157 static unsigned long
2158 xfs_inodegc_shrinker_scan(
2159 struct shrinker *shrink,
2160 struct shrink_control *sc)
2162 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2163 m_inodegc_shrinker);
2164 struct xfs_inodegc *gc;
2166 bool no_items = true;
2168 if (!xfs_is_inodegc_enabled(mp))
2171 trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2173 for_each_online_cpu(cpu) {
2174 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2175 if (!llist_empty(&gc->list)) {
2176 unsigned int h = READ_ONCE(gc->shrinker_hits);
2178 WRITE_ONCE(gc->shrinker_hits, h + 1);
2179 queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
2185 * If there are no inodes to inactivate, we don't want the shrinker
2186 * to think there's deferred work to call us back about.
2194 /* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2196 xfs_inodegc_register_shrinker(
2197 struct xfs_mount *mp)
2199 struct shrinker *shrink = &mp->m_inodegc_shrinker;
2201 shrink->count_objects = xfs_inodegc_shrinker_count;
2202 shrink->scan_objects = xfs_inodegc_shrinker_scan;
2204 shrink->flags = SHRINKER_NONSLAB;
2205 shrink->batch = XFS_INODEGC_SHRINKER_BATCH;
2207 return register_shrinker(shrink);