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_zone, GFP_KERNEL | __GFP_NOFAIL);
82 if (inode_init_always(mp->m_super, VFS_I(ip))) {
83 kmem_cache_free(xfs_inode_zone, 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;
91 XFS_STATS_INC(mp, vn_active);
92 ASSERT(atomic_read(&ip->i_pincount) == 0);
93 ASSERT(ip->i_ino == 0);
95 /* initialise the xfs inode */
98 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
101 memset(&ip->i_df, 0, sizeof(ip->i_df));
103 ip->i_delayed_blks = 0;
104 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
109 INIT_WORK(&ip->i_ioend_work, xfs_end_io);
110 INIT_LIST_HEAD(&ip->i_ioend_list);
111 spin_lock_init(&ip->i_ioend_lock);
117 xfs_inode_free_callback(
118 struct rcu_head *head)
120 struct inode *inode = container_of(head, struct inode, i_rcu);
121 struct xfs_inode *ip = XFS_I(inode);
123 switch (VFS_I(ip)->i_mode & S_IFMT) {
127 xfs_idestroy_fork(&ip->i_df);
132 xfs_idestroy_fork(ip->i_afp);
133 kmem_cache_free(xfs_ifork_zone, ip->i_afp);
136 xfs_idestroy_fork(ip->i_cowfp);
137 kmem_cache_free(xfs_ifork_zone, ip->i_cowfp);
140 ASSERT(!test_bit(XFS_LI_IN_AIL,
141 &ip->i_itemp->ili_item.li_flags));
142 xfs_inode_item_destroy(ip);
146 kmem_cache_free(xfs_inode_zone, ip);
151 struct xfs_inode *ip)
153 /* asserts to verify all state is correct here */
154 ASSERT(atomic_read(&ip->i_pincount) == 0);
155 ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
156 XFS_STATS_DEC(ip->i_mount, vn_active);
158 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
163 struct xfs_inode *ip)
165 ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
168 * Because we use RCU freeing we need to ensure the inode always
169 * appears to be reclaimed with an invalid inode number when in the
170 * free state. The ip->i_flags_lock provides the barrier against lookup
173 spin_lock(&ip->i_flags_lock);
174 ip->i_flags = XFS_IRECLAIM;
176 spin_unlock(&ip->i_flags_lock);
178 __xfs_inode_free(ip);
182 * Queue background inode reclaim work if there are reclaimable inodes and there
183 * isn't reclaim work already scheduled or in progress.
186 xfs_reclaim_work_queue(
187 struct xfs_mount *mp)
191 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
192 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
193 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
199 * Background scanning to trim preallocated space. This is queued based on the
200 * 'speculative_prealloc_lifetime' tunable (5m by default).
204 struct xfs_perag *pag)
206 struct xfs_mount *mp = pag->pag_mount;
208 if (!xfs_is_blockgc_enabled(mp))
212 if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
213 queue_delayed_work(pag->pag_mount->m_blockgc_wq,
214 &pag->pag_blockgc_work,
215 msecs_to_jiffies(xfs_blockgc_secs * 1000));
219 /* Set a tag on both the AG incore inode tree and the AG radix tree. */
221 xfs_perag_set_inode_tag(
222 struct xfs_perag *pag,
226 struct xfs_mount *mp = pag->pag_mount;
229 lockdep_assert_held(&pag->pag_ici_lock);
231 was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
232 radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
234 if (tag == XFS_ICI_RECLAIM_TAG)
235 pag->pag_ici_reclaimable++;
240 /* propagate the tag up into the perag radix tree */
241 spin_lock(&mp->m_perag_lock);
242 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno, tag);
243 spin_unlock(&mp->m_perag_lock);
245 /* start background work */
247 case XFS_ICI_RECLAIM_TAG:
248 xfs_reclaim_work_queue(mp);
250 case XFS_ICI_BLOCKGC_TAG:
251 xfs_blockgc_queue(pag);
255 trace_xfs_perag_set_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
258 /* Clear a tag on both the AG incore inode tree and the AG radix tree. */
260 xfs_perag_clear_inode_tag(
261 struct xfs_perag *pag,
265 struct xfs_mount *mp = pag->pag_mount;
267 lockdep_assert_held(&pag->pag_ici_lock);
270 * Reclaim can signal (with a null agino) that it cleared its own tag
271 * by removing the inode from the radix tree.
273 if (agino != NULLAGINO)
274 radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
276 ASSERT(tag == XFS_ICI_RECLAIM_TAG);
278 if (tag == XFS_ICI_RECLAIM_TAG)
279 pag->pag_ici_reclaimable--;
281 if (radix_tree_tagged(&pag->pag_ici_root, tag))
284 /* clear the tag from the perag radix tree */
285 spin_lock(&mp->m_perag_lock);
286 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno, tag);
287 spin_unlock(&mp->m_perag_lock);
289 trace_xfs_perag_clear_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
294 struct xfs_inode *ip)
296 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_INEW_BIT);
297 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_INEW_BIT);
300 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
301 if (!xfs_iflags_test(ip, XFS_INEW))
305 finish_wait(wq, &wait.wq_entry);
309 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
310 * part of the structure. This is made more complex by the fact we store
311 * information about the on-disk values in the VFS inode and so we can't just
312 * overwrite the values unconditionally. Hence we save the parameters we
313 * need to retain across reinitialisation, and rewrite them into the VFS inode
314 * after reinitialisation even if it fails.
318 struct xfs_mount *mp,
322 uint32_t nlink = inode->i_nlink;
323 uint32_t generation = inode->i_generation;
324 uint64_t version = inode_peek_iversion(inode);
325 umode_t mode = inode->i_mode;
326 dev_t dev = inode->i_rdev;
327 kuid_t uid = inode->i_uid;
328 kgid_t gid = inode->i_gid;
330 error = inode_init_always(mp->m_super, inode);
332 set_nlink(inode, nlink);
333 inode->i_generation = generation;
334 inode_set_iversion_queried(inode, version);
335 inode->i_mode = mode;
343 * Carefully nudge an inode whose VFS state has been torn down back into a
344 * usable state. Drops the i_flags_lock and the rcu read lock.
348 struct xfs_perag *pag,
349 struct xfs_inode *ip) __releases(&ip->i_flags_lock)
351 struct xfs_mount *mp = ip->i_mount;
352 struct inode *inode = VFS_I(ip);
355 trace_xfs_iget_recycle(ip);
358 * We need to make it look like the inode is being reclaimed to prevent
359 * the actual reclaim workers from stomping over us while we recycle
360 * the inode. We can't clear the radix tree tag yet as it requires
361 * pag_ici_lock to be held exclusive.
363 ip->i_flags |= XFS_IRECLAIM;
365 spin_unlock(&ip->i_flags_lock);
368 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
369 error = xfs_reinit_inode(mp, inode);
374 * Re-initializing the inode failed, and we are in deep
375 * trouble. Try to re-add it to the reclaim list.
378 spin_lock(&ip->i_flags_lock);
379 wake = !!__xfs_iflags_test(ip, XFS_INEW);
380 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
382 wake_up_bit(&ip->i_flags, __XFS_INEW_BIT);
383 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
384 spin_unlock(&ip->i_flags_lock);
387 trace_xfs_iget_recycle_fail(ip);
391 spin_lock(&pag->pag_ici_lock);
392 spin_lock(&ip->i_flags_lock);
395 * Clear the per-lifetime state in the inode as we are now effectively
396 * a new inode and need to return to the initial state before reuse
399 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
400 ip->i_flags |= XFS_INEW;
401 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
402 XFS_ICI_RECLAIM_TAG);
403 inode->i_state = I_NEW;
404 spin_unlock(&ip->i_flags_lock);
405 spin_unlock(&pag->pag_ici_lock);
411 * If we are allocating a new inode, then check what was returned is
412 * actually a free, empty inode. If we are not allocating an inode,
413 * then check we didn't find a free inode.
416 * 0 if the inode free state matches the lookup context
417 * -ENOENT if the inode is free and we are not allocating
418 * -EFSCORRUPTED if there is any state mismatch at all
421 xfs_iget_check_free_state(
422 struct xfs_inode *ip,
425 if (flags & XFS_IGET_CREATE) {
426 /* should be a free inode */
427 if (VFS_I(ip)->i_mode != 0) {
428 xfs_warn(ip->i_mount,
429 "Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
430 ip->i_ino, VFS_I(ip)->i_mode);
431 return -EFSCORRUPTED;
434 if (ip->i_nblocks != 0) {
435 xfs_warn(ip->i_mount,
436 "Corruption detected! Free inode 0x%llx has blocks allocated!",
438 return -EFSCORRUPTED;
443 /* should be an allocated inode */
444 if (VFS_I(ip)->i_mode == 0)
450 /* Make all pending inactivation work start immediately. */
452 xfs_inodegc_queue_all(
453 struct xfs_mount *mp)
455 struct xfs_inodegc *gc;
459 for_each_online_cpu(cpu) {
460 gc = per_cpu_ptr(mp->m_inodegc, cpu);
461 if (!llist_empty(&gc->list)) {
462 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
471 * Check the validity of the inode we just found it the cache
475 struct xfs_perag *pag,
476 struct xfs_inode *ip,
479 int lock_flags) __releases(RCU)
481 struct inode *inode = VFS_I(ip);
482 struct xfs_mount *mp = ip->i_mount;
486 * check for re-use of an inode within an RCU grace period due to the
487 * radix tree nodes not being updated yet. We monitor for this by
488 * setting the inode number to zero before freeing the inode structure.
489 * If the inode has been reallocated and set up, then the inode number
490 * will not match, so check for that, too.
492 spin_lock(&ip->i_flags_lock);
493 if (ip->i_ino != ino)
497 * If we are racing with another cache hit that is currently
498 * instantiating this inode or currently recycling it out of
499 * reclaimable state, wait for the initialisation to complete
502 * If we're racing with the inactivation worker we also want to wait.
503 * If we're creating a new file, it's possible that the worker
504 * previously marked the inode as free on disk but hasn't finished
505 * updating the incore state yet. The AGI buffer will be dirty and
506 * locked to the icreate transaction, so a synchronous push of the
507 * inodegc workers would result in deadlock. For a regular iget, the
508 * worker is running already, so we might as well wait.
510 * XXX(hch): eventually we should do something equivalent to
511 * wait_on_inode to wait for these flags to be cleared
512 * instead of polling for it.
514 if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
517 if (ip->i_flags & XFS_NEED_INACTIVE) {
518 /* Unlinked inodes cannot be re-grabbed. */
519 if (VFS_I(ip)->i_nlink == 0) {
523 goto out_inodegc_flush;
527 * Check the inode free state is valid. This also detects lookup
528 * racing with unlinks.
530 error = xfs_iget_check_free_state(ip, flags);
534 /* Skip inodes that have no vfs state. */
535 if ((flags & XFS_IGET_INCORE) &&
536 (ip->i_flags & XFS_IRECLAIMABLE))
539 /* The inode fits the selection criteria; process it. */
540 if (ip->i_flags & XFS_IRECLAIMABLE) {
541 /* Drops i_flags_lock and RCU read lock. */
542 error = xfs_iget_recycle(pag, ip);
546 /* If the VFS inode is being torn down, pause and try again. */
550 /* We've got a live one. */
551 spin_unlock(&ip->i_flags_lock);
553 trace_xfs_iget_hit(ip);
557 xfs_ilock(ip, lock_flags);
559 if (!(flags & XFS_IGET_INCORE))
560 xfs_iflags_clear(ip, XFS_ISTALE);
561 XFS_STATS_INC(mp, xs_ig_found);
566 trace_xfs_iget_skip(ip);
567 XFS_STATS_INC(mp, xs_ig_frecycle);
570 spin_unlock(&ip->i_flags_lock);
575 spin_unlock(&ip->i_flags_lock);
578 * Do not wait for the workers, because the caller could hold an AGI
579 * buffer lock. We're just going to sleep in a loop anyway.
581 if (xfs_is_inodegc_enabled(mp))
582 xfs_inodegc_queue_all(mp);
588 struct xfs_mount *mp,
589 struct xfs_perag *pag,
592 struct xfs_inode **ipp,
596 struct xfs_inode *ip;
598 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
601 ip = xfs_inode_alloc(mp, ino);
605 error = xfs_imap(mp, tp, ip->i_ino, &ip->i_imap, flags);
610 * For version 5 superblocks, if we are initialising a new inode and we
611 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can
612 * simply build the new inode core with a random generation number.
614 * For version 4 (and older) superblocks, log recovery is dependent on
615 * the i_flushiter field being initialised from the current on-disk
616 * value and hence we must also read the inode off disk even when
617 * initializing new inodes.
619 if (xfs_has_v3inodes(mp) &&
620 (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
621 VFS_I(ip)->i_generation = prandom_u32();
625 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
629 error = xfs_inode_from_disk(ip,
630 xfs_buf_offset(bp, ip->i_imap.im_boffset));
632 xfs_buf_set_ref(bp, XFS_INO_REF);
633 xfs_trans_brelse(tp, bp);
639 trace_xfs_iget_miss(ip);
642 * Check the inode free state is valid. This also detects lookup
643 * racing with unlinks.
645 error = xfs_iget_check_free_state(ip, flags);
650 * Preload the radix tree so we can insert safely under the
651 * write spinlock. Note that we cannot sleep inside the preload
652 * region. Since we can be called from transaction context, don't
653 * recurse into the file system.
655 if (radix_tree_preload(GFP_NOFS)) {
661 * Because the inode hasn't been added to the radix-tree yet it can't
662 * be found by another thread, so we can do the non-sleeping lock here.
665 if (!xfs_ilock_nowait(ip, lock_flags))
670 * These values must be set before inserting the inode into the radix
671 * tree as the moment it is inserted a concurrent lookup (allowed by the
672 * RCU locking mechanism) can find it and that lookup must see that this
673 * is an inode currently under construction (i.e. that XFS_INEW is set).
674 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
675 * memory barrier that ensures this detection works correctly at lookup
679 if (flags & XFS_IGET_DONTCACHE)
680 d_mark_dontcache(VFS_I(ip));
684 xfs_iflags_set(ip, iflags);
686 /* insert the new inode */
687 spin_lock(&pag->pag_ici_lock);
688 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
689 if (unlikely(error)) {
690 WARN_ON(error != -EEXIST);
691 XFS_STATS_INC(mp, xs_ig_dup);
693 goto out_preload_end;
695 spin_unlock(&pag->pag_ici_lock);
696 radix_tree_preload_end();
702 spin_unlock(&pag->pag_ici_lock);
703 radix_tree_preload_end();
705 xfs_iunlock(ip, lock_flags);
707 __destroy_inode(VFS_I(ip));
713 * Look up an inode by number in the given file system. The inode is looked up
714 * in the cache held in each AG. If the inode is found in the cache, initialise
715 * the vfs inode if necessary.
717 * If it is not in core, read it in from the file system's device, add it to the
718 * cache and initialise the vfs inode.
720 * The inode is locked according to the value of the lock_flags parameter.
721 * Inode lookup is only done during metadata operations and not as part of the
722 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
726 struct xfs_mount *mp,
727 struct xfs_trans *tp,
731 struct xfs_inode **ipp)
733 struct xfs_inode *ip;
734 struct xfs_perag *pag;
738 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
740 /* reject inode numbers outside existing AGs */
741 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
744 XFS_STATS_INC(mp, xs_ig_attempts);
746 /* get the perag structure and ensure that it's inode capable */
747 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
748 agino = XFS_INO_TO_AGINO(mp, ino);
753 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
756 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
758 goto out_error_or_again;
761 if (flags & XFS_IGET_INCORE) {
763 goto out_error_or_again;
765 XFS_STATS_INC(mp, xs_ig_missed);
767 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
770 goto out_error_or_again;
777 * If we have a real type for an on-disk inode, we can setup the inode
778 * now. If it's a new inode being created, xfs_ialloc will handle it.
780 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
781 xfs_setup_existing_inode(ip);
785 if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
794 * "Is this a cached inode that's also allocated?"
796 * Look up an inode by number in the given file system. If the inode is
797 * in cache and isn't in purgatory, return 1 if the inode is allocated
798 * and 0 if it is not. For all other cases (not in cache, being torn
799 * down, etc.), return a negative error code.
801 * The caller has to prevent inode allocation and freeing activity,
802 * presumably by locking the AGI buffer. This is to ensure that an
803 * inode cannot transition from allocated to freed until the caller is
804 * ready to allow that. If the inode is in an intermediate state (new,
805 * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
806 * inode is not in the cache, -ENOENT will be returned. The caller must
807 * deal with these scenarios appropriately.
809 * This is a specialized use case for the online scrubber; if you're
810 * reading this, you probably want xfs_iget.
813 xfs_icache_inode_is_allocated(
814 struct xfs_mount *mp,
815 struct xfs_trans *tp,
819 struct xfs_inode *ip;
822 error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
826 *inuse = !!(VFS_I(ip)->i_mode);
832 * Grab the inode for reclaim exclusively.
834 * We have found this inode via a lookup under RCU, so the inode may have
835 * already been freed, or it may be in the process of being recycled by
836 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
837 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
838 * will not be set. Hence we need to check for both these flag conditions to
839 * avoid inodes that are no longer reclaim candidates.
841 * Note: checking for other state flags here, under the i_flags_lock or not, is
842 * racy and should be avoided. Those races should be resolved only after we have
843 * ensured that we are able to reclaim this inode and the world can see that we
844 * are going to reclaim it.
846 * Return true if we grabbed it, false otherwise.
850 struct xfs_inode *ip,
851 struct xfs_icwalk *icw)
853 ASSERT(rcu_read_lock_held());
855 spin_lock(&ip->i_flags_lock);
856 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
857 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
858 /* not a reclaim candidate. */
859 spin_unlock(&ip->i_flags_lock);
863 /* Don't reclaim a sick inode unless the caller asked for it. */
865 (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
866 spin_unlock(&ip->i_flags_lock);
870 __xfs_iflags_set(ip, XFS_IRECLAIM);
871 spin_unlock(&ip->i_flags_lock);
876 * Inode reclaim is non-blocking, so the default action if progress cannot be
877 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
878 * XFS_IRECLAIM flag. If we are in a shutdown state, we don't care about
879 * blocking anymore and hence we can wait for the inode to be able to reclaim
882 * We do no IO here - if callers require inodes to be cleaned they must push the
883 * AIL first to trigger writeback of dirty inodes. This enables writeback to be
884 * done in the background in a non-blocking manner, and enables memory reclaim
885 * to make progress without blocking.
889 struct xfs_inode *ip,
890 struct xfs_perag *pag)
892 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
894 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
896 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
899 if (xfs_is_shutdown(ip->i_mount)) {
901 xfs_iflush_abort(ip);
904 if (xfs_ipincount(ip))
905 goto out_clear_flush;
906 if (!xfs_inode_clean(ip))
907 goto out_clear_flush;
909 xfs_iflags_clear(ip, XFS_IFLUSHING);
911 trace_xfs_inode_reclaiming(ip);
914 * Because we use RCU freeing we need to ensure the inode always appears
915 * to be reclaimed with an invalid inode number when in the free state.
916 * We do this as early as possible under the ILOCK so that
917 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
918 * detect races with us here. By doing this, we guarantee that once
919 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
920 * it will see either a valid inode that will serialise correctly, or it
921 * will see an invalid inode that it can skip.
923 spin_lock(&ip->i_flags_lock);
924 ip->i_flags = XFS_IRECLAIM;
928 spin_unlock(&ip->i_flags_lock);
930 xfs_iunlock(ip, XFS_ILOCK_EXCL);
932 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
934 * Remove the inode from the per-AG radix tree.
936 * Because radix_tree_delete won't complain even if the item was never
937 * added to the tree assert that it's been there before to catch
938 * problems with the inode life time early on.
940 spin_lock(&pag->pag_ici_lock);
941 if (!radix_tree_delete(&pag->pag_ici_root,
942 XFS_INO_TO_AGINO(ip->i_mount, ino)))
944 xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
945 spin_unlock(&pag->pag_ici_lock);
948 * Here we do an (almost) spurious inode lock in order to coordinate
949 * with inode cache radix tree lookups. This is because the lookup
950 * can reference the inodes in the cache without taking references.
952 * We make that OK here by ensuring that we wait until the inode is
953 * unlocked after the lookup before we go ahead and free it.
955 xfs_ilock(ip, XFS_ILOCK_EXCL);
956 ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
957 xfs_iunlock(ip, XFS_ILOCK_EXCL);
958 ASSERT(xfs_inode_clean(ip));
960 __xfs_inode_free(ip);
964 xfs_iflags_clear(ip, XFS_IFLUSHING);
966 xfs_iunlock(ip, XFS_ILOCK_EXCL);
968 xfs_iflags_clear(ip, XFS_IRECLAIM);
971 /* Reclaim sick inodes if we're unmounting or the fs went down. */
973 xfs_want_reclaim_sick(
974 struct xfs_mount *mp)
976 return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
982 struct xfs_mount *mp)
984 struct xfs_icwalk icw = {
988 if (xfs_want_reclaim_sick(mp))
989 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
991 while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
992 xfs_ail_push_all_sync(mp->m_ail);
993 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
998 * The shrinker infrastructure determines how many inodes we should scan for
999 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
1000 * push the AIL here. We also want to proactively free up memory if we can to
1001 * minimise the amount of work memory reclaim has to do so we kick the
1002 * background reclaim if it isn't already scheduled.
1005 xfs_reclaim_inodes_nr(
1006 struct xfs_mount *mp,
1007 unsigned long nr_to_scan)
1009 struct xfs_icwalk icw = {
1010 .icw_flags = XFS_ICWALK_FLAG_SCAN_LIMIT,
1011 .icw_scan_limit = min_t(unsigned long, LONG_MAX, nr_to_scan),
1014 if (xfs_want_reclaim_sick(mp))
1015 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1017 /* kick background reclaimer and push the AIL */
1018 xfs_reclaim_work_queue(mp);
1019 xfs_ail_push_all(mp->m_ail);
1021 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1026 * Return the number of reclaimable inodes in the filesystem for
1027 * the shrinker to determine how much to reclaim.
1030 xfs_reclaim_inodes_count(
1031 struct xfs_mount *mp)
1033 struct xfs_perag *pag;
1034 xfs_agnumber_t ag = 0;
1035 long reclaimable = 0;
1037 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1038 ag = pag->pag_agno + 1;
1039 reclaimable += pag->pag_ici_reclaimable;
1046 xfs_icwalk_match_id(
1047 struct xfs_inode *ip,
1048 struct xfs_icwalk *icw)
1050 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1051 !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1054 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1055 !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1058 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1059 ip->i_projid != icw->icw_prid)
1066 * A union-based inode filtering algorithm. Process the inode if any of the
1067 * criteria match. This is for global/internal scans only.
1070 xfs_icwalk_match_id_union(
1071 struct xfs_inode *ip,
1072 struct xfs_icwalk *icw)
1074 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1075 uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1078 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1079 gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1082 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1083 ip->i_projid == icw->icw_prid)
1090 * Is this inode @ip eligible for eof/cow block reclamation, given some
1091 * filtering parameters @icw? The inode is eligible if @icw is null or
1092 * if the predicate functions match.
1096 struct xfs_inode *ip,
1097 struct xfs_icwalk *icw)
1104 if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1105 match = xfs_icwalk_match_id_union(ip, icw);
1107 match = xfs_icwalk_match_id(ip, icw);
1111 /* skip the inode if the file size is too small */
1112 if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1113 XFS_ISIZE(ip) < icw->icw_min_file_size)
1120 * This is a fast pass over the inode cache to try to get reclaim moving on as
1121 * many inodes as possible in a short period of time. It kicks itself every few
1122 * seconds, as well as being kicked by the inode cache shrinker when memory
1127 struct work_struct *work)
1129 struct xfs_mount *mp = container_of(to_delayed_work(work),
1130 struct xfs_mount, m_reclaim_work);
1132 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1133 xfs_reclaim_work_queue(mp);
1137 xfs_inode_free_eofblocks(
1138 struct xfs_inode *ip,
1139 struct xfs_icwalk *icw,
1140 unsigned int *lockflags)
1144 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1146 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1150 * If the mapping is dirty the operation can block and wait for some
1151 * time. Unless we are waiting, skip it.
1153 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1156 if (!xfs_icwalk_match(ip, icw))
1160 * If the caller is waiting, return -EAGAIN to keep the background
1161 * scanner moving and revisit the inode in a subsequent pass.
1163 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1168 *lockflags |= XFS_IOLOCK_EXCL;
1170 if (xfs_can_free_eofblocks(ip, false))
1171 return xfs_free_eofblocks(ip);
1173 /* inode could be preallocated or append-only */
1174 trace_xfs_inode_free_eofblocks_invalid(ip);
1175 xfs_inode_clear_eofblocks_tag(ip);
1180 xfs_blockgc_set_iflag(
1181 struct xfs_inode *ip,
1182 unsigned long iflag)
1184 struct xfs_mount *mp = ip->i_mount;
1185 struct xfs_perag *pag;
1187 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1190 * Don't bother locking the AG and looking up in the radix trees
1191 * if we already know that we have the tag set.
1193 if (ip->i_flags & iflag)
1195 spin_lock(&ip->i_flags_lock);
1196 ip->i_flags |= iflag;
1197 spin_unlock(&ip->i_flags_lock);
1199 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1200 spin_lock(&pag->pag_ici_lock);
1202 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1203 XFS_ICI_BLOCKGC_TAG);
1205 spin_unlock(&pag->pag_ici_lock);
1210 xfs_inode_set_eofblocks_tag(
1213 trace_xfs_inode_set_eofblocks_tag(ip);
1214 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1218 xfs_blockgc_clear_iflag(
1219 struct xfs_inode *ip,
1220 unsigned long iflag)
1222 struct xfs_mount *mp = ip->i_mount;
1223 struct xfs_perag *pag;
1226 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1228 spin_lock(&ip->i_flags_lock);
1229 ip->i_flags &= ~iflag;
1230 clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1231 spin_unlock(&ip->i_flags_lock);
1236 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1237 spin_lock(&pag->pag_ici_lock);
1239 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1240 XFS_ICI_BLOCKGC_TAG);
1242 spin_unlock(&pag->pag_ici_lock);
1247 xfs_inode_clear_eofblocks_tag(
1250 trace_xfs_inode_clear_eofblocks_tag(ip);
1251 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1255 * Set ourselves up to free CoW blocks from this file. If it's already clean
1256 * then we can bail out quickly, but otherwise we must back off if the file
1257 * is undergoing some kind of write.
1260 xfs_prep_free_cowblocks(
1261 struct xfs_inode *ip)
1264 * Just clear the tag if we have an empty cow fork or none at all. It's
1265 * possible the inode was fully unshared since it was originally tagged.
1267 if (!xfs_inode_has_cow_data(ip)) {
1268 trace_xfs_inode_free_cowblocks_invalid(ip);
1269 xfs_inode_clear_cowblocks_tag(ip);
1274 * If the mapping is dirty or under writeback we cannot touch the
1275 * CoW fork. Leave it alone if we're in the midst of a directio.
1277 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1278 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1279 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1280 atomic_read(&VFS_I(ip)->i_dio_count))
1287 * Automatic CoW Reservation Freeing
1289 * These functions automatically garbage collect leftover CoW reservations
1290 * that were made on behalf of a cowextsize hint when we start to run out
1291 * of quota or when the reservations sit around for too long. If the file
1292 * has dirty pages or is undergoing writeback, its CoW reservations will
1295 * The actual garbage collection piggybacks off the same code that runs
1296 * the speculative EOF preallocation garbage collector.
1299 xfs_inode_free_cowblocks(
1300 struct xfs_inode *ip,
1301 struct xfs_icwalk *icw,
1302 unsigned int *lockflags)
1307 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1309 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1312 if (!xfs_prep_free_cowblocks(ip))
1315 if (!xfs_icwalk_match(ip, icw))
1319 * If the caller is waiting, return -EAGAIN to keep the background
1320 * scanner moving and revisit the inode in a subsequent pass.
1322 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1323 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1328 *lockflags |= XFS_IOLOCK_EXCL;
1330 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1335 *lockflags |= XFS_MMAPLOCK_EXCL;
1338 * Check again, nobody else should be able to dirty blocks or change
1339 * the reflink iflag now that we have the first two locks held.
1341 if (xfs_prep_free_cowblocks(ip))
1342 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1347 xfs_inode_set_cowblocks_tag(
1350 trace_xfs_inode_set_cowblocks_tag(ip);
1351 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1355 xfs_inode_clear_cowblocks_tag(
1358 trace_xfs_inode_clear_cowblocks_tag(ip);
1359 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1362 /* Disable post-EOF and CoW block auto-reclamation. */
1365 struct xfs_mount *mp)
1367 struct xfs_perag *pag;
1368 xfs_agnumber_t agno;
1370 if (!xfs_clear_blockgc_enabled(mp))
1373 for_each_perag(mp, agno, pag)
1374 cancel_delayed_work_sync(&pag->pag_blockgc_work);
1375 trace_xfs_blockgc_stop(mp, __return_address);
1378 /* Enable post-EOF and CoW block auto-reclamation. */
1381 struct xfs_mount *mp)
1383 struct xfs_perag *pag;
1384 xfs_agnumber_t agno;
1386 if (xfs_set_blockgc_enabled(mp))
1389 trace_xfs_blockgc_start(mp, __return_address);
1390 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1391 xfs_blockgc_queue(pag);
1394 /* Don't try to run block gc on an inode that's in any of these states. */
1395 #define XFS_BLOCKGC_NOGRAB_IFLAGS (XFS_INEW | \
1396 XFS_NEED_INACTIVE | \
1397 XFS_INACTIVATING | \
1398 XFS_IRECLAIMABLE | \
1401 * Decide if the given @ip is eligible for garbage collection of speculative
1402 * preallocations, and grab it if so. Returns true if it's ready to go or
1403 * false if we should just ignore it.
1407 struct xfs_inode *ip)
1409 struct inode *inode = VFS_I(ip);
1411 ASSERT(rcu_read_lock_held());
1413 /* Check for stale RCU freed inode */
1414 spin_lock(&ip->i_flags_lock);
1416 goto out_unlock_noent;
1418 if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1419 goto out_unlock_noent;
1420 spin_unlock(&ip->i_flags_lock);
1422 /* nothing to sync during shutdown */
1423 if (xfs_is_shutdown(ip->i_mount))
1426 /* If we can't grab the inode, it must on it's way to reclaim. */
1430 /* inode is valid */
1434 spin_unlock(&ip->i_flags_lock);
1438 /* Scan one incore inode for block preallocations that we can remove. */
1440 xfs_blockgc_scan_inode(
1441 struct xfs_inode *ip,
1442 struct xfs_icwalk *icw)
1444 unsigned int lockflags = 0;
1447 error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1451 error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1454 xfs_iunlock(ip, lockflags);
1459 /* Background worker that trims preallocated space. */
1462 struct work_struct *work)
1464 struct xfs_perag *pag = container_of(to_delayed_work(work),
1465 struct xfs_perag, pag_blockgc_work);
1466 struct xfs_mount *mp = pag->pag_mount;
1469 trace_xfs_blockgc_worker(mp, __return_address);
1471 error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1473 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1474 pag->pag_agno, error);
1475 xfs_blockgc_queue(pag);
1479 * Try to free space in the filesystem by purging inactive inodes, eofblocks
1483 xfs_blockgc_free_space(
1484 struct xfs_mount *mp,
1485 struct xfs_icwalk *icw)
1489 trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1491 error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1495 xfs_inodegc_flush(mp);
1500 * Reclaim all the free space that we can by scheduling the background blockgc
1501 * and inodegc workers immediately and waiting for them all to clear.
1504 xfs_blockgc_flush_all(
1505 struct xfs_mount *mp)
1507 struct xfs_perag *pag;
1508 xfs_agnumber_t agno;
1510 trace_xfs_blockgc_flush_all(mp, __return_address);
1513 * For each blockgc worker, move its queue time up to now. If it
1514 * wasn't queued, it will not be requeued. Then flush whatever's
1517 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1518 mod_delayed_work(pag->pag_mount->m_blockgc_wq,
1519 &pag->pag_blockgc_work, 0);
1521 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1522 flush_delayed_work(&pag->pag_blockgc_work);
1524 xfs_inodegc_flush(mp);
1528 * Run cow/eofblocks scans on the supplied dquots. We don't know exactly which
1529 * quota caused an allocation failure, so we make a best effort by including
1530 * each quota under low free space conditions (less than 1% free space) in the
1533 * Callers must not hold any inode's ILOCK. If requesting a synchronous scan
1534 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1538 xfs_blockgc_free_dquots(
1539 struct xfs_mount *mp,
1540 struct xfs_dquot *udqp,
1541 struct xfs_dquot *gdqp,
1542 struct xfs_dquot *pdqp,
1543 unsigned int iwalk_flags)
1545 struct xfs_icwalk icw = {0};
1546 bool do_work = false;
1548 if (!udqp && !gdqp && !pdqp)
1552 * Run a scan to free blocks using the union filter to cover all
1553 * applicable quotas in a single scan.
1555 icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1557 if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1558 icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1559 icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1563 if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1564 icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1565 icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1569 if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1570 icw.icw_prid = pdqp->q_id;
1571 icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1578 return xfs_blockgc_free_space(mp, &icw);
1581 /* Run cow/eofblocks scans on the quotas attached to the inode. */
1583 xfs_blockgc_free_quota(
1584 struct xfs_inode *ip,
1585 unsigned int iwalk_flags)
1587 return xfs_blockgc_free_dquots(ip->i_mount,
1588 xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1589 xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1590 xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1593 /* XFS Inode Cache Walking Code */
1596 * The inode lookup is done in batches to keep the amount of lock traffic and
1597 * radix tree lookups to a minimum. The batch size is a trade off between
1598 * lookup reduction and stack usage. This is in the reclaim path, so we can't
1601 #define XFS_LOOKUP_BATCH 32
1605 * Decide if we want to grab this inode in anticipation of doing work towards
1610 enum xfs_icwalk_goal goal,
1611 struct xfs_inode *ip,
1612 struct xfs_icwalk *icw)
1615 case XFS_ICWALK_BLOCKGC:
1616 return xfs_blockgc_igrab(ip);
1617 case XFS_ICWALK_RECLAIM:
1618 return xfs_reclaim_igrab(ip, icw);
1625 * Process an inode. Each processing function must handle any state changes
1626 * made by the icwalk igrab function. Return -EAGAIN to skip an inode.
1629 xfs_icwalk_process_inode(
1630 enum xfs_icwalk_goal goal,
1631 struct xfs_inode *ip,
1632 struct xfs_perag *pag,
1633 struct xfs_icwalk *icw)
1638 case XFS_ICWALK_BLOCKGC:
1639 error = xfs_blockgc_scan_inode(ip, icw);
1641 case XFS_ICWALK_RECLAIM:
1642 xfs_reclaim_inode(ip, pag);
1649 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1650 * process them in some manner.
1654 struct xfs_perag *pag,
1655 enum xfs_icwalk_goal goal,
1656 struct xfs_icwalk *icw)
1658 struct xfs_mount *mp = pag->pag_mount;
1659 uint32_t first_index;
1668 if (goal == XFS_ICWALK_RECLAIM)
1669 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1674 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1680 nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1681 (void **) batch, first_index,
1682 XFS_LOOKUP_BATCH, goal);
1690 * Grab the inodes before we drop the lock. if we found
1691 * nothing, nr == 0 and the loop will be skipped.
1693 for (i = 0; i < nr_found; i++) {
1694 struct xfs_inode *ip = batch[i];
1696 if (done || !xfs_icwalk_igrab(goal, ip, icw))
1700 * Update the index for the next lookup. Catch
1701 * overflows into the next AG range which can occur if
1702 * we have inodes in the last block of the AG and we
1703 * are currently pointing to the last inode.
1705 * Because we may see inodes that are from the wrong AG
1706 * due to RCU freeing and reallocation, only update the
1707 * index if it lies in this AG. It was a race that lead
1708 * us to see this inode, so another lookup from the
1709 * same index will not find it again.
1711 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
1713 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1714 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1718 /* unlock now we've grabbed the inodes. */
1721 for (i = 0; i < nr_found; i++) {
1724 error = xfs_icwalk_process_inode(goal, batch[i], pag,
1726 if (error == -EAGAIN) {
1730 if (error && last_error != -EFSCORRUPTED)
1734 /* bail out if the filesystem is corrupted. */
1735 if (error == -EFSCORRUPTED)
1740 if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1741 icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1742 if (icw->icw_scan_limit <= 0)
1745 } while (nr_found && !done);
1747 if (goal == XFS_ICWALK_RECLAIM) {
1750 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1760 /* Walk all incore inodes to achieve a given goal. */
1763 struct xfs_mount *mp,
1764 enum xfs_icwalk_goal goal,
1765 struct xfs_icwalk *icw)
1767 struct xfs_perag *pag;
1770 xfs_agnumber_t agno;
1772 for_each_perag_tag(mp, agno, pag, goal) {
1773 error = xfs_icwalk_ag(pag, goal, icw);
1776 if (error == -EFSCORRUPTED) {
1783 BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1789 struct xfs_inode *ip,
1792 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, whichfork);
1793 struct xfs_bmbt_irec got;
1794 struct xfs_iext_cursor icur;
1796 if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1799 if (isnullstartblock(got.br_startblock)) {
1800 xfs_warn(ip->i_mount,
1801 "ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",
1803 whichfork == XFS_DATA_FORK ? "data" : "cow",
1804 got.br_startoff, got.br_blockcount);
1806 } while (xfs_iext_next_extent(ifp, &icur, &got));
1809 #define xfs_check_delalloc(ip, whichfork) do { } while (0)
1812 /* Schedule the inode for reclaim. */
1814 xfs_inodegc_set_reclaimable(
1815 struct xfs_inode *ip)
1817 struct xfs_mount *mp = ip->i_mount;
1818 struct xfs_perag *pag;
1820 if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {
1821 xfs_check_delalloc(ip, XFS_DATA_FORK);
1822 xfs_check_delalloc(ip, XFS_COW_FORK);
1826 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1827 spin_lock(&pag->pag_ici_lock);
1828 spin_lock(&ip->i_flags_lock);
1830 trace_xfs_inode_set_reclaimable(ip);
1831 ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);
1832 ip->i_flags |= XFS_IRECLAIMABLE;
1833 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1834 XFS_ICI_RECLAIM_TAG);
1836 spin_unlock(&ip->i_flags_lock);
1837 spin_unlock(&pag->pag_ici_lock);
1842 * Free all speculative preallocations and possibly even the inode itself.
1843 * This is the last chance to make changes to an otherwise unreferenced file
1844 * before incore reclamation happens.
1847 xfs_inodegc_inactivate(
1848 struct xfs_inode *ip)
1850 trace_xfs_inode_inactivating(ip);
1852 xfs_inodegc_set_reclaimable(ip);
1857 struct work_struct *work)
1859 struct xfs_inodegc *gc = container_of(to_delayed_work(work),
1860 struct xfs_inodegc, work);
1861 struct llist_node *node = llist_del_all(&gc->list);
1862 struct xfs_inode *ip, *n;
1864 ASSERT(gc->cpu == smp_processor_id());
1866 WRITE_ONCE(gc->items, 0);
1871 ip = llist_entry(node, struct xfs_inode, i_gclist);
1872 trace_xfs_inodegc_worker(ip->i_mount, READ_ONCE(gc->shrinker_hits));
1874 WRITE_ONCE(gc->shrinker_hits, 0);
1875 llist_for_each_entry_safe(ip, n, node, i_gclist) {
1876 xfs_iflags_set(ip, XFS_INACTIVATING);
1877 xfs_inodegc_inactivate(ip);
1882 * Expedite all pending inodegc work to run immediately. This does not wait for
1883 * completion of the work.
1887 struct xfs_mount *mp)
1889 if (!xfs_is_inodegc_enabled(mp))
1891 trace_xfs_inodegc_push(mp, __return_address);
1892 xfs_inodegc_queue_all(mp);
1896 * Force all currently queued inode inactivation work to run immediately and
1897 * wait for the work to finish.
1901 struct xfs_mount *mp)
1903 xfs_inodegc_push(mp);
1904 trace_xfs_inodegc_flush(mp, __return_address);
1905 flush_workqueue(mp->m_inodegc_wq);
1909 * Flush all the pending work and then disable the inode inactivation background
1910 * workers and wait for them to stop. Caller must hold sb->s_umount to
1911 * coordinate changes in the inodegc_enabled state.
1915 struct xfs_mount *mp)
1919 if (!xfs_clear_inodegc_enabled(mp))
1923 * Drain all pending inodegc work, including inodes that could be
1924 * queued by racing xfs_inodegc_queue or xfs_inodegc_shrinker_scan
1925 * threads that sample the inodegc state just prior to us clearing it.
1926 * The inodegc flag state prevents new threads from queuing more
1927 * inodes, so we queue pending work items and flush the workqueue until
1928 * all inodegc lists are empty. IOWs, we cannot use drain_workqueue
1929 * here because it does not allow other unserialized mechanisms to
1930 * reschedule inodegc work while this draining is in progress.
1932 xfs_inodegc_queue_all(mp);
1934 flush_workqueue(mp->m_inodegc_wq);
1935 rerun = xfs_inodegc_queue_all(mp);
1938 trace_xfs_inodegc_stop(mp, __return_address);
1942 * Enable the inode inactivation background workers and schedule deferred inode
1943 * inactivation work if there is any. Caller must hold sb->s_umount to
1944 * coordinate changes in the inodegc_enabled state.
1948 struct xfs_mount *mp)
1950 if (xfs_set_inodegc_enabled(mp))
1953 trace_xfs_inodegc_start(mp, __return_address);
1954 xfs_inodegc_queue_all(mp);
1957 #ifdef CONFIG_XFS_RT
1959 xfs_inodegc_want_queue_rt_file(
1960 struct xfs_inode *ip)
1962 struct xfs_mount *mp = ip->i_mount;
1965 if (!XFS_IS_REALTIME_INODE(ip))
1968 freertx = READ_ONCE(mp->m_sb.sb_frextents);
1969 return freertx < mp->m_low_rtexts[XFS_LOWSP_5_PCNT];
1972 # define xfs_inodegc_want_queue_rt_file(ip) (false)
1973 #endif /* CONFIG_XFS_RT */
1976 * Schedule the inactivation worker when:
1978 * - We've accumulated more than one inode cluster buffer's worth of inodes.
1979 * - There is less than 5% free space left.
1980 * - Any of the quotas for this inode are near an enforcement limit.
1983 xfs_inodegc_want_queue_work(
1984 struct xfs_inode *ip,
1987 struct xfs_mount *mp = ip->i_mount;
1989 if (items > mp->m_ino_geo.inodes_per_cluster)
1992 if (__percpu_counter_compare(&mp->m_fdblocks,
1993 mp->m_low_space[XFS_LOWSP_5_PCNT],
1994 XFS_FDBLOCKS_BATCH) < 0)
1997 if (xfs_inodegc_want_queue_rt_file(ip))
2000 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
2003 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
2006 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
2013 * Upper bound on the number of inodes in each AG that can be queued for
2014 * inactivation at any given time, to avoid monopolizing the workqueue.
2016 #define XFS_INODEGC_MAX_BACKLOG (4 * XFS_INODES_PER_CHUNK)
2019 * Make the frontend wait for inactivations when:
2021 * - Memory shrinkers queued the inactivation worker and it hasn't finished.
2022 * - The queue depth exceeds the maximum allowable percpu backlog.
2024 * Note: If the current thread is running a transaction, we don't ever want to
2025 * wait for other transactions because that could introduce a deadlock.
2028 xfs_inodegc_want_flush_work(
2029 struct xfs_inode *ip,
2031 unsigned int shrinker_hits)
2033 if (current->journal_info)
2036 if (shrinker_hits > 0)
2039 if (items > XFS_INODEGC_MAX_BACKLOG)
2046 * Queue a background inactivation worker if there are inodes that need to be
2047 * inactivated and higher level xfs code hasn't disabled the background
2052 struct xfs_inode *ip)
2054 struct xfs_mount *mp = ip->i_mount;
2055 struct xfs_inodegc *gc;
2057 unsigned int shrinker_hits;
2058 unsigned long queue_delay = 1;
2060 trace_xfs_inode_set_need_inactive(ip);
2061 spin_lock(&ip->i_flags_lock);
2062 ip->i_flags |= XFS_NEED_INACTIVE;
2063 spin_unlock(&ip->i_flags_lock);
2065 gc = get_cpu_ptr(mp->m_inodegc);
2066 llist_add(&ip->i_gclist, &gc->list);
2067 items = READ_ONCE(gc->items);
2068 WRITE_ONCE(gc->items, items + 1);
2069 shrinker_hits = READ_ONCE(gc->shrinker_hits);
2072 * We queue the work while holding the current CPU so that the work
2073 * is scheduled to run on this CPU.
2075 if (!xfs_is_inodegc_enabled(mp)) {
2080 if (xfs_inodegc_want_queue_work(ip, items))
2083 trace_xfs_inodegc_queue(mp, __return_address);
2084 mod_delayed_work_on(current_cpu(), mp->m_inodegc_wq, &gc->work,
2088 if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
2089 trace_xfs_inodegc_throttle(mp, __return_address);
2090 flush_delayed_work(&gc->work);
2095 * Fold the dead CPU inodegc queue into the current CPUs queue.
2098 xfs_inodegc_cpu_dead(
2099 struct xfs_mount *mp,
2100 unsigned int dead_cpu)
2102 struct xfs_inodegc *dead_gc, *gc;
2103 struct llist_node *first, *last;
2104 unsigned int count = 0;
2106 dead_gc = per_cpu_ptr(mp->m_inodegc, dead_cpu);
2107 cancel_delayed_work_sync(&dead_gc->work);
2109 if (llist_empty(&dead_gc->list))
2112 first = dead_gc->list.first;
2114 while (last->next) {
2118 dead_gc->list.first = NULL;
2121 /* Add pending work to current CPU */
2122 gc = get_cpu_ptr(mp->m_inodegc);
2123 llist_add_batch(first, last, &gc->list);
2124 count += READ_ONCE(gc->items);
2125 WRITE_ONCE(gc->items, count);
2127 if (xfs_is_inodegc_enabled(mp)) {
2128 trace_xfs_inodegc_queue(mp, __return_address);
2129 mod_delayed_work_on(current_cpu(), mp->m_inodegc_wq, &gc->work,
2136 * We set the inode flag atomically with the radix tree tag. Once we get tag
2137 * lookups on the radix tree, this inode flag can go away.
2139 * We always use background reclaim here because even if the inode is clean, it
2140 * still may be under IO and hence we have wait for IO completion to occur
2141 * before we can reclaim the inode. The background reclaim path handles this
2142 * more efficiently than we can here, so simply let background reclaim tear down
2146 xfs_inode_mark_reclaimable(
2147 struct xfs_inode *ip)
2149 struct xfs_mount *mp = ip->i_mount;
2152 XFS_STATS_INC(mp, vn_reclaim);
2155 * We should never get here with any of the reclaim flags already set.
2157 ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2159 need_inactive = xfs_inode_needs_inactive(ip);
2160 if (need_inactive) {
2161 xfs_inodegc_queue(ip);
2165 /* Going straight to reclaim, so drop the dquots. */
2166 xfs_qm_dqdetach(ip);
2167 xfs_inodegc_set_reclaimable(ip);
2171 * Register a phony shrinker so that we can run background inodegc sooner when
2172 * there's memory pressure. Inactivation does not itself free any memory but
2173 * it does make inodes reclaimable, which eventually frees memory.
2175 * The count function, seek value, and batch value are crafted to trigger the
2176 * scan function during the second round of scanning. Hopefully this means
2177 * that we reclaimed enough memory that initiating metadata transactions won't
2178 * make things worse.
2180 #define XFS_INODEGC_SHRINKER_COUNT (1UL << DEF_PRIORITY)
2181 #define XFS_INODEGC_SHRINKER_BATCH ((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2183 static unsigned long
2184 xfs_inodegc_shrinker_count(
2185 struct shrinker *shrink,
2186 struct shrink_control *sc)
2188 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2189 m_inodegc_shrinker);
2190 struct xfs_inodegc *gc;
2193 if (!xfs_is_inodegc_enabled(mp))
2196 for_each_online_cpu(cpu) {
2197 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2198 if (!llist_empty(&gc->list))
2199 return XFS_INODEGC_SHRINKER_COUNT;
2205 static unsigned long
2206 xfs_inodegc_shrinker_scan(
2207 struct shrinker *shrink,
2208 struct shrink_control *sc)
2210 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2211 m_inodegc_shrinker);
2212 struct xfs_inodegc *gc;
2214 bool no_items = true;
2216 if (!xfs_is_inodegc_enabled(mp))
2219 trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2221 for_each_online_cpu(cpu) {
2222 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2223 if (!llist_empty(&gc->list)) {
2224 unsigned int h = READ_ONCE(gc->shrinker_hits);
2226 WRITE_ONCE(gc->shrinker_hits, h + 1);
2227 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
2233 * If there are no inodes to inactivate, we don't want the shrinker
2234 * to think there's deferred work to call us back about.
2242 /* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2244 xfs_inodegc_register_shrinker(
2245 struct xfs_mount *mp)
2247 struct shrinker *shrink = &mp->m_inodegc_shrinker;
2249 shrink->count_objects = xfs_inodegc_shrinker_count;
2250 shrink->scan_objects = xfs_inodegc_shrinker_scan;
2252 shrink->flags = SHRINKER_NONSLAB;
2253 shrink->batch = XFS_INODEGC_SHRINKER_BATCH;
2255 return register_shrinker(shrink);