2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_format.h"
21 #include "xfs_log_format.h"
22 #include "xfs_trans_resv.h"
24 #include "xfs_mount.h"
25 #include "xfs_inode.h"
26 #include "xfs_error.h"
27 #include "xfs_trans.h"
28 #include "xfs_trans_priv.h"
29 #include "xfs_inode_item.h"
30 #include "xfs_quota.h"
31 #include "xfs_trace.h"
32 #include "xfs_icache.h"
33 #include "xfs_bmap_util.h"
34 #include "xfs_dquot_item.h"
35 #include "xfs_dquot.h"
36 #include "xfs_reflink.h"
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
42 * Allocate and initialise an xfs_inode.
52 * if this didn't occur in transactions, we could use
53 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
54 * code up to do this anyway.
56 ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
59 if (inode_init_always(mp->m_super, VFS_I(ip))) {
60 kmem_zone_free(xfs_inode_zone, ip);
64 /* VFS doesn't initialise i_mode! */
65 VFS_I(ip)->i_mode = 0;
67 XFS_STATS_INC(mp, vn_active);
68 ASSERT(atomic_read(&ip->i_pincount) == 0);
69 ASSERT(!xfs_isiflocked(ip));
70 ASSERT(ip->i_ino == 0);
72 /* initialise the xfs inode */
75 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
79 ip->i_cformat = XFS_DINODE_FMT_EXTENTS;
80 memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
82 ip->i_delayed_blks = 0;
83 memset(&ip->i_d, 0, sizeof(ip->i_d));
89 xfs_inode_free_callback(
90 struct rcu_head *head)
92 struct inode *inode = container_of(head, struct inode, i_rcu);
93 struct xfs_inode *ip = XFS_I(inode);
95 switch (VFS_I(ip)->i_mode & S_IFMT) {
99 xfs_idestroy_fork(ip, XFS_DATA_FORK);
104 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
106 xfs_idestroy_fork(ip, XFS_COW_FORK);
109 ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
110 xfs_inode_item_destroy(ip);
114 kmem_zone_free(xfs_inode_zone, ip);
119 struct xfs_inode *ip)
121 /* asserts to verify all state is correct here */
122 ASSERT(atomic_read(&ip->i_pincount) == 0);
123 XFS_STATS_DEC(ip->i_mount, vn_active);
125 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
130 struct xfs_inode *ip)
132 ASSERT(!xfs_isiflocked(ip));
135 * Because we use RCU freeing we need to ensure the inode always
136 * appears to be reclaimed with an invalid inode number when in the
137 * free state. The ip->i_flags_lock provides the barrier against lookup
140 spin_lock(&ip->i_flags_lock);
141 ip->i_flags = XFS_IRECLAIM;
143 spin_unlock(&ip->i_flags_lock);
145 __xfs_inode_free(ip);
149 * Queue a new inode reclaim pass if there are reclaimable inodes and there
150 * isn't a reclaim pass already in progress. By default it runs every 5s based
151 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
152 * tunable, but that can be done if this method proves to be ineffective or too
156 xfs_reclaim_work_queue(
157 struct xfs_mount *mp)
161 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
162 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
163 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
169 * This is a fast pass over the inode cache to try to get reclaim moving on as
170 * many inodes as possible in a short period of time. It kicks itself every few
171 * seconds, as well as being kicked by the inode cache shrinker when memory
172 * goes low. It scans as quickly as possible avoiding locked inodes or those
173 * already being flushed, and once done schedules a future pass.
177 struct work_struct *work)
179 struct xfs_mount *mp = container_of(to_delayed_work(work),
180 struct xfs_mount, m_reclaim_work);
182 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
183 xfs_reclaim_work_queue(mp);
187 xfs_perag_set_reclaim_tag(
188 struct xfs_perag *pag)
190 struct xfs_mount *mp = pag->pag_mount;
192 lockdep_assert_held(&pag->pag_ici_lock);
193 if (pag->pag_ici_reclaimable++)
196 /* propagate the reclaim tag up into the perag radix tree */
197 spin_lock(&mp->m_perag_lock);
198 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno,
199 XFS_ICI_RECLAIM_TAG);
200 spin_unlock(&mp->m_perag_lock);
202 /* schedule periodic background inode reclaim */
203 xfs_reclaim_work_queue(mp);
205 trace_xfs_perag_set_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
209 xfs_perag_clear_reclaim_tag(
210 struct xfs_perag *pag)
212 struct xfs_mount *mp = pag->pag_mount;
214 lockdep_assert_held(&pag->pag_ici_lock);
215 if (--pag->pag_ici_reclaimable)
218 /* clear the reclaim tag from the perag radix tree */
219 spin_lock(&mp->m_perag_lock);
220 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno,
221 XFS_ICI_RECLAIM_TAG);
222 spin_unlock(&mp->m_perag_lock);
223 trace_xfs_perag_clear_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
228 * We set the inode flag atomically with the radix tree tag.
229 * Once we get tag lookups on the radix tree, this inode flag
233 xfs_inode_set_reclaim_tag(
234 struct xfs_inode *ip)
236 struct xfs_mount *mp = ip->i_mount;
237 struct xfs_perag *pag;
239 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
240 spin_lock(&pag->pag_ici_lock);
241 spin_lock(&ip->i_flags_lock);
243 radix_tree_tag_set(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, ip->i_ino),
244 XFS_ICI_RECLAIM_TAG);
245 xfs_perag_set_reclaim_tag(pag);
246 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
248 spin_unlock(&ip->i_flags_lock);
249 spin_unlock(&pag->pag_ici_lock);
254 xfs_inode_clear_reclaim_tag(
255 struct xfs_perag *pag,
258 radix_tree_tag_clear(&pag->pag_ici_root,
259 XFS_INO_TO_AGINO(pag->pag_mount, ino),
260 XFS_ICI_RECLAIM_TAG);
261 xfs_perag_clear_reclaim_tag(pag);
266 struct xfs_inode *ip)
268 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_INEW_BIT);
269 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_INEW_BIT);
272 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
273 if (!xfs_iflags_test(ip, XFS_INEW))
277 finish_wait(wq, &wait.wait);
281 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
282 * part of the structure. This is made more complex by the fact we store
283 * information about the on-disk values in the VFS inode and so we can't just
284 * overwrite the values unconditionally. Hence we save the parameters we
285 * need to retain across reinitialisation, and rewrite them into the VFS inode
286 * after reinitialisation even if it fails.
290 struct xfs_mount *mp,
294 uint32_t nlink = inode->i_nlink;
295 uint32_t generation = inode->i_generation;
296 uint64_t version = inode->i_version;
297 umode_t mode = inode->i_mode;
299 error = inode_init_always(mp->m_super, inode);
301 set_nlink(inode, nlink);
302 inode->i_generation = generation;
303 inode->i_version = version;
304 inode->i_mode = mode;
309 * Check the validity of the inode we just found it the cache
313 struct xfs_perag *pag,
314 struct xfs_inode *ip,
317 int lock_flags) __releases(RCU)
319 struct inode *inode = VFS_I(ip);
320 struct xfs_mount *mp = ip->i_mount;
324 * check for re-use of an inode within an RCU grace period due to the
325 * radix tree nodes not being updated yet. We monitor for this by
326 * setting the inode number to zero before freeing the inode structure.
327 * If the inode has been reallocated and set up, then the inode number
328 * will not match, so check for that, too.
330 spin_lock(&ip->i_flags_lock);
331 if (ip->i_ino != ino) {
332 trace_xfs_iget_skip(ip);
333 XFS_STATS_INC(mp, xs_ig_frecycle);
340 * If we are racing with another cache hit that is currently
341 * instantiating this inode or currently recycling it out of
342 * reclaimabe state, wait for the initialisation to complete
345 * XXX(hch): eventually we should do something equivalent to
346 * wait_on_inode to wait for these flags to be cleared
347 * instead of polling for it.
349 if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
350 trace_xfs_iget_skip(ip);
351 XFS_STATS_INC(mp, xs_ig_frecycle);
357 * If lookup is racing with unlink return an error immediately.
359 if (VFS_I(ip)->i_mode == 0 && !(flags & XFS_IGET_CREATE)) {
365 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
366 * Need to carefully get it back into useable state.
368 if (ip->i_flags & XFS_IRECLAIMABLE) {
369 trace_xfs_iget_reclaim(ip);
372 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
373 * from stomping over us while we recycle the inode. We can't
374 * clear the radix tree reclaimable tag yet as it requires
375 * pag_ici_lock to be held exclusive.
377 ip->i_flags |= XFS_IRECLAIM;
379 spin_unlock(&ip->i_flags_lock);
382 error = xfs_reinit_inode(mp, inode);
386 * Re-initializing the inode failed, and we are in deep
387 * trouble. Try to re-add it to the reclaim list.
390 spin_lock(&ip->i_flags_lock);
391 wake = !!__xfs_iflags_test(ip, XFS_INEW);
392 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
394 wake_up_bit(&ip->i_flags, __XFS_INEW_BIT);
395 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
396 trace_xfs_iget_reclaim_fail(ip);
400 spin_lock(&pag->pag_ici_lock);
401 spin_lock(&ip->i_flags_lock);
404 * Clear the per-lifetime state in the inode as we are now
405 * effectively a new inode and need to return to the initial
406 * state before reuse occurs.
408 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
409 ip->i_flags |= XFS_INEW;
410 xfs_inode_clear_reclaim_tag(pag, ip->i_ino);
411 inode->i_state = I_NEW;
413 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
414 init_rwsem(&inode->i_rwsem);
416 spin_unlock(&ip->i_flags_lock);
417 spin_unlock(&pag->pag_ici_lock);
419 /* If the VFS inode is being torn down, pause and try again. */
421 trace_xfs_iget_skip(ip);
426 /* We've got a live one. */
427 spin_unlock(&ip->i_flags_lock);
429 trace_xfs_iget_hit(ip);
433 xfs_ilock(ip, lock_flags);
435 xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
436 XFS_STATS_INC(mp, xs_ig_found);
441 spin_unlock(&ip->i_flags_lock);
449 struct xfs_mount *mp,
450 struct xfs_perag *pag,
453 struct xfs_inode **ipp,
457 struct xfs_inode *ip;
459 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
462 ip = xfs_inode_alloc(mp, ino);
466 error = xfs_iread(mp, tp, ip, flags);
470 trace_xfs_iget_miss(ip);
472 if ((VFS_I(ip)->i_mode == 0) && !(flags & XFS_IGET_CREATE)) {
478 * Preload the radix tree so we can insert safely under the
479 * write spinlock. Note that we cannot sleep inside the preload
480 * region. Since we can be called from transaction context, don't
481 * recurse into the file system.
483 if (radix_tree_preload(GFP_NOFS)) {
489 * Because the inode hasn't been added to the radix-tree yet it can't
490 * be found by another thread, so we can do the non-sleeping lock here.
493 if (!xfs_ilock_nowait(ip, lock_flags))
498 * These values must be set before inserting the inode into the radix
499 * tree as the moment it is inserted a concurrent lookup (allowed by the
500 * RCU locking mechanism) can find it and that lookup must see that this
501 * is an inode currently under construction (i.e. that XFS_INEW is set).
502 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
503 * memory barrier that ensures this detection works correctly at lookup
507 if (flags & XFS_IGET_DONTCACHE)
508 iflags |= XFS_IDONTCACHE;
512 xfs_iflags_set(ip, iflags);
514 /* insert the new inode */
515 spin_lock(&pag->pag_ici_lock);
516 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
517 if (unlikely(error)) {
518 WARN_ON(error != -EEXIST);
519 XFS_STATS_INC(mp, xs_ig_dup);
521 goto out_preload_end;
523 spin_unlock(&pag->pag_ici_lock);
524 radix_tree_preload_end();
530 spin_unlock(&pag->pag_ici_lock);
531 radix_tree_preload_end();
533 xfs_iunlock(ip, lock_flags);
535 __destroy_inode(VFS_I(ip));
541 * Look up an inode by number in the given file system.
542 * The inode is looked up in the cache held in each AG.
543 * If the inode is found in the cache, initialise the vfs inode
546 * If it is not in core, read it in from the file system's device,
547 * add it to the cache and initialise the vfs inode.
549 * The inode is locked according to the value of the lock_flags parameter.
550 * This flag parameter indicates how and if the inode's IO lock and inode lock
553 * mp -- the mount point structure for the current file system. It points
554 * to the inode hash table.
555 * tp -- a pointer to the current transaction if there is one. This is
556 * simply passed through to the xfs_iread() call.
557 * ino -- the number of the inode desired. This is the unique identifier
558 * within the file system for the inode being requested.
559 * lock_flags -- flags indicating how to lock the inode. See the comment
560 * for xfs_ilock() for a list of valid values.
577 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
578 * doesn't get freed while it's being referenced during a
579 * radix tree traversal here. It assumes this function
580 * aqcuires only the ILOCK (and therefore it has no need to
581 * involve the IOLOCK in this synchronization).
583 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
585 /* reject inode numbers outside existing AGs */
586 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
589 XFS_STATS_INC(mp, xs_ig_attempts);
591 /* get the perag structure and ensure that it's inode capable */
592 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
593 agino = XFS_INO_TO_AGINO(mp, ino);
598 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
601 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
603 goto out_error_or_again;
606 XFS_STATS_INC(mp, xs_ig_missed);
608 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
611 goto out_error_or_again;
618 * If we have a real type for an on-disk inode, we can setup the inode
619 * now. If it's a new inode being created, xfs_ialloc will handle it.
621 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
622 xfs_setup_existing_inode(ip);
626 if (error == -EAGAIN) {
635 * The inode lookup is done in batches to keep the amount of lock traffic and
636 * radix tree lookups to a minimum. The batch size is a trade off between
637 * lookup reduction and stack usage. This is in the reclaim path, so we can't
640 #define XFS_LOOKUP_BATCH 32
643 xfs_inode_ag_walk_grab(
644 struct xfs_inode *ip,
647 struct inode *inode = VFS_I(ip);
648 bool newinos = !!(flags & XFS_AGITER_INEW_WAIT);
650 ASSERT(rcu_read_lock_held());
653 * check for stale RCU freed inode
655 * If the inode has been reallocated, it doesn't matter if it's not in
656 * the AG we are walking - we are walking for writeback, so if it
657 * passes all the "valid inode" checks and is dirty, then we'll write
658 * it back anyway. If it has been reallocated and still being
659 * initialised, the XFS_INEW check below will catch it.
661 spin_lock(&ip->i_flags_lock);
663 goto out_unlock_noent;
665 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
666 if ((!newinos && __xfs_iflags_test(ip, XFS_INEW)) ||
667 __xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM))
668 goto out_unlock_noent;
669 spin_unlock(&ip->i_flags_lock);
671 /* nothing to sync during shutdown */
672 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
673 return -EFSCORRUPTED;
675 /* If we can't grab the inode, it must on it's way to reclaim. */
683 spin_unlock(&ip->i_flags_lock);
689 struct xfs_mount *mp,
690 struct xfs_perag *pag,
691 int (*execute)(struct xfs_inode *ip, int flags,
698 uint32_t first_index;
710 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
717 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
718 (void **)batch, first_index,
721 nr_found = radix_tree_gang_lookup_tag(
723 (void **) batch, first_index,
724 XFS_LOOKUP_BATCH, tag);
732 * Grab the inodes before we drop the lock. if we found
733 * nothing, nr == 0 and the loop will be skipped.
735 for (i = 0; i < nr_found; i++) {
736 struct xfs_inode *ip = batch[i];
738 if (done || xfs_inode_ag_walk_grab(ip, iter_flags))
742 * Update the index for the next lookup. Catch
743 * overflows into the next AG range which can occur if
744 * we have inodes in the last block of the AG and we
745 * are currently pointing to the last inode.
747 * Because we may see inodes that are from the wrong AG
748 * due to RCU freeing and reallocation, only update the
749 * index if it lies in this AG. It was a race that lead
750 * us to see this inode, so another lookup from the
751 * same index will not find it again.
753 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
755 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
756 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
760 /* unlock now we've grabbed the inodes. */
763 for (i = 0; i < nr_found; i++) {
766 if ((iter_flags & XFS_AGITER_INEW_WAIT) &&
767 xfs_iflags_test(batch[i], XFS_INEW))
768 xfs_inew_wait(batch[i]);
769 error = execute(batch[i], flags, args);
771 if (error == -EAGAIN) {
775 if (error && last_error != -EFSCORRUPTED)
779 /* bail out if the filesystem is corrupted. */
780 if (error == -EFSCORRUPTED)
785 } while (nr_found && !done);
795 * Background scanning to trim post-EOF preallocated space. This is queued
796 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
800 struct xfs_mount *mp)
803 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG))
804 queue_delayed_work(mp->m_eofblocks_workqueue,
805 &mp->m_eofblocks_work,
806 msecs_to_jiffies(xfs_eofb_secs * 1000));
811 xfs_eofblocks_worker(
812 struct work_struct *work)
814 struct xfs_mount *mp = container_of(to_delayed_work(work),
815 struct xfs_mount, m_eofblocks_work);
816 xfs_icache_free_eofblocks(mp, NULL);
817 xfs_queue_eofblocks(mp);
821 * Background scanning to trim preallocated CoW space. This is queued
822 * based on the 'speculative_cow_prealloc_lifetime' tunable (5m by default).
823 * (We'll just piggyback on the post-EOF prealloc space workqueue.)
827 struct xfs_mount *mp)
830 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_COWBLOCKS_TAG))
831 queue_delayed_work(mp->m_eofblocks_workqueue,
832 &mp->m_cowblocks_work,
833 msecs_to_jiffies(xfs_cowb_secs * 1000));
838 xfs_cowblocks_worker(
839 struct work_struct *work)
841 struct xfs_mount *mp = container_of(to_delayed_work(work),
842 struct xfs_mount, m_cowblocks_work);
843 xfs_icache_free_cowblocks(mp, NULL);
844 xfs_queue_cowblocks(mp);
848 xfs_inode_ag_iterator_flags(
849 struct xfs_mount *mp,
850 int (*execute)(struct xfs_inode *ip, int flags,
856 struct xfs_perag *pag;
862 while ((pag = xfs_perag_get(mp, ag))) {
863 ag = pag->pag_agno + 1;
864 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1,
869 if (error == -EFSCORRUPTED)
877 xfs_inode_ag_iterator(
878 struct xfs_mount *mp,
879 int (*execute)(struct xfs_inode *ip, int flags,
884 return xfs_inode_ag_iterator_flags(mp, execute, flags, args, 0);
888 xfs_inode_ag_iterator_tag(
889 struct xfs_mount *mp,
890 int (*execute)(struct xfs_inode *ip, int flags,
896 struct xfs_perag *pag;
902 while ((pag = xfs_perag_get_tag(mp, ag, tag))) {
903 ag = pag->pag_agno + 1;
904 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag,
909 if (error == -EFSCORRUPTED)
917 * Grab the inode for reclaim exclusively.
918 * Return 0 if we grabbed it, non-zero otherwise.
921 xfs_reclaim_inode_grab(
922 struct xfs_inode *ip,
925 ASSERT(rcu_read_lock_held());
927 /* quick check for stale RCU freed inode */
932 * If we are asked for non-blocking operation, do unlocked checks to
933 * see if the inode already is being flushed or in reclaim to avoid
936 if ((flags & SYNC_TRYLOCK) &&
937 __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
941 * The radix tree lock here protects a thread in xfs_iget from racing
942 * with us starting reclaim on the inode. Once we have the
943 * XFS_IRECLAIM flag set it will not touch us.
945 * Due to RCU lookup, we may find inodes that have been freed and only
946 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
947 * aren't candidates for reclaim at all, so we must check the
948 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
950 spin_lock(&ip->i_flags_lock);
951 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
952 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
953 /* not a reclaim candidate. */
954 spin_unlock(&ip->i_flags_lock);
957 __xfs_iflags_set(ip, XFS_IRECLAIM);
958 spin_unlock(&ip->i_flags_lock);
963 * Inodes in different states need to be treated differently. The following
964 * table lists the inode states and the reclaim actions necessary:
966 * inode state iflush ret required action
967 * --------------- ---------- ---------------
969 * shutdown EIO unpin and reclaim
970 * clean, unpinned 0 reclaim
971 * stale, unpinned 0 reclaim
972 * clean, pinned(*) 0 requeue
973 * stale, pinned EAGAIN requeue
974 * dirty, async - requeue
975 * dirty, sync 0 reclaim
977 * (*) dgc: I don't think the clean, pinned state is possible but it gets
978 * handled anyway given the order of checks implemented.
980 * Also, because we get the flush lock first, we know that any inode that has
981 * been flushed delwri has had the flush completed by the time we check that
982 * the inode is clean.
984 * Note that because the inode is flushed delayed write by AIL pushing, the
985 * flush lock may already be held here and waiting on it can result in very
986 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
987 * the caller should push the AIL first before trying to reclaim inodes to
988 * minimise the amount of time spent waiting. For background relaim, we only
989 * bother to reclaim clean inodes anyway.
991 * Hence the order of actions after gaining the locks should be:
993 * shutdown => unpin and reclaim
994 * pinned, async => requeue
995 * pinned, sync => unpin
998 * dirty, async => requeue
999 * dirty, sync => flush, wait and reclaim
1003 struct xfs_inode *ip,
1004 struct xfs_perag *pag,
1007 struct xfs_buf *bp = NULL;
1008 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
1013 xfs_ilock(ip, XFS_ILOCK_EXCL);
1014 if (!xfs_iflock_nowait(ip)) {
1015 if (!(sync_mode & SYNC_WAIT))
1020 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1021 xfs_iunpin_wait(ip);
1022 /* xfs_iflush_abort() drops the flush lock */
1023 xfs_iflush_abort(ip, false);
1026 if (xfs_ipincount(ip)) {
1027 if (!(sync_mode & SYNC_WAIT))
1029 xfs_iunpin_wait(ip);
1031 if (xfs_iflags_test(ip, XFS_ISTALE) || xfs_inode_clean(ip)) {
1037 * Never flush out dirty data during non-blocking reclaim, as it would
1038 * just contend with AIL pushing trying to do the same job.
1040 if (!(sync_mode & SYNC_WAIT))
1044 * Now we have an inode that needs flushing.
1046 * Note that xfs_iflush will never block on the inode buffer lock, as
1047 * xfs_ifree_cluster() can lock the inode buffer before it locks the
1048 * ip->i_lock, and we are doing the exact opposite here. As a result,
1049 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
1050 * result in an ABBA deadlock with xfs_ifree_cluster().
1052 * As xfs_ifree_cluser() must gather all inodes that are active in the
1053 * cache to mark them stale, if we hit this case we don't actually want
1054 * to do IO here - we want the inode marked stale so we can simply
1055 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
1056 * inode, back off and try again. Hopefully the next pass through will
1057 * see the stale flag set on the inode.
1059 error = xfs_iflush(ip, &bp);
1060 if (error == -EAGAIN) {
1061 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1062 /* backoff longer than in xfs_ifree_cluster */
1068 error = xfs_bwrite(bp);
1073 ASSERT(!xfs_isiflocked(ip));
1076 * Because we use RCU freeing we need to ensure the inode always appears
1077 * to be reclaimed with an invalid inode number when in the free state.
1078 * We do this as early as possible under the ILOCK so that
1079 * xfs_iflush_cluster() can be guaranteed to detect races with us here.
1080 * By doing this, we guarantee that once xfs_iflush_cluster has locked
1081 * XFS_ILOCK that it will see either a valid, flushable inode that will
1082 * serialise correctly, or it will see a clean (and invalid) inode that
1085 spin_lock(&ip->i_flags_lock);
1086 ip->i_flags = XFS_IRECLAIM;
1088 spin_unlock(&ip->i_flags_lock);
1090 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1092 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1094 * Remove the inode from the per-AG radix tree.
1096 * Because radix_tree_delete won't complain even if the item was never
1097 * added to the tree assert that it's been there before to catch
1098 * problems with the inode life time early on.
1100 spin_lock(&pag->pag_ici_lock);
1101 if (!radix_tree_delete(&pag->pag_ici_root,
1102 XFS_INO_TO_AGINO(ip->i_mount, ino)))
1104 xfs_perag_clear_reclaim_tag(pag);
1105 spin_unlock(&pag->pag_ici_lock);
1108 * Here we do an (almost) spurious inode lock in order to coordinate
1109 * with inode cache radix tree lookups. This is because the lookup
1110 * can reference the inodes in the cache without taking references.
1112 * We make that OK here by ensuring that we wait until the inode is
1113 * unlocked after the lookup before we go ahead and free it.
1115 xfs_ilock(ip, XFS_ILOCK_EXCL);
1116 xfs_qm_dqdetach(ip);
1117 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1119 __xfs_inode_free(ip);
1125 xfs_iflags_clear(ip, XFS_IRECLAIM);
1126 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1128 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1129 * a short while. However, this just burns CPU time scanning the tree
1130 * waiting for IO to complete and the reclaim work never goes back to
1131 * the idle state. Instead, return 0 to let the next scheduled
1132 * background reclaim attempt to reclaim the inode again.
1138 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1139 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1140 * then a shut down during filesystem unmount reclaim walk leak all the
1141 * unreclaimed inodes.
1144 xfs_reclaim_inodes_ag(
1145 struct xfs_mount *mp,
1149 struct xfs_perag *pag;
1153 int trylock = flags & SYNC_TRYLOCK;
1159 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1160 unsigned long first_index = 0;
1164 ag = pag->pag_agno + 1;
1167 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
1172 first_index = pag->pag_ici_reclaim_cursor;
1174 mutex_lock(&pag->pag_ici_reclaim_lock);
1177 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1181 nr_found = radix_tree_gang_lookup_tag(
1183 (void **)batch, first_index,
1185 XFS_ICI_RECLAIM_TAG);
1193 * Grab the inodes before we drop the lock. if we found
1194 * nothing, nr == 0 and the loop will be skipped.
1196 for (i = 0; i < nr_found; i++) {
1197 struct xfs_inode *ip = batch[i];
1199 if (done || xfs_reclaim_inode_grab(ip, flags))
1203 * Update the index for the next lookup. Catch
1204 * overflows into the next AG range which can
1205 * occur if we have inodes in the last block of
1206 * the AG and we are currently pointing to the
1209 * Because we may see inodes that are from the
1210 * wrong AG due to RCU freeing and
1211 * reallocation, only update the index if it
1212 * lies in this AG. It was a race that lead us
1213 * to see this inode, so another lookup from
1214 * the same index will not find it again.
1216 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1219 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1220 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1224 /* unlock now we've grabbed the inodes. */
1227 for (i = 0; i < nr_found; i++) {
1230 error = xfs_reclaim_inode(batch[i], pag, flags);
1231 if (error && last_error != -EFSCORRUPTED)
1235 *nr_to_scan -= XFS_LOOKUP_BATCH;
1239 } while (nr_found && !done && *nr_to_scan > 0);
1241 if (trylock && !done)
1242 pag->pag_ici_reclaim_cursor = first_index;
1244 pag->pag_ici_reclaim_cursor = 0;
1245 mutex_unlock(&pag->pag_ici_reclaim_lock);
1250 * if we skipped any AG, and we still have scan count remaining, do
1251 * another pass this time using blocking reclaim semantics (i.e
1252 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1253 * ensure that when we get more reclaimers than AGs we block rather
1254 * than spin trying to execute reclaim.
1256 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1268 int nr_to_scan = INT_MAX;
1270 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1274 * Scan a certain number of inodes for reclaim.
1276 * When called we make sure that there is a background (fast) inode reclaim in
1277 * progress, while we will throttle the speed of reclaim via doing synchronous
1278 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1279 * them to be cleaned, which we hope will not be very long due to the
1280 * background walker having already kicked the IO off on those dirty inodes.
1283 xfs_reclaim_inodes_nr(
1284 struct xfs_mount *mp,
1287 /* kick background reclaimer and push the AIL */
1288 xfs_reclaim_work_queue(mp);
1289 xfs_ail_push_all(mp->m_ail);
1291 return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1295 * Return the number of reclaimable inodes in the filesystem for
1296 * the shrinker to determine how much to reclaim.
1299 xfs_reclaim_inodes_count(
1300 struct xfs_mount *mp)
1302 struct xfs_perag *pag;
1303 xfs_agnumber_t ag = 0;
1304 int reclaimable = 0;
1306 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1307 ag = pag->pag_agno + 1;
1308 reclaimable += pag->pag_ici_reclaimable;
1316 struct xfs_inode *ip,
1317 struct xfs_eofblocks *eofb)
1319 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1320 !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1323 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1324 !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1327 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1328 xfs_get_projid(ip) != eofb->eof_prid)
1335 * A union-based inode filtering algorithm. Process the inode if any of the
1336 * criteria match. This is for global/internal scans only.
1339 xfs_inode_match_id_union(
1340 struct xfs_inode *ip,
1341 struct xfs_eofblocks *eofb)
1343 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1344 uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1347 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1348 gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1351 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1352 xfs_get_projid(ip) == eofb->eof_prid)
1359 xfs_inode_free_eofblocks(
1360 struct xfs_inode *ip,
1365 struct xfs_eofblocks *eofb = args;
1368 if (!xfs_can_free_eofblocks(ip, false)) {
1369 /* inode could be preallocated or append-only */
1370 trace_xfs_inode_free_eofblocks_invalid(ip);
1371 xfs_inode_clear_eofblocks_tag(ip);
1376 * If the mapping is dirty the operation can block and wait for some
1377 * time. Unless we are waiting, skip it.
1379 if (!(flags & SYNC_WAIT) &&
1380 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1384 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1385 match = xfs_inode_match_id_union(ip, eofb);
1387 match = xfs_inode_match_id(ip, eofb);
1391 /* skip the inode if the file size is too small */
1392 if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1393 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1398 * If the caller is waiting, return -EAGAIN to keep the background
1399 * scanner moving and revisit the inode in a subsequent pass.
1401 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1402 if (flags & SYNC_WAIT)
1406 ret = xfs_free_eofblocks(ip);
1407 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1413 __xfs_icache_free_eofblocks(
1414 struct xfs_mount *mp,
1415 struct xfs_eofblocks *eofb,
1416 int (*execute)(struct xfs_inode *ip, int flags,
1420 int flags = SYNC_TRYLOCK;
1422 if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
1425 return xfs_inode_ag_iterator_tag(mp, execute, flags,
1430 xfs_icache_free_eofblocks(
1431 struct xfs_mount *mp,
1432 struct xfs_eofblocks *eofb)
1434 return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_eofblocks,
1435 XFS_ICI_EOFBLOCKS_TAG);
1439 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1440 * multiple quotas, we don't know exactly which quota caused an allocation
1441 * failure. We make a best effort by including each quota under low free space
1442 * conditions (less than 1% free space) in the scan.
1445 __xfs_inode_free_quota_eofblocks(
1446 struct xfs_inode *ip,
1447 int (*execute)(struct xfs_mount *mp,
1448 struct xfs_eofblocks *eofb))
1451 struct xfs_eofblocks eofb = {0};
1452 struct xfs_dquot *dq;
1455 * Run a sync scan to increase effectiveness and use the union filter to
1456 * cover all applicable quotas in a single scan.
1458 eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC;
1460 if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) {
1461 dq = xfs_inode_dquot(ip, XFS_DQ_USER);
1462 if (dq && xfs_dquot_lowsp(dq)) {
1463 eofb.eof_uid = VFS_I(ip)->i_uid;
1464 eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1469 if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) {
1470 dq = xfs_inode_dquot(ip, XFS_DQ_GROUP);
1471 if (dq && xfs_dquot_lowsp(dq)) {
1472 eofb.eof_gid = VFS_I(ip)->i_gid;
1473 eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1479 execute(ip->i_mount, &eofb);
1485 xfs_inode_free_quota_eofblocks(
1486 struct xfs_inode *ip)
1488 return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_eofblocks);
1492 __xfs_inode_set_eofblocks_tag(
1494 void (*execute)(struct xfs_mount *mp),
1495 void (*set_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1496 int error, unsigned long caller_ip),
1499 struct xfs_mount *mp = ip->i_mount;
1500 struct xfs_perag *pag;
1504 * Don't bother locking the AG and looking up in the radix trees
1505 * if we already know that we have the tag set.
1507 if (ip->i_flags & XFS_IEOFBLOCKS)
1509 spin_lock(&ip->i_flags_lock);
1510 ip->i_flags |= XFS_IEOFBLOCKS;
1511 spin_unlock(&ip->i_flags_lock);
1513 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1514 spin_lock(&pag->pag_ici_lock);
1516 tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
1517 radix_tree_tag_set(&pag->pag_ici_root,
1518 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
1520 /* propagate the eofblocks tag up into the perag radix tree */
1521 spin_lock(&ip->i_mount->m_perag_lock);
1522 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
1523 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1525 spin_unlock(&ip->i_mount->m_perag_lock);
1527 /* kick off background trimming */
1528 execute(ip->i_mount);
1530 set_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
1533 spin_unlock(&pag->pag_ici_lock);
1538 xfs_inode_set_eofblocks_tag(
1541 trace_xfs_inode_set_eofblocks_tag(ip);
1542 return __xfs_inode_set_eofblocks_tag(ip, xfs_queue_eofblocks,
1543 trace_xfs_perag_set_eofblocks,
1544 XFS_ICI_EOFBLOCKS_TAG);
1548 __xfs_inode_clear_eofblocks_tag(
1550 void (*clear_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1551 int error, unsigned long caller_ip),
1554 struct xfs_mount *mp = ip->i_mount;
1555 struct xfs_perag *pag;
1557 spin_lock(&ip->i_flags_lock);
1558 ip->i_flags &= ~XFS_IEOFBLOCKS;
1559 spin_unlock(&ip->i_flags_lock);
1561 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1562 spin_lock(&pag->pag_ici_lock);
1564 radix_tree_tag_clear(&pag->pag_ici_root,
1565 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
1566 if (!radix_tree_tagged(&pag->pag_ici_root, tag)) {
1567 /* clear the eofblocks tag from the perag radix tree */
1568 spin_lock(&ip->i_mount->m_perag_lock);
1569 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1570 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1572 spin_unlock(&ip->i_mount->m_perag_lock);
1573 clear_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
1576 spin_unlock(&pag->pag_ici_lock);
1581 xfs_inode_clear_eofblocks_tag(
1584 trace_xfs_inode_clear_eofblocks_tag(ip);
1585 return __xfs_inode_clear_eofblocks_tag(ip,
1586 trace_xfs_perag_clear_eofblocks, XFS_ICI_EOFBLOCKS_TAG);
1590 * Automatic CoW Reservation Freeing
1592 * These functions automatically garbage collect leftover CoW reservations
1593 * that were made on behalf of a cowextsize hint when we start to run out
1594 * of quota or when the reservations sit around for too long. If the file
1595 * has dirty pages or is undergoing writeback, its CoW reservations will
1598 * The actual garbage collection piggybacks off the same code that runs
1599 * the speculative EOF preallocation garbage collector.
1602 xfs_inode_free_cowblocks(
1603 struct xfs_inode *ip,
1608 struct xfs_eofblocks *eofb = args;
1610 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1613 * Just clear the tag if we have an empty cow fork or none at all. It's
1614 * possible the inode was fully unshared since it was originally tagged.
1616 if (!xfs_is_reflink_inode(ip) || !ifp->if_bytes) {
1617 trace_xfs_inode_free_cowblocks_invalid(ip);
1618 xfs_inode_clear_cowblocks_tag(ip);
1623 * If the mapping is dirty or under writeback we cannot touch the
1624 * CoW fork. Leave it alone if we're in the midst of a directio.
1626 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1627 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1628 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1629 atomic_read(&VFS_I(ip)->i_dio_count))
1633 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1634 match = xfs_inode_match_id_union(ip, eofb);
1636 match = xfs_inode_match_id(ip, eofb);
1640 /* skip the inode if the file size is too small */
1641 if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1642 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1646 /* Free the CoW blocks */
1647 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1648 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1650 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1652 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
1653 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1659 xfs_icache_free_cowblocks(
1660 struct xfs_mount *mp,
1661 struct xfs_eofblocks *eofb)
1663 return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_cowblocks,
1664 XFS_ICI_COWBLOCKS_TAG);
1668 xfs_inode_free_quota_cowblocks(
1669 struct xfs_inode *ip)
1671 return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_cowblocks);
1675 xfs_inode_set_cowblocks_tag(
1678 trace_xfs_inode_set_cowblocks_tag(ip);
1679 return __xfs_inode_set_eofblocks_tag(ip, xfs_queue_cowblocks,
1680 trace_xfs_perag_set_cowblocks,
1681 XFS_ICI_COWBLOCKS_TAG);
1685 xfs_inode_clear_cowblocks_tag(
1688 trace_xfs_inode_clear_cowblocks_tag(ip);
1689 return __xfs_inode_clear_eofblocks_tag(ip,
1690 trace_xfs_perag_clear_cowblocks, XFS_ICI_COWBLOCKS_TAG);