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(!spin_is_locked(&ip->i_flags_lock));
70 ASSERT(!xfs_isiflocked(ip));
71 ASSERT(ip->i_ino == 0);
73 /* initialise the xfs inode */
76 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
80 ip->i_cformat = XFS_DINODE_FMT_EXTENTS;
81 memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
83 ip->i_delayed_blks = 0;
84 memset(&ip->i_d, 0, sizeof(ip->i_d));
90 xfs_inode_free_callback(
91 struct rcu_head *head)
93 struct inode *inode = container_of(head, struct inode, i_rcu);
94 struct xfs_inode *ip = XFS_I(inode);
96 switch (VFS_I(ip)->i_mode & S_IFMT) {
100 xfs_idestroy_fork(ip, XFS_DATA_FORK);
105 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
107 xfs_idestroy_fork(ip, XFS_COW_FORK);
110 ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
111 xfs_inode_item_destroy(ip);
115 kmem_zone_free(xfs_inode_zone, ip);
120 struct xfs_inode *ip)
122 /* asserts to verify all state is correct here */
123 ASSERT(atomic_read(&ip->i_pincount) == 0);
124 XFS_STATS_DEC(ip->i_mount, vn_active);
126 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
131 struct xfs_inode *ip)
133 ASSERT(!xfs_isiflocked(ip));
136 * Because we use RCU freeing we need to ensure the inode always
137 * appears to be reclaimed with an invalid inode number when in the
138 * free state. The ip->i_flags_lock provides the barrier against lookup
141 spin_lock(&ip->i_flags_lock);
142 ip->i_flags = XFS_IRECLAIM;
144 spin_unlock(&ip->i_flags_lock);
146 __xfs_inode_free(ip);
150 * Queue a new inode reclaim pass if there are reclaimable inodes and there
151 * isn't a reclaim pass already in progress. By default it runs every 5s based
152 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
153 * tunable, but that can be done if this method proves to be ineffective or too
157 xfs_reclaim_work_queue(
158 struct xfs_mount *mp)
162 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
163 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
164 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
170 * This is a fast pass over the inode cache to try to get reclaim moving on as
171 * many inodes as possible in a short period of time. It kicks itself every few
172 * seconds, as well as being kicked by the inode cache shrinker when memory
173 * goes low. It scans as quickly as possible avoiding locked inodes or those
174 * already being flushed, and once done schedules a future pass.
178 struct work_struct *work)
180 struct xfs_mount *mp = container_of(to_delayed_work(work),
181 struct xfs_mount, m_reclaim_work);
183 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
184 xfs_reclaim_work_queue(mp);
188 xfs_perag_set_reclaim_tag(
189 struct xfs_perag *pag)
191 struct xfs_mount *mp = pag->pag_mount;
193 ASSERT(spin_is_locked(&pag->pag_ici_lock));
194 if (pag->pag_ici_reclaimable++)
197 /* propagate the reclaim tag up into the perag radix tree */
198 spin_lock(&mp->m_perag_lock);
199 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno,
200 XFS_ICI_RECLAIM_TAG);
201 spin_unlock(&mp->m_perag_lock);
203 /* schedule periodic background inode reclaim */
204 xfs_reclaim_work_queue(mp);
206 trace_xfs_perag_set_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
210 xfs_perag_clear_reclaim_tag(
211 struct xfs_perag *pag)
213 struct xfs_mount *mp = pag->pag_mount;
215 ASSERT(spin_is_locked(&pag->pag_ici_lock));
216 if (--pag->pag_ici_reclaimable)
219 /* clear the reclaim tag from the perag radix tree */
220 spin_lock(&mp->m_perag_lock);
221 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno,
222 XFS_ICI_RECLAIM_TAG);
223 spin_unlock(&mp->m_perag_lock);
224 trace_xfs_perag_clear_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
229 * We set the inode flag atomically with the radix tree tag.
230 * Once we get tag lookups on the radix tree, this inode flag
234 xfs_inode_set_reclaim_tag(
235 struct xfs_inode *ip)
237 struct xfs_mount *mp = ip->i_mount;
238 struct xfs_perag *pag;
240 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
241 spin_lock(&pag->pag_ici_lock);
242 spin_lock(&ip->i_flags_lock);
244 radix_tree_tag_set(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, ip->i_ino),
245 XFS_ICI_RECLAIM_TAG);
246 xfs_perag_set_reclaim_tag(pag);
247 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
249 spin_unlock(&ip->i_flags_lock);
250 spin_unlock(&pag->pag_ici_lock);
255 xfs_inode_clear_reclaim_tag(
256 struct xfs_perag *pag,
259 radix_tree_tag_clear(&pag->pag_ici_root,
260 XFS_INO_TO_AGINO(pag->pag_mount, ino),
261 XFS_ICI_RECLAIM_TAG);
262 xfs_perag_clear_reclaim_tag(pag);
267 struct xfs_inode *ip)
269 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_INEW_BIT);
270 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_INEW_BIT);
273 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
274 if (!xfs_iflags_test(ip, XFS_INEW))
278 finish_wait(wq, &wait.wait);
282 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
283 * part of the structure. This is made more complex by the fact we store
284 * information about the on-disk values in the VFS inode and so we can't just
285 * overwrite the values unconditionally. Hence we save the parameters we
286 * need to retain across reinitialisation, and rewrite them into the VFS inode
287 * after reinitialisation even if it fails.
291 struct xfs_mount *mp,
295 uint32_t nlink = inode->i_nlink;
296 uint32_t generation = inode->i_generation;
297 uint64_t version = inode->i_version;
298 umode_t mode = inode->i_mode;
300 error = inode_init_always(mp->m_super, inode);
302 set_nlink(inode, nlink);
303 inode->i_generation = generation;
304 inode->i_version = version;
305 inode->i_mode = mode;
310 * Check the validity of the inode we just found it the cache
314 struct xfs_perag *pag,
315 struct xfs_inode *ip,
318 int lock_flags) __releases(RCU)
320 struct inode *inode = VFS_I(ip);
321 struct xfs_mount *mp = ip->i_mount;
325 * check for re-use of an inode within an RCU grace period due to the
326 * radix tree nodes not being updated yet. We monitor for this by
327 * setting the inode number to zero before freeing the inode structure.
328 * If the inode has been reallocated and set up, then the inode number
329 * will not match, so check for that, too.
331 spin_lock(&ip->i_flags_lock);
332 if (ip->i_ino != ino) {
333 trace_xfs_iget_skip(ip);
334 XFS_STATS_INC(mp, xs_ig_frecycle);
341 * If we are racing with another cache hit that is currently
342 * instantiating this inode or currently recycling it out of
343 * reclaimabe state, wait for the initialisation to complete
346 * XXX(hch): eventually we should do something equivalent to
347 * wait_on_inode to wait for these flags to be cleared
348 * instead of polling for it.
350 if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
351 trace_xfs_iget_skip(ip);
352 XFS_STATS_INC(mp, xs_ig_frecycle);
358 * If lookup is racing with unlink return an error immediately.
360 if (VFS_I(ip)->i_mode == 0 && !(flags & XFS_IGET_CREATE)) {
366 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
367 * Need to carefully get it back into useable state.
369 if (ip->i_flags & XFS_IRECLAIMABLE) {
370 trace_xfs_iget_reclaim(ip);
373 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
374 * from stomping over us while we recycle the inode. We can't
375 * clear the radix tree reclaimable 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 error = xfs_reinit_inode(mp, inode);
387 * Re-initializing the inode failed, and we are in deep
388 * trouble. Try to re-add it to the reclaim list.
391 spin_lock(&ip->i_flags_lock);
392 wake = !!__xfs_iflags_test(ip, XFS_INEW);
393 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
395 wake_up_bit(&ip->i_flags, __XFS_INEW_BIT);
396 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
397 trace_xfs_iget_reclaim_fail(ip);
401 spin_lock(&pag->pag_ici_lock);
402 spin_lock(&ip->i_flags_lock);
405 * Clear the per-lifetime state in the inode as we are now
406 * effectively a new inode and need to return to the initial
407 * state before reuse occurs.
409 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
410 ip->i_flags |= XFS_INEW;
411 xfs_inode_clear_reclaim_tag(pag, ip->i_ino);
412 inode->i_state = I_NEW;
414 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
415 init_rwsem(&inode->i_rwsem);
417 spin_unlock(&ip->i_flags_lock);
418 spin_unlock(&pag->pag_ici_lock);
420 /* If the VFS inode is being torn down, pause and try again. */
422 trace_xfs_iget_skip(ip);
427 /* We've got a live one. */
428 spin_unlock(&ip->i_flags_lock);
430 trace_xfs_iget_hit(ip);
434 xfs_ilock(ip, lock_flags);
436 xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
437 XFS_STATS_INC(mp, xs_ig_found);
442 spin_unlock(&ip->i_flags_lock);
450 struct xfs_mount *mp,
451 struct xfs_perag *pag,
454 struct xfs_inode **ipp,
458 struct xfs_inode *ip;
460 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
463 ip = xfs_inode_alloc(mp, ino);
467 error = xfs_iread(mp, tp, ip, flags);
471 trace_xfs_iget_miss(ip);
473 if ((VFS_I(ip)->i_mode == 0) && !(flags & XFS_IGET_CREATE)) {
479 * Preload the radix tree so we can insert safely under the
480 * write spinlock. Note that we cannot sleep inside the preload
481 * region. Since we can be called from transaction context, don't
482 * recurse into the file system.
484 if (radix_tree_preload(GFP_NOFS)) {
490 * Because the inode hasn't been added to the radix-tree yet it can't
491 * be found by another thread, so we can do the non-sleeping lock here.
494 if (!xfs_ilock_nowait(ip, lock_flags))
499 * These values must be set before inserting the inode into the radix
500 * tree as the moment it is inserted a concurrent lookup (allowed by the
501 * RCU locking mechanism) can find it and that lookup must see that this
502 * is an inode currently under construction (i.e. that XFS_INEW is set).
503 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
504 * memory barrier that ensures this detection works correctly at lookup
508 if (flags & XFS_IGET_DONTCACHE)
509 iflags |= XFS_IDONTCACHE;
513 xfs_iflags_set(ip, iflags);
515 /* insert the new inode */
516 spin_lock(&pag->pag_ici_lock);
517 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
518 if (unlikely(error)) {
519 WARN_ON(error != -EEXIST);
520 XFS_STATS_INC(mp, xs_ig_dup);
522 goto out_preload_end;
524 spin_unlock(&pag->pag_ici_lock);
525 radix_tree_preload_end();
531 spin_unlock(&pag->pag_ici_lock);
532 radix_tree_preload_end();
534 xfs_iunlock(ip, lock_flags);
536 __destroy_inode(VFS_I(ip));
542 * Look up an inode by number in the given file system.
543 * The inode is looked up in the cache held in each AG.
544 * If the inode is found in the cache, initialise the vfs inode
547 * If it is not in core, read it in from the file system's device,
548 * add it to the cache and initialise the vfs inode.
550 * The inode is locked according to the value of the lock_flags parameter.
551 * This flag parameter indicates how and if the inode's IO lock and inode lock
554 * mp -- the mount point structure for the current file system. It points
555 * to the inode hash table.
556 * tp -- a pointer to the current transaction if there is one. This is
557 * simply passed through to the xfs_iread() call.
558 * ino -- the number of the inode desired. This is the unique identifier
559 * within the file system for the inode being requested.
560 * lock_flags -- flags indicating how to lock the inode. See the comment
561 * for xfs_ilock() for a list of valid values.
578 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
579 * doesn't get freed while it's being referenced during a
580 * radix tree traversal here. It assumes this function
581 * aqcuires only the ILOCK (and therefore it has no need to
582 * involve the IOLOCK in this synchronization).
584 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
586 /* reject inode numbers outside existing AGs */
587 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
590 XFS_STATS_INC(mp, xs_ig_attempts);
592 /* get the perag structure and ensure that it's inode capable */
593 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
594 agino = XFS_INO_TO_AGINO(mp, ino);
599 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
602 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
604 goto out_error_or_again;
607 XFS_STATS_INC(mp, xs_ig_missed);
609 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
612 goto out_error_or_again;
619 * If we have a real type for an on-disk inode, we can setup the inode
620 * now. If it's a new inode being created, xfs_ialloc will handle it.
622 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
623 xfs_setup_existing_inode(ip);
627 if (error == -EAGAIN) {
636 * The inode lookup is done in batches to keep the amount of lock traffic and
637 * radix tree lookups to a minimum. The batch size is a trade off between
638 * lookup reduction and stack usage. This is in the reclaim path, so we can't
641 #define XFS_LOOKUP_BATCH 32
644 xfs_inode_ag_walk_grab(
645 struct xfs_inode *ip,
648 struct inode *inode = VFS_I(ip);
649 bool newinos = !!(flags & XFS_AGITER_INEW_WAIT);
651 ASSERT(rcu_read_lock_held());
654 * check for stale RCU freed inode
656 * If the inode has been reallocated, it doesn't matter if it's not in
657 * the AG we are walking - we are walking for writeback, so if it
658 * passes all the "valid inode" checks and is dirty, then we'll write
659 * it back anyway. If it has been reallocated and still being
660 * initialised, the XFS_INEW check below will catch it.
662 spin_lock(&ip->i_flags_lock);
664 goto out_unlock_noent;
666 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
667 if ((!newinos && __xfs_iflags_test(ip, XFS_INEW)) ||
668 __xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM))
669 goto out_unlock_noent;
670 spin_unlock(&ip->i_flags_lock);
672 /* nothing to sync during shutdown */
673 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
674 return -EFSCORRUPTED;
676 /* If we can't grab the inode, it must on it's way to reclaim. */
684 spin_unlock(&ip->i_flags_lock);
690 struct xfs_mount *mp,
691 struct xfs_perag *pag,
692 int (*execute)(struct xfs_inode *ip, int flags,
699 uint32_t first_index;
711 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
718 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
719 (void **)batch, first_index,
722 nr_found = radix_tree_gang_lookup_tag(
724 (void **) batch, first_index,
725 XFS_LOOKUP_BATCH, tag);
733 * Grab the inodes before we drop the lock. if we found
734 * nothing, nr == 0 and the loop will be skipped.
736 for (i = 0; i < nr_found; i++) {
737 struct xfs_inode *ip = batch[i];
739 if (done || xfs_inode_ag_walk_grab(ip, iter_flags))
743 * Update the index for the next lookup. Catch
744 * overflows into the next AG range which can occur if
745 * we have inodes in the last block of the AG and we
746 * are currently pointing to the last inode.
748 * Because we may see inodes that are from the wrong AG
749 * due to RCU freeing and reallocation, only update the
750 * index if it lies in this AG. It was a race that lead
751 * us to see this inode, so another lookup from the
752 * same index will not find it again.
754 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
756 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
757 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
761 /* unlock now we've grabbed the inodes. */
764 for (i = 0; i < nr_found; i++) {
767 if ((iter_flags & XFS_AGITER_INEW_WAIT) &&
768 xfs_iflags_test(batch[i], XFS_INEW))
769 xfs_inew_wait(batch[i]);
770 error = execute(batch[i], flags, args);
772 if (error == -EAGAIN) {
776 if (error && last_error != -EFSCORRUPTED)
780 /* bail out if the filesystem is corrupted. */
781 if (error == -EFSCORRUPTED)
786 } while (nr_found && !done);
796 * Background scanning to trim post-EOF preallocated space. This is queued
797 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
801 struct xfs_mount *mp)
804 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG))
805 queue_delayed_work(mp->m_eofblocks_workqueue,
806 &mp->m_eofblocks_work,
807 msecs_to_jiffies(xfs_eofb_secs * 1000));
812 xfs_eofblocks_worker(
813 struct work_struct *work)
815 struct xfs_mount *mp = container_of(to_delayed_work(work),
816 struct xfs_mount, m_eofblocks_work);
817 xfs_icache_free_eofblocks(mp, NULL);
818 xfs_queue_eofblocks(mp);
822 * Background scanning to trim preallocated CoW space. This is queued
823 * based on the 'speculative_cow_prealloc_lifetime' tunable (5m by default).
824 * (We'll just piggyback on the post-EOF prealloc space workqueue.)
828 struct xfs_mount *mp)
831 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_COWBLOCKS_TAG))
832 queue_delayed_work(mp->m_eofblocks_workqueue,
833 &mp->m_cowblocks_work,
834 msecs_to_jiffies(xfs_cowb_secs * 1000));
839 xfs_cowblocks_worker(
840 struct work_struct *work)
842 struct xfs_mount *mp = container_of(to_delayed_work(work),
843 struct xfs_mount, m_cowblocks_work);
844 xfs_icache_free_cowblocks(mp, NULL);
845 xfs_queue_cowblocks(mp);
849 xfs_inode_ag_iterator_flags(
850 struct xfs_mount *mp,
851 int (*execute)(struct xfs_inode *ip, int flags,
857 struct xfs_perag *pag;
863 while ((pag = xfs_perag_get(mp, ag))) {
864 ag = pag->pag_agno + 1;
865 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1,
870 if (error == -EFSCORRUPTED)
878 xfs_inode_ag_iterator(
879 struct xfs_mount *mp,
880 int (*execute)(struct xfs_inode *ip, int flags,
885 return xfs_inode_ag_iterator_flags(mp, execute, flags, args, 0);
889 xfs_inode_ag_iterator_tag(
890 struct xfs_mount *mp,
891 int (*execute)(struct xfs_inode *ip, int flags,
897 struct xfs_perag *pag;
903 while ((pag = xfs_perag_get_tag(mp, ag, tag))) {
904 ag = pag->pag_agno + 1;
905 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag,
910 if (error == -EFSCORRUPTED)
918 * Grab the inode for reclaim exclusively.
919 * Return 0 if we grabbed it, non-zero otherwise.
922 xfs_reclaim_inode_grab(
923 struct xfs_inode *ip,
926 ASSERT(rcu_read_lock_held());
928 /* quick check for stale RCU freed inode */
933 * If we are asked for non-blocking operation, do unlocked checks to
934 * see if the inode already is being flushed or in reclaim to avoid
937 if ((flags & SYNC_TRYLOCK) &&
938 __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
942 * The radix tree lock here protects a thread in xfs_iget from racing
943 * with us starting reclaim on the inode. Once we have the
944 * XFS_IRECLAIM flag set it will not touch us.
946 * Due to RCU lookup, we may find inodes that have been freed and only
947 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
948 * aren't candidates for reclaim at all, so we must check the
949 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
951 spin_lock(&ip->i_flags_lock);
952 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
953 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
954 /* not a reclaim candidate. */
955 spin_unlock(&ip->i_flags_lock);
958 __xfs_iflags_set(ip, XFS_IRECLAIM);
959 spin_unlock(&ip->i_flags_lock);
964 * Inodes in different states need to be treated differently. The following
965 * table lists the inode states and the reclaim actions necessary:
967 * inode state iflush ret required action
968 * --------------- ---------- ---------------
970 * shutdown EIO unpin and reclaim
971 * clean, unpinned 0 reclaim
972 * stale, unpinned 0 reclaim
973 * clean, pinned(*) 0 requeue
974 * stale, pinned EAGAIN requeue
975 * dirty, async - requeue
976 * dirty, sync 0 reclaim
978 * (*) dgc: I don't think the clean, pinned state is possible but it gets
979 * handled anyway given the order of checks implemented.
981 * Also, because we get the flush lock first, we know that any inode that has
982 * been flushed delwri has had the flush completed by the time we check that
983 * the inode is clean.
985 * Note that because the inode is flushed delayed write by AIL pushing, the
986 * flush lock may already be held here and waiting on it can result in very
987 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
988 * the caller should push the AIL first before trying to reclaim inodes to
989 * minimise the amount of time spent waiting. For background relaim, we only
990 * bother to reclaim clean inodes anyway.
992 * Hence the order of actions after gaining the locks should be:
994 * shutdown => unpin and reclaim
995 * pinned, async => requeue
996 * pinned, sync => unpin
999 * dirty, async => requeue
1000 * dirty, sync => flush, wait and reclaim
1004 struct xfs_inode *ip,
1005 struct xfs_perag *pag,
1008 struct xfs_buf *bp = NULL;
1009 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
1014 xfs_ilock(ip, XFS_ILOCK_EXCL);
1015 if (!xfs_iflock_nowait(ip)) {
1016 if (!(sync_mode & SYNC_WAIT))
1021 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1022 xfs_iunpin_wait(ip);
1023 /* xfs_iflush_abort() drops the flush lock */
1024 xfs_iflush_abort(ip, false);
1027 if (xfs_ipincount(ip)) {
1028 if (!(sync_mode & SYNC_WAIT))
1030 xfs_iunpin_wait(ip);
1032 if (xfs_iflags_test(ip, XFS_ISTALE) || xfs_inode_clean(ip)) {
1038 * Never flush out dirty data during non-blocking reclaim, as it would
1039 * just contend with AIL pushing trying to do the same job.
1041 if (!(sync_mode & SYNC_WAIT))
1045 * Now we have an inode that needs flushing.
1047 * Note that xfs_iflush will never block on the inode buffer lock, as
1048 * xfs_ifree_cluster() can lock the inode buffer before it locks the
1049 * ip->i_lock, and we are doing the exact opposite here. As a result,
1050 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
1051 * result in an ABBA deadlock with xfs_ifree_cluster().
1053 * As xfs_ifree_cluser() must gather all inodes that are active in the
1054 * cache to mark them stale, if we hit this case we don't actually want
1055 * to do IO here - we want the inode marked stale so we can simply
1056 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
1057 * inode, back off and try again. Hopefully the next pass through will
1058 * see the stale flag set on the inode.
1060 error = xfs_iflush(ip, &bp);
1061 if (error == -EAGAIN) {
1062 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1063 /* backoff longer than in xfs_ifree_cluster */
1069 error = xfs_bwrite(bp);
1074 ASSERT(!xfs_isiflocked(ip));
1077 * Because we use RCU freeing we need to ensure the inode always appears
1078 * to be reclaimed with an invalid inode number when in the free state.
1079 * We do this as early as possible under the ILOCK so that
1080 * xfs_iflush_cluster() can be guaranteed to detect races with us here.
1081 * By doing this, we guarantee that once xfs_iflush_cluster has locked
1082 * XFS_ILOCK that it will see either a valid, flushable inode that will
1083 * serialise correctly, or it will see a clean (and invalid) inode that
1086 spin_lock(&ip->i_flags_lock);
1087 ip->i_flags = XFS_IRECLAIM;
1089 spin_unlock(&ip->i_flags_lock);
1091 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1093 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1095 * Remove the inode from the per-AG radix tree.
1097 * Because radix_tree_delete won't complain even if the item was never
1098 * added to the tree assert that it's been there before to catch
1099 * problems with the inode life time early on.
1101 spin_lock(&pag->pag_ici_lock);
1102 if (!radix_tree_delete(&pag->pag_ici_root,
1103 XFS_INO_TO_AGINO(ip->i_mount, ino)))
1105 xfs_perag_clear_reclaim_tag(pag);
1106 spin_unlock(&pag->pag_ici_lock);
1109 * Here we do an (almost) spurious inode lock in order to coordinate
1110 * with inode cache radix tree lookups. This is because the lookup
1111 * can reference the inodes in the cache without taking references.
1113 * We make that OK here by ensuring that we wait until the inode is
1114 * unlocked after the lookup before we go ahead and free it.
1116 xfs_ilock(ip, XFS_ILOCK_EXCL);
1117 xfs_qm_dqdetach(ip);
1118 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1120 __xfs_inode_free(ip);
1126 xfs_iflags_clear(ip, XFS_IRECLAIM);
1127 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1129 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1130 * a short while. However, this just burns CPU time scanning the tree
1131 * waiting for IO to complete and the reclaim work never goes back to
1132 * the idle state. Instead, return 0 to let the next scheduled
1133 * background reclaim attempt to reclaim the inode again.
1139 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1140 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1141 * then a shut down during filesystem unmount reclaim walk leak all the
1142 * unreclaimed inodes.
1145 xfs_reclaim_inodes_ag(
1146 struct xfs_mount *mp,
1150 struct xfs_perag *pag;
1154 int trylock = flags & SYNC_TRYLOCK;
1160 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1161 unsigned long first_index = 0;
1165 ag = pag->pag_agno + 1;
1168 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
1173 first_index = pag->pag_ici_reclaim_cursor;
1175 mutex_lock(&pag->pag_ici_reclaim_lock);
1178 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1182 nr_found = radix_tree_gang_lookup_tag(
1184 (void **)batch, first_index,
1186 XFS_ICI_RECLAIM_TAG);
1194 * Grab the inodes before we drop the lock. if we found
1195 * nothing, nr == 0 and the loop will be skipped.
1197 for (i = 0; i < nr_found; i++) {
1198 struct xfs_inode *ip = batch[i];
1200 if (done || xfs_reclaim_inode_grab(ip, flags))
1204 * Update the index for the next lookup. Catch
1205 * overflows into the next AG range which can
1206 * occur if we have inodes in the last block of
1207 * the AG and we are currently pointing to the
1210 * Because we may see inodes that are from the
1211 * wrong AG due to RCU freeing and
1212 * reallocation, only update the index if it
1213 * lies in this AG. It was a race that lead us
1214 * to see this inode, so another lookup from
1215 * the same index will not find it again.
1217 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1220 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1221 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1225 /* unlock now we've grabbed the inodes. */
1228 for (i = 0; i < nr_found; i++) {
1231 error = xfs_reclaim_inode(batch[i], pag, flags);
1232 if (error && last_error != -EFSCORRUPTED)
1236 *nr_to_scan -= XFS_LOOKUP_BATCH;
1240 } while (nr_found && !done && *nr_to_scan > 0);
1242 if (trylock && !done)
1243 pag->pag_ici_reclaim_cursor = first_index;
1245 pag->pag_ici_reclaim_cursor = 0;
1246 mutex_unlock(&pag->pag_ici_reclaim_lock);
1251 * if we skipped any AG, and we still have scan count remaining, do
1252 * another pass this time using blocking reclaim semantics (i.e
1253 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1254 * ensure that when we get more reclaimers than AGs we block rather
1255 * than spin trying to execute reclaim.
1257 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1269 int nr_to_scan = INT_MAX;
1271 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1275 * Scan a certain number of inodes for reclaim.
1277 * When called we make sure that there is a background (fast) inode reclaim in
1278 * progress, while we will throttle the speed of reclaim via doing synchronous
1279 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1280 * them to be cleaned, which we hope will not be very long due to the
1281 * background walker having already kicked the IO off on those dirty inodes.
1284 xfs_reclaim_inodes_nr(
1285 struct xfs_mount *mp,
1288 /* kick background reclaimer and push the AIL */
1289 xfs_reclaim_work_queue(mp);
1290 xfs_ail_push_all(mp->m_ail);
1292 return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1296 * Return the number of reclaimable inodes in the filesystem for
1297 * the shrinker to determine how much to reclaim.
1300 xfs_reclaim_inodes_count(
1301 struct xfs_mount *mp)
1303 struct xfs_perag *pag;
1304 xfs_agnumber_t ag = 0;
1305 int reclaimable = 0;
1307 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1308 ag = pag->pag_agno + 1;
1309 reclaimable += pag->pag_ici_reclaimable;
1317 struct xfs_inode *ip,
1318 struct xfs_eofblocks *eofb)
1320 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1321 !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1324 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1325 !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1328 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1329 xfs_get_projid(ip) != eofb->eof_prid)
1336 * A union-based inode filtering algorithm. Process the inode if any of the
1337 * criteria match. This is for global/internal scans only.
1340 xfs_inode_match_id_union(
1341 struct xfs_inode *ip,
1342 struct xfs_eofblocks *eofb)
1344 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1345 uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1348 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1349 gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1352 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1353 xfs_get_projid(ip) == eofb->eof_prid)
1360 xfs_inode_free_eofblocks(
1361 struct xfs_inode *ip,
1366 struct xfs_eofblocks *eofb = args;
1369 if (!xfs_can_free_eofblocks(ip, false)) {
1370 /* inode could be preallocated or append-only */
1371 trace_xfs_inode_free_eofblocks_invalid(ip);
1372 xfs_inode_clear_eofblocks_tag(ip);
1377 * If the mapping is dirty the operation can block and wait for some
1378 * time. Unless we are waiting, skip it.
1380 if (!(flags & SYNC_WAIT) &&
1381 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1385 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1386 match = xfs_inode_match_id_union(ip, eofb);
1388 match = xfs_inode_match_id(ip, eofb);
1392 /* skip the inode if the file size is too small */
1393 if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1394 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1399 * If the caller is waiting, return -EAGAIN to keep the background
1400 * scanner moving and revisit the inode in a subsequent pass.
1402 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1403 if (flags & SYNC_WAIT)
1407 ret = xfs_free_eofblocks(ip);
1408 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1414 __xfs_icache_free_eofblocks(
1415 struct xfs_mount *mp,
1416 struct xfs_eofblocks *eofb,
1417 int (*execute)(struct xfs_inode *ip, int flags,
1421 int flags = SYNC_TRYLOCK;
1423 if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
1426 return xfs_inode_ag_iterator_tag(mp, execute, flags,
1431 xfs_icache_free_eofblocks(
1432 struct xfs_mount *mp,
1433 struct xfs_eofblocks *eofb)
1435 return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_eofblocks,
1436 XFS_ICI_EOFBLOCKS_TAG);
1440 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1441 * multiple quotas, we don't know exactly which quota caused an allocation
1442 * failure. We make a best effort by including each quota under low free space
1443 * conditions (less than 1% free space) in the scan.
1446 __xfs_inode_free_quota_eofblocks(
1447 struct xfs_inode *ip,
1448 int (*execute)(struct xfs_mount *mp,
1449 struct xfs_eofblocks *eofb))
1452 struct xfs_eofblocks eofb = {0};
1453 struct xfs_dquot *dq;
1456 * Run a sync scan to increase effectiveness and use the union filter to
1457 * cover all applicable quotas in a single scan.
1459 eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC;
1461 if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) {
1462 dq = xfs_inode_dquot(ip, XFS_DQ_USER);
1463 if (dq && xfs_dquot_lowsp(dq)) {
1464 eofb.eof_uid = VFS_I(ip)->i_uid;
1465 eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1470 if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) {
1471 dq = xfs_inode_dquot(ip, XFS_DQ_GROUP);
1472 if (dq && xfs_dquot_lowsp(dq)) {
1473 eofb.eof_gid = VFS_I(ip)->i_gid;
1474 eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1480 execute(ip->i_mount, &eofb);
1486 xfs_inode_free_quota_eofblocks(
1487 struct xfs_inode *ip)
1489 return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_eofblocks);
1493 __xfs_inode_set_eofblocks_tag(
1495 void (*execute)(struct xfs_mount *mp),
1496 void (*set_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1497 int error, unsigned long caller_ip),
1500 struct xfs_mount *mp = ip->i_mount;
1501 struct xfs_perag *pag;
1505 * Don't bother locking the AG and looking up in the radix trees
1506 * if we already know that we have the tag set.
1508 if (ip->i_flags & XFS_IEOFBLOCKS)
1510 spin_lock(&ip->i_flags_lock);
1511 ip->i_flags |= XFS_IEOFBLOCKS;
1512 spin_unlock(&ip->i_flags_lock);
1514 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1515 spin_lock(&pag->pag_ici_lock);
1517 tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
1518 radix_tree_tag_set(&pag->pag_ici_root,
1519 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
1521 /* propagate the eofblocks tag up into the perag radix tree */
1522 spin_lock(&ip->i_mount->m_perag_lock);
1523 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
1524 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1526 spin_unlock(&ip->i_mount->m_perag_lock);
1528 /* kick off background trimming */
1529 execute(ip->i_mount);
1531 set_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
1534 spin_unlock(&pag->pag_ici_lock);
1539 xfs_inode_set_eofblocks_tag(
1542 trace_xfs_inode_set_eofblocks_tag(ip);
1543 return __xfs_inode_set_eofblocks_tag(ip, xfs_queue_eofblocks,
1544 trace_xfs_perag_set_eofblocks,
1545 XFS_ICI_EOFBLOCKS_TAG);
1549 __xfs_inode_clear_eofblocks_tag(
1551 void (*clear_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1552 int error, unsigned long caller_ip),
1555 struct xfs_mount *mp = ip->i_mount;
1556 struct xfs_perag *pag;
1558 spin_lock(&ip->i_flags_lock);
1559 ip->i_flags &= ~XFS_IEOFBLOCKS;
1560 spin_unlock(&ip->i_flags_lock);
1562 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1563 spin_lock(&pag->pag_ici_lock);
1565 radix_tree_tag_clear(&pag->pag_ici_root,
1566 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
1567 if (!radix_tree_tagged(&pag->pag_ici_root, tag)) {
1568 /* clear the eofblocks tag from the perag radix tree */
1569 spin_lock(&ip->i_mount->m_perag_lock);
1570 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1571 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1573 spin_unlock(&ip->i_mount->m_perag_lock);
1574 clear_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
1577 spin_unlock(&pag->pag_ici_lock);
1582 xfs_inode_clear_eofblocks_tag(
1585 trace_xfs_inode_clear_eofblocks_tag(ip);
1586 return __xfs_inode_clear_eofblocks_tag(ip,
1587 trace_xfs_perag_clear_eofblocks, XFS_ICI_EOFBLOCKS_TAG);
1591 * Automatic CoW Reservation Freeing
1593 * These functions automatically garbage collect leftover CoW reservations
1594 * that were made on behalf of a cowextsize hint when we start to run out
1595 * of quota or when the reservations sit around for too long. If the file
1596 * has dirty pages or is undergoing writeback, its CoW reservations will
1599 * The actual garbage collection piggybacks off the same code that runs
1600 * the speculative EOF preallocation garbage collector.
1603 xfs_inode_free_cowblocks(
1604 struct xfs_inode *ip,
1609 struct xfs_eofblocks *eofb = args;
1611 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1614 * Just clear the tag if we have an empty cow fork or none at all. It's
1615 * possible the inode was fully unshared since it was originally tagged.
1617 if (!xfs_is_reflink_inode(ip) || !ifp->if_bytes) {
1618 trace_xfs_inode_free_cowblocks_invalid(ip);
1619 xfs_inode_clear_cowblocks_tag(ip);
1624 * If the mapping is dirty or under writeback we cannot touch the
1625 * CoW fork. Leave it alone if we're in the midst of a directio.
1627 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1628 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1629 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1630 atomic_read(&VFS_I(ip)->i_dio_count))
1634 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1635 match = xfs_inode_match_id_union(ip, eofb);
1637 match = xfs_inode_match_id(ip, eofb);
1641 /* skip the inode if the file size is too small */
1642 if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1643 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1647 /* Free the CoW blocks */
1648 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1649 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1651 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1653 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
1654 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1660 xfs_icache_free_cowblocks(
1661 struct xfs_mount *mp,
1662 struct xfs_eofblocks *eofb)
1664 return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_cowblocks,
1665 XFS_ICI_COWBLOCKS_TAG);
1669 xfs_inode_free_quota_cowblocks(
1670 struct xfs_inode *ip)
1672 return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_cowblocks);
1676 xfs_inode_set_cowblocks_tag(
1679 trace_xfs_inode_set_cowblocks_tag(ip);
1680 return __xfs_inode_set_eofblocks_tag(ip, xfs_queue_cowblocks,
1681 trace_xfs_perag_set_cowblocks,
1682 XFS_ICI_COWBLOCKS_TAG);
1686 xfs_inode_clear_cowblocks_tag(
1689 trace_xfs_inode_clear_cowblocks_tag(ip);
1690 return __xfs_inode_clear_eofblocks_tag(ip,
1691 trace_xfs_perag_clear_cowblocks, XFS_ICI_COWBLOCKS_TAG);