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;
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_cache, ip->i_afp);
136 xfs_idestroy_fork(ip->i_cowfp);
137 kmem_cache_free(xfs_ifork_cache, 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_cache, 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_);
293 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
294 * part of the structure. This is made more complex by the fact we store
295 * information about the on-disk values in the VFS inode and so we can't just
296 * overwrite the values unconditionally. Hence we save the parameters we
297 * need to retain across reinitialisation, and rewrite them into the VFS inode
298 * after reinitialisation even if it fails.
302 struct xfs_mount *mp,
306 uint32_t nlink = inode->i_nlink;
307 uint32_t generation = inode->i_generation;
308 uint64_t version = inode_peek_iversion(inode);
309 umode_t mode = inode->i_mode;
310 dev_t dev = inode->i_rdev;
311 kuid_t uid = inode->i_uid;
312 kgid_t gid = inode->i_gid;
314 error = inode_init_always(mp->m_super, inode);
316 set_nlink(inode, nlink);
317 inode->i_generation = generation;
318 inode_set_iversion_queried(inode, version);
319 inode->i_mode = mode;
327 * Carefully nudge an inode whose VFS state has been torn down back into a
328 * usable state. Drops the i_flags_lock and the rcu read lock.
332 struct xfs_perag *pag,
333 struct xfs_inode *ip) __releases(&ip->i_flags_lock)
335 struct xfs_mount *mp = ip->i_mount;
336 struct inode *inode = VFS_I(ip);
339 trace_xfs_iget_recycle(ip);
342 * We need to make it look like the inode is being reclaimed to prevent
343 * the actual reclaim workers from stomping over us while we recycle
344 * the inode. We can't clear the radix tree tag yet as it requires
345 * pag_ici_lock to be held exclusive.
347 ip->i_flags |= XFS_IRECLAIM;
349 spin_unlock(&ip->i_flags_lock);
352 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
353 error = xfs_reinit_inode(mp, inode);
356 * Re-initializing the inode failed, and we are in deep
357 * trouble. Try to re-add it to the reclaim list.
360 spin_lock(&ip->i_flags_lock);
361 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
362 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
363 spin_unlock(&ip->i_flags_lock);
366 trace_xfs_iget_recycle_fail(ip);
370 spin_lock(&pag->pag_ici_lock);
371 spin_lock(&ip->i_flags_lock);
374 * Clear the per-lifetime state in the inode as we are now effectively
375 * a new inode and need to return to the initial state before reuse
378 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
379 ip->i_flags |= XFS_INEW;
380 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
381 XFS_ICI_RECLAIM_TAG);
382 inode->i_state = I_NEW;
383 spin_unlock(&ip->i_flags_lock);
384 spin_unlock(&pag->pag_ici_lock);
390 * If we are allocating a new inode, then check what was returned is
391 * actually a free, empty inode. If we are not allocating an inode,
392 * then check we didn't find a free inode.
395 * 0 if the inode free state matches the lookup context
396 * -ENOENT if the inode is free and we are not allocating
397 * -EFSCORRUPTED if there is any state mismatch at all
400 xfs_iget_check_free_state(
401 struct xfs_inode *ip,
404 if (flags & XFS_IGET_CREATE) {
405 /* should be a free inode */
406 if (VFS_I(ip)->i_mode != 0) {
407 xfs_warn(ip->i_mount,
408 "Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
409 ip->i_ino, VFS_I(ip)->i_mode);
410 return -EFSCORRUPTED;
413 if (ip->i_nblocks != 0) {
414 xfs_warn(ip->i_mount,
415 "Corruption detected! Free inode 0x%llx has blocks allocated!",
417 return -EFSCORRUPTED;
422 /* should be an allocated inode */
423 if (VFS_I(ip)->i_mode == 0)
429 /* Make all pending inactivation work start immediately. */
431 xfs_inodegc_queue_all(
432 struct xfs_mount *mp)
434 struct xfs_inodegc *gc;
437 for_each_online_cpu(cpu) {
438 gc = per_cpu_ptr(mp->m_inodegc, cpu);
439 if (!llist_empty(&gc->list))
440 queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
445 * Check the validity of the inode we just found it the cache
449 struct xfs_perag *pag,
450 struct xfs_inode *ip,
453 int lock_flags) __releases(RCU)
455 struct inode *inode = VFS_I(ip);
456 struct xfs_mount *mp = ip->i_mount;
460 * check for re-use of an inode within an RCU grace period due to the
461 * radix tree nodes not being updated yet. We monitor for this by
462 * setting the inode number to zero before freeing the inode structure.
463 * If the inode has been reallocated and set up, then the inode number
464 * will not match, so check for that, too.
466 spin_lock(&ip->i_flags_lock);
467 if (ip->i_ino != ino)
471 * If we are racing with another cache hit that is currently
472 * instantiating this inode or currently recycling it out of
473 * reclaimable state, wait for the initialisation to complete
476 * If we're racing with the inactivation worker we also want to wait.
477 * If we're creating a new file, it's possible that the worker
478 * previously marked the inode as free on disk but hasn't finished
479 * updating the incore state yet. The AGI buffer will be dirty and
480 * locked to the icreate transaction, so a synchronous push of the
481 * inodegc workers would result in deadlock. For a regular iget, the
482 * worker is running already, so we might as well wait.
484 * XXX(hch): eventually we should do something equivalent to
485 * wait_on_inode to wait for these flags to be cleared
486 * instead of polling for it.
488 if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
491 if (ip->i_flags & XFS_NEED_INACTIVE) {
492 /* Unlinked inodes cannot be re-grabbed. */
493 if (VFS_I(ip)->i_nlink == 0) {
497 goto out_inodegc_flush;
501 * Check the inode free state is valid. This also detects lookup
502 * racing with unlinks.
504 error = xfs_iget_check_free_state(ip, flags);
508 /* Skip inodes that have no vfs state. */
509 if ((flags & XFS_IGET_INCORE) &&
510 (ip->i_flags & XFS_IRECLAIMABLE))
513 /* The inode fits the selection criteria; process it. */
514 if (ip->i_flags & XFS_IRECLAIMABLE) {
515 /* Drops i_flags_lock and RCU read lock. */
516 error = xfs_iget_recycle(pag, ip);
520 /* If the VFS inode is being torn down, pause and try again. */
524 /* We've got a live one. */
525 spin_unlock(&ip->i_flags_lock);
527 trace_xfs_iget_hit(ip);
531 xfs_ilock(ip, lock_flags);
533 if (!(flags & XFS_IGET_INCORE))
534 xfs_iflags_clear(ip, XFS_ISTALE);
535 XFS_STATS_INC(mp, xs_ig_found);
540 trace_xfs_iget_skip(ip);
541 XFS_STATS_INC(mp, xs_ig_frecycle);
544 spin_unlock(&ip->i_flags_lock);
549 spin_unlock(&ip->i_flags_lock);
552 * Do not wait for the workers, because the caller could hold an AGI
553 * buffer lock. We're just going to sleep in a loop anyway.
555 if (xfs_is_inodegc_enabled(mp))
556 xfs_inodegc_queue_all(mp);
562 struct xfs_mount *mp,
563 struct xfs_perag *pag,
566 struct xfs_inode **ipp,
570 struct xfs_inode *ip;
572 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
575 ip = xfs_inode_alloc(mp, ino);
579 error = xfs_imap(mp, tp, ip->i_ino, &ip->i_imap, flags);
584 * For version 5 superblocks, if we are initialising a new inode and we
585 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can
586 * simply build the new inode core with a random generation number.
588 * For version 4 (and older) superblocks, log recovery is dependent on
589 * the i_flushiter field being initialised from the current on-disk
590 * value and hence we must also read the inode off disk even when
591 * initializing new inodes.
593 if (xfs_has_v3inodes(mp) &&
594 (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
595 VFS_I(ip)->i_generation = prandom_u32();
599 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
603 error = xfs_inode_from_disk(ip,
604 xfs_buf_offset(bp, ip->i_imap.im_boffset));
606 xfs_buf_set_ref(bp, XFS_INO_REF);
607 xfs_trans_brelse(tp, bp);
613 trace_xfs_iget_miss(ip);
616 * Check the inode free state is valid. This also detects lookup
617 * racing with unlinks.
619 error = xfs_iget_check_free_state(ip, flags);
624 * Preload the radix tree so we can insert safely under the
625 * write spinlock. Note that we cannot sleep inside the preload
626 * region. Since we can be called from transaction context, don't
627 * recurse into the file system.
629 if (radix_tree_preload(GFP_NOFS)) {
635 * Because the inode hasn't been added to the radix-tree yet it can't
636 * be found by another thread, so we can do the non-sleeping lock here.
639 if (!xfs_ilock_nowait(ip, lock_flags))
644 * These values must be set before inserting the inode into the radix
645 * tree as the moment it is inserted a concurrent lookup (allowed by the
646 * RCU locking mechanism) can find it and that lookup must see that this
647 * is an inode currently under construction (i.e. that XFS_INEW is set).
648 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
649 * memory barrier that ensures this detection works correctly at lookup
653 if (flags & XFS_IGET_DONTCACHE)
654 d_mark_dontcache(VFS_I(ip));
658 xfs_iflags_set(ip, iflags);
660 /* insert the new inode */
661 spin_lock(&pag->pag_ici_lock);
662 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
663 if (unlikely(error)) {
664 WARN_ON(error != -EEXIST);
665 XFS_STATS_INC(mp, xs_ig_dup);
667 goto out_preload_end;
669 spin_unlock(&pag->pag_ici_lock);
670 radix_tree_preload_end();
676 spin_unlock(&pag->pag_ici_lock);
677 radix_tree_preload_end();
679 xfs_iunlock(ip, lock_flags);
681 __destroy_inode(VFS_I(ip));
687 * Look up an inode by number in the given file system. The inode is looked up
688 * in the cache held in each AG. If the inode is found in the cache, initialise
689 * the vfs inode if necessary.
691 * If it is not in core, read it in from the file system's device, add it to the
692 * cache and initialise the vfs inode.
694 * The inode is locked according to the value of the lock_flags parameter.
695 * Inode lookup is only done during metadata operations and not as part of the
696 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
700 struct xfs_mount *mp,
701 struct xfs_trans *tp,
705 struct xfs_inode **ipp)
707 struct xfs_inode *ip;
708 struct xfs_perag *pag;
712 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
714 /* reject inode numbers outside existing AGs */
715 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
718 XFS_STATS_INC(mp, xs_ig_attempts);
720 /* get the perag structure and ensure that it's inode capable */
721 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
722 agino = XFS_INO_TO_AGINO(mp, ino);
727 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
730 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
732 goto out_error_or_again;
735 if (flags & XFS_IGET_INCORE) {
737 goto out_error_or_again;
739 XFS_STATS_INC(mp, xs_ig_missed);
741 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
744 goto out_error_or_again;
751 * If we have a real type for an on-disk inode, we can setup the inode
752 * now. If it's a new inode being created, xfs_ialloc will handle it.
754 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
755 xfs_setup_existing_inode(ip);
759 if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
768 * "Is this a cached inode that's also allocated?"
770 * Look up an inode by number in the given file system. If the inode is
771 * in cache and isn't in purgatory, return 1 if the inode is allocated
772 * and 0 if it is not. For all other cases (not in cache, being torn
773 * down, etc.), return a negative error code.
775 * The caller has to prevent inode allocation and freeing activity,
776 * presumably by locking the AGI buffer. This is to ensure that an
777 * inode cannot transition from allocated to freed until the caller is
778 * ready to allow that. If the inode is in an intermediate state (new,
779 * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
780 * inode is not in the cache, -ENOENT will be returned. The caller must
781 * deal with these scenarios appropriately.
783 * This is a specialized use case for the online scrubber; if you're
784 * reading this, you probably want xfs_iget.
787 xfs_icache_inode_is_allocated(
788 struct xfs_mount *mp,
789 struct xfs_trans *tp,
793 struct xfs_inode *ip;
796 error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
800 *inuse = !!(VFS_I(ip)->i_mode);
806 * Grab the inode for reclaim exclusively.
808 * We have found this inode via a lookup under RCU, so the inode may have
809 * already been freed, or it may be in the process of being recycled by
810 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
811 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
812 * will not be set. Hence we need to check for both these flag conditions to
813 * avoid inodes that are no longer reclaim candidates.
815 * Note: checking for other state flags here, under the i_flags_lock or not, is
816 * racy and should be avoided. Those races should be resolved only after we have
817 * ensured that we are able to reclaim this inode and the world can see that we
818 * are going to reclaim it.
820 * Return true if we grabbed it, false otherwise.
824 struct xfs_inode *ip,
825 struct xfs_icwalk *icw)
827 ASSERT(rcu_read_lock_held());
829 spin_lock(&ip->i_flags_lock);
830 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
831 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
832 /* not a reclaim candidate. */
833 spin_unlock(&ip->i_flags_lock);
837 /* Don't reclaim a sick inode unless the caller asked for it. */
839 (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
840 spin_unlock(&ip->i_flags_lock);
844 __xfs_iflags_set(ip, XFS_IRECLAIM);
845 spin_unlock(&ip->i_flags_lock);
850 * Inode reclaim is non-blocking, so the default action if progress cannot be
851 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
852 * XFS_IRECLAIM flag. If we are in a shutdown state, we don't care about
853 * blocking anymore and hence we can wait for the inode to be able to reclaim
856 * We do no IO here - if callers require inodes to be cleaned they must push the
857 * AIL first to trigger writeback of dirty inodes. This enables writeback to be
858 * done in the background in a non-blocking manner, and enables memory reclaim
859 * to make progress without blocking.
863 struct xfs_inode *ip,
864 struct xfs_perag *pag)
866 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
868 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
870 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
873 if (xfs_is_shutdown(ip->i_mount)) {
875 xfs_iflush_abort(ip);
878 if (xfs_ipincount(ip))
879 goto out_clear_flush;
880 if (!xfs_inode_clean(ip))
881 goto out_clear_flush;
883 xfs_iflags_clear(ip, XFS_IFLUSHING);
885 trace_xfs_inode_reclaiming(ip);
888 * Because we use RCU freeing we need to ensure the inode always appears
889 * to be reclaimed with an invalid inode number when in the free state.
890 * We do this as early as possible under the ILOCK so that
891 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
892 * detect races with us here. By doing this, we guarantee that once
893 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
894 * it will see either a valid inode that will serialise correctly, or it
895 * will see an invalid inode that it can skip.
897 spin_lock(&ip->i_flags_lock);
898 ip->i_flags = XFS_IRECLAIM;
902 spin_unlock(&ip->i_flags_lock);
904 xfs_iunlock(ip, XFS_ILOCK_EXCL);
906 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
908 * Remove the inode from the per-AG radix tree.
910 * Because radix_tree_delete won't complain even if the item was never
911 * added to the tree assert that it's been there before to catch
912 * problems with the inode life time early on.
914 spin_lock(&pag->pag_ici_lock);
915 if (!radix_tree_delete(&pag->pag_ici_root,
916 XFS_INO_TO_AGINO(ip->i_mount, ino)))
918 xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
919 spin_unlock(&pag->pag_ici_lock);
922 * Here we do an (almost) spurious inode lock in order to coordinate
923 * with inode cache radix tree lookups. This is because the lookup
924 * can reference the inodes in the cache without taking references.
926 * We make that OK here by ensuring that we wait until the inode is
927 * unlocked after the lookup before we go ahead and free it.
929 xfs_ilock(ip, XFS_ILOCK_EXCL);
930 ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
931 xfs_iunlock(ip, XFS_ILOCK_EXCL);
932 ASSERT(xfs_inode_clean(ip));
934 __xfs_inode_free(ip);
938 xfs_iflags_clear(ip, XFS_IFLUSHING);
940 xfs_iunlock(ip, XFS_ILOCK_EXCL);
942 xfs_iflags_clear(ip, XFS_IRECLAIM);
945 /* Reclaim sick inodes if we're unmounting or the fs went down. */
947 xfs_want_reclaim_sick(
948 struct xfs_mount *mp)
950 return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
956 struct xfs_mount *mp)
958 struct xfs_icwalk icw = {
962 if (xfs_want_reclaim_sick(mp))
963 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
965 while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
966 xfs_ail_push_all_sync(mp->m_ail);
967 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
972 * The shrinker infrastructure determines how many inodes we should scan for
973 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
974 * push the AIL here. We also want to proactively free up memory if we can to
975 * minimise the amount of work memory reclaim has to do so we kick the
976 * background reclaim if it isn't already scheduled.
979 xfs_reclaim_inodes_nr(
980 struct xfs_mount *mp,
981 unsigned long nr_to_scan)
983 struct xfs_icwalk icw = {
984 .icw_flags = XFS_ICWALK_FLAG_SCAN_LIMIT,
985 .icw_scan_limit = min_t(unsigned long, LONG_MAX, nr_to_scan),
988 if (xfs_want_reclaim_sick(mp))
989 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
991 /* kick background reclaimer and push the AIL */
992 xfs_reclaim_work_queue(mp);
993 xfs_ail_push_all(mp->m_ail);
995 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1000 * Return the number of reclaimable inodes in the filesystem for
1001 * the shrinker to determine how much to reclaim.
1004 xfs_reclaim_inodes_count(
1005 struct xfs_mount *mp)
1007 struct xfs_perag *pag;
1008 xfs_agnumber_t ag = 0;
1009 long reclaimable = 0;
1011 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1012 ag = pag->pag_agno + 1;
1013 reclaimable += pag->pag_ici_reclaimable;
1020 xfs_icwalk_match_id(
1021 struct xfs_inode *ip,
1022 struct xfs_icwalk *icw)
1024 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1025 !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1028 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1029 !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1032 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1033 ip->i_projid != icw->icw_prid)
1040 * A union-based inode filtering algorithm. Process the inode if any of the
1041 * criteria match. This is for global/internal scans only.
1044 xfs_icwalk_match_id_union(
1045 struct xfs_inode *ip,
1046 struct xfs_icwalk *icw)
1048 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1049 uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1052 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1053 gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1056 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1057 ip->i_projid == icw->icw_prid)
1064 * Is this inode @ip eligible for eof/cow block reclamation, given some
1065 * filtering parameters @icw? The inode is eligible if @icw is null or
1066 * if the predicate functions match.
1070 struct xfs_inode *ip,
1071 struct xfs_icwalk *icw)
1078 if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1079 match = xfs_icwalk_match_id_union(ip, icw);
1081 match = xfs_icwalk_match_id(ip, icw);
1085 /* skip the inode if the file size is too small */
1086 if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1087 XFS_ISIZE(ip) < icw->icw_min_file_size)
1094 * This is a fast pass over the inode cache to try to get reclaim moving on as
1095 * many inodes as possible in a short period of time. It kicks itself every few
1096 * seconds, as well as being kicked by the inode cache shrinker when memory
1101 struct work_struct *work)
1103 struct xfs_mount *mp = container_of(to_delayed_work(work),
1104 struct xfs_mount, m_reclaim_work);
1106 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1107 xfs_reclaim_work_queue(mp);
1111 xfs_inode_free_eofblocks(
1112 struct xfs_inode *ip,
1113 struct xfs_icwalk *icw,
1114 unsigned int *lockflags)
1118 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1120 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1124 * If the mapping is dirty the operation can block and wait for some
1125 * time. Unless we are waiting, skip it.
1127 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1130 if (!xfs_icwalk_match(ip, icw))
1134 * If the caller is waiting, return -EAGAIN to keep the background
1135 * scanner moving and revisit the inode in a subsequent pass.
1137 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1142 *lockflags |= XFS_IOLOCK_EXCL;
1144 if (xfs_can_free_eofblocks(ip, false))
1145 return xfs_free_eofblocks(ip);
1147 /* inode could be preallocated or append-only */
1148 trace_xfs_inode_free_eofblocks_invalid(ip);
1149 xfs_inode_clear_eofblocks_tag(ip);
1154 xfs_blockgc_set_iflag(
1155 struct xfs_inode *ip,
1156 unsigned long iflag)
1158 struct xfs_mount *mp = ip->i_mount;
1159 struct xfs_perag *pag;
1161 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1164 * Don't bother locking the AG and looking up in the radix trees
1165 * if we already know that we have the tag set.
1167 if (ip->i_flags & iflag)
1169 spin_lock(&ip->i_flags_lock);
1170 ip->i_flags |= iflag;
1171 spin_unlock(&ip->i_flags_lock);
1173 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1174 spin_lock(&pag->pag_ici_lock);
1176 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1177 XFS_ICI_BLOCKGC_TAG);
1179 spin_unlock(&pag->pag_ici_lock);
1184 xfs_inode_set_eofblocks_tag(
1187 trace_xfs_inode_set_eofblocks_tag(ip);
1188 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1192 xfs_blockgc_clear_iflag(
1193 struct xfs_inode *ip,
1194 unsigned long iflag)
1196 struct xfs_mount *mp = ip->i_mount;
1197 struct xfs_perag *pag;
1200 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1202 spin_lock(&ip->i_flags_lock);
1203 ip->i_flags &= ~iflag;
1204 clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1205 spin_unlock(&ip->i_flags_lock);
1210 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1211 spin_lock(&pag->pag_ici_lock);
1213 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1214 XFS_ICI_BLOCKGC_TAG);
1216 spin_unlock(&pag->pag_ici_lock);
1221 xfs_inode_clear_eofblocks_tag(
1224 trace_xfs_inode_clear_eofblocks_tag(ip);
1225 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1229 * Set ourselves up to free CoW blocks from this file. If it's already clean
1230 * then we can bail out quickly, but otherwise we must back off if the file
1231 * is undergoing some kind of write.
1234 xfs_prep_free_cowblocks(
1235 struct xfs_inode *ip)
1238 * Just clear the tag if we have an empty cow fork or none at all. It's
1239 * possible the inode was fully unshared since it was originally tagged.
1241 if (!xfs_inode_has_cow_data(ip)) {
1242 trace_xfs_inode_free_cowblocks_invalid(ip);
1243 xfs_inode_clear_cowblocks_tag(ip);
1248 * If the mapping is dirty or under writeback we cannot touch the
1249 * CoW fork. Leave it alone if we're in the midst of a directio.
1251 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1252 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1253 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1254 atomic_read(&VFS_I(ip)->i_dio_count))
1261 * Automatic CoW Reservation Freeing
1263 * These functions automatically garbage collect leftover CoW reservations
1264 * that were made on behalf of a cowextsize hint when we start to run out
1265 * of quota or when the reservations sit around for too long. If the file
1266 * has dirty pages or is undergoing writeback, its CoW reservations will
1269 * The actual garbage collection piggybacks off the same code that runs
1270 * the speculative EOF preallocation garbage collector.
1273 xfs_inode_free_cowblocks(
1274 struct xfs_inode *ip,
1275 struct xfs_icwalk *icw,
1276 unsigned int *lockflags)
1281 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1283 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1286 if (!xfs_prep_free_cowblocks(ip))
1289 if (!xfs_icwalk_match(ip, icw))
1293 * If the caller is waiting, return -EAGAIN to keep the background
1294 * scanner moving and revisit the inode in a subsequent pass.
1296 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1297 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1302 *lockflags |= XFS_IOLOCK_EXCL;
1304 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1309 *lockflags |= XFS_MMAPLOCK_EXCL;
1312 * Check again, nobody else should be able to dirty blocks or change
1313 * the reflink iflag now that we have the first two locks held.
1315 if (xfs_prep_free_cowblocks(ip))
1316 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1321 xfs_inode_set_cowblocks_tag(
1324 trace_xfs_inode_set_cowblocks_tag(ip);
1325 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1329 xfs_inode_clear_cowblocks_tag(
1332 trace_xfs_inode_clear_cowblocks_tag(ip);
1333 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1336 /* Disable post-EOF and CoW block auto-reclamation. */
1339 struct xfs_mount *mp)
1341 struct xfs_perag *pag;
1342 xfs_agnumber_t agno;
1344 if (!xfs_clear_blockgc_enabled(mp))
1347 for_each_perag(mp, agno, pag)
1348 cancel_delayed_work_sync(&pag->pag_blockgc_work);
1349 trace_xfs_blockgc_stop(mp, __return_address);
1352 /* Enable post-EOF and CoW block auto-reclamation. */
1355 struct xfs_mount *mp)
1357 struct xfs_perag *pag;
1358 xfs_agnumber_t agno;
1360 if (xfs_set_blockgc_enabled(mp))
1363 trace_xfs_blockgc_start(mp, __return_address);
1364 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1365 xfs_blockgc_queue(pag);
1368 /* Don't try to run block gc on an inode that's in any of these states. */
1369 #define XFS_BLOCKGC_NOGRAB_IFLAGS (XFS_INEW | \
1370 XFS_NEED_INACTIVE | \
1371 XFS_INACTIVATING | \
1372 XFS_IRECLAIMABLE | \
1375 * Decide if the given @ip is eligible for garbage collection of speculative
1376 * preallocations, and grab it if so. Returns true if it's ready to go or
1377 * false if we should just ignore it.
1381 struct xfs_inode *ip)
1383 struct inode *inode = VFS_I(ip);
1385 ASSERT(rcu_read_lock_held());
1387 /* Check for stale RCU freed inode */
1388 spin_lock(&ip->i_flags_lock);
1390 goto out_unlock_noent;
1392 if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1393 goto out_unlock_noent;
1394 spin_unlock(&ip->i_flags_lock);
1396 /* nothing to sync during shutdown */
1397 if (xfs_is_shutdown(ip->i_mount))
1400 /* If we can't grab the inode, it must on it's way to reclaim. */
1404 /* inode is valid */
1408 spin_unlock(&ip->i_flags_lock);
1412 /* Scan one incore inode for block preallocations that we can remove. */
1414 xfs_blockgc_scan_inode(
1415 struct xfs_inode *ip,
1416 struct xfs_icwalk *icw)
1418 unsigned int lockflags = 0;
1421 error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1425 error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1428 xfs_iunlock(ip, lockflags);
1433 /* Background worker that trims preallocated space. */
1436 struct work_struct *work)
1438 struct xfs_perag *pag = container_of(to_delayed_work(work),
1439 struct xfs_perag, pag_blockgc_work);
1440 struct xfs_mount *mp = pag->pag_mount;
1443 trace_xfs_blockgc_worker(mp, __return_address);
1445 error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1447 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1448 pag->pag_agno, error);
1449 xfs_blockgc_queue(pag);
1453 * Try to free space in the filesystem by purging inactive inodes, eofblocks
1457 xfs_blockgc_free_space(
1458 struct xfs_mount *mp,
1459 struct xfs_icwalk *icw)
1463 trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1465 error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1469 xfs_inodegc_flush(mp);
1474 * Reclaim all the free space that we can by scheduling the background blockgc
1475 * and inodegc workers immediately and waiting for them all to clear.
1478 xfs_blockgc_flush_all(
1479 struct xfs_mount *mp)
1481 struct xfs_perag *pag;
1482 xfs_agnumber_t agno;
1484 trace_xfs_blockgc_flush_all(mp, __return_address);
1487 * For each blockgc worker, move its queue time up to now. If it
1488 * wasn't queued, it will not be requeued. Then flush whatever's
1491 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1492 mod_delayed_work(pag->pag_mount->m_blockgc_wq,
1493 &pag->pag_blockgc_work, 0);
1495 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1496 flush_delayed_work(&pag->pag_blockgc_work);
1498 xfs_inodegc_flush(mp);
1502 * Run cow/eofblocks scans on the supplied dquots. We don't know exactly which
1503 * quota caused an allocation failure, so we make a best effort by including
1504 * each quota under low free space conditions (less than 1% free space) in the
1507 * Callers must not hold any inode's ILOCK. If requesting a synchronous scan
1508 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1512 xfs_blockgc_free_dquots(
1513 struct xfs_mount *mp,
1514 struct xfs_dquot *udqp,
1515 struct xfs_dquot *gdqp,
1516 struct xfs_dquot *pdqp,
1517 unsigned int iwalk_flags)
1519 struct xfs_icwalk icw = {0};
1520 bool do_work = false;
1522 if (!udqp && !gdqp && !pdqp)
1526 * Run a scan to free blocks using the union filter to cover all
1527 * applicable quotas in a single scan.
1529 icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1531 if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1532 icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1533 icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1537 if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1538 icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1539 icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1543 if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1544 icw.icw_prid = pdqp->q_id;
1545 icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1552 return xfs_blockgc_free_space(mp, &icw);
1555 /* Run cow/eofblocks scans on the quotas attached to the inode. */
1557 xfs_blockgc_free_quota(
1558 struct xfs_inode *ip,
1559 unsigned int iwalk_flags)
1561 return xfs_blockgc_free_dquots(ip->i_mount,
1562 xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1563 xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1564 xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1567 /* XFS Inode Cache Walking Code */
1570 * The inode lookup is done in batches to keep the amount of lock traffic and
1571 * radix tree lookups to a minimum. The batch size is a trade off between
1572 * lookup reduction and stack usage. This is in the reclaim path, so we can't
1575 #define XFS_LOOKUP_BATCH 32
1579 * Decide if we want to grab this inode in anticipation of doing work towards
1584 enum xfs_icwalk_goal goal,
1585 struct xfs_inode *ip,
1586 struct xfs_icwalk *icw)
1589 case XFS_ICWALK_BLOCKGC:
1590 return xfs_blockgc_igrab(ip);
1591 case XFS_ICWALK_RECLAIM:
1592 return xfs_reclaim_igrab(ip, icw);
1599 * Process an inode. Each processing function must handle any state changes
1600 * made by the icwalk igrab function. Return -EAGAIN to skip an inode.
1603 xfs_icwalk_process_inode(
1604 enum xfs_icwalk_goal goal,
1605 struct xfs_inode *ip,
1606 struct xfs_perag *pag,
1607 struct xfs_icwalk *icw)
1612 case XFS_ICWALK_BLOCKGC:
1613 error = xfs_blockgc_scan_inode(ip, icw);
1615 case XFS_ICWALK_RECLAIM:
1616 xfs_reclaim_inode(ip, pag);
1623 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1624 * process them in some manner.
1628 struct xfs_perag *pag,
1629 enum xfs_icwalk_goal goal,
1630 struct xfs_icwalk *icw)
1632 struct xfs_mount *mp = pag->pag_mount;
1633 uint32_t first_index;
1642 if (goal == XFS_ICWALK_RECLAIM)
1643 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1648 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1654 nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1655 (void **) batch, first_index,
1656 XFS_LOOKUP_BATCH, goal);
1664 * Grab the inodes before we drop the lock. if we found
1665 * nothing, nr == 0 and the loop will be skipped.
1667 for (i = 0; i < nr_found; i++) {
1668 struct xfs_inode *ip = batch[i];
1670 if (done || !xfs_icwalk_igrab(goal, ip, icw))
1674 * Update the index for the next lookup. Catch
1675 * overflows into the next AG range which can occur if
1676 * we have inodes in the last block of the AG and we
1677 * are currently pointing to the last inode.
1679 * Because we may see inodes that are from the wrong AG
1680 * due to RCU freeing and reallocation, only update the
1681 * index if it lies in this AG. It was a race that lead
1682 * us to see this inode, so another lookup from the
1683 * same index will not find it again.
1685 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
1687 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1688 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1692 /* unlock now we've grabbed the inodes. */
1695 for (i = 0; i < nr_found; i++) {
1698 error = xfs_icwalk_process_inode(goal, batch[i], pag,
1700 if (error == -EAGAIN) {
1704 if (error && last_error != -EFSCORRUPTED)
1708 /* bail out if the filesystem is corrupted. */
1709 if (error == -EFSCORRUPTED)
1714 if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1715 icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1716 if (icw->icw_scan_limit <= 0)
1719 } while (nr_found && !done);
1721 if (goal == XFS_ICWALK_RECLAIM) {
1724 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1734 /* Walk all incore inodes to achieve a given goal. */
1737 struct xfs_mount *mp,
1738 enum xfs_icwalk_goal goal,
1739 struct xfs_icwalk *icw)
1741 struct xfs_perag *pag;
1744 xfs_agnumber_t agno;
1746 for_each_perag_tag(mp, agno, pag, goal) {
1747 error = xfs_icwalk_ag(pag, goal, icw);
1750 if (error == -EFSCORRUPTED) {
1757 BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1763 struct xfs_inode *ip,
1766 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, whichfork);
1767 struct xfs_bmbt_irec got;
1768 struct xfs_iext_cursor icur;
1770 if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1773 if (isnullstartblock(got.br_startblock)) {
1774 xfs_warn(ip->i_mount,
1775 "ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",
1777 whichfork == XFS_DATA_FORK ? "data" : "cow",
1778 got.br_startoff, got.br_blockcount);
1780 } while (xfs_iext_next_extent(ifp, &icur, &got));
1783 #define xfs_check_delalloc(ip, whichfork) do { } while (0)
1786 /* Schedule the inode for reclaim. */
1788 xfs_inodegc_set_reclaimable(
1789 struct xfs_inode *ip)
1791 struct xfs_mount *mp = ip->i_mount;
1792 struct xfs_perag *pag;
1794 if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {
1795 xfs_check_delalloc(ip, XFS_DATA_FORK);
1796 xfs_check_delalloc(ip, XFS_COW_FORK);
1800 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1801 spin_lock(&pag->pag_ici_lock);
1802 spin_lock(&ip->i_flags_lock);
1804 trace_xfs_inode_set_reclaimable(ip);
1805 ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);
1806 ip->i_flags |= XFS_IRECLAIMABLE;
1807 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1808 XFS_ICI_RECLAIM_TAG);
1810 spin_unlock(&ip->i_flags_lock);
1811 spin_unlock(&pag->pag_ici_lock);
1816 * Free all speculative preallocations and possibly even the inode itself.
1817 * This is the last chance to make changes to an otherwise unreferenced file
1818 * before incore reclamation happens.
1821 xfs_inodegc_inactivate(
1822 struct xfs_inode *ip)
1824 trace_xfs_inode_inactivating(ip);
1826 xfs_inodegc_set_reclaimable(ip);
1831 struct work_struct *work)
1833 struct xfs_inodegc *gc = container_of(work, struct xfs_inodegc,
1835 struct llist_node *node = llist_del_all(&gc->list);
1836 struct xfs_inode *ip, *n;
1838 WRITE_ONCE(gc->items, 0);
1843 ip = llist_entry(node, struct xfs_inode, i_gclist);
1844 trace_xfs_inodegc_worker(ip->i_mount, READ_ONCE(gc->shrinker_hits));
1846 WRITE_ONCE(gc->shrinker_hits, 0);
1847 llist_for_each_entry_safe(ip, n, node, i_gclist) {
1848 xfs_iflags_set(ip, XFS_INACTIVATING);
1849 xfs_inodegc_inactivate(ip);
1854 * Force all currently queued inode inactivation work to run immediately, and
1855 * wait for the work to finish. Two pass - queue all the work first pass, wait
1856 * for it in a second pass.
1860 struct xfs_mount *mp)
1862 struct xfs_inodegc *gc;
1865 if (!xfs_is_inodegc_enabled(mp))
1868 trace_xfs_inodegc_flush(mp, __return_address);
1870 xfs_inodegc_queue_all(mp);
1872 for_each_online_cpu(cpu) {
1873 gc = per_cpu_ptr(mp->m_inodegc, cpu);
1874 flush_work(&gc->work);
1879 * Flush all the pending work and then disable the inode inactivation background
1880 * workers and wait for them to stop.
1884 struct xfs_mount *mp)
1886 struct xfs_inodegc *gc;
1889 if (!xfs_clear_inodegc_enabled(mp))
1892 xfs_inodegc_queue_all(mp);
1894 for_each_online_cpu(cpu) {
1895 gc = per_cpu_ptr(mp->m_inodegc, cpu);
1896 cancel_work_sync(&gc->work);
1898 trace_xfs_inodegc_stop(mp, __return_address);
1902 * Enable the inode inactivation background workers and schedule deferred inode
1903 * inactivation work if there is any.
1907 struct xfs_mount *mp)
1909 if (xfs_set_inodegc_enabled(mp))
1912 trace_xfs_inodegc_start(mp, __return_address);
1913 xfs_inodegc_queue_all(mp);
1916 #ifdef CONFIG_XFS_RT
1918 xfs_inodegc_want_queue_rt_file(
1919 struct xfs_inode *ip)
1921 struct xfs_mount *mp = ip->i_mount;
1924 if (!XFS_IS_REALTIME_INODE(ip))
1927 freertx = READ_ONCE(mp->m_sb.sb_frextents);
1928 return freertx < mp->m_low_rtexts[XFS_LOWSP_5_PCNT];
1931 # define xfs_inodegc_want_queue_rt_file(ip) (false)
1932 #endif /* CONFIG_XFS_RT */
1935 * Schedule the inactivation worker when:
1937 * - We've accumulated more than one inode cluster buffer's worth of inodes.
1938 * - There is less than 5% free space left.
1939 * - Any of the quotas for this inode are near an enforcement limit.
1942 xfs_inodegc_want_queue_work(
1943 struct xfs_inode *ip,
1946 struct xfs_mount *mp = ip->i_mount;
1948 if (items > mp->m_ino_geo.inodes_per_cluster)
1951 if (__percpu_counter_compare(&mp->m_fdblocks,
1952 mp->m_low_space[XFS_LOWSP_5_PCNT],
1953 XFS_FDBLOCKS_BATCH) < 0)
1956 if (xfs_inodegc_want_queue_rt_file(ip))
1959 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
1962 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
1965 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
1972 * Upper bound on the number of inodes in each AG that can be queued for
1973 * inactivation at any given time, to avoid monopolizing the workqueue.
1975 #define XFS_INODEGC_MAX_BACKLOG (4 * XFS_INODES_PER_CHUNK)
1978 * Make the frontend wait for inactivations when:
1980 * - Memory shrinkers queued the inactivation worker and it hasn't finished.
1981 * - The queue depth exceeds the maximum allowable percpu backlog.
1983 * Note: If the current thread is running a transaction, we don't ever want to
1984 * wait for other transactions because that could introduce a deadlock.
1987 xfs_inodegc_want_flush_work(
1988 struct xfs_inode *ip,
1990 unsigned int shrinker_hits)
1992 if (current->journal_info)
1995 if (shrinker_hits > 0)
1998 if (items > XFS_INODEGC_MAX_BACKLOG)
2005 * Queue a background inactivation worker if there are inodes that need to be
2006 * inactivated and higher level xfs code hasn't disabled the background
2011 struct xfs_inode *ip)
2013 struct xfs_mount *mp = ip->i_mount;
2014 struct xfs_inodegc *gc;
2016 unsigned int shrinker_hits;
2018 trace_xfs_inode_set_need_inactive(ip);
2019 spin_lock(&ip->i_flags_lock);
2020 ip->i_flags |= XFS_NEED_INACTIVE;
2021 spin_unlock(&ip->i_flags_lock);
2023 gc = get_cpu_ptr(mp->m_inodegc);
2024 llist_add(&ip->i_gclist, &gc->list);
2025 items = READ_ONCE(gc->items);
2026 WRITE_ONCE(gc->items, items + 1);
2027 shrinker_hits = READ_ONCE(gc->shrinker_hits);
2030 if (!xfs_is_inodegc_enabled(mp))
2033 if (xfs_inodegc_want_queue_work(ip, items)) {
2034 trace_xfs_inodegc_queue(mp, __return_address);
2035 queue_work(mp->m_inodegc_wq, &gc->work);
2038 if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
2039 trace_xfs_inodegc_throttle(mp, __return_address);
2040 flush_work(&gc->work);
2045 * Fold the dead CPU inodegc queue into the current CPUs queue.
2048 xfs_inodegc_cpu_dead(
2049 struct xfs_mount *mp,
2050 unsigned int dead_cpu)
2052 struct xfs_inodegc *dead_gc, *gc;
2053 struct llist_node *first, *last;
2054 unsigned int count = 0;
2056 dead_gc = per_cpu_ptr(mp->m_inodegc, dead_cpu);
2057 cancel_work_sync(&dead_gc->work);
2059 if (llist_empty(&dead_gc->list))
2062 first = dead_gc->list.first;
2064 while (last->next) {
2068 dead_gc->list.first = NULL;
2071 /* Add pending work to current CPU */
2072 gc = get_cpu_ptr(mp->m_inodegc);
2073 llist_add_batch(first, last, &gc->list);
2074 count += READ_ONCE(gc->items);
2075 WRITE_ONCE(gc->items, count);
2078 if (xfs_is_inodegc_enabled(mp)) {
2079 trace_xfs_inodegc_queue(mp, __return_address);
2080 queue_work(mp->m_inodegc_wq, &gc->work);
2085 * We set the inode flag atomically with the radix tree tag. Once we get tag
2086 * lookups on the radix tree, this inode flag can go away.
2088 * We always use background reclaim here because even if the inode is clean, it
2089 * still may be under IO and hence we have wait for IO completion to occur
2090 * before we can reclaim the inode. The background reclaim path handles this
2091 * more efficiently than we can here, so simply let background reclaim tear down
2095 xfs_inode_mark_reclaimable(
2096 struct xfs_inode *ip)
2098 struct xfs_mount *mp = ip->i_mount;
2101 XFS_STATS_INC(mp, vn_reclaim);
2104 * We should never get here with any of the reclaim flags already set.
2106 ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2108 need_inactive = xfs_inode_needs_inactive(ip);
2109 if (need_inactive) {
2110 xfs_inodegc_queue(ip);
2114 /* Going straight to reclaim, so drop the dquots. */
2115 xfs_qm_dqdetach(ip);
2116 xfs_inodegc_set_reclaimable(ip);
2120 * Register a phony shrinker so that we can run background inodegc sooner when
2121 * there's memory pressure. Inactivation does not itself free any memory but
2122 * it does make inodes reclaimable, which eventually frees memory.
2124 * The count function, seek value, and batch value are crafted to trigger the
2125 * scan function during the second round of scanning. Hopefully this means
2126 * that we reclaimed enough memory that initiating metadata transactions won't
2127 * make things worse.
2129 #define XFS_INODEGC_SHRINKER_COUNT (1UL << DEF_PRIORITY)
2130 #define XFS_INODEGC_SHRINKER_BATCH ((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2132 static unsigned long
2133 xfs_inodegc_shrinker_count(
2134 struct shrinker *shrink,
2135 struct shrink_control *sc)
2137 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2138 m_inodegc_shrinker);
2139 struct xfs_inodegc *gc;
2142 if (!xfs_is_inodegc_enabled(mp))
2145 for_each_online_cpu(cpu) {
2146 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2147 if (!llist_empty(&gc->list))
2148 return XFS_INODEGC_SHRINKER_COUNT;
2154 static unsigned long
2155 xfs_inodegc_shrinker_scan(
2156 struct shrinker *shrink,
2157 struct shrink_control *sc)
2159 struct xfs_mount *mp = container_of(shrink, struct xfs_mount,
2160 m_inodegc_shrinker);
2161 struct xfs_inodegc *gc;
2163 bool no_items = true;
2165 if (!xfs_is_inodegc_enabled(mp))
2168 trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2170 for_each_online_cpu(cpu) {
2171 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2172 if (!llist_empty(&gc->list)) {
2173 unsigned int h = READ_ONCE(gc->shrinker_hits);
2175 WRITE_ONCE(gc->shrinker_hits, h + 1);
2176 queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
2182 * If there are no inodes to inactivate, we don't want the shrinker
2183 * to think there's deferred work to call us back about.
2191 /* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2193 xfs_inodegc_register_shrinker(
2194 struct xfs_mount *mp)
2196 struct shrinker *shrink = &mp->m_inodegc_shrinker;
2198 shrink->count_objects = xfs_inodegc_shrinker_count;
2199 shrink->scan_objects = xfs_inodegc_shrinker_scan;
2201 shrink->flags = SHRINKER_NONSLAB;
2202 shrink->batch = XFS_INODEGC_SHRINKER_BATCH;
2204 return register_shrinker(shrink);