1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_inode.h"
14 #include "xfs_trans.h"
15 #include "xfs_trans_priv.h"
16 #include "xfs_inode_item.h"
17 #include "xfs_quota.h"
18 #include "xfs_trace.h"
19 #include "xfs_icache.h"
20 #include "xfs_bmap_util.h"
21 #include "xfs_dquot_item.h"
22 #include "xfs_dquot.h"
23 #include "xfs_reflink.h"
24 #include "xfs_ialloc.h"
26 #include "xfs_log_priv.h"
27 #include "xfs_health.h"
28 #include "xfs_da_format.h"
30 #include "xfs_metafile.h"
32 #include <linux/iversion.h>
34 /* Radix tree tags for incore inode tree. */
36 /* inode is to be reclaimed */
37 #define XFS_ICI_RECLAIM_TAG 0
38 /* Inode has speculative preallocations (posteof or cow) to clean. */
39 #define XFS_ICI_BLOCKGC_TAG 1
42 * The goal for walking incore inodes. These can correspond with incore inode
43 * radix tree tags when convenient. Avoid existing XFS_IWALK namespace.
45 enum xfs_icwalk_goal {
46 /* Goals directly associated with tagged inodes. */
47 XFS_ICWALK_BLOCKGC = XFS_ICI_BLOCKGC_TAG,
48 XFS_ICWALK_RECLAIM = XFS_ICI_RECLAIM_TAG,
51 static int xfs_icwalk(struct xfs_mount *mp,
52 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
53 static int xfs_icwalk_ag(struct xfs_perag *pag,
54 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
57 * Private inode cache walk flags for struct xfs_icwalk. Must not
58 * coincide with XFS_ICWALK_FLAGS_VALID.
61 /* Stop scanning after icw_scan_limit inodes. */
62 #define XFS_ICWALK_FLAG_SCAN_LIMIT (1U << 28)
64 #define XFS_ICWALK_FLAG_RECLAIM_SICK (1U << 27)
65 #define XFS_ICWALK_FLAG_UNION (1U << 26) /* union filter algorithm */
67 #define XFS_ICWALK_PRIVATE_FLAGS (XFS_ICWALK_FLAG_SCAN_LIMIT | \
68 XFS_ICWALK_FLAG_RECLAIM_SICK | \
69 XFS_ICWALK_FLAG_UNION)
71 /* Marks for the perag xarray */
72 #define XFS_PERAG_RECLAIM_MARK XA_MARK_0
73 #define XFS_PERAG_BLOCKGC_MARK XA_MARK_1
75 static inline xa_mark_t ici_tag_to_mark(unsigned int tag)
77 if (tag == XFS_ICI_RECLAIM_TAG)
78 return XFS_PERAG_RECLAIM_MARK;
79 ASSERT(tag == XFS_ICI_BLOCKGC_TAG);
80 return XFS_PERAG_BLOCKGC_MARK;
84 * Allocate and initialise an xfs_inode.
94 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL
95 * and return NULL here on ENOMEM.
97 ip = alloc_inode_sb(mp->m_super, xfs_inode_cache, GFP_KERNEL | __GFP_NOFAIL);
99 if (inode_init_always(mp->m_super, VFS_I(ip))) {
100 kmem_cache_free(xfs_inode_cache, ip);
104 /* VFS doesn't initialise i_mode! */
105 VFS_I(ip)->i_mode = 0;
106 mapping_set_folio_min_order(VFS_I(ip)->i_mapping,
107 M_IGEO(mp)->min_folio_order);
109 XFS_STATS_INC(mp, vn_active);
110 ASSERT(atomic_read(&ip->i_pincount) == 0);
111 ASSERT(ip->i_ino == 0);
113 /* initialise the xfs inode */
116 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
118 memset(&ip->i_af, 0, sizeof(ip->i_af));
119 ip->i_af.if_format = XFS_DINODE_FMT_EXTENTS;
120 memset(&ip->i_df, 0, sizeof(ip->i_df));
122 ip->i_delayed_blks = 0;
123 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
128 INIT_WORK(&ip->i_ioend_work, xfs_end_io);
129 INIT_LIST_HEAD(&ip->i_ioend_list);
130 spin_lock_init(&ip->i_ioend_lock);
131 ip->i_next_unlinked = NULLAGINO;
132 ip->i_prev_unlinked = 0;
138 xfs_inode_free_callback(
139 struct rcu_head *head)
141 struct inode *inode = container_of(head, struct inode, i_rcu);
142 struct xfs_inode *ip = XFS_I(inode);
144 switch (VFS_I(ip)->i_mode & S_IFMT) {
148 xfs_idestroy_fork(&ip->i_df);
152 xfs_ifork_zap_attr(ip);
155 xfs_idestroy_fork(ip->i_cowfp);
156 kmem_cache_free(xfs_ifork_cache, ip->i_cowfp);
159 ASSERT(!test_bit(XFS_LI_IN_AIL,
160 &ip->i_itemp->ili_item.li_flags));
161 xfs_inode_item_destroy(ip);
165 kmem_cache_free(xfs_inode_cache, ip);
170 struct xfs_inode *ip)
172 /* asserts to verify all state is correct here */
173 ASSERT(atomic_read(&ip->i_pincount) == 0);
174 ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
175 XFS_STATS_DEC(ip->i_mount, vn_active);
177 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
182 struct xfs_inode *ip)
184 ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
187 * Because we use RCU freeing we need to ensure the inode always
188 * appears to be reclaimed with an invalid inode number when in the
189 * free state. The ip->i_flags_lock provides the barrier against lookup
192 spin_lock(&ip->i_flags_lock);
193 ip->i_flags = XFS_IRECLAIM;
195 spin_unlock(&ip->i_flags_lock);
197 __xfs_inode_free(ip);
201 * Queue background inode reclaim work if there are reclaimable inodes and there
202 * isn't reclaim work already scheduled or in progress.
205 xfs_reclaim_work_queue(
206 struct xfs_mount *mp)
210 if (xfs_group_marked(mp, XG_TYPE_AG, XFS_PERAG_RECLAIM_MARK)) {
211 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
212 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
218 * Background scanning to trim preallocated space. This is queued based on the
219 * 'speculative_prealloc_lifetime' tunable (5m by default).
223 struct xfs_perag *pag)
225 struct xfs_mount *mp = pag_mount(pag);
227 if (!xfs_is_blockgc_enabled(mp))
231 if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
232 queue_delayed_work(mp->m_blockgc_wq, &pag->pag_blockgc_work,
233 msecs_to_jiffies(xfs_blockgc_secs * 1000));
237 /* Set a tag on both the AG incore inode tree and the AG radix tree. */
239 xfs_perag_set_inode_tag(
240 struct xfs_perag *pag,
246 lockdep_assert_held(&pag->pag_ici_lock);
248 was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
249 radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
251 if (tag == XFS_ICI_RECLAIM_TAG)
252 pag->pag_ici_reclaimable++;
257 /* propagate the tag up into the pag xarray tree */
258 xfs_group_set_mark(pag_group(pag), ici_tag_to_mark(tag));
260 /* start background work */
262 case XFS_ICI_RECLAIM_TAG:
263 xfs_reclaim_work_queue(pag_mount(pag));
265 case XFS_ICI_BLOCKGC_TAG:
266 xfs_blockgc_queue(pag);
270 trace_xfs_perag_set_inode_tag(pag, _RET_IP_);
273 /* Clear a tag on both the AG incore inode tree and the AG radix tree. */
275 xfs_perag_clear_inode_tag(
276 struct xfs_perag *pag,
280 lockdep_assert_held(&pag->pag_ici_lock);
283 * Reclaim can signal (with a null agino) that it cleared its own tag
284 * by removing the inode from the radix tree.
286 if (agino != NULLAGINO)
287 radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
289 ASSERT(tag == XFS_ICI_RECLAIM_TAG);
291 if (tag == XFS_ICI_RECLAIM_TAG)
292 pag->pag_ici_reclaimable--;
294 if (radix_tree_tagged(&pag->pag_ici_root, tag))
297 /* clear the tag from the pag xarray */
298 xfs_group_clear_mark(pag_group(pag), ici_tag_to_mark(tag));
299 trace_xfs_perag_clear_inode_tag(pag, _RET_IP_);
303 * Find the next AG after @pag, or the first AG if @pag is NULL.
305 static struct xfs_perag *
306 xfs_perag_grab_next_tag(
307 struct xfs_mount *mp,
308 struct xfs_perag *pag,
311 return to_perag(xfs_group_grab_next_mark(mp,
312 pag ? pag_group(pag) : NULL,
313 ici_tag_to_mark(tag), XG_TYPE_AG));
317 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
318 * part of the structure. This is made more complex by the fact we store
319 * information about the on-disk values in the VFS inode and so we can't just
320 * overwrite the values unconditionally. Hence we save the parameters we
321 * need to retain across reinitialisation, and rewrite them into the VFS inode
322 * after reinitialisation even if it fails.
326 struct xfs_mount *mp,
330 uint32_t nlink = inode->i_nlink;
331 uint32_t generation = inode->i_generation;
332 uint64_t version = inode_peek_iversion(inode);
333 umode_t mode = inode->i_mode;
334 dev_t dev = inode->i_rdev;
335 kuid_t uid = inode->i_uid;
336 kgid_t gid = inode->i_gid;
337 unsigned long state = inode->i_state;
339 error = inode_init_always(mp->m_super, inode);
341 set_nlink(inode, nlink);
342 inode->i_generation = generation;
343 inode_set_iversion_queried(inode, version);
344 inode->i_mode = mode;
348 inode->i_state = state;
349 mapping_set_folio_min_order(inode->i_mapping,
350 M_IGEO(mp)->min_folio_order);
355 * Carefully nudge an inode whose VFS state has been torn down back into a
356 * usable state. Drops the i_flags_lock and the rcu read lock.
360 struct xfs_perag *pag,
361 struct xfs_inode *ip) __releases(&ip->i_flags_lock)
363 struct xfs_mount *mp = ip->i_mount;
364 struct inode *inode = VFS_I(ip);
367 trace_xfs_iget_recycle(ip);
369 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
373 * We need to make it look like the inode is being reclaimed to prevent
374 * the actual reclaim workers from stomping over us while we recycle
375 * the inode. We can't clear the radix tree tag yet as it requires
376 * pag_ici_lock to be held exclusive.
378 ip->i_flags |= XFS_IRECLAIM;
380 spin_unlock(&ip->i_flags_lock);
383 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
384 error = xfs_reinit_inode(mp, inode);
385 xfs_iunlock(ip, XFS_ILOCK_EXCL);
388 * Re-initializing the inode failed, and we are in deep
389 * trouble. Try to re-add it to the reclaim list.
392 spin_lock(&ip->i_flags_lock);
393 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
394 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
395 spin_unlock(&ip->i_flags_lock);
398 trace_xfs_iget_recycle_fail(ip);
402 spin_lock(&pag->pag_ici_lock);
403 spin_lock(&ip->i_flags_lock);
406 * Clear the per-lifetime state in the inode as we are now effectively
407 * a new inode and need to return to the initial state before reuse
410 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
411 ip->i_flags |= XFS_INEW;
412 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
413 XFS_ICI_RECLAIM_TAG);
414 inode->i_state = I_NEW;
415 spin_unlock(&ip->i_flags_lock);
416 spin_unlock(&pag->pag_ici_lock);
422 * If we are allocating a new inode, then check what was returned is
423 * actually a free, empty inode. If we are not allocating an inode,
424 * then check we didn't find a free inode.
427 * 0 if the inode free state matches the lookup context
428 * -ENOENT if the inode is free and we are not allocating
429 * -EFSCORRUPTED if there is any state mismatch at all
432 xfs_iget_check_free_state(
433 struct xfs_inode *ip,
436 if (flags & XFS_IGET_CREATE) {
437 /* should be a free inode */
438 if (VFS_I(ip)->i_mode != 0) {
439 xfs_warn(ip->i_mount,
440 "Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
441 ip->i_ino, VFS_I(ip)->i_mode);
442 xfs_agno_mark_sick(ip->i_mount,
443 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
445 return -EFSCORRUPTED;
448 if (ip->i_nblocks != 0) {
449 xfs_warn(ip->i_mount,
450 "Corruption detected! Free inode 0x%llx has blocks allocated!",
452 xfs_agno_mark_sick(ip->i_mount,
453 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
455 return -EFSCORRUPTED;
460 /* should be an allocated inode */
461 if (VFS_I(ip)->i_mode == 0)
467 /* Make all pending inactivation work start immediately. */
469 xfs_inodegc_queue_all(
470 struct xfs_mount *mp)
472 struct xfs_inodegc *gc;
476 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
477 gc = per_cpu_ptr(mp->m_inodegc, cpu);
478 if (!llist_empty(&gc->list)) {
479 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
487 /* Wait for all queued work and collect errors */
489 xfs_inodegc_wait_all(
490 struct xfs_mount *mp)
495 flush_workqueue(mp->m_inodegc_wq);
496 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
497 struct xfs_inodegc *gc;
499 gc = per_cpu_ptr(mp->m_inodegc, cpu);
500 if (gc->error && !error)
509 * Check the validity of the inode we just found it the cache
513 struct xfs_perag *pag,
514 struct xfs_inode *ip,
517 int lock_flags) __releases(RCU)
519 struct inode *inode = VFS_I(ip);
520 struct xfs_mount *mp = ip->i_mount;
524 * check for re-use of an inode within an RCU grace period due to the
525 * radix tree nodes not being updated yet. We monitor for this by
526 * setting the inode number to zero before freeing the inode structure.
527 * If the inode has been reallocated and set up, then the inode number
528 * will not match, so check for that, too.
530 spin_lock(&ip->i_flags_lock);
531 if (ip->i_ino != ino)
535 * If we are racing with another cache hit that is currently
536 * instantiating this inode or currently recycling it out of
537 * reclaimable state, wait for the initialisation to complete
540 * If we're racing with the inactivation worker we also want to wait.
541 * If we're creating a new file, it's possible that the worker
542 * previously marked the inode as free on disk but hasn't finished
543 * updating the incore state yet. The AGI buffer will be dirty and
544 * locked to the icreate transaction, so a synchronous push of the
545 * inodegc workers would result in deadlock. For a regular iget, the
546 * worker is running already, so we might as well wait.
548 * XXX(hch): eventually we should do something equivalent to
549 * wait_on_inode to wait for these flags to be cleared
550 * instead of polling for it.
552 if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
555 if (ip->i_flags & XFS_NEED_INACTIVE) {
556 /* Unlinked inodes cannot be re-grabbed. */
557 if (VFS_I(ip)->i_nlink == 0) {
561 goto out_inodegc_flush;
565 * Check the inode free state is valid. This also detects lookup
566 * racing with unlinks.
568 error = xfs_iget_check_free_state(ip, flags);
572 /* Skip inodes that have no vfs state. */
573 if ((flags & XFS_IGET_INCORE) &&
574 (ip->i_flags & XFS_IRECLAIMABLE))
577 /* The inode fits the selection criteria; process it. */
578 if (ip->i_flags & XFS_IRECLAIMABLE) {
579 /* Drops i_flags_lock and RCU read lock. */
580 error = xfs_iget_recycle(pag, ip);
581 if (error == -EAGAIN)
586 /* If the VFS inode is being torn down, pause and try again. */
590 /* We've got a live one. */
591 spin_unlock(&ip->i_flags_lock);
593 trace_xfs_iget_hit(ip);
597 xfs_ilock(ip, lock_flags);
599 if (!(flags & XFS_IGET_INCORE))
600 xfs_iflags_clear(ip, XFS_ISTALE);
601 XFS_STATS_INC(mp, xs_ig_found);
606 trace_xfs_iget_skip(ip);
607 XFS_STATS_INC(mp, xs_ig_frecycle);
610 spin_unlock(&ip->i_flags_lock);
615 spin_unlock(&ip->i_flags_lock);
618 * Do not wait for the workers, because the caller could hold an AGI
619 * buffer lock. We're just going to sleep in a loop anyway.
621 if (xfs_is_inodegc_enabled(mp))
622 xfs_inodegc_queue_all(mp);
628 struct xfs_mount *mp,
629 struct xfs_perag *pag,
632 struct xfs_inode **ipp,
636 struct xfs_inode *ip;
638 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
640 ip = xfs_inode_alloc(mp, ino);
644 error = xfs_imap(pag, tp, ip->i_ino, &ip->i_imap, flags);
649 * For version 5 superblocks, if we are initialising a new inode and we
650 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can
651 * simply build the new inode core with a random generation number.
653 * For version 4 (and older) superblocks, log recovery is dependent on
654 * the i_flushiter field being initialised from the current on-disk
655 * value and hence we must also read the inode off disk even when
656 * initializing new inodes.
658 if (xfs_has_v3inodes(mp) &&
659 (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
660 VFS_I(ip)->i_generation = get_random_u32();
664 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
668 error = xfs_inode_from_disk(ip,
669 xfs_buf_offset(bp, ip->i_imap.im_boffset));
671 xfs_buf_set_ref(bp, XFS_INO_REF);
673 xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
674 xfs_trans_brelse(tp, bp);
680 trace_xfs_iget_miss(ip);
683 * Check the inode free state is valid. This also detects lookup
684 * racing with unlinks.
686 error = xfs_iget_check_free_state(ip, flags);
691 * Preload the radix tree so we can insert safely under the
692 * write spinlock. Note that we cannot sleep inside the preload
695 if (radix_tree_preload(GFP_KERNEL | __GFP_NOLOCKDEP)) {
701 * Because the inode hasn't been added to the radix-tree yet it can't
702 * be found by another thread, so we can do the non-sleeping lock here.
705 if (!xfs_ilock_nowait(ip, lock_flags))
710 * These values must be set before inserting the inode into the radix
711 * tree as the moment it is inserted a concurrent lookup (allowed by the
712 * RCU locking mechanism) can find it and that lookup must see that this
713 * is an inode currently under construction (i.e. that XFS_INEW is set).
714 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
715 * memory barrier that ensures this detection works correctly at lookup
718 if (flags & XFS_IGET_DONTCACHE)
719 d_mark_dontcache(VFS_I(ip));
723 xfs_iflags_set(ip, XFS_INEW);
725 /* insert the new inode */
726 spin_lock(&pag->pag_ici_lock);
727 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
728 if (unlikely(error)) {
729 WARN_ON(error != -EEXIST);
730 XFS_STATS_INC(mp, xs_ig_dup);
732 goto out_preload_end;
734 spin_unlock(&pag->pag_ici_lock);
735 radix_tree_preload_end();
741 spin_unlock(&pag->pag_ici_lock);
742 radix_tree_preload_end();
744 xfs_iunlock(ip, lock_flags);
746 __destroy_inode(VFS_I(ip));
752 * Look up an inode by number in the given file system. The inode is looked up
753 * in the cache held in each AG. If the inode is found in the cache, initialise
754 * the vfs inode if necessary.
756 * If it is not in core, read it in from the file system's device, add it to the
757 * cache and initialise the vfs inode.
759 * The inode is locked according to the value of the lock_flags parameter.
760 * Inode lookup is only done during metadata operations and not as part of the
761 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
765 struct xfs_mount *mp,
766 struct xfs_trans *tp,
770 struct xfs_inode **ipp)
772 struct xfs_inode *ip;
773 struct xfs_perag *pag;
777 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
779 /* reject inode numbers outside existing AGs */
780 if (!xfs_verify_ino(mp, ino))
783 XFS_STATS_INC(mp, xs_ig_attempts);
785 /* get the perag structure and ensure that it's inode capable */
786 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
787 agino = XFS_INO_TO_AGINO(mp, ino);
792 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
795 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
797 goto out_error_or_again;
800 if (flags & XFS_IGET_INCORE) {
802 goto out_error_or_again;
804 XFS_STATS_INC(mp, xs_ig_missed);
806 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
809 goto out_error_or_again;
816 * If we have a real type for an on-disk inode, we can setup the inode
817 * now. If it's a new inode being created, xfs_init_new_inode will
820 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
821 xfs_setup_existing_inode(ip);
825 if (!(flags & (XFS_IGET_INCORE | XFS_IGET_NORETRY)) &&
835 * Get a metadata inode.
837 * The metafile type must match the file mode exactly, and for files in the
838 * metadata directory tree, it must match the inode's metatype exactly.
841 xfs_trans_metafile_iget(
842 struct xfs_trans *tp,
844 enum xfs_metafile_type metafile_type,
845 struct xfs_inode **ipp)
847 struct xfs_mount *mp = tp->t_mountp;
848 struct xfs_inode *ip;
852 error = xfs_iget(mp, tp, ino, 0, 0, &ip);
853 if (error == -EFSCORRUPTED || error == -EINVAL)
858 if (VFS_I(ip)->i_nlink == 0)
861 if (metafile_type == XFS_METAFILE_DIR)
865 if (inode_wrong_type(VFS_I(ip), mode))
867 if (xfs_has_metadir(mp)) {
868 if (!xfs_is_metadir_inode(ip))
870 if (metafile_type != ip->i_metatype)
879 xfs_err(mp, "metadata inode 0x%llx type %u is corrupt", ino,
881 xfs_fs_mark_sick(mp, XFS_SICK_FS_METADIR);
882 return -EFSCORRUPTED;
885 /* Grab a metadata file if the caller doesn't already have a transaction. */
888 struct xfs_mount *mp,
890 enum xfs_metafile_type metafile_type,
891 struct xfs_inode **ipp)
893 struct xfs_trans *tp;
896 error = xfs_trans_alloc_empty(mp, &tp);
900 error = xfs_trans_metafile_iget(tp, ino, metafile_type, ipp);
901 xfs_trans_cancel(tp);
906 * Grab the inode for reclaim exclusively.
908 * We have found this inode via a lookup under RCU, so the inode may have
909 * already been freed, or it may be in the process of being recycled by
910 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
911 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
912 * will not be set. Hence we need to check for both these flag conditions to
913 * avoid inodes that are no longer reclaim candidates.
915 * Note: checking for other state flags here, under the i_flags_lock or not, is
916 * racy and should be avoided. Those races should be resolved only after we have
917 * ensured that we are able to reclaim this inode and the world can see that we
918 * are going to reclaim it.
920 * Return true if we grabbed it, false otherwise.
924 struct xfs_inode *ip,
925 struct xfs_icwalk *icw)
927 ASSERT(rcu_read_lock_held());
929 spin_lock(&ip->i_flags_lock);
930 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
931 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
932 /* not a reclaim candidate. */
933 spin_unlock(&ip->i_flags_lock);
937 /* Don't reclaim a sick inode unless the caller asked for it. */
939 (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
940 spin_unlock(&ip->i_flags_lock);
944 __xfs_iflags_set(ip, XFS_IRECLAIM);
945 spin_unlock(&ip->i_flags_lock);
950 * Inode reclaim is non-blocking, so the default action if progress cannot be
951 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
952 * XFS_IRECLAIM flag. If we are in a shutdown state, we don't care about
953 * blocking anymore and hence we can wait for the inode to be able to reclaim
956 * We do no IO here - if callers require inodes to be cleaned they must push the
957 * AIL first to trigger writeback of dirty inodes. This enables writeback to be
958 * done in the background in a non-blocking manner, and enables memory reclaim
959 * to make progress without blocking.
963 struct xfs_inode *ip,
964 struct xfs_perag *pag)
966 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
968 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
970 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
974 * Check for log shutdown because aborting the inode can move the log
975 * tail and corrupt in memory state. This is fine if the log is shut
976 * down, but if the log is still active and only the mount is shut down
977 * then the in-memory log tail movement caused by the abort can be
978 * incorrectly propagated to disk.
980 if (xlog_is_shutdown(ip->i_mount->m_log)) {
982 xfs_iflush_shutdown_abort(ip);
985 if (xfs_ipincount(ip))
986 goto out_clear_flush;
987 if (!xfs_inode_clean(ip))
988 goto out_clear_flush;
990 xfs_iflags_clear(ip, XFS_IFLUSHING);
992 trace_xfs_inode_reclaiming(ip);
995 * Because we use RCU freeing we need to ensure the inode always appears
996 * to be reclaimed with an invalid inode number when in the free state.
997 * We do this as early as possible under the ILOCK so that
998 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
999 * detect races with us here. By doing this, we guarantee that once
1000 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
1001 * it will see either a valid inode that will serialise correctly, or it
1002 * will see an invalid inode that it can skip.
1004 spin_lock(&ip->i_flags_lock);
1005 ip->i_flags = XFS_IRECLAIM;
1009 spin_unlock(&ip->i_flags_lock);
1011 ASSERT(!ip->i_itemp || ip->i_itemp->ili_item.li_buf == NULL);
1012 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1014 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1016 * Remove the inode from the per-AG radix tree.
1018 * Because radix_tree_delete won't complain even if the item was never
1019 * added to the tree assert that it's been there before to catch
1020 * problems with the inode life time early on.
1022 spin_lock(&pag->pag_ici_lock);
1023 if (!radix_tree_delete(&pag->pag_ici_root,
1024 XFS_INO_TO_AGINO(ip->i_mount, ino)))
1026 xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
1027 spin_unlock(&pag->pag_ici_lock);
1030 * Here we do an (almost) spurious inode lock in order to coordinate
1031 * with inode cache radix tree lookups. This is because the lookup
1032 * can reference the inodes in the cache without taking references.
1034 * We make that OK here by ensuring that we wait until the inode is
1035 * unlocked after the lookup before we go ahead and free it.
1037 xfs_ilock(ip, XFS_ILOCK_EXCL);
1038 ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
1039 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1040 ASSERT(xfs_inode_clean(ip));
1042 __xfs_inode_free(ip);
1046 xfs_iflags_clear(ip, XFS_IFLUSHING);
1048 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1050 xfs_iflags_clear(ip, XFS_IRECLAIM);
1053 /* Reclaim sick inodes if we're unmounting or the fs went down. */
1055 xfs_want_reclaim_sick(
1056 struct xfs_mount *mp)
1058 return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
1059 xfs_is_shutdown(mp);
1064 struct xfs_mount *mp)
1066 struct xfs_icwalk icw = {
1070 if (xfs_want_reclaim_sick(mp))
1071 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1073 while (xfs_group_marked(mp, XG_TYPE_AG, XFS_PERAG_RECLAIM_MARK)) {
1074 xfs_ail_push_all_sync(mp->m_ail);
1075 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1080 * The shrinker infrastructure determines how many inodes we should scan for
1081 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
1082 * push the AIL here. We also want to proactively free up memory if we can to
1083 * minimise the amount of work memory reclaim has to do so we kick the
1084 * background reclaim if it isn't already scheduled.
1087 xfs_reclaim_inodes_nr(
1088 struct xfs_mount *mp,
1089 unsigned long nr_to_scan)
1091 struct xfs_icwalk icw = {
1092 .icw_flags = XFS_ICWALK_FLAG_SCAN_LIMIT,
1093 .icw_scan_limit = min_t(unsigned long, LONG_MAX, nr_to_scan),
1096 if (xfs_want_reclaim_sick(mp))
1097 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1099 /* kick background reclaimer and push the AIL */
1100 xfs_reclaim_work_queue(mp);
1101 xfs_ail_push_all(mp->m_ail);
1103 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1108 * Return the number of reclaimable inodes in the filesystem for
1109 * the shrinker to determine how much to reclaim.
1112 xfs_reclaim_inodes_count(
1113 struct xfs_mount *mp)
1115 XA_STATE (xas, &mp->m_groups[XG_TYPE_AG].xa, 0);
1116 long reclaimable = 0;
1117 struct xfs_perag *pag;
1120 xas_for_each_marked(&xas, pag, ULONG_MAX, XFS_PERAG_RECLAIM_MARK) {
1121 trace_xfs_reclaim_inodes_count(pag, _THIS_IP_);
1122 reclaimable += pag->pag_ici_reclaimable;
1130 xfs_icwalk_match_id(
1131 struct xfs_inode *ip,
1132 struct xfs_icwalk *icw)
1134 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1135 !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1138 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1139 !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1142 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1143 ip->i_projid != icw->icw_prid)
1150 * A union-based inode filtering algorithm. Process the inode if any of the
1151 * criteria match. This is for global/internal scans only.
1154 xfs_icwalk_match_id_union(
1155 struct xfs_inode *ip,
1156 struct xfs_icwalk *icw)
1158 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1159 uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1162 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1163 gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1166 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1167 ip->i_projid == icw->icw_prid)
1174 * Is this inode @ip eligible for eof/cow block reclamation, given some
1175 * filtering parameters @icw? The inode is eligible if @icw is null or
1176 * if the predicate functions match.
1180 struct xfs_inode *ip,
1181 struct xfs_icwalk *icw)
1188 if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1189 match = xfs_icwalk_match_id_union(ip, icw);
1191 match = xfs_icwalk_match_id(ip, icw);
1195 /* skip the inode if the file size is too small */
1196 if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1197 XFS_ISIZE(ip) < icw->icw_min_file_size)
1204 * This is a fast pass over the inode cache to try to get reclaim moving on as
1205 * many inodes as possible in a short period of time. It kicks itself every few
1206 * seconds, as well as being kicked by the inode cache shrinker when memory
1211 struct work_struct *work)
1213 struct xfs_mount *mp = container_of(to_delayed_work(work),
1214 struct xfs_mount, m_reclaim_work);
1216 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1217 xfs_reclaim_work_queue(mp);
1221 xfs_inode_free_eofblocks(
1222 struct xfs_inode *ip,
1223 struct xfs_icwalk *icw,
1224 unsigned int *lockflags)
1228 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1230 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1234 * If the mapping is dirty the operation can block and wait for some
1235 * time. Unless we are waiting, skip it.
1237 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1240 if (!xfs_icwalk_match(ip, icw))
1244 * If the caller is waiting, return -EAGAIN to keep the background
1245 * scanner moving and revisit the inode in a subsequent pass.
1247 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1252 *lockflags |= XFS_IOLOCK_EXCL;
1254 if (xfs_can_free_eofblocks(ip))
1255 return xfs_free_eofblocks(ip);
1257 /* inode could be preallocated */
1258 trace_xfs_inode_free_eofblocks_invalid(ip);
1259 xfs_inode_clear_eofblocks_tag(ip);
1264 xfs_blockgc_set_iflag(
1265 struct xfs_inode *ip,
1266 unsigned long iflag)
1268 struct xfs_mount *mp = ip->i_mount;
1269 struct xfs_perag *pag;
1271 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1274 * Don't bother locking the AG and looking up in the radix trees
1275 * if we already know that we have the tag set.
1277 if (ip->i_flags & iflag)
1279 spin_lock(&ip->i_flags_lock);
1280 ip->i_flags |= iflag;
1281 spin_unlock(&ip->i_flags_lock);
1283 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1284 spin_lock(&pag->pag_ici_lock);
1286 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1287 XFS_ICI_BLOCKGC_TAG);
1289 spin_unlock(&pag->pag_ici_lock);
1294 xfs_inode_set_eofblocks_tag(
1297 trace_xfs_inode_set_eofblocks_tag(ip);
1298 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1302 xfs_blockgc_clear_iflag(
1303 struct xfs_inode *ip,
1304 unsigned long iflag)
1306 struct xfs_mount *mp = ip->i_mount;
1307 struct xfs_perag *pag;
1310 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1312 spin_lock(&ip->i_flags_lock);
1313 ip->i_flags &= ~iflag;
1314 clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1315 spin_unlock(&ip->i_flags_lock);
1320 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1321 spin_lock(&pag->pag_ici_lock);
1323 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1324 XFS_ICI_BLOCKGC_TAG);
1326 spin_unlock(&pag->pag_ici_lock);
1331 xfs_inode_clear_eofblocks_tag(
1334 trace_xfs_inode_clear_eofblocks_tag(ip);
1335 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1339 * Prepare to free COW fork blocks from an inode.
1342 xfs_prep_free_cowblocks(
1343 struct xfs_inode *ip,
1344 struct xfs_icwalk *icw)
1348 sync = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1351 * Just clear the tag if we have an empty cow fork or none at all. It's
1352 * possible the inode was fully unshared since it was originally tagged.
1354 if (!xfs_inode_has_cow_data(ip)) {
1355 trace_xfs_inode_free_cowblocks_invalid(ip);
1356 xfs_inode_clear_cowblocks_tag(ip);
1361 * A cowblocks trim of an inode can have a significant effect on
1362 * fragmentation even when a reasonable COW extent size hint is set.
1363 * Therefore, we prefer to not process cowblocks unless they are clean
1364 * and idle. We can never process a cowblocks inode that is dirty or has
1365 * in-flight I/O under any circumstances, because outstanding writeback
1366 * or dio expects targeted COW fork blocks exist through write
1367 * completion where they can be remapped into the data fork.
1369 * Therefore, the heuristic used here is to never process inodes
1370 * currently opened for write from background (i.e. non-sync) scans. For
1371 * sync scans, use the pagecache/dio state of the inode to ensure we
1372 * never free COW fork blocks out from under pending I/O.
1374 if (!sync && inode_is_open_for_write(VFS_I(ip)))
1376 return xfs_can_free_cowblocks(ip);
1380 * Automatic CoW Reservation Freeing
1382 * These functions automatically garbage collect leftover CoW reservations
1383 * that were made on behalf of a cowextsize hint when we start to run out
1384 * of quota or when the reservations sit around for too long. If the file
1385 * has dirty pages or is undergoing writeback, its CoW reservations will
1388 * The actual garbage collection piggybacks off the same code that runs
1389 * the speculative EOF preallocation garbage collector.
1392 xfs_inode_free_cowblocks(
1393 struct xfs_inode *ip,
1394 struct xfs_icwalk *icw,
1395 unsigned int *lockflags)
1400 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1402 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1405 if (!xfs_prep_free_cowblocks(ip, icw))
1408 if (!xfs_icwalk_match(ip, icw))
1412 * If the caller is waiting, return -EAGAIN to keep the background
1413 * scanner moving and revisit the inode in a subsequent pass.
1415 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1416 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1421 *lockflags |= XFS_IOLOCK_EXCL;
1423 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1428 *lockflags |= XFS_MMAPLOCK_EXCL;
1431 * Check again, nobody else should be able to dirty blocks or change
1432 * the reflink iflag now that we have the first two locks held.
1434 if (xfs_prep_free_cowblocks(ip, icw))
1435 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1440 xfs_inode_set_cowblocks_tag(
1443 trace_xfs_inode_set_cowblocks_tag(ip);
1444 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1448 xfs_inode_clear_cowblocks_tag(
1451 trace_xfs_inode_clear_cowblocks_tag(ip);
1452 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1455 /* Disable post-EOF and CoW block auto-reclamation. */
1458 struct xfs_mount *mp)
1460 struct xfs_perag *pag = NULL;
1462 if (!xfs_clear_blockgc_enabled(mp))
1465 while ((pag = xfs_perag_next(mp, pag)))
1466 cancel_delayed_work_sync(&pag->pag_blockgc_work);
1467 trace_xfs_blockgc_stop(mp, __return_address);
1470 /* Enable post-EOF and CoW block auto-reclamation. */
1473 struct xfs_mount *mp)
1475 struct xfs_perag *pag = NULL;
1477 if (xfs_set_blockgc_enabled(mp))
1480 trace_xfs_blockgc_start(mp, __return_address);
1481 while ((pag = xfs_perag_grab_next_tag(mp, pag, XFS_ICI_BLOCKGC_TAG)))
1482 xfs_blockgc_queue(pag);
1485 /* Don't try to run block gc on an inode that's in any of these states. */
1486 #define XFS_BLOCKGC_NOGRAB_IFLAGS (XFS_INEW | \
1487 XFS_NEED_INACTIVE | \
1488 XFS_INACTIVATING | \
1489 XFS_IRECLAIMABLE | \
1492 * Decide if the given @ip is eligible for garbage collection of speculative
1493 * preallocations, and grab it if so. Returns true if it's ready to go or
1494 * false if we should just ignore it.
1498 struct xfs_inode *ip)
1500 struct inode *inode = VFS_I(ip);
1502 ASSERT(rcu_read_lock_held());
1504 /* Check for stale RCU freed inode */
1505 spin_lock(&ip->i_flags_lock);
1507 goto out_unlock_noent;
1509 if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1510 goto out_unlock_noent;
1511 spin_unlock(&ip->i_flags_lock);
1513 /* nothing to sync during shutdown */
1514 if (xfs_is_shutdown(ip->i_mount))
1517 /* If we can't grab the inode, it must on it's way to reclaim. */
1521 /* inode is valid */
1525 spin_unlock(&ip->i_flags_lock);
1529 /* Scan one incore inode for block preallocations that we can remove. */
1531 xfs_blockgc_scan_inode(
1532 struct xfs_inode *ip,
1533 struct xfs_icwalk *icw)
1535 unsigned int lockflags = 0;
1538 error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1542 error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1545 xfs_iunlock(ip, lockflags);
1550 /* Background worker that trims preallocated space. */
1553 struct work_struct *work)
1555 struct xfs_perag *pag = container_of(to_delayed_work(work),
1556 struct xfs_perag, pag_blockgc_work);
1557 struct xfs_mount *mp = pag_mount(pag);
1560 trace_xfs_blockgc_worker(mp, __return_address);
1562 error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1564 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1565 pag_agno(pag), error);
1566 xfs_blockgc_queue(pag);
1570 * Try to free space in the filesystem by purging inactive inodes, eofblocks
1574 xfs_blockgc_free_space(
1575 struct xfs_mount *mp,
1576 struct xfs_icwalk *icw)
1580 trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1582 error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1586 return xfs_inodegc_flush(mp);
1590 * Reclaim all the free space that we can by scheduling the background blockgc
1591 * and inodegc workers immediately and waiting for them all to clear.
1594 xfs_blockgc_flush_all(
1595 struct xfs_mount *mp)
1597 struct xfs_perag *pag = NULL;
1599 trace_xfs_blockgc_flush_all(mp, __return_address);
1602 * For each blockgc worker, move its queue time up to now. If it wasn't
1603 * queued, it will not be requeued. Then flush whatever is left.
1605 while ((pag = xfs_perag_grab_next_tag(mp, pag, XFS_ICI_BLOCKGC_TAG)))
1606 mod_delayed_work(mp->m_blockgc_wq, &pag->pag_blockgc_work, 0);
1608 while ((pag = xfs_perag_grab_next_tag(mp, pag, XFS_ICI_BLOCKGC_TAG)))
1609 flush_delayed_work(&pag->pag_blockgc_work);
1611 return xfs_inodegc_flush(mp);
1615 * Run cow/eofblocks scans on the supplied dquots. We don't know exactly which
1616 * quota caused an allocation failure, so we make a best effort by including
1617 * each quota under low free space conditions (less than 1% free space) in the
1620 * Callers must not hold any inode's ILOCK. If requesting a synchronous scan
1621 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1625 xfs_blockgc_free_dquots(
1626 struct xfs_mount *mp,
1627 struct xfs_dquot *udqp,
1628 struct xfs_dquot *gdqp,
1629 struct xfs_dquot *pdqp,
1630 unsigned int iwalk_flags)
1632 struct xfs_icwalk icw = {0};
1633 bool do_work = false;
1635 if (!udqp && !gdqp && !pdqp)
1639 * Run a scan to free blocks using the union filter to cover all
1640 * applicable quotas in a single scan.
1642 icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1644 if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1645 icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1646 icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1650 if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1651 icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1652 icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1656 if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1657 icw.icw_prid = pdqp->q_id;
1658 icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1665 return xfs_blockgc_free_space(mp, &icw);
1668 /* Run cow/eofblocks scans on the quotas attached to the inode. */
1670 xfs_blockgc_free_quota(
1671 struct xfs_inode *ip,
1672 unsigned int iwalk_flags)
1674 return xfs_blockgc_free_dquots(ip->i_mount,
1675 xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1676 xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1677 xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1680 /* XFS Inode Cache Walking Code */
1683 * The inode lookup is done in batches to keep the amount of lock traffic and
1684 * radix tree lookups to a minimum. The batch size is a trade off between
1685 * lookup reduction and stack usage. This is in the reclaim path, so we can't
1688 #define XFS_LOOKUP_BATCH 32
1692 * Decide if we want to grab this inode in anticipation of doing work towards
1697 enum xfs_icwalk_goal goal,
1698 struct xfs_inode *ip,
1699 struct xfs_icwalk *icw)
1702 case XFS_ICWALK_BLOCKGC:
1703 return xfs_blockgc_igrab(ip);
1704 case XFS_ICWALK_RECLAIM:
1705 return xfs_reclaim_igrab(ip, icw);
1712 * Process an inode. Each processing function must handle any state changes
1713 * made by the icwalk igrab function. Return -EAGAIN to skip an inode.
1716 xfs_icwalk_process_inode(
1717 enum xfs_icwalk_goal goal,
1718 struct xfs_inode *ip,
1719 struct xfs_perag *pag,
1720 struct xfs_icwalk *icw)
1725 case XFS_ICWALK_BLOCKGC:
1726 error = xfs_blockgc_scan_inode(ip, icw);
1728 case XFS_ICWALK_RECLAIM:
1729 xfs_reclaim_inode(ip, pag);
1736 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1737 * process them in some manner.
1741 struct xfs_perag *pag,
1742 enum xfs_icwalk_goal goal,
1743 struct xfs_icwalk *icw)
1745 struct xfs_mount *mp = pag_mount(pag);
1746 uint32_t first_index;
1755 if (goal == XFS_ICWALK_RECLAIM)
1756 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1761 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1767 nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1768 (void **) batch, first_index,
1769 XFS_LOOKUP_BATCH, goal);
1777 * Grab the inodes before we drop the lock. if we found
1778 * nothing, nr == 0 and the loop will be skipped.
1780 for (i = 0; i < nr_found; i++) {
1781 struct xfs_inode *ip = batch[i];
1783 if (done || !xfs_icwalk_igrab(goal, ip, icw))
1787 * Update the index for the next lookup. Catch
1788 * overflows into the next AG range which can occur if
1789 * we have inodes in the last block of the AG and we
1790 * are currently pointing to the last inode.
1792 * Because we may see inodes that are from the wrong AG
1793 * due to RCU freeing and reallocation, only update the
1794 * index if it lies in this AG. It was a race that lead
1795 * us to see this inode, so another lookup from the
1796 * same index will not find it again.
1798 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag_agno(pag))
1800 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1801 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1805 /* unlock now we've grabbed the inodes. */
1808 for (i = 0; i < nr_found; i++) {
1811 error = xfs_icwalk_process_inode(goal, batch[i], pag,
1813 if (error == -EAGAIN) {
1817 if (error && last_error != -EFSCORRUPTED)
1821 /* bail out if the filesystem is corrupted. */
1822 if (error == -EFSCORRUPTED)
1827 if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1828 icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1829 if (icw->icw_scan_limit <= 0)
1832 } while (nr_found && !done);
1834 if (goal == XFS_ICWALK_RECLAIM) {
1837 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1847 /* Walk all incore inodes to achieve a given goal. */
1850 struct xfs_mount *mp,
1851 enum xfs_icwalk_goal goal,
1852 struct xfs_icwalk *icw)
1854 struct xfs_perag *pag = NULL;
1858 while ((pag = xfs_perag_grab_next_tag(mp, pag, goal))) {
1859 error = xfs_icwalk_ag(pag, goal, icw);
1862 if (error == -EFSCORRUPTED) {
1863 xfs_perag_rele(pag);
1869 BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1875 struct xfs_inode *ip,
1878 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, whichfork);
1879 struct xfs_bmbt_irec got;
1880 struct xfs_iext_cursor icur;
1882 if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1885 if (isnullstartblock(got.br_startblock)) {
1886 xfs_warn(ip->i_mount,
1887 "ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",
1889 whichfork == XFS_DATA_FORK ? "data" : "cow",
1890 got.br_startoff, got.br_blockcount);
1892 } while (xfs_iext_next_extent(ifp, &icur, &got));
1895 #define xfs_check_delalloc(ip, whichfork) do { } while (0)
1898 /* Schedule the inode for reclaim. */
1900 xfs_inodegc_set_reclaimable(
1901 struct xfs_inode *ip)
1903 struct xfs_mount *mp = ip->i_mount;
1904 struct xfs_perag *pag;
1906 if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {
1907 xfs_check_delalloc(ip, XFS_DATA_FORK);
1908 xfs_check_delalloc(ip, XFS_COW_FORK);
1912 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1913 spin_lock(&pag->pag_ici_lock);
1914 spin_lock(&ip->i_flags_lock);
1916 trace_xfs_inode_set_reclaimable(ip);
1917 ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);
1918 ip->i_flags |= XFS_IRECLAIMABLE;
1919 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1920 XFS_ICI_RECLAIM_TAG);
1922 spin_unlock(&ip->i_flags_lock);
1923 spin_unlock(&pag->pag_ici_lock);
1928 * Free all speculative preallocations and possibly even the inode itself.
1929 * This is the last chance to make changes to an otherwise unreferenced file
1930 * before incore reclamation happens.
1933 xfs_inodegc_inactivate(
1934 struct xfs_inode *ip)
1938 trace_xfs_inode_inactivating(ip);
1939 error = xfs_inactive(ip);
1940 xfs_inodegc_set_reclaimable(ip);
1947 struct work_struct *work)
1949 struct xfs_inodegc *gc = container_of(to_delayed_work(work),
1950 struct xfs_inodegc, work);
1951 struct llist_node *node = llist_del_all(&gc->list);
1952 struct xfs_inode *ip, *n;
1953 struct xfs_mount *mp = gc->mp;
1954 unsigned int nofs_flag;
1957 * Clear the cpu mask bit and ensure that we have seen the latest
1958 * update of the gc structure associated with this CPU. This matches
1959 * with the release semantics used when setting the cpumask bit in
1960 * xfs_inodegc_queue.
1962 cpumask_clear_cpu(gc->cpu, &mp->m_inodegc_cpumask);
1963 smp_mb__after_atomic();
1965 WRITE_ONCE(gc->items, 0);
1971 * We can allocate memory here while doing writeback on behalf of
1972 * memory reclaim. To avoid memory allocation deadlocks set the
1973 * task-wide nofs context for the following operations.
1975 nofs_flag = memalloc_nofs_save();
1977 ip = llist_entry(node, struct xfs_inode, i_gclist);
1978 trace_xfs_inodegc_worker(mp, READ_ONCE(gc->shrinker_hits));
1980 WRITE_ONCE(gc->shrinker_hits, 0);
1981 llist_for_each_entry_safe(ip, n, node, i_gclist) {
1984 xfs_iflags_set(ip, XFS_INACTIVATING);
1985 error = xfs_inodegc_inactivate(ip);
1986 if (error && !gc->error)
1990 memalloc_nofs_restore(nofs_flag);
1994 * Expedite all pending inodegc work to run immediately. This does not wait for
1995 * completion of the work.
1999 struct xfs_mount *mp)
2001 if (!xfs_is_inodegc_enabled(mp))
2003 trace_xfs_inodegc_push(mp, __return_address);
2004 xfs_inodegc_queue_all(mp);
2008 * Force all currently queued inode inactivation work to run immediately and
2009 * wait for the work to finish.
2013 struct xfs_mount *mp)
2015 xfs_inodegc_push(mp);
2016 trace_xfs_inodegc_flush(mp, __return_address);
2017 return xfs_inodegc_wait_all(mp);
2021 * Flush all the pending work and then disable the inode inactivation background
2022 * workers and wait for them to stop. Caller must hold sb->s_umount to
2023 * coordinate changes in the inodegc_enabled state.
2027 struct xfs_mount *mp)
2031 if (!xfs_clear_inodegc_enabled(mp))
2035 * Drain all pending inodegc work, including inodes that could be
2036 * queued by racing xfs_inodegc_queue or xfs_inodegc_shrinker_scan
2037 * threads that sample the inodegc state just prior to us clearing it.
2038 * The inodegc flag state prevents new threads from queuing more
2039 * inodes, so we queue pending work items and flush the workqueue until
2040 * all inodegc lists are empty. IOWs, we cannot use drain_workqueue
2041 * here because it does not allow other unserialized mechanisms to
2042 * reschedule inodegc work while this draining is in progress.
2044 xfs_inodegc_queue_all(mp);
2046 flush_workqueue(mp->m_inodegc_wq);
2047 rerun = xfs_inodegc_queue_all(mp);
2050 trace_xfs_inodegc_stop(mp, __return_address);
2054 * Enable the inode inactivation background workers and schedule deferred inode
2055 * inactivation work if there is any. Caller must hold sb->s_umount to
2056 * coordinate changes in the inodegc_enabled state.
2060 struct xfs_mount *mp)
2062 if (xfs_set_inodegc_enabled(mp))
2065 trace_xfs_inodegc_start(mp, __return_address);
2066 xfs_inodegc_queue_all(mp);
2069 #ifdef CONFIG_XFS_RT
2071 xfs_inodegc_want_queue_rt_file(
2072 struct xfs_inode *ip)
2074 struct xfs_mount *mp = ip->i_mount;
2076 if (!XFS_IS_REALTIME_INODE(ip))
2079 if (__percpu_counter_compare(&mp->m_frextents,
2080 mp->m_low_rtexts[XFS_LOWSP_5_PCNT],
2081 XFS_FDBLOCKS_BATCH) < 0)
2087 # define xfs_inodegc_want_queue_rt_file(ip) (false)
2088 #endif /* CONFIG_XFS_RT */
2091 * Schedule the inactivation worker when:
2093 * - We've accumulated more than one inode cluster buffer's worth of inodes.
2094 * - There is less than 5% free space left.
2095 * - Any of the quotas for this inode are near an enforcement limit.
2098 xfs_inodegc_want_queue_work(
2099 struct xfs_inode *ip,
2102 struct xfs_mount *mp = ip->i_mount;
2104 if (items > mp->m_ino_geo.inodes_per_cluster)
2107 if (__percpu_counter_compare(&mp->m_fdblocks,
2108 mp->m_low_space[XFS_LOWSP_5_PCNT],
2109 XFS_FDBLOCKS_BATCH) < 0)
2112 if (xfs_inodegc_want_queue_rt_file(ip))
2115 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
2118 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
2121 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
2128 * Upper bound on the number of inodes in each AG that can be queued for
2129 * inactivation at any given time, to avoid monopolizing the workqueue.
2131 #define XFS_INODEGC_MAX_BACKLOG (4 * XFS_INODES_PER_CHUNK)
2134 * Make the frontend wait for inactivations when:
2136 * - Memory shrinkers queued the inactivation worker and it hasn't finished.
2137 * - The queue depth exceeds the maximum allowable percpu backlog.
2139 * Note: If we are in a NOFS context here (e.g. current thread is running a
2140 * transaction) the we don't want to block here as inodegc progress may require
2141 * filesystem resources we hold to make progress and that could result in a
2142 * deadlock. Hence we skip out of here if we are in a scoped NOFS context.
2145 xfs_inodegc_want_flush_work(
2146 struct xfs_inode *ip,
2148 unsigned int shrinker_hits)
2150 if (current->flags & PF_MEMALLOC_NOFS)
2153 if (shrinker_hits > 0)
2156 if (items > XFS_INODEGC_MAX_BACKLOG)
2163 * Queue a background inactivation worker if there are inodes that need to be
2164 * inactivated and higher level xfs code hasn't disabled the background
2169 struct xfs_inode *ip)
2171 struct xfs_mount *mp = ip->i_mount;
2172 struct xfs_inodegc *gc;
2174 unsigned int shrinker_hits;
2175 unsigned int cpu_nr;
2176 unsigned long queue_delay = 1;
2178 trace_xfs_inode_set_need_inactive(ip);
2179 spin_lock(&ip->i_flags_lock);
2180 ip->i_flags |= XFS_NEED_INACTIVE;
2181 spin_unlock(&ip->i_flags_lock);
2184 gc = this_cpu_ptr(mp->m_inodegc);
2185 llist_add(&ip->i_gclist, &gc->list);
2186 items = READ_ONCE(gc->items);
2187 WRITE_ONCE(gc->items, items + 1);
2188 shrinker_hits = READ_ONCE(gc->shrinker_hits);
2191 * Ensure the list add is always seen by anyone who finds the cpumask
2192 * bit set. This effectively gives the cpumask bit set operation
2193 * release ordering semantics.
2195 smp_mb__before_atomic();
2196 if (!cpumask_test_cpu(cpu_nr, &mp->m_inodegc_cpumask))
2197 cpumask_test_and_set_cpu(cpu_nr, &mp->m_inodegc_cpumask);
2200 * We queue the work while holding the current CPU so that the work
2201 * is scheduled to run on this CPU.
2203 if (!xfs_is_inodegc_enabled(mp)) {
2208 if (xfs_inodegc_want_queue_work(ip, items))
2211 trace_xfs_inodegc_queue(mp, __return_address);
2212 mod_delayed_work_on(current_cpu(), mp->m_inodegc_wq, &gc->work,
2216 if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
2217 trace_xfs_inodegc_throttle(mp, __return_address);
2218 flush_delayed_work(&gc->work);
2223 * We set the inode flag atomically with the radix tree tag. Once we get tag
2224 * lookups on the radix tree, this inode flag can go away.
2226 * We always use background reclaim here because even if the inode is clean, it
2227 * still may be under IO and hence we have wait for IO completion to occur
2228 * before we can reclaim the inode. The background reclaim path handles this
2229 * more efficiently than we can here, so simply let background reclaim tear down
2233 xfs_inode_mark_reclaimable(
2234 struct xfs_inode *ip)
2236 struct xfs_mount *mp = ip->i_mount;
2239 XFS_STATS_INC(mp, vn_reclaim);
2242 * We should never get here with any of the reclaim flags already set.
2244 ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2246 need_inactive = xfs_inode_needs_inactive(ip);
2247 if (need_inactive) {
2248 xfs_inodegc_queue(ip);
2252 /* Going straight to reclaim, so drop the dquots. */
2253 xfs_qm_dqdetach(ip);
2254 xfs_inodegc_set_reclaimable(ip);
2258 * Register a phony shrinker so that we can run background inodegc sooner when
2259 * there's memory pressure. Inactivation does not itself free any memory but
2260 * it does make inodes reclaimable, which eventually frees memory.
2262 * The count function, seek value, and batch value are crafted to trigger the
2263 * scan function during the second round of scanning. Hopefully this means
2264 * that we reclaimed enough memory that initiating metadata transactions won't
2265 * make things worse.
2267 #define XFS_INODEGC_SHRINKER_COUNT (1UL << DEF_PRIORITY)
2268 #define XFS_INODEGC_SHRINKER_BATCH ((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2270 static unsigned long
2271 xfs_inodegc_shrinker_count(
2272 struct shrinker *shrink,
2273 struct shrink_control *sc)
2275 struct xfs_mount *mp = shrink->private_data;
2276 struct xfs_inodegc *gc;
2279 if (!xfs_is_inodegc_enabled(mp))
2282 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
2283 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2284 if (!llist_empty(&gc->list))
2285 return XFS_INODEGC_SHRINKER_COUNT;
2291 static unsigned long
2292 xfs_inodegc_shrinker_scan(
2293 struct shrinker *shrink,
2294 struct shrink_control *sc)
2296 struct xfs_mount *mp = shrink->private_data;
2297 struct xfs_inodegc *gc;
2299 bool no_items = true;
2301 if (!xfs_is_inodegc_enabled(mp))
2304 trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2306 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
2307 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2308 if (!llist_empty(&gc->list)) {
2309 unsigned int h = READ_ONCE(gc->shrinker_hits);
2311 WRITE_ONCE(gc->shrinker_hits, h + 1);
2312 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
2318 * If there are no inodes to inactivate, we don't want the shrinker
2319 * to think there's deferred work to call us back about.
2327 /* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2329 xfs_inodegc_register_shrinker(
2330 struct xfs_mount *mp)
2332 mp->m_inodegc_shrinker = shrinker_alloc(SHRINKER_NONSLAB,
2335 if (!mp->m_inodegc_shrinker)
2338 mp->m_inodegc_shrinker->count_objects = xfs_inodegc_shrinker_count;
2339 mp->m_inodegc_shrinker->scan_objects = xfs_inodegc_shrinker_scan;
2340 mp->m_inodegc_shrinker->seeks = 0;
2341 mp->m_inodegc_shrinker->batch = XFS_INODEGC_SHRINKER_BATCH;
2342 mp->m_inodegc_shrinker->private_data = mp;
2344 shrinker_register(mp->m_inodegc_shrinker);