Loading...
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
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.
8 *
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.
13 *
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
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_format.h"
21#include "xfs_log_format.h"
22#include "xfs_trans_resv.h"
23#include "xfs_sb.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"
37
38#include <linux/kthread.h>
39#include <linux/freezer.h>
40#include <linux/iversion.h>
41
42/*
43 * Allocate and initialise an xfs_inode.
44 */
45struct xfs_inode *
46xfs_inode_alloc(
47 struct xfs_mount *mp,
48 xfs_ino_t ino)
49{
50 struct xfs_inode *ip;
51
52 /*
53 * if this didn't occur in transactions, we could use
54 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
55 * code up to do this anyway.
56 */
57 ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
58 if (!ip)
59 return NULL;
60 if (inode_init_always(mp->m_super, VFS_I(ip))) {
61 kmem_zone_free(xfs_inode_zone, ip);
62 return NULL;
63 }
64
65 /* VFS doesn't initialise i_mode! */
66 VFS_I(ip)->i_mode = 0;
67
68 XFS_STATS_INC(mp, vn_active);
69 ASSERT(atomic_read(&ip->i_pincount) == 0);
70 ASSERT(!xfs_isiflocked(ip));
71 ASSERT(ip->i_ino == 0);
72
73 /* initialise the xfs inode */
74 ip->i_ino = ino;
75 ip->i_mount = mp;
76 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
77 ip->i_afp = NULL;
78 ip->i_cowfp = NULL;
79 ip->i_cnextents = 0;
80 ip->i_cformat = XFS_DINODE_FMT_EXTENTS;
81 memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
82 ip->i_flags = 0;
83 ip->i_delayed_blks = 0;
84 memset(&ip->i_d, 0, sizeof(ip->i_d));
85
86 return ip;
87}
88
89STATIC void
90xfs_inode_free_callback(
91 struct rcu_head *head)
92{
93 struct inode *inode = container_of(head, struct inode, i_rcu);
94 struct xfs_inode *ip = XFS_I(inode);
95
96 switch (VFS_I(ip)->i_mode & S_IFMT) {
97 case S_IFREG:
98 case S_IFDIR:
99 case S_IFLNK:
100 xfs_idestroy_fork(ip, XFS_DATA_FORK);
101 break;
102 }
103
104 if (ip->i_afp)
105 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
106 if (ip->i_cowfp)
107 xfs_idestroy_fork(ip, XFS_COW_FORK);
108
109 if (ip->i_itemp) {
110 ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
111 xfs_inode_item_destroy(ip);
112 ip->i_itemp = NULL;
113 }
114
115 kmem_zone_free(xfs_inode_zone, ip);
116}
117
118static void
119__xfs_inode_free(
120 struct xfs_inode *ip)
121{
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);
125
126 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
127}
128
129void
130xfs_inode_free(
131 struct xfs_inode *ip)
132{
133 ASSERT(!xfs_isiflocked(ip));
134
135 /*
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
139 * races.
140 */
141 spin_lock(&ip->i_flags_lock);
142 ip->i_flags = XFS_IRECLAIM;
143 ip->i_ino = 0;
144 spin_unlock(&ip->i_flags_lock);
145
146 __xfs_inode_free(ip);
147}
148
149/*
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
154 * aggressive.
155 */
156static void
157xfs_reclaim_work_queue(
158 struct xfs_mount *mp)
159{
160
161 rcu_read_lock();
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));
165 }
166 rcu_read_unlock();
167}
168
169/*
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.
175 */
176void
177xfs_reclaim_worker(
178 struct work_struct *work)
179{
180 struct xfs_mount *mp = container_of(to_delayed_work(work),
181 struct xfs_mount, m_reclaim_work);
182
183 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
184 xfs_reclaim_work_queue(mp);
185}
186
187static void
188xfs_perag_set_reclaim_tag(
189 struct xfs_perag *pag)
190{
191 struct xfs_mount *mp = pag->pag_mount;
192
193 lockdep_assert_held(&pag->pag_ici_lock);
194 if (pag->pag_ici_reclaimable++)
195 return;
196
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);
202
203 /* schedule periodic background inode reclaim */
204 xfs_reclaim_work_queue(mp);
205
206 trace_xfs_perag_set_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
207}
208
209static void
210xfs_perag_clear_reclaim_tag(
211 struct xfs_perag *pag)
212{
213 struct xfs_mount *mp = pag->pag_mount;
214
215 lockdep_assert_held(&pag->pag_ici_lock);
216 if (--pag->pag_ici_reclaimable)
217 return;
218
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_);
225}
226
227
228/*
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
231 * can go away.
232 */
233void
234xfs_inode_set_reclaim_tag(
235 struct xfs_inode *ip)
236{
237 struct xfs_mount *mp = ip->i_mount;
238 struct xfs_perag *pag;
239
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);
243
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);
248
249 spin_unlock(&ip->i_flags_lock);
250 spin_unlock(&pag->pag_ici_lock);
251 xfs_perag_put(pag);
252}
253
254STATIC void
255xfs_inode_clear_reclaim_tag(
256 struct xfs_perag *pag,
257 xfs_ino_t ino)
258{
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);
263}
264
265static void
266xfs_inew_wait(
267 struct xfs_inode *ip)
268{
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);
271
272 do {
273 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
274 if (!xfs_iflags_test(ip, XFS_INEW))
275 break;
276 schedule();
277 } while (true);
278 finish_wait(wq, &wait.wq_entry);
279}
280
281/*
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.
288 */
289static int
290xfs_reinit_inode(
291 struct xfs_mount *mp,
292 struct inode *inode)
293{
294 int error;
295 uint32_t nlink = inode->i_nlink;
296 uint32_t generation = inode->i_generation;
297 uint64_t version = inode_peek_iversion(inode);
298 umode_t mode = inode->i_mode;
299 dev_t dev = inode->i_rdev;
300
301 error = inode_init_always(mp->m_super, inode);
302
303 set_nlink(inode, nlink);
304 inode->i_generation = generation;
305 inode_set_iversion_queried(inode, version);
306 inode->i_mode = mode;
307 inode->i_rdev = dev;
308 return error;
309}
310
311/*
312 * Check the validity of the inode we just found it the cache
313 */
314static int
315xfs_iget_cache_hit(
316 struct xfs_perag *pag,
317 struct xfs_inode *ip,
318 xfs_ino_t ino,
319 int flags,
320 int lock_flags) __releases(RCU)
321{
322 struct inode *inode = VFS_I(ip);
323 struct xfs_mount *mp = ip->i_mount;
324 int error;
325
326 /*
327 * check for re-use of an inode within an RCU grace period due to the
328 * radix tree nodes not being updated yet. We monitor for this by
329 * setting the inode number to zero before freeing the inode structure.
330 * If the inode has been reallocated and set up, then the inode number
331 * will not match, so check for that, too.
332 */
333 spin_lock(&ip->i_flags_lock);
334 if (ip->i_ino != ino) {
335 trace_xfs_iget_skip(ip);
336 XFS_STATS_INC(mp, xs_ig_frecycle);
337 error = -EAGAIN;
338 goto out_error;
339 }
340
341
342 /*
343 * If we are racing with another cache hit that is currently
344 * instantiating this inode or currently recycling it out of
345 * reclaimabe state, wait for the initialisation to complete
346 * before continuing.
347 *
348 * XXX(hch): eventually we should do something equivalent to
349 * wait_on_inode to wait for these flags to be cleared
350 * instead of polling for it.
351 */
352 if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
353 trace_xfs_iget_skip(ip);
354 XFS_STATS_INC(mp, xs_ig_frecycle);
355 error = -EAGAIN;
356 goto out_error;
357 }
358
359 /*
360 * If lookup is racing with unlink return an error immediately.
361 */
362 if (VFS_I(ip)->i_mode == 0 && !(flags & XFS_IGET_CREATE)) {
363 error = -ENOENT;
364 goto out_error;
365 }
366
367 /*
368 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
369 * Need to carefully get it back into useable state.
370 */
371 if (ip->i_flags & XFS_IRECLAIMABLE) {
372 trace_xfs_iget_reclaim(ip);
373
374 if (flags & XFS_IGET_INCORE) {
375 error = -EAGAIN;
376 goto out_error;
377 }
378
379 /*
380 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
381 * from stomping over us while we recycle the inode. We can't
382 * clear the radix tree reclaimable tag yet as it requires
383 * pag_ici_lock to be held exclusive.
384 */
385 ip->i_flags |= XFS_IRECLAIM;
386
387 spin_unlock(&ip->i_flags_lock);
388 rcu_read_unlock();
389
390 error = xfs_reinit_inode(mp, inode);
391 if (error) {
392 bool wake;
393 /*
394 * Re-initializing the inode failed, and we are in deep
395 * trouble. Try to re-add it to the reclaim list.
396 */
397 rcu_read_lock();
398 spin_lock(&ip->i_flags_lock);
399 wake = !!__xfs_iflags_test(ip, XFS_INEW);
400 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
401 if (wake)
402 wake_up_bit(&ip->i_flags, __XFS_INEW_BIT);
403 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
404 trace_xfs_iget_reclaim_fail(ip);
405 goto out_error;
406 }
407
408 spin_lock(&pag->pag_ici_lock);
409 spin_lock(&ip->i_flags_lock);
410
411 /*
412 * Clear the per-lifetime state in the inode as we are now
413 * effectively a new inode and need to return to the initial
414 * state before reuse occurs.
415 */
416 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
417 ip->i_flags |= XFS_INEW;
418 xfs_inode_clear_reclaim_tag(pag, ip->i_ino);
419 inode->i_state = I_NEW;
420
421 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
422 init_rwsem(&inode->i_rwsem);
423
424 spin_unlock(&ip->i_flags_lock);
425 spin_unlock(&pag->pag_ici_lock);
426 } else {
427 /* If the VFS inode is being torn down, pause and try again. */
428 if (!igrab(inode)) {
429 trace_xfs_iget_skip(ip);
430 error = -EAGAIN;
431 goto out_error;
432 }
433
434 /* We've got a live one. */
435 spin_unlock(&ip->i_flags_lock);
436 rcu_read_unlock();
437 trace_xfs_iget_hit(ip);
438 }
439
440 if (lock_flags != 0)
441 xfs_ilock(ip, lock_flags);
442
443 if (!(flags & XFS_IGET_INCORE))
444 xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
445 XFS_STATS_INC(mp, xs_ig_found);
446
447 return 0;
448
449out_error:
450 spin_unlock(&ip->i_flags_lock);
451 rcu_read_unlock();
452 return error;
453}
454
455
456static int
457xfs_iget_cache_miss(
458 struct xfs_mount *mp,
459 struct xfs_perag *pag,
460 xfs_trans_t *tp,
461 xfs_ino_t ino,
462 struct xfs_inode **ipp,
463 int flags,
464 int lock_flags)
465{
466 struct xfs_inode *ip;
467 int error;
468 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
469 int iflags;
470
471 ip = xfs_inode_alloc(mp, ino);
472 if (!ip)
473 return -ENOMEM;
474
475 error = xfs_iread(mp, tp, ip, flags);
476 if (error)
477 goto out_destroy;
478
479 if (!xfs_inode_verify_forks(ip)) {
480 error = -EFSCORRUPTED;
481 goto out_destroy;
482 }
483
484 trace_xfs_iget_miss(ip);
485
486
487 /*
488 * If we are allocating a new inode, then check what was returned is
489 * actually a free, empty inode. If we are not allocating an inode,
490 * the check we didn't find a free inode.
491 */
492 if (flags & XFS_IGET_CREATE) {
493 if (VFS_I(ip)->i_mode != 0) {
494 xfs_warn(mp,
495"Corruption detected! Free inode 0x%llx not marked free on disk",
496 ino);
497 error = -EFSCORRUPTED;
498 goto out_destroy;
499 }
500 if (ip->i_d.di_nblocks != 0) {
501 xfs_warn(mp,
502"Corruption detected! Free inode 0x%llx has blocks allocated!",
503 ino);
504 error = -EFSCORRUPTED;
505 goto out_destroy;
506 }
507 } else if (VFS_I(ip)->i_mode == 0) {
508 error = -ENOENT;
509 goto out_destroy;
510 }
511
512 /*
513 * Preload the radix tree so we can insert safely under the
514 * write spinlock. Note that we cannot sleep inside the preload
515 * region. Since we can be called from transaction context, don't
516 * recurse into the file system.
517 */
518 if (radix_tree_preload(GFP_NOFS)) {
519 error = -EAGAIN;
520 goto out_destroy;
521 }
522
523 /*
524 * Because the inode hasn't been added to the radix-tree yet it can't
525 * be found by another thread, so we can do the non-sleeping lock here.
526 */
527 if (lock_flags) {
528 if (!xfs_ilock_nowait(ip, lock_flags))
529 BUG();
530 }
531
532 /*
533 * These values must be set before inserting the inode into the radix
534 * tree as the moment it is inserted a concurrent lookup (allowed by the
535 * RCU locking mechanism) can find it and that lookup must see that this
536 * is an inode currently under construction (i.e. that XFS_INEW is set).
537 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
538 * memory barrier that ensures this detection works correctly at lookup
539 * time.
540 */
541 iflags = XFS_INEW;
542 if (flags & XFS_IGET_DONTCACHE)
543 iflags |= XFS_IDONTCACHE;
544 ip->i_udquot = NULL;
545 ip->i_gdquot = NULL;
546 ip->i_pdquot = NULL;
547 xfs_iflags_set(ip, iflags);
548
549 /* insert the new inode */
550 spin_lock(&pag->pag_ici_lock);
551 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
552 if (unlikely(error)) {
553 WARN_ON(error != -EEXIST);
554 XFS_STATS_INC(mp, xs_ig_dup);
555 error = -EAGAIN;
556 goto out_preload_end;
557 }
558 spin_unlock(&pag->pag_ici_lock);
559 radix_tree_preload_end();
560
561 *ipp = ip;
562 return 0;
563
564out_preload_end:
565 spin_unlock(&pag->pag_ici_lock);
566 radix_tree_preload_end();
567 if (lock_flags)
568 xfs_iunlock(ip, lock_flags);
569out_destroy:
570 __destroy_inode(VFS_I(ip));
571 xfs_inode_free(ip);
572 return error;
573}
574
575/*
576 * Look up an inode by number in the given file system.
577 * The inode is looked up in the cache held in each AG.
578 * If the inode is found in the cache, initialise the vfs inode
579 * if necessary.
580 *
581 * If it is not in core, read it in from the file system's device,
582 * add it to the cache and initialise the vfs inode.
583 *
584 * The inode is locked according to the value of the lock_flags parameter.
585 * This flag parameter indicates how and if the inode's IO lock and inode lock
586 * should be taken.
587 *
588 * mp -- the mount point structure for the current file system. It points
589 * to the inode hash table.
590 * tp -- a pointer to the current transaction if there is one. This is
591 * simply passed through to the xfs_iread() call.
592 * ino -- the number of the inode desired. This is the unique identifier
593 * within the file system for the inode being requested.
594 * lock_flags -- flags indicating how to lock the inode. See the comment
595 * for xfs_ilock() for a list of valid values.
596 */
597int
598xfs_iget(
599 xfs_mount_t *mp,
600 xfs_trans_t *tp,
601 xfs_ino_t ino,
602 uint flags,
603 uint lock_flags,
604 xfs_inode_t **ipp)
605{
606 xfs_inode_t *ip;
607 int error;
608 xfs_perag_t *pag;
609 xfs_agino_t agino;
610
611 /*
612 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
613 * doesn't get freed while it's being referenced during a
614 * radix tree traversal here. It assumes this function
615 * aqcuires only the ILOCK (and therefore it has no need to
616 * involve the IOLOCK in this synchronization).
617 */
618 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
619
620 /* reject inode numbers outside existing AGs */
621 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
622 return -EINVAL;
623
624 XFS_STATS_INC(mp, xs_ig_attempts);
625
626 /* get the perag structure and ensure that it's inode capable */
627 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
628 agino = XFS_INO_TO_AGINO(mp, ino);
629
630again:
631 error = 0;
632 rcu_read_lock();
633 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
634
635 if (ip) {
636 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
637 if (error)
638 goto out_error_or_again;
639 } else {
640 rcu_read_unlock();
641 if (flags & XFS_IGET_INCORE) {
642 error = -ENODATA;
643 goto out_error_or_again;
644 }
645 XFS_STATS_INC(mp, xs_ig_missed);
646
647 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
648 flags, lock_flags);
649 if (error)
650 goto out_error_or_again;
651 }
652 xfs_perag_put(pag);
653
654 *ipp = ip;
655
656 /*
657 * If we have a real type for an on-disk inode, we can setup the inode
658 * now. If it's a new inode being created, xfs_ialloc will handle it.
659 */
660 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
661 xfs_setup_existing_inode(ip);
662 return 0;
663
664out_error_or_again:
665 if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
666 delay(1);
667 goto again;
668 }
669 xfs_perag_put(pag);
670 return error;
671}
672
673/*
674 * "Is this a cached inode that's also allocated?"
675 *
676 * Look up an inode by number in the given file system. If the inode is
677 * in cache and isn't in purgatory, return 1 if the inode is allocated
678 * and 0 if it is not. For all other cases (not in cache, being torn
679 * down, etc.), return a negative error code.
680 *
681 * The caller has to prevent inode allocation and freeing activity,
682 * presumably by locking the AGI buffer. This is to ensure that an
683 * inode cannot transition from allocated to freed until the caller is
684 * ready to allow that. If the inode is in an intermediate state (new,
685 * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
686 * inode is not in the cache, -ENOENT will be returned. The caller must
687 * deal with these scenarios appropriately.
688 *
689 * This is a specialized use case for the online scrubber; if you're
690 * reading this, you probably want xfs_iget.
691 */
692int
693xfs_icache_inode_is_allocated(
694 struct xfs_mount *mp,
695 struct xfs_trans *tp,
696 xfs_ino_t ino,
697 bool *inuse)
698{
699 struct xfs_inode *ip;
700 int error;
701
702 error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
703 if (error)
704 return error;
705
706 *inuse = !!(VFS_I(ip)->i_mode);
707 IRELE(ip);
708 return 0;
709}
710
711/*
712 * The inode lookup is done in batches to keep the amount of lock traffic and
713 * radix tree lookups to a minimum. The batch size is a trade off between
714 * lookup reduction and stack usage. This is in the reclaim path, so we can't
715 * be too greedy.
716 */
717#define XFS_LOOKUP_BATCH 32
718
719STATIC int
720xfs_inode_ag_walk_grab(
721 struct xfs_inode *ip,
722 int flags)
723{
724 struct inode *inode = VFS_I(ip);
725 bool newinos = !!(flags & XFS_AGITER_INEW_WAIT);
726
727 ASSERT(rcu_read_lock_held());
728
729 /*
730 * check for stale RCU freed inode
731 *
732 * If the inode has been reallocated, it doesn't matter if it's not in
733 * the AG we are walking - we are walking for writeback, so if it
734 * passes all the "valid inode" checks and is dirty, then we'll write
735 * it back anyway. If it has been reallocated and still being
736 * initialised, the XFS_INEW check below will catch it.
737 */
738 spin_lock(&ip->i_flags_lock);
739 if (!ip->i_ino)
740 goto out_unlock_noent;
741
742 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
743 if ((!newinos && __xfs_iflags_test(ip, XFS_INEW)) ||
744 __xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM))
745 goto out_unlock_noent;
746 spin_unlock(&ip->i_flags_lock);
747
748 /* nothing to sync during shutdown */
749 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
750 return -EFSCORRUPTED;
751
752 /* If we can't grab the inode, it must on it's way to reclaim. */
753 if (!igrab(inode))
754 return -ENOENT;
755
756 /* inode is valid */
757 return 0;
758
759out_unlock_noent:
760 spin_unlock(&ip->i_flags_lock);
761 return -ENOENT;
762}
763
764STATIC int
765xfs_inode_ag_walk(
766 struct xfs_mount *mp,
767 struct xfs_perag *pag,
768 int (*execute)(struct xfs_inode *ip, int flags,
769 void *args),
770 int flags,
771 void *args,
772 int tag,
773 int iter_flags)
774{
775 uint32_t first_index;
776 int last_error = 0;
777 int skipped;
778 int done;
779 int nr_found;
780
781restart:
782 done = 0;
783 skipped = 0;
784 first_index = 0;
785 nr_found = 0;
786 do {
787 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
788 int error = 0;
789 int i;
790
791 rcu_read_lock();
792
793 if (tag == -1)
794 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
795 (void **)batch, first_index,
796 XFS_LOOKUP_BATCH);
797 else
798 nr_found = radix_tree_gang_lookup_tag(
799 &pag->pag_ici_root,
800 (void **) batch, first_index,
801 XFS_LOOKUP_BATCH, tag);
802
803 if (!nr_found) {
804 rcu_read_unlock();
805 break;
806 }
807
808 /*
809 * Grab the inodes before we drop the lock. if we found
810 * nothing, nr == 0 and the loop will be skipped.
811 */
812 for (i = 0; i < nr_found; i++) {
813 struct xfs_inode *ip = batch[i];
814
815 if (done || xfs_inode_ag_walk_grab(ip, iter_flags))
816 batch[i] = NULL;
817
818 /*
819 * Update the index for the next lookup. Catch
820 * overflows into the next AG range which can occur if
821 * we have inodes in the last block of the AG and we
822 * are currently pointing to the last inode.
823 *
824 * Because we may see inodes that are from the wrong AG
825 * due to RCU freeing and reallocation, only update the
826 * index if it lies in this AG. It was a race that lead
827 * us to see this inode, so another lookup from the
828 * same index will not find it again.
829 */
830 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
831 continue;
832 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
833 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
834 done = 1;
835 }
836
837 /* unlock now we've grabbed the inodes. */
838 rcu_read_unlock();
839
840 for (i = 0; i < nr_found; i++) {
841 if (!batch[i])
842 continue;
843 if ((iter_flags & XFS_AGITER_INEW_WAIT) &&
844 xfs_iflags_test(batch[i], XFS_INEW))
845 xfs_inew_wait(batch[i]);
846 error = execute(batch[i], flags, args);
847 IRELE(batch[i]);
848 if (error == -EAGAIN) {
849 skipped++;
850 continue;
851 }
852 if (error && last_error != -EFSCORRUPTED)
853 last_error = error;
854 }
855
856 /* bail out if the filesystem is corrupted. */
857 if (error == -EFSCORRUPTED)
858 break;
859
860 cond_resched();
861
862 } while (nr_found && !done);
863
864 if (skipped) {
865 delay(1);
866 goto restart;
867 }
868 return last_error;
869}
870
871/*
872 * Background scanning to trim post-EOF preallocated space. This is queued
873 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
874 */
875void
876xfs_queue_eofblocks(
877 struct xfs_mount *mp)
878{
879 rcu_read_lock();
880 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG))
881 queue_delayed_work(mp->m_eofblocks_workqueue,
882 &mp->m_eofblocks_work,
883 msecs_to_jiffies(xfs_eofb_secs * 1000));
884 rcu_read_unlock();
885}
886
887void
888xfs_eofblocks_worker(
889 struct work_struct *work)
890{
891 struct xfs_mount *mp = container_of(to_delayed_work(work),
892 struct xfs_mount, m_eofblocks_work);
893 xfs_icache_free_eofblocks(mp, NULL);
894 xfs_queue_eofblocks(mp);
895}
896
897/*
898 * Background scanning to trim preallocated CoW space. This is queued
899 * based on the 'speculative_cow_prealloc_lifetime' tunable (5m by default).
900 * (We'll just piggyback on the post-EOF prealloc space workqueue.)
901 */
902void
903xfs_queue_cowblocks(
904 struct xfs_mount *mp)
905{
906 rcu_read_lock();
907 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_COWBLOCKS_TAG))
908 queue_delayed_work(mp->m_eofblocks_workqueue,
909 &mp->m_cowblocks_work,
910 msecs_to_jiffies(xfs_cowb_secs * 1000));
911 rcu_read_unlock();
912}
913
914void
915xfs_cowblocks_worker(
916 struct work_struct *work)
917{
918 struct xfs_mount *mp = container_of(to_delayed_work(work),
919 struct xfs_mount, m_cowblocks_work);
920 xfs_icache_free_cowblocks(mp, NULL);
921 xfs_queue_cowblocks(mp);
922}
923
924int
925xfs_inode_ag_iterator_flags(
926 struct xfs_mount *mp,
927 int (*execute)(struct xfs_inode *ip, int flags,
928 void *args),
929 int flags,
930 void *args,
931 int iter_flags)
932{
933 struct xfs_perag *pag;
934 int error = 0;
935 int last_error = 0;
936 xfs_agnumber_t ag;
937
938 ag = 0;
939 while ((pag = xfs_perag_get(mp, ag))) {
940 ag = pag->pag_agno + 1;
941 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1,
942 iter_flags);
943 xfs_perag_put(pag);
944 if (error) {
945 last_error = error;
946 if (error == -EFSCORRUPTED)
947 break;
948 }
949 }
950 return last_error;
951}
952
953int
954xfs_inode_ag_iterator(
955 struct xfs_mount *mp,
956 int (*execute)(struct xfs_inode *ip, int flags,
957 void *args),
958 int flags,
959 void *args)
960{
961 return xfs_inode_ag_iterator_flags(mp, execute, flags, args, 0);
962}
963
964int
965xfs_inode_ag_iterator_tag(
966 struct xfs_mount *mp,
967 int (*execute)(struct xfs_inode *ip, int flags,
968 void *args),
969 int flags,
970 void *args,
971 int tag)
972{
973 struct xfs_perag *pag;
974 int error = 0;
975 int last_error = 0;
976 xfs_agnumber_t ag;
977
978 ag = 0;
979 while ((pag = xfs_perag_get_tag(mp, ag, tag))) {
980 ag = pag->pag_agno + 1;
981 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag,
982 0);
983 xfs_perag_put(pag);
984 if (error) {
985 last_error = error;
986 if (error == -EFSCORRUPTED)
987 break;
988 }
989 }
990 return last_error;
991}
992
993/*
994 * Grab the inode for reclaim exclusively.
995 * Return 0 if we grabbed it, non-zero otherwise.
996 */
997STATIC int
998xfs_reclaim_inode_grab(
999 struct xfs_inode *ip,
1000 int flags)
1001{
1002 ASSERT(rcu_read_lock_held());
1003
1004 /* quick check for stale RCU freed inode */
1005 if (!ip->i_ino)
1006 return 1;
1007
1008 /*
1009 * If we are asked for non-blocking operation, do unlocked checks to
1010 * see if the inode already is being flushed or in reclaim to avoid
1011 * lock traffic.
1012 */
1013 if ((flags & SYNC_TRYLOCK) &&
1014 __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
1015 return 1;
1016
1017 /*
1018 * The radix tree lock here protects a thread in xfs_iget from racing
1019 * with us starting reclaim on the inode. Once we have the
1020 * XFS_IRECLAIM flag set it will not touch us.
1021 *
1022 * Due to RCU lookup, we may find inodes that have been freed and only
1023 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
1024 * aren't candidates for reclaim at all, so we must check the
1025 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
1026 */
1027 spin_lock(&ip->i_flags_lock);
1028 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
1029 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
1030 /* not a reclaim candidate. */
1031 spin_unlock(&ip->i_flags_lock);
1032 return 1;
1033 }
1034 __xfs_iflags_set(ip, XFS_IRECLAIM);
1035 spin_unlock(&ip->i_flags_lock);
1036 return 0;
1037}
1038
1039/*
1040 * Inodes in different states need to be treated differently. The following
1041 * table lists the inode states and the reclaim actions necessary:
1042 *
1043 * inode state iflush ret required action
1044 * --------------- ---------- ---------------
1045 * bad - reclaim
1046 * shutdown EIO unpin and reclaim
1047 * clean, unpinned 0 reclaim
1048 * stale, unpinned 0 reclaim
1049 * clean, pinned(*) 0 requeue
1050 * stale, pinned EAGAIN requeue
1051 * dirty, async - requeue
1052 * dirty, sync 0 reclaim
1053 *
1054 * (*) dgc: I don't think the clean, pinned state is possible but it gets
1055 * handled anyway given the order of checks implemented.
1056 *
1057 * Also, because we get the flush lock first, we know that any inode that has
1058 * been flushed delwri has had the flush completed by the time we check that
1059 * the inode is clean.
1060 *
1061 * Note that because the inode is flushed delayed write by AIL pushing, the
1062 * flush lock may already be held here and waiting on it can result in very
1063 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
1064 * the caller should push the AIL first before trying to reclaim inodes to
1065 * minimise the amount of time spent waiting. For background relaim, we only
1066 * bother to reclaim clean inodes anyway.
1067 *
1068 * Hence the order of actions after gaining the locks should be:
1069 * bad => reclaim
1070 * shutdown => unpin and reclaim
1071 * pinned, async => requeue
1072 * pinned, sync => unpin
1073 * stale => reclaim
1074 * clean => reclaim
1075 * dirty, async => requeue
1076 * dirty, sync => flush, wait and reclaim
1077 */
1078STATIC int
1079xfs_reclaim_inode(
1080 struct xfs_inode *ip,
1081 struct xfs_perag *pag,
1082 int sync_mode)
1083{
1084 struct xfs_buf *bp = NULL;
1085 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
1086 int error;
1087
1088restart:
1089 error = 0;
1090 xfs_ilock(ip, XFS_ILOCK_EXCL);
1091 if (!xfs_iflock_nowait(ip)) {
1092 if (!(sync_mode & SYNC_WAIT))
1093 goto out;
1094 xfs_iflock(ip);
1095 }
1096
1097 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1098 xfs_iunpin_wait(ip);
1099 /* xfs_iflush_abort() drops the flush lock */
1100 xfs_iflush_abort(ip, false);
1101 goto reclaim;
1102 }
1103 if (xfs_ipincount(ip)) {
1104 if (!(sync_mode & SYNC_WAIT))
1105 goto out_ifunlock;
1106 xfs_iunpin_wait(ip);
1107 }
1108 if (xfs_iflags_test(ip, XFS_ISTALE) || xfs_inode_clean(ip)) {
1109 xfs_ifunlock(ip);
1110 goto reclaim;
1111 }
1112
1113 /*
1114 * Never flush out dirty data during non-blocking reclaim, as it would
1115 * just contend with AIL pushing trying to do the same job.
1116 */
1117 if (!(sync_mode & SYNC_WAIT))
1118 goto out_ifunlock;
1119
1120 /*
1121 * Now we have an inode that needs flushing.
1122 *
1123 * Note that xfs_iflush will never block on the inode buffer lock, as
1124 * xfs_ifree_cluster() can lock the inode buffer before it locks the
1125 * ip->i_lock, and we are doing the exact opposite here. As a result,
1126 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
1127 * result in an ABBA deadlock with xfs_ifree_cluster().
1128 *
1129 * As xfs_ifree_cluser() must gather all inodes that are active in the
1130 * cache to mark them stale, if we hit this case we don't actually want
1131 * to do IO here - we want the inode marked stale so we can simply
1132 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
1133 * inode, back off and try again. Hopefully the next pass through will
1134 * see the stale flag set on the inode.
1135 */
1136 error = xfs_iflush(ip, &bp);
1137 if (error == -EAGAIN) {
1138 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1139 /* backoff longer than in xfs_ifree_cluster */
1140 delay(2);
1141 goto restart;
1142 }
1143
1144 if (!error) {
1145 error = xfs_bwrite(bp);
1146 xfs_buf_relse(bp);
1147 }
1148
1149reclaim:
1150 ASSERT(!xfs_isiflocked(ip));
1151
1152 /*
1153 * Because we use RCU freeing we need to ensure the inode always appears
1154 * to be reclaimed with an invalid inode number when in the free state.
1155 * We do this as early as possible under the ILOCK so that
1156 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
1157 * detect races with us here. By doing this, we guarantee that once
1158 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
1159 * it will see either a valid inode that will serialise correctly, or it
1160 * will see an invalid inode that it can skip.
1161 */
1162 spin_lock(&ip->i_flags_lock);
1163 ip->i_flags = XFS_IRECLAIM;
1164 ip->i_ino = 0;
1165 spin_unlock(&ip->i_flags_lock);
1166
1167 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1168
1169 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1170 /*
1171 * Remove the inode from the per-AG radix tree.
1172 *
1173 * Because radix_tree_delete won't complain even if the item was never
1174 * added to the tree assert that it's been there before to catch
1175 * problems with the inode life time early on.
1176 */
1177 spin_lock(&pag->pag_ici_lock);
1178 if (!radix_tree_delete(&pag->pag_ici_root,
1179 XFS_INO_TO_AGINO(ip->i_mount, ino)))
1180 ASSERT(0);
1181 xfs_perag_clear_reclaim_tag(pag);
1182 spin_unlock(&pag->pag_ici_lock);
1183
1184 /*
1185 * Here we do an (almost) spurious inode lock in order to coordinate
1186 * with inode cache radix tree lookups. This is because the lookup
1187 * can reference the inodes in the cache without taking references.
1188 *
1189 * We make that OK here by ensuring that we wait until the inode is
1190 * unlocked after the lookup before we go ahead and free it.
1191 */
1192 xfs_ilock(ip, XFS_ILOCK_EXCL);
1193 xfs_qm_dqdetach(ip);
1194 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1195
1196 __xfs_inode_free(ip);
1197 return error;
1198
1199out_ifunlock:
1200 xfs_ifunlock(ip);
1201out:
1202 xfs_iflags_clear(ip, XFS_IRECLAIM);
1203 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1204 /*
1205 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1206 * a short while. However, this just burns CPU time scanning the tree
1207 * waiting for IO to complete and the reclaim work never goes back to
1208 * the idle state. Instead, return 0 to let the next scheduled
1209 * background reclaim attempt to reclaim the inode again.
1210 */
1211 return 0;
1212}
1213
1214/*
1215 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1216 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1217 * then a shut down during filesystem unmount reclaim walk leak all the
1218 * unreclaimed inodes.
1219 */
1220STATIC int
1221xfs_reclaim_inodes_ag(
1222 struct xfs_mount *mp,
1223 int flags,
1224 int *nr_to_scan)
1225{
1226 struct xfs_perag *pag;
1227 int error = 0;
1228 int last_error = 0;
1229 xfs_agnumber_t ag;
1230 int trylock = flags & SYNC_TRYLOCK;
1231 int skipped;
1232
1233restart:
1234 ag = 0;
1235 skipped = 0;
1236 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1237 unsigned long first_index = 0;
1238 int done = 0;
1239 int nr_found = 0;
1240
1241 ag = pag->pag_agno + 1;
1242
1243 if (trylock) {
1244 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
1245 skipped++;
1246 xfs_perag_put(pag);
1247 continue;
1248 }
1249 first_index = pag->pag_ici_reclaim_cursor;
1250 } else
1251 mutex_lock(&pag->pag_ici_reclaim_lock);
1252
1253 do {
1254 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1255 int i;
1256
1257 rcu_read_lock();
1258 nr_found = radix_tree_gang_lookup_tag(
1259 &pag->pag_ici_root,
1260 (void **)batch, first_index,
1261 XFS_LOOKUP_BATCH,
1262 XFS_ICI_RECLAIM_TAG);
1263 if (!nr_found) {
1264 done = 1;
1265 rcu_read_unlock();
1266 break;
1267 }
1268
1269 /*
1270 * Grab the inodes before we drop the lock. if we found
1271 * nothing, nr == 0 and the loop will be skipped.
1272 */
1273 for (i = 0; i < nr_found; i++) {
1274 struct xfs_inode *ip = batch[i];
1275
1276 if (done || xfs_reclaim_inode_grab(ip, flags))
1277 batch[i] = NULL;
1278
1279 /*
1280 * Update the index for the next lookup. Catch
1281 * overflows into the next AG range which can
1282 * occur if we have inodes in the last block of
1283 * the AG and we are currently pointing to the
1284 * last inode.
1285 *
1286 * Because we may see inodes that are from the
1287 * wrong AG due to RCU freeing and
1288 * reallocation, only update the index if it
1289 * lies in this AG. It was a race that lead us
1290 * to see this inode, so another lookup from
1291 * the same index will not find it again.
1292 */
1293 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1294 pag->pag_agno)
1295 continue;
1296 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1297 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1298 done = 1;
1299 }
1300
1301 /* unlock now we've grabbed the inodes. */
1302 rcu_read_unlock();
1303
1304 for (i = 0; i < nr_found; i++) {
1305 if (!batch[i])
1306 continue;
1307 error = xfs_reclaim_inode(batch[i], pag, flags);
1308 if (error && last_error != -EFSCORRUPTED)
1309 last_error = error;
1310 }
1311
1312 *nr_to_scan -= XFS_LOOKUP_BATCH;
1313
1314 cond_resched();
1315
1316 } while (nr_found && !done && *nr_to_scan > 0);
1317
1318 if (trylock && !done)
1319 pag->pag_ici_reclaim_cursor = first_index;
1320 else
1321 pag->pag_ici_reclaim_cursor = 0;
1322 mutex_unlock(&pag->pag_ici_reclaim_lock);
1323 xfs_perag_put(pag);
1324 }
1325
1326 /*
1327 * if we skipped any AG, and we still have scan count remaining, do
1328 * another pass this time using blocking reclaim semantics (i.e
1329 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1330 * ensure that when we get more reclaimers than AGs we block rather
1331 * than spin trying to execute reclaim.
1332 */
1333 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1334 trylock = 0;
1335 goto restart;
1336 }
1337 return last_error;
1338}
1339
1340int
1341xfs_reclaim_inodes(
1342 xfs_mount_t *mp,
1343 int mode)
1344{
1345 int nr_to_scan = INT_MAX;
1346
1347 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1348}
1349
1350/*
1351 * Scan a certain number of inodes for reclaim.
1352 *
1353 * When called we make sure that there is a background (fast) inode reclaim in
1354 * progress, while we will throttle the speed of reclaim via doing synchronous
1355 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1356 * them to be cleaned, which we hope will not be very long due to the
1357 * background walker having already kicked the IO off on those dirty inodes.
1358 */
1359long
1360xfs_reclaim_inodes_nr(
1361 struct xfs_mount *mp,
1362 int nr_to_scan)
1363{
1364 /* kick background reclaimer and push the AIL */
1365 xfs_reclaim_work_queue(mp);
1366 xfs_ail_push_all(mp->m_ail);
1367
1368 return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1369}
1370
1371/*
1372 * Return the number of reclaimable inodes in the filesystem for
1373 * the shrinker to determine how much to reclaim.
1374 */
1375int
1376xfs_reclaim_inodes_count(
1377 struct xfs_mount *mp)
1378{
1379 struct xfs_perag *pag;
1380 xfs_agnumber_t ag = 0;
1381 int reclaimable = 0;
1382
1383 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1384 ag = pag->pag_agno + 1;
1385 reclaimable += pag->pag_ici_reclaimable;
1386 xfs_perag_put(pag);
1387 }
1388 return reclaimable;
1389}
1390
1391STATIC int
1392xfs_inode_match_id(
1393 struct xfs_inode *ip,
1394 struct xfs_eofblocks *eofb)
1395{
1396 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1397 !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1398 return 0;
1399
1400 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1401 !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1402 return 0;
1403
1404 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1405 xfs_get_projid(ip) != eofb->eof_prid)
1406 return 0;
1407
1408 return 1;
1409}
1410
1411/*
1412 * A union-based inode filtering algorithm. Process the inode if any of the
1413 * criteria match. This is for global/internal scans only.
1414 */
1415STATIC int
1416xfs_inode_match_id_union(
1417 struct xfs_inode *ip,
1418 struct xfs_eofblocks *eofb)
1419{
1420 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1421 uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1422 return 1;
1423
1424 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1425 gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1426 return 1;
1427
1428 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1429 xfs_get_projid(ip) == eofb->eof_prid)
1430 return 1;
1431
1432 return 0;
1433}
1434
1435STATIC int
1436xfs_inode_free_eofblocks(
1437 struct xfs_inode *ip,
1438 int flags,
1439 void *args)
1440{
1441 int ret = 0;
1442 struct xfs_eofblocks *eofb = args;
1443 int match;
1444
1445 if (!xfs_can_free_eofblocks(ip, false)) {
1446 /* inode could be preallocated or append-only */
1447 trace_xfs_inode_free_eofblocks_invalid(ip);
1448 xfs_inode_clear_eofblocks_tag(ip);
1449 return 0;
1450 }
1451
1452 /*
1453 * If the mapping is dirty the operation can block and wait for some
1454 * time. Unless we are waiting, skip it.
1455 */
1456 if (!(flags & SYNC_WAIT) &&
1457 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1458 return 0;
1459
1460 if (eofb) {
1461 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1462 match = xfs_inode_match_id_union(ip, eofb);
1463 else
1464 match = xfs_inode_match_id(ip, eofb);
1465 if (!match)
1466 return 0;
1467
1468 /* skip the inode if the file size is too small */
1469 if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1470 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1471 return 0;
1472 }
1473
1474 /*
1475 * If the caller is waiting, return -EAGAIN to keep the background
1476 * scanner moving and revisit the inode in a subsequent pass.
1477 */
1478 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1479 if (flags & SYNC_WAIT)
1480 ret = -EAGAIN;
1481 return ret;
1482 }
1483 ret = xfs_free_eofblocks(ip);
1484 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1485
1486 return ret;
1487}
1488
1489static int
1490__xfs_icache_free_eofblocks(
1491 struct xfs_mount *mp,
1492 struct xfs_eofblocks *eofb,
1493 int (*execute)(struct xfs_inode *ip, int flags,
1494 void *args),
1495 int tag)
1496{
1497 int flags = SYNC_TRYLOCK;
1498
1499 if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
1500 flags = SYNC_WAIT;
1501
1502 return xfs_inode_ag_iterator_tag(mp, execute, flags,
1503 eofb, tag);
1504}
1505
1506int
1507xfs_icache_free_eofblocks(
1508 struct xfs_mount *mp,
1509 struct xfs_eofblocks *eofb)
1510{
1511 return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_eofblocks,
1512 XFS_ICI_EOFBLOCKS_TAG);
1513}
1514
1515/*
1516 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1517 * multiple quotas, we don't know exactly which quota caused an allocation
1518 * failure. We make a best effort by including each quota under low free space
1519 * conditions (less than 1% free space) in the scan.
1520 */
1521static int
1522__xfs_inode_free_quota_eofblocks(
1523 struct xfs_inode *ip,
1524 int (*execute)(struct xfs_mount *mp,
1525 struct xfs_eofblocks *eofb))
1526{
1527 int scan = 0;
1528 struct xfs_eofblocks eofb = {0};
1529 struct xfs_dquot *dq;
1530
1531 /*
1532 * Run a sync scan to increase effectiveness and use the union filter to
1533 * cover all applicable quotas in a single scan.
1534 */
1535 eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC;
1536
1537 if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) {
1538 dq = xfs_inode_dquot(ip, XFS_DQ_USER);
1539 if (dq && xfs_dquot_lowsp(dq)) {
1540 eofb.eof_uid = VFS_I(ip)->i_uid;
1541 eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1542 scan = 1;
1543 }
1544 }
1545
1546 if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) {
1547 dq = xfs_inode_dquot(ip, XFS_DQ_GROUP);
1548 if (dq && xfs_dquot_lowsp(dq)) {
1549 eofb.eof_gid = VFS_I(ip)->i_gid;
1550 eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1551 scan = 1;
1552 }
1553 }
1554
1555 if (scan)
1556 execute(ip->i_mount, &eofb);
1557
1558 return scan;
1559}
1560
1561int
1562xfs_inode_free_quota_eofblocks(
1563 struct xfs_inode *ip)
1564{
1565 return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_eofblocks);
1566}
1567
1568static inline unsigned long
1569xfs_iflag_for_tag(
1570 int tag)
1571{
1572 switch (tag) {
1573 case XFS_ICI_EOFBLOCKS_TAG:
1574 return XFS_IEOFBLOCKS;
1575 case XFS_ICI_COWBLOCKS_TAG:
1576 return XFS_ICOWBLOCKS;
1577 default:
1578 ASSERT(0);
1579 return 0;
1580 }
1581}
1582
1583static void
1584__xfs_inode_set_blocks_tag(
1585 xfs_inode_t *ip,
1586 void (*execute)(struct xfs_mount *mp),
1587 void (*set_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1588 int error, unsigned long caller_ip),
1589 int tag)
1590{
1591 struct xfs_mount *mp = ip->i_mount;
1592 struct xfs_perag *pag;
1593 int tagged;
1594
1595 /*
1596 * Don't bother locking the AG and looking up in the radix trees
1597 * if we already know that we have the tag set.
1598 */
1599 if (ip->i_flags & xfs_iflag_for_tag(tag))
1600 return;
1601 spin_lock(&ip->i_flags_lock);
1602 ip->i_flags |= xfs_iflag_for_tag(tag);
1603 spin_unlock(&ip->i_flags_lock);
1604
1605 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1606 spin_lock(&pag->pag_ici_lock);
1607
1608 tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
1609 radix_tree_tag_set(&pag->pag_ici_root,
1610 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
1611 if (!tagged) {
1612 /* propagate the eofblocks tag up into the perag radix tree */
1613 spin_lock(&ip->i_mount->m_perag_lock);
1614 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
1615 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1616 tag);
1617 spin_unlock(&ip->i_mount->m_perag_lock);
1618
1619 /* kick off background trimming */
1620 execute(ip->i_mount);
1621
1622 set_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
1623 }
1624
1625 spin_unlock(&pag->pag_ici_lock);
1626 xfs_perag_put(pag);
1627}
1628
1629void
1630xfs_inode_set_eofblocks_tag(
1631 xfs_inode_t *ip)
1632{
1633 trace_xfs_inode_set_eofblocks_tag(ip);
1634 return __xfs_inode_set_blocks_tag(ip, xfs_queue_eofblocks,
1635 trace_xfs_perag_set_eofblocks,
1636 XFS_ICI_EOFBLOCKS_TAG);
1637}
1638
1639static void
1640__xfs_inode_clear_blocks_tag(
1641 xfs_inode_t *ip,
1642 void (*clear_tp)(struct xfs_mount *mp, xfs_agnumber_t agno,
1643 int error, unsigned long caller_ip),
1644 int tag)
1645{
1646 struct xfs_mount *mp = ip->i_mount;
1647 struct xfs_perag *pag;
1648
1649 spin_lock(&ip->i_flags_lock);
1650 ip->i_flags &= ~xfs_iflag_for_tag(tag);
1651 spin_unlock(&ip->i_flags_lock);
1652
1653 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1654 spin_lock(&pag->pag_ici_lock);
1655
1656 radix_tree_tag_clear(&pag->pag_ici_root,
1657 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), tag);
1658 if (!radix_tree_tagged(&pag->pag_ici_root, tag)) {
1659 /* clear the eofblocks tag from the perag radix tree */
1660 spin_lock(&ip->i_mount->m_perag_lock);
1661 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1662 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1663 tag);
1664 spin_unlock(&ip->i_mount->m_perag_lock);
1665 clear_tp(ip->i_mount, pag->pag_agno, -1, _RET_IP_);
1666 }
1667
1668 spin_unlock(&pag->pag_ici_lock);
1669 xfs_perag_put(pag);
1670}
1671
1672void
1673xfs_inode_clear_eofblocks_tag(
1674 xfs_inode_t *ip)
1675{
1676 trace_xfs_inode_clear_eofblocks_tag(ip);
1677 return __xfs_inode_clear_blocks_tag(ip,
1678 trace_xfs_perag_clear_eofblocks, XFS_ICI_EOFBLOCKS_TAG);
1679}
1680
1681/*
1682 * Set ourselves up to free CoW blocks from this file. If it's already clean
1683 * then we can bail out quickly, but otherwise we must back off if the file
1684 * is undergoing some kind of write.
1685 */
1686static bool
1687xfs_prep_free_cowblocks(
1688 struct xfs_inode *ip,
1689 struct xfs_ifork *ifp)
1690{
1691 /*
1692 * Just clear the tag if we have an empty cow fork or none at all. It's
1693 * possible the inode was fully unshared since it was originally tagged.
1694 */
1695 if (!xfs_is_reflink_inode(ip) || !ifp->if_bytes) {
1696 trace_xfs_inode_free_cowblocks_invalid(ip);
1697 xfs_inode_clear_cowblocks_tag(ip);
1698 return false;
1699 }
1700
1701 /*
1702 * If the mapping is dirty or under writeback we cannot touch the
1703 * CoW fork. Leave it alone if we're in the midst of a directio.
1704 */
1705 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1706 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1707 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1708 atomic_read(&VFS_I(ip)->i_dio_count))
1709 return false;
1710
1711 return true;
1712}
1713
1714/*
1715 * Automatic CoW Reservation Freeing
1716 *
1717 * These functions automatically garbage collect leftover CoW reservations
1718 * that were made on behalf of a cowextsize hint when we start to run out
1719 * of quota or when the reservations sit around for too long. If the file
1720 * has dirty pages or is undergoing writeback, its CoW reservations will
1721 * be retained.
1722 *
1723 * The actual garbage collection piggybacks off the same code that runs
1724 * the speculative EOF preallocation garbage collector.
1725 */
1726STATIC int
1727xfs_inode_free_cowblocks(
1728 struct xfs_inode *ip,
1729 int flags,
1730 void *args)
1731{
1732 struct xfs_eofblocks *eofb = args;
1733 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1734 int match;
1735 int ret = 0;
1736
1737 if (!xfs_prep_free_cowblocks(ip, ifp))
1738 return 0;
1739
1740 if (eofb) {
1741 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1742 match = xfs_inode_match_id_union(ip, eofb);
1743 else
1744 match = xfs_inode_match_id(ip, eofb);
1745 if (!match)
1746 return 0;
1747
1748 /* skip the inode if the file size is too small */
1749 if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1750 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1751 return 0;
1752 }
1753
1754 /* Free the CoW blocks */
1755 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1756 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1757
1758 /*
1759 * Check again, nobody else should be able to dirty blocks or change
1760 * the reflink iflag now that we have the first two locks held.
1761 */
1762 if (xfs_prep_free_cowblocks(ip, ifp))
1763 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1764
1765 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
1766 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1767
1768 return ret;
1769}
1770
1771int
1772xfs_icache_free_cowblocks(
1773 struct xfs_mount *mp,
1774 struct xfs_eofblocks *eofb)
1775{
1776 return __xfs_icache_free_eofblocks(mp, eofb, xfs_inode_free_cowblocks,
1777 XFS_ICI_COWBLOCKS_TAG);
1778}
1779
1780int
1781xfs_inode_free_quota_cowblocks(
1782 struct xfs_inode *ip)
1783{
1784 return __xfs_inode_free_quota_eofblocks(ip, xfs_icache_free_cowblocks);
1785}
1786
1787void
1788xfs_inode_set_cowblocks_tag(
1789 xfs_inode_t *ip)
1790{
1791 trace_xfs_inode_set_cowblocks_tag(ip);
1792 return __xfs_inode_set_blocks_tag(ip, xfs_queue_cowblocks,
1793 trace_xfs_perag_set_cowblocks,
1794 XFS_ICI_COWBLOCKS_TAG);
1795}
1796
1797void
1798xfs_inode_clear_cowblocks_tag(
1799 xfs_inode_t *ip)
1800{
1801 trace_xfs_inode_clear_cowblocks_tag(ip);
1802 return __xfs_inode_clear_blocks_tag(ip,
1803 trace_xfs_perag_clear_cowblocks, XFS_ICI_COWBLOCKS_TAG);
1804}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
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"
25#include "xfs_ag.h"
26#include "xfs_log_priv.h"
27#include "xfs_health.h"
28#include "xfs_da_format.h"
29#include "xfs_dir2.h"
30#include "xfs_metafile.h"
31
32#include <linux/iversion.h>
33
34/* Radix tree tags for incore inode tree. */
35
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
40
41/*
42 * The goal for walking incore inodes. These can correspond with incore inode
43 * radix tree tags when convenient. Avoid existing XFS_IWALK namespace.
44 */
45enum 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,
49};
50
51static int xfs_icwalk(struct xfs_mount *mp,
52 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
53static int xfs_icwalk_ag(struct xfs_perag *pag,
54 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
55
56/*
57 * Private inode cache walk flags for struct xfs_icwalk. Must not
58 * coincide with XFS_ICWALK_FLAGS_VALID.
59 */
60
61/* Stop scanning after icw_scan_limit inodes. */
62#define XFS_ICWALK_FLAG_SCAN_LIMIT (1U << 28)
63
64#define XFS_ICWALK_FLAG_RECLAIM_SICK (1U << 27)
65#define XFS_ICWALK_FLAG_UNION (1U << 26) /* union filter algorithm */
66
67#define XFS_ICWALK_PRIVATE_FLAGS (XFS_ICWALK_FLAG_SCAN_LIMIT | \
68 XFS_ICWALK_FLAG_RECLAIM_SICK | \
69 XFS_ICWALK_FLAG_UNION)
70
71/* Marks for the perag xarray */
72#define XFS_PERAG_RECLAIM_MARK XA_MARK_0
73#define XFS_PERAG_BLOCKGC_MARK XA_MARK_1
74
75static inline xa_mark_t ici_tag_to_mark(unsigned int tag)
76{
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;
81}
82
83/*
84 * Allocate and initialise an xfs_inode.
85 */
86struct xfs_inode *
87xfs_inode_alloc(
88 struct xfs_mount *mp,
89 xfs_ino_t ino)
90{
91 struct xfs_inode *ip;
92
93 /*
94 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL
95 * and return NULL here on ENOMEM.
96 */
97 ip = alloc_inode_sb(mp->m_super, xfs_inode_cache, GFP_KERNEL | __GFP_NOFAIL);
98
99 if (inode_init_always(mp->m_super, VFS_I(ip))) {
100 kmem_cache_free(xfs_inode_cache, ip);
101 return NULL;
102 }
103
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);
108
109 XFS_STATS_INC(mp, vn_active);
110 ASSERT(atomic_read(&ip->i_pincount) == 0);
111 ASSERT(ip->i_ino == 0);
112
113 /* initialise the xfs inode */
114 ip->i_ino = ino;
115 ip->i_mount = mp;
116 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
117 ip->i_cowfp = NULL;
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));
121 ip->i_flags = 0;
122 ip->i_delayed_blks = 0;
123 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
124 ip->i_nblocks = 0;
125 ip->i_forkoff = 0;
126 ip->i_sick = 0;
127 ip->i_checked = 0;
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;
133
134 return ip;
135}
136
137STATIC void
138xfs_inode_free_callback(
139 struct rcu_head *head)
140{
141 struct inode *inode = container_of(head, struct inode, i_rcu);
142 struct xfs_inode *ip = XFS_I(inode);
143
144 switch (VFS_I(ip)->i_mode & S_IFMT) {
145 case S_IFREG:
146 case S_IFDIR:
147 case S_IFLNK:
148 xfs_idestroy_fork(&ip->i_df);
149 break;
150 }
151
152 xfs_ifork_zap_attr(ip);
153
154 if (ip->i_cowfp) {
155 xfs_idestroy_fork(ip->i_cowfp);
156 kmem_cache_free(xfs_ifork_cache, ip->i_cowfp);
157 }
158 if (ip->i_itemp) {
159 ASSERT(!test_bit(XFS_LI_IN_AIL,
160 &ip->i_itemp->ili_item.li_flags));
161 xfs_inode_item_destroy(ip);
162 ip->i_itemp = NULL;
163 }
164
165 kmem_cache_free(xfs_inode_cache, ip);
166}
167
168static void
169__xfs_inode_free(
170 struct xfs_inode *ip)
171{
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);
176
177 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
178}
179
180void
181xfs_inode_free(
182 struct xfs_inode *ip)
183{
184 ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
185
186 /*
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
190 * races.
191 */
192 spin_lock(&ip->i_flags_lock);
193 ip->i_flags = XFS_IRECLAIM;
194 ip->i_ino = 0;
195 spin_unlock(&ip->i_flags_lock);
196
197 __xfs_inode_free(ip);
198}
199
200/*
201 * Queue background inode reclaim work if there are reclaimable inodes and there
202 * isn't reclaim work already scheduled or in progress.
203 */
204static void
205xfs_reclaim_work_queue(
206 struct xfs_mount *mp)
207{
208
209 rcu_read_lock();
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));
213 }
214 rcu_read_unlock();
215}
216
217/*
218 * Background scanning to trim preallocated space. This is queued based on the
219 * 'speculative_prealloc_lifetime' tunable (5m by default).
220 */
221static inline void
222xfs_blockgc_queue(
223 struct xfs_perag *pag)
224{
225 struct xfs_mount *mp = pag_mount(pag);
226
227 if (!xfs_is_blockgc_enabled(mp))
228 return;
229
230 rcu_read_lock();
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));
234 rcu_read_unlock();
235}
236
237/* Set a tag on both the AG incore inode tree and the AG radix tree. */
238static void
239xfs_perag_set_inode_tag(
240 struct xfs_perag *pag,
241 xfs_agino_t agino,
242 unsigned int tag)
243{
244 bool was_tagged;
245
246 lockdep_assert_held(&pag->pag_ici_lock);
247
248 was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
249 radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
250
251 if (tag == XFS_ICI_RECLAIM_TAG)
252 pag->pag_ici_reclaimable++;
253
254 if (was_tagged)
255 return;
256
257 /* propagate the tag up into the pag xarray tree */
258 xfs_group_set_mark(pag_group(pag), ici_tag_to_mark(tag));
259
260 /* start background work */
261 switch (tag) {
262 case XFS_ICI_RECLAIM_TAG:
263 xfs_reclaim_work_queue(pag_mount(pag));
264 break;
265 case XFS_ICI_BLOCKGC_TAG:
266 xfs_blockgc_queue(pag);
267 break;
268 }
269
270 trace_xfs_perag_set_inode_tag(pag, _RET_IP_);
271}
272
273/* Clear a tag on both the AG incore inode tree and the AG radix tree. */
274static void
275xfs_perag_clear_inode_tag(
276 struct xfs_perag *pag,
277 xfs_agino_t agino,
278 unsigned int tag)
279{
280 lockdep_assert_held(&pag->pag_ici_lock);
281
282 /*
283 * Reclaim can signal (with a null agino) that it cleared its own tag
284 * by removing the inode from the radix tree.
285 */
286 if (agino != NULLAGINO)
287 radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
288 else
289 ASSERT(tag == XFS_ICI_RECLAIM_TAG);
290
291 if (tag == XFS_ICI_RECLAIM_TAG)
292 pag->pag_ici_reclaimable--;
293
294 if (radix_tree_tagged(&pag->pag_ici_root, tag))
295 return;
296
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_);
300}
301
302/*
303 * Find the next AG after @pag, or the first AG if @pag is NULL.
304 */
305static struct xfs_perag *
306xfs_perag_grab_next_tag(
307 struct xfs_mount *mp,
308 struct xfs_perag *pag,
309 int tag)
310{
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));
314}
315
316/*
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.
323 */
324static int
325xfs_reinit_inode(
326 struct xfs_mount *mp,
327 struct inode *inode)
328{
329 int error;
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;
338
339 error = inode_init_always(mp->m_super, inode);
340
341 set_nlink(inode, nlink);
342 inode->i_generation = generation;
343 inode_set_iversion_queried(inode, version);
344 inode->i_mode = mode;
345 inode->i_rdev = dev;
346 inode->i_uid = uid;
347 inode->i_gid = gid;
348 inode->i_state = state;
349 mapping_set_folio_min_order(inode->i_mapping,
350 M_IGEO(mp)->min_folio_order);
351 return error;
352}
353
354/*
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.
357 */
358static int
359xfs_iget_recycle(
360 struct xfs_perag *pag,
361 struct xfs_inode *ip) __releases(&ip->i_flags_lock)
362{
363 struct xfs_mount *mp = ip->i_mount;
364 struct inode *inode = VFS_I(ip);
365 int error;
366
367 trace_xfs_iget_recycle(ip);
368
369 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
370 return -EAGAIN;
371
372 /*
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.
377 */
378 ip->i_flags |= XFS_IRECLAIM;
379
380 spin_unlock(&ip->i_flags_lock);
381 rcu_read_unlock();
382
383 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
384 error = xfs_reinit_inode(mp, inode);
385 xfs_iunlock(ip, XFS_ILOCK_EXCL);
386 if (error) {
387 /*
388 * Re-initializing the inode failed, and we are in deep
389 * trouble. Try to re-add it to the reclaim list.
390 */
391 rcu_read_lock();
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);
396 rcu_read_unlock();
397
398 trace_xfs_iget_recycle_fail(ip);
399 return error;
400 }
401
402 spin_lock(&pag->pag_ici_lock);
403 spin_lock(&ip->i_flags_lock);
404
405 /*
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
408 * occurs.
409 */
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);
417
418 return 0;
419}
420
421/*
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.
425 *
426 * Returns:
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
430 */
431static int
432xfs_iget_check_free_state(
433 struct xfs_inode *ip,
434 int flags)
435{
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),
444 XFS_SICK_AG_INOBT);
445 return -EFSCORRUPTED;
446 }
447
448 if (ip->i_nblocks != 0) {
449 xfs_warn(ip->i_mount,
450"Corruption detected! Free inode 0x%llx has blocks allocated!",
451 ip->i_ino);
452 xfs_agno_mark_sick(ip->i_mount,
453 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
454 XFS_SICK_AG_INOBT);
455 return -EFSCORRUPTED;
456 }
457 return 0;
458 }
459
460 /* should be an allocated inode */
461 if (VFS_I(ip)->i_mode == 0)
462 return -ENOENT;
463
464 return 0;
465}
466
467/* Make all pending inactivation work start immediately. */
468static bool
469xfs_inodegc_queue_all(
470 struct xfs_mount *mp)
471{
472 struct xfs_inodegc *gc;
473 int cpu;
474 bool ret = false;
475
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);
480 ret = true;
481 }
482 }
483
484 return ret;
485}
486
487/* Wait for all queued work and collect errors */
488static int
489xfs_inodegc_wait_all(
490 struct xfs_mount *mp)
491{
492 int cpu;
493 int error = 0;
494
495 flush_workqueue(mp->m_inodegc_wq);
496 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
497 struct xfs_inodegc *gc;
498
499 gc = per_cpu_ptr(mp->m_inodegc, cpu);
500 if (gc->error && !error)
501 error = gc->error;
502 gc->error = 0;
503 }
504
505 return error;
506}
507
508/*
509 * Check the validity of the inode we just found it the cache
510 */
511static int
512xfs_iget_cache_hit(
513 struct xfs_perag *pag,
514 struct xfs_inode *ip,
515 xfs_ino_t ino,
516 int flags,
517 int lock_flags) __releases(RCU)
518{
519 struct inode *inode = VFS_I(ip);
520 struct xfs_mount *mp = ip->i_mount;
521 int error;
522
523 /*
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.
529 */
530 spin_lock(&ip->i_flags_lock);
531 if (ip->i_ino != ino)
532 goto out_skip;
533
534 /*
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
538 * before continuing.
539 *
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.
547 *
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.
551 */
552 if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
553 goto out_skip;
554
555 if (ip->i_flags & XFS_NEED_INACTIVE) {
556 /* Unlinked inodes cannot be re-grabbed. */
557 if (VFS_I(ip)->i_nlink == 0) {
558 error = -ENOENT;
559 goto out_error;
560 }
561 goto out_inodegc_flush;
562 }
563
564 /*
565 * Check the inode free state is valid. This also detects lookup
566 * racing with unlinks.
567 */
568 error = xfs_iget_check_free_state(ip, flags);
569 if (error)
570 goto out_error;
571
572 /* Skip inodes that have no vfs state. */
573 if ((flags & XFS_IGET_INCORE) &&
574 (ip->i_flags & XFS_IRECLAIMABLE))
575 goto out_skip;
576
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)
582 goto out_skip;
583 if (error)
584 return error;
585 } else {
586 /* If the VFS inode is being torn down, pause and try again. */
587 if (!igrab(inode))
588 goto out_skip;
589
590 /* We've got a live one. */
591 spin_unlock(&ip->i_flags_lock);
592 rcu_read_unlock();
593 trace_xfs_iget_hit(ip);
594 }
595
596 if (lock_flags != 0)
597 xfs_ilock(ip, lock_flags);
598
599 if (!(flags & XFS_IGET_INCORE))
600 xfs_iflags_clear(ip, XFS_ISTALE);
601 XFS_STATS_INC(mp, xs_ig_found);
602
603 return 0;
604
605out_skip:
606 trace_xfs_iget_skip(ip);
607 XFS_STATS_INC(mp, xs_ig_frecycle);
608 error = -EAGAIN;
609out_error:
610 spin_unlock(&ip->i_flags_lock);
611 rcu_read_unlock();
612 return error;
613
614out_inodegc_flush:
615 spin_unlock(&ip->i_flags_lock);
616 rcu_read_unlock();
617 /*
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.
620 */
621 if (xfs_is_inodegc_enabled(mp))
622 xfs_inodegc_queue_all(mp);
623 return -EAGAIN;
624}
625
626static int
627xfs_iget_cache_miss(
628 struct xfs_mount *mp,
629 struct xfs_perag *pag,
630 xfs_trans_t *tp,
631 xfs_ino_t ino,
632 struct xfs_inode **ipp,
633 int flags,
634 int lock_flags)
635{
636 struct xfs_inode *ip;
637 int error;
638 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
639
640 ip = xfs_inode_alloc(mp, ino);
641 if (!ip)
642 return -ENOMEM;
643
644 error = xfs_imap(pag, tp, ip->i_ino, &ip->i_imap, flags);
645 if (error)
646 goto out_destroy;
647
648 /*
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.
652 *
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.
657 */
658 if (xfs_has_v3inodes(mp) &&
659 (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
660 VFS_I(ip)->i_generation = get_random_u32();
661 } else {
662 struct xfs_buf *bp;
663
664 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
665 if (error)
666 goto out_destroy;
667
668 error = xfs_inode_from_disk(ip,
669 xfs_buf_offset(bp, ip->i_imap.im_boffset));
670 if (!error)
671 xfs_buf_set_ref(bp, XFS_INO_REF);
672 else
673 xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
674 xfs_trans_brelse(tp, bp);
675
676 if (error)
677 goto out_destroy;
678 }
679
680 trace_xfs_iget_miss(ip);
681
682 /*
683 * Check the inode free state is valid. This also detects lookup
684 * racing with unlinks.
685 */
686 error = xfs_iget_check_free_state(ip, flags);
687 if (error)
688 goto out_destroy;
689
690 /*
691 * Preload the radix tree so we can insert safely under the
692 * write spinlock. Note that we cannot sleep inside the preload
693 * region.
694 */
695 if (radix_tree_preload(GFP_KERNEL | __GFP_NOLOCKDEP)) {
696 error = -EAGAIN;
697 goto out_destroy;
698 }
699
700 /*
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.
703 */
704 if (lock_flags) {
705 if (!xfs_ilock_nowait(ip, lock_flags))
706 BUG();
707 }
708
709 /*
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
716 * time.
717 */
718 if (flags & XFS_IGET_DONTCACHE)
719 d_mark_dontcache(VFS_I(ip));
720 ip->i_udquot = NULL;
721 ip->i_gdquot = NULL;
722 ip->i_pdquot = NULL;
723 xfs_iflags_set(ip, XFS_INEW);
724
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);
731 error = -EAGAIN;
732 goto out_preload_end;
733 }
734 spin_unlock(&pag->pag_ici_lock);
735 radix_tree_preload_end();
736
737 *ipp = ip;
738 return 0;
739
740out_preload_end:
741 spin_unlock(&pag->pag_ici_lock);
742 radix_tree_preload_end();
743 if (lock_flags)
744 xfs_iunlock(ip, lock_flags);
745out_destroy:
746 __destroy_inode(VFS_I(ip));
747 xfs_inode_free(ip);
748 return error;
749}
750
751/*
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.
755 *
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.
758 *
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.
762 */
763int
764xfs_iget(
765 struct xfs_mount *mp,
766 struct xfs_trans *tp,
767 xfs_ino_t ino,
768 uint flags,
769 uint lock_flags,
770 struct xfs_inode **ipp)
771{
772 struct xfs_inode *ip;
773 struct xfs_perag *pag;
774 xfs_agino_t agino;
775 int error;
776
777 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
778
779 /* reject inode numbers outside existing AGs */
780 if (!xfs_verify_ino(mp, ino))
781 return -EINVAL;
782
783 XFS_STATS_INC(mp, xs_ig_attempts);
784
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);
788
789again:
790 error = 0;
791 rcu_read_lock();
792 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
793
794 if (ip) {
795 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
796 if (error)
797 goto out_error_or_again;
798 } else {
799 rcu_read_unlock();
800 if (flags & XFS_IGET_INCORE) {
801 error = -ENODATA;
802 goto out_error_or_again;
803 }
804 XFS_STATS_INC(mp, xs_ig_missed);
805
806 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
807 flags, lock_flags);
808 if (error)
809 goto out_error_or_again;
810 }
811 xfs_perag_put(pag);
812
813 *ipp = ip;
814
815 /*
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
818 * handle it.
819 */
820 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
821 xfs_setup_existing_inode(ip);
822 return 0;
823
824out_error_or_again:
825 if (!(flags & (XFS_IGET_INCORE | XFS_IGET_NORETRY)) &&
826 error == -EAGAIN) {
827 delay(1);
828 goto again;
829 }
830 xfs_perag_put(pag);
831 return error;
832}
833
834/*
835 * Get a metadata inode.
836 *
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.
839 */
840int
841xfs_trans_metafile_iget(
842 struct xfs_trans *tp,
843 xfs_ino_t ino,
844 enum xfs_metafile_type metafile_type,
845 struct xfs_inode **ipp)
846{
847 struct xfs_mount *mp = tp->t_mountp;
848 struct xfs_inode *ip;
849 umode_t mode;
850 int error;
851
852 error = xfs_iget(mp, tp, ino, 0, 0, &ip);
853 if (error == -EFSCORRUPTED || error == -EINVAL)
854 goto whine;
855 if (error)
856 return error;
857
858 if (VFS_I(ip)->i_nlink == 0)
859 goto bad_rele;
860
861 if (metafile_type == XFS_METAFILE_DIR)
862 mode = S_IFDIR;
863 else
864 mode = S_IFREG;
865 if (inode_wrong_type(VFS_I(ip), mode))
866 goto bad_rele;
867 if (xfs_has_metadir(mp)) {
868 if (!xfs_is_metadir_inode(ip))
869 goto bad_rele;
870 if (metafile_type != ip->i_metatype)
871 goto bad_rele;
872 }
873
874 *ipp = ip;
875 return 0;
876bad_rele:
877 xfs_irele(ip);
878whine:
879 xfs_err(mp, "metadata inode 0x%llx type %u is corrupt", ino,
880 metafile_type);
881 xfs_fs_mark_sick(mp, XFS_SICK_FS_METADIR);
882 return -EFSCORRUPTED;
883}
884
885/* Grab a metadata file if the caller doesn't already have a transaction. */
886int
887xfs_metafile_iget(
888 struct xfs_mount *mp,
889 xfs_ino_t ino,
890 enum xfs_metafile_type metafile_type,
891 struct xfs_inode **ipp)
892{
893 struct xfs_trans *tp;
894 int error;
895
896 error = xfs_trans_alloc_empty(mp, &tp);
897 if (error)
898 return error;
899
900 error = xfs_trans_metafile_iget(tp, ino, metafile_type, ipp);
901 xfs_trans_cancel(tp);
902 return error;
903}
904
905/*
906 * Grab the inode for reclaim exclusively.
907 *
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.
914 *
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.
919 *
920 * Return true if we grabbed it, false otherwise.
921 */
922static bool
923xfs_reclaim_igrab(
924 struct xfs_inode *ip,
925 struct xfs_icwalk *icw)
926{
927 ASSERT(rcu_read_lock_held());
928
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);
934 return false;
935 }
936
937 /* Don't reclaim a sick inode unless the caller asked for it. */
938 if (ip->i_sick &&
939 (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
940 spin_unlock(&ip->i_flags_lock);
941 return false;
942 }
943
944 __xfs_iflags_set(ip, XFS_IRECLAIM);
945 spin_unlock(&ip->i_flags_lock);
946 return true;
947}
948
949/*
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
954 * it.
955 *
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.
960 */
961static void
962xfs_reclaim_inode(
963 struct xfs_inode *ip,
964 struct xfs_perag *pag)
965{
966 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
967
968 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
969 goto out;
970 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
971 goto out_iunlock;
972
973 /*
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.
979 */
980 if (xlog_is_shutdown(ip->i_mount->m_log)) {
981 xfs_iunpin_wait(ip);
982 xfs_iflush_shutdown_abort(ip);
983 goto reclaim;
984 }
985 if (xfs_ipincount(ip))
986 goto out_clear_flush;
987 if (!xfs_inode_clean(ip))
988 goto out_clear_flush;
989
990 xfs_iflags_clear(ip, XFS_IFLUSHING);
991reclaim:
992 trace_xfs_inode_reclaiming(ip);
993
994 /*
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.
1003 */
1004 spin_lock(&ip->i_flags_lock);
1005 ip->i_flags = XFS_IRECLAIM;
1006 ip->i_ino = 0;
1007 ip->i_sick = 0;
1008 ip->i_checked = 0;
1009 spin_unlock(&ip->i_flags_lock);
1010
1011 ASSERT(!ip->i_itemp || ip->i_itemp->ili_item.li_buf == NULL);
1012 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1013
1014 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1015 /*
1016 * Remove the inode from the per-AG radix tree.
1017 *
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.
1021 */
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)))
1025 ASSERT(0);
1026 xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
1027 spin_unlock(&pag->pag_ici_lock);
1028
1029 /*
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.
1033 *
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.
1036 */
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));
1041
1042 __xfs_inode_free(ip);
1043 return;
1044
1045out_clear_flush:
1046 xfs_iflags_clear(ip, XFS_IFLUSHING);
1047out_iunlock:
1048 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1049out:
1050 xfs_iflags_clear(ip, XFS_IRECLAIM);
1051}
1052
1053/* Reclaim sick inodes if we're unmounting or the fs went down. */
1054static inline bool
1055xfs_want_reclaim_sick(
1056 struct xfs_mount *mp)
1057{
1058 return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
1059 xfs_is_shutdown(mp);
1060}
1061
1062void
1063xfs_reclaim_inodes(
1064 struct xfs_mount *mp)
1065{
1066 struct xfs_icwalk icw = {
1067 .icw_flags = 0,
1068 };
1069
1070 if (xfs_want_reclaim_sick(mp))
1071 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1072
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);
1076 }
1077}
1078
1079/*
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.
1085 */
1086long
1087xfs_reclaim_inodes_nr(
1088 struct xfs_mount *mp,
1089 unsigned long nr_to_scan)
1090{
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),
1094 };
1095
1096 if (xfs_want_reclaim_sick(mp))
1097 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1098
1099 /* kick background reclaimer and push the AIL */
1100 xfs_reclaim_work_queue(mp);
1101 xfs_ail_push_all(mp->m_ail);
1102
1103 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1104 return 0;
1105}
1106
1107/*
1108 * Return the number of reclaimable inodes in the filesystem for
1109 * the shrinker to determine how much to reclaim.
1110 */
1111long
1112xfs_reclaim_inodes_count(
1113 struct xfs_mount *mp)
1114{
1115 XA_STATE (xas, &mp->m_groups[XG_TYPE_AG].xa, 0);
1116 long reclaimable = 0;
1117 struct xfs_perag *pag;
1118
1119 rcu_read_lock();
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;
1123 }
1124 rcu_read_unlock();
1125
1126 return reclaimable;
1127}
1128
1129STATIC bool
1130xfs_icwalk_match_id(
1131 struct xfs_inode *ip,
1132 struct xfs_icwalk *icw)
1133{
1134 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1135 !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1136 return false;
1137
1138 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1139 !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1140 return false;
1141
1142 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1143 ip->i_projid != icw->icw_prid)
1144 return false;
1145
1146 return true;
1147}
1148
1149/*
1150 * A union-based inode filtering algorithm. Process the inode if any of the
1151 * criteria match. This is for global/internal scans only.
1152 */
1153STATIC bool
1154xfs_icwalk_match_id_union(
1155 struct xfs_inode *ip,
1156 struct xfs_icwalk *icw)
1157{
1158 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1159 uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1160 return true;
1161
1162 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1163 gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1164 return true;
1165
1166 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1167 ip->i_projid == icw->icw_prid)
1168 return true;
1169
1170 return false;
1171}
1172
1173/*
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.
1177 */
1178static bool
1179xfs_icwalk_match(
1180 struct xfs_inode *ip,
1181 struct xfs_icwalk *icw)
1182{
1183 bool match;
1184
1185 if (!icw)
1186 return true;
1187
1188 if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1189 match = xfs_icwalk_match_id_union(ip, icw);
1190 else
1191 match = xfs_icwalk_match_id(ip, icw);
1192 if (!match)
1193 return false;
1194
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)
1198 return false;
1199
1200 return true;
1201}
1202
1203/*
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
1207 * goes low.
1208 */
1209void
1210xfs_reclaim_worker(
1211 struct work_struct *work)
1212{
1213 struct xfs_mount *mp = container_of(to_delayed_work(work),
1214 struct xfs_mount, m_reclaim_work);
1215
1216 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1217 xfs_reclaim_work_queue(mp);
1218}
1219
1220STATIC int
1221xfs_inode_free_eofblocks(
1222 struct xfs_inode *ip,
1223 struct xfs_icwalk *icw,
1224 unsigned int *lockflags)
1225{
1226 bool wait;
1227
1228 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1229
1230 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1231 return 0;
1232
1233 /*
1234 * If the mapping is dirty the operation can block and wait for some
1235 * time. Unless we are waiting, skip it.
1236 */
1237 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1238 return 0;
1239
1240 if (!xfs_icwalk_match(ip, icw))
1241 return 0;
1242
1243 /*
1244 * If the caller is waiting, return -EAGAIN to keep the background
1245 * scanner moving and revisit the inode in a subsequent pass.
1246 */
1247 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1248 if (wait)
1249 return -EAGAIN;
1250 return 0;
1251 }
1252 *lockflags |= XFS_IOLOCK_EXCL;
1253
1254 if (xfs_can_free_eofblocks(ip))
1255 return xfs_free_eofblocks(ip);
1256
1257 /* inode could be preallocated */
1258 trace_xfs_inode_free_eofblocks_invalid(ip);
1259 xfs_inode_clear_eofblocks_tag(ip);
1260 return 0;
1261}
1262
1263static void
1264xfs_blockgc_set_iflag(
1265 struct xfs_inode *ip,
1266 unsigned long iflag)
1267{
1268 struct xfs_mount *mp = ip->i_mount;
1269 struct xfs_perag *pag;
1270
1271 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1272
1273 /*
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.
1276 */
1277 if (ip->i_flags & iflag)
1278 return;
1279 spin_lock(&ip->i_flags_lock);
1280 ip->i_flags |= iflag;
1281 spin_unlock(&ip->i_flags_lock);
1282
1283 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1284 spin_lock(&pag->pag_ici_lock);
1285
1286 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1287 XFS_ICI_BLOCKGC_TAG);
1288
1289 spin_unlock(&pag->pag_ici_lock);
1290 xfs_perag_put(pag);
1291}
1292
1293void
1294xfs_inode_set_eofblocks_tag(
1295 xfs_inode_t *ip)
1296{
1297 trace_xfs_inode_set_eofblocks_tag(ip);
1298 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1299}
1300
1301static void
1302xfs_blockgc_clear_iflag(
1303 struct xfs_inode *ip,
1304 unsigned long iflag)
1305{
1306 struct xfs_mount *mp = ip->i_mount;
1307 struct xfs_perag *pag;
1308 bool clear_tag;
1309
1310 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1311
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);
1316
1317 if (!clear_tag)
1318 return;
1319
1320 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1321 spin_lock(&pag->pag_ici_lock);
1322
1323 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1324 XFS_ICI_BLOCKGC_TAG);
1325
1326 spin_unlock(&pag->pag_ici_lock);
1327 xfs_perag_put(pag);
1328}
1329
1330void
1331xfs_inode_clear_eofblocks_tag(
1332 xfs_inode_t *ip)
1333{
1334 trace_xfs_inode_clear_eofblocks_tag(ip);
1335 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1336}
1337
1338/*
1339 * Prepare to free COW fork blocks from an inode.
1340 */
1341static bool
1342xfs_prep_free_cowblocks(
1343 struct xfs_inode *ip,
1344 struct xfs_icwalk *icw)
1345{
1346 bool sync;
1347
1348 sync = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1349
1350 /*
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.
1353 */
1354 if (!xfs_inode_has_cow_data(ip)) {
1355 trace_xfs_inode_free_cowblocks_invalid(ip);
1356 xfs_inode_clear_cowblocks_tag(ip);
1357 return false;
1358 }
1359
1360 /*
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.
1368 *
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.
1373 */
1374 if (!sync && inode_is_open_for_write(VFS_I(ip)))
1375 return false;
1376 return xfs_can_free_cowblocks(ip);
1377}
1378
1379/*
1380 * Automatic CoW Reservation Freeing
1381 *
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
1386 * be retained.
1387 *
1388 * The actual garbage collection piggybacks off the same code that runs
1389 * the speculative EOF preallocation garbage collector.
1390 */
1391STATIC int
1392xfs_inode_free_cowblocks(
1393 struct xfs_inode *ip,
1394 struct xfs_icwalk *icw,
1395 unsigned int *lockflags)
1396{
1397 bool wait;
1398 int ret = 0;
1399
1400 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1401
1402 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1403 return 0;
1404
1405 if (!xfs_prep_free_cowblocks(ip, icw))
1406 return 0;
1407
1408 if (!xfs_icwalk_match(ip, icw))
1409 return 0;
1410
1411 /*
1412 * If the caller is waiting, return -EAGAIN to keep the background
1413 * scanner moving and revisit the inode in a subsequent pass.
1414 */
1415 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1416 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1417 if (wait)
1418 return -EAGAIN;
1419 return 0;
1420 }
1421 *lockflags |= XFS_IOLOCK_EXCL;
1422
1423 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1424 if (wait)
1425 return -EAGAIN;
1426 return 0;
1427 }
1428 *lockflags |= XFS_MMAPLOCK_EXCL;
1429
1430 /*
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.
1433 */
1434 if (xfs_prep_free_cowblocks(ip, icw))
1435 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1436 return ret;
1437}
1438
1439void
1440xfs_inode_set_cowblocks_tag(
1441 xfs_inode_t *ip)
1442{
1443 trace_xfs_inode_set_cowblocks_tag(ip);
1444 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1445}
1446
1447void
1448xfs_inode_clear_cowblocks_tag(
1449 xfs_inode_t *ip)
1450{
1451 trace_xfs_inode_clear_cowblocks_tag(ip);
1452 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1453}
1454
1455/* Disable post-EOF and CoW block auto-reclamation. */
1456void
1457xfs_blockgc_stop(
1458 struct xfs_mount *mp)
1459{
1460 struct xfs_perag *pag = NULL;
1461
1462 if (!xfs_clear_blockgc_enabled(mp))
1463 return;
1464
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);
1468}
1469
1470/* Enable post-EOF and CoW block auto-reclamation. */
1471void
1472xfs_blockgc_start(
1473 struct xfs_mount *mp)
1474{
1475 struct xfs_perag *pag = NULL;
1476
1477 if (xfs_set_blockgc_enabled(mp))
1478 return;
1479
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);
1483}
1484
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 | \
1490 XFS_IRECLAIM)
1491/*
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.
1495 */
1496static bool
1497xfs_blockgc_igrab(
1498 struct xfs_inode *ip)
1499{
1500 struct inode *inode = VFS_I(ip);
1501
1502 ASSERT(rcu_read_lock_held());
1503
1504 /* Check for stale RCU freed inode */
1505 spin_lock(&ip->i_flags_lock);
1506 if (!ip->i_ino)
1507 goto out_unlock_noent;
1508
1509 if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1510 goto out_unlock_noent;
1511 spin_unlock(&ip->i_flags_lock);
1512
1513 /* nothing to sync during shutdown */
1514 if (xfs_is_shutdown(ip->i_mount))
1515 return false;
1516
1517 /* If we can't grab the inode, it must on it's way to reclaim. */
1518 if (!igrab(inode))
1519 return false;
1520
1521 /* inode is valid */
1522 return true;
1523
1524out_unlock_noent:
1525 spin_unlock(&ip->i_flags_lock);
1526 return false;
1527}
1528
1529/* Scan one incore inode for block preallocations that we can remove. */
1530static int
1531xfs_blockgc_scan_inode(
1532 struct xfs_inode *ip,
1533 struct xfs_icwalk *icw)
1534{
1535 unsigned int lockflags = 0;
1536 int error;
1537
1538 error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1539 if (error)
1540 goto unlock;
1541
1542 error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1543unlock:
1544 if (lockflags)
1545 xfs_iunlock(ip, lockflags);
1546 xfs_irele(ip);
1547 return error;
1548}
1549
1550/* Background worker that trims preallocated space. */
1551void
1552xfs_blockgc_worker(
1553 struct work_struct *work)
1554{
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);
1558 int error;
1559
1560 trace_xfs_blockgc_worker(mp, __return_address);
1561
1562 error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1563 if (error)
1564 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1565 pag_agno(pag), error);
1566 xfs_blockgc_queue(pag);
1567}
1568
1569/*
1570 * Try to free space in the filesystem by purging inactive inodes, eofblocks
1571 * and cowblocks.
1572 */
1573int
1574xfs_blockgc_free_space(
1575 struct xfs_mount *mp,
1576 struct xfs_icwalk *icw)
1577{
1578 int error;
1579
1580 trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1581
1582 error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1583 if (error)
1584 return error;
1585
1586 return xfs_inodegc_flush(mp);
1587}
1588
1589/*
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.
1592 */
1593int
1594xfs_blockgc_flush_all(
1595 struct xfs_mount *mp)
1596{
1597 struct xfs_perag *pag = NULL;
1598
1599 trace_xfs_blockgc_flush_all(mp, __return_address);
1600
1601 /*
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.
1604 */
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);
1607
1608 while ((pag = xfs_perag_grab_next_tag(mp, pag, XFS_ICI_BLOCKGC_TAG)))
1609 flush_delayed_work(&pag->pag_blockgc_work);
1610
1611 return xfs_inodegc_flush(mp);
1612}
1613
1614/*
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
1618 * scan.
1619 *
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
1622 * MMAPLOCK.
1623 */
1624int
1625xfs_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)
1631{
1632 struct xfs_icwalk icw = {0};
1633 bool do_work = false;
1634
1635 if (!udqp && !gdqp && !pdqp)
1636 return 0;
1637
1638 /*
1639 * Run a scan to free blocks using the union filter to cover all
1640 * applicable quotas in a single scan.
1641 */
1642 icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1643
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;
1647 do_work = true;
1648 }
1649
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;
1653 do_work = true;
1654 }
1655
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;
1659 do_work = true;
1660 }
1661
1662 if (!do_work)
1663 return 0;
1664
1665 return xfs_blockgc_free_space(mp, &icw);
1666}
1667
1668/* Run cow/eofblocks scans on the quotas attached to the inode. */
1669int
1670xfs_blockgc_free_quota(
1671 struct xfs_inode *ip,
1672 unsigned int iwalk_flags)
1673{
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);
1678}
1679
1680/* XFS Inode Cache Walking Code */
1681
1682/*
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
1686 * be too greedy.
1687 */
1688#define XFS_LOOKUP_BATCH 32
1689
1690
1691/*
1692 * Decide if we want to grab this inode in anticipation of doing work towards
1693 * the goal.
1694 */
1695static inline bool
1696xfs_icwalk_igrab(
1697 enum xfs_icwalk_goal goal,
1698 struct xfs_inode *ip,
1699 struct xfs_icwalk *icw)
1700{
1701 switch (goal) {
1702 case XFS_ICWALK_BLOCKGC:
1703 return xfs_blockgc_igrab(ip);
1704 case XFS_ICWALK_RECLAIM:
1705 return xfs_reclaim_igrab(ip, icw);
1706 default:
1707 return false;
1708 }
1709}
1710
1711/*
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.
1714 */
1715static inline int
1716xfs_icwalk_process_inode(
1717 enum xfs_icwalk_goal goal,
1718 struct xfs_inode *ip,
1719 struct xfs_perag *pag,
1720 struct xfs_icwalk *icw)
1721{
1722 int error = 0;
1723
1724 switch (goal) {
1725 case XFS_ICWALK_BLOCKGC:
1726 error = xfs_blockgc_scan_inode(ip, icw);
1727 break;
1728 case XFS_ICWALK_RECLAIM:
1729 xfs_reclaim_inode(ip, pag);
1730 break;
1731 }
1732 return error;
1733}
1734
1735/*
1736 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1737 * process them in some manner.
1738 */
1739static int
1740xfs_icwalk_ag(
1741 struct xfs_perag *pag,
1742 enum xfs_icwalk_goal goal,
1743 struct xfs_icwalk *icw)
1744{
1745 struct xfs_mount *mp = pag_mount(pag);
1746 uint32_t first_index;
1747 int last_error = 0;
1748 int skipped;
1749 bool done;
1750 int nr_found;
1751
1752restart:
1753 done = false;
1754 skipped = 0;
1755 if (goal == XFS_ICWALK_RECLAIM)
1756 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1757 else
1758 first_index = 0;
1759 nr_found = 0;
1760 do {
1761 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1762 int error = 0;
1763 int i;
1764
1765 rcu_read_lock();
1766
1767 nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1768 (void **) batch, first_index,
1769 XFS_LOOKUP_BATCH, goal);
1770 if (!nr_found) {
1771 done = true;
1772 rcu_read_unlock();
1773 break;
1774 }
1775
1776 /*
1777 * Grab the inodes before we drop the lock. if we found
1778 * nothing, nr == 0 and the loop will be skipped.
1779 */
1780 for (i = 0; i < nr_found; i++) {
1781 struct xfs_inode *ip = batch[i];
1782
1783 if (done || !xfs_icwalk_igrab(goal, ip, icw))
1784 batch[i] = NULL;
1785
1786 /*
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.
1791 *
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.
1797 */
1798 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag_agno(pag))
1799 continue;
1800 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1801 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1802 done = true;
1803 }
1804
1805 /* unlock now we've grabbed the inodes. */
1806 rcu_read_unlock();
1807
1808 for (i = 0; i < nr_found; i++) {
1809 if (!batch[i])
1810 continue;
1811 error = xfs_icwalk_process_inode(goal, batch[i], pag,
1812 icw);
1813 if (error == -EAGAIN) {
1814 skipped++;
1815 continue;
1816 }
1817 if (error && last_error != -EFSCORRUPTED)
1818 last_error = error;
1819 }
1820
1821 /* bail out if the filesystem is corrupted. */
1822 if (error == -EFSCORRUPTED)
1823 break;
1824
1825 cond_resched();
1826
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)
1830 break;
1831 }
1832 } while (nr_found && !done);
1833
1834 if (goal == XFS_ICWALK_RECLAIM) {
1835 if (done)
1836 first_index = 0;
1837 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1838 }
1839
1840 if (skipped) {
1841 delay(1);
1842 goto restart;
1843 }
1844 return last_error;
1845}
1846
1847/* Walk all incore inodes to achieve a given goal. */
1848static int
1849xfs_icwalk(
1850 struct xfs_mount *mp,
1851 enum xfs_icwalk_goal goal,
1852 struct xfs_icwalk *icw)
1853{
1854 struct xfs_perag *pag = NULL;
1855 int error = 0;
1856 int last_error = 0;
1857
1858 while ((pag = xfs_perag_grab_next_tag(mp, pag, goal))) {
1859 error = xfs_icwalk_ag(pag, goal, icw);
1860 if (error) {
1861 last_error = error;
1862 if (error == -EFSCORRUPTED) {
1863 xfs_perag_rele(pag);
1864 break;
1865 }
1866 }
1867 }
1868 return last_error;
1869 BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1870}
1871
1872#ifdef DEBUG
1873static void
1874xfs_check_delalloc(
1875 struct xfs_inode *ip,
1876 int whichfork)
1877{
1878 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, whichfork);
1879 struct xfs_bmbt_irec got;
1880 struct xfs_iext_cursor icur;
1881
1882 if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1883 return;
1884 do {
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]",
1888 ip->i_ino,
1889 whichfork == XFS_DATA_FORK ? "data" : "cow",
1890 got.br_startoff, got.br_blockcount);
1891 }
1892 } while (xfs_iext_next_extent(ifp, &icur, &got));
1893}
1894#else
1895#define xfs_check_delalloc(ip, whichfork) do { } while (0)
1896#endif
1897
1898/* Schedule the inode for reclaim. */
1899static void
1900xfs_inodegc_set_reclaimable(
1901 struct xfs_inode *ip)
1902{
1903 struct xfs_mount *mp = ip->i_mount;
1904 struct xfs_perag *pag;
1905
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);
1909 ASSERT(0);
1910 }
1911
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);
1915
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);
1921
1922 spin_unlock(&ip->i_flags_lock);
1923 spin_unlock(&pag->pag_ici_lock);
1924 xfs_perag_put(pag);
1925}
1926
1927/*
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.
1931 */
1932static int
1933xfs_inodegc_inactivate(
1934 struct xfs_inode *ip)
1935{
1936 int error;
1937
1938 trace_xfs_inode_inactivating(ip);
1939 error = xfs_inactive(ip);
1940 xfs_inodegc_set_reclaimable(ip);
1941 return error;
1942
1943}
1944
1945void
1946xfs_inodegc_worker(
1947 struct work_struct *work)
1948{
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;
1955
1956 /*
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.
1961 */
1962 cpumask_clear_cpu(gc->cpu, &mp->m_inodegc_cpumask);
1963 smp_mb__after_atomic();
1964
1965 WRITE_ONCE(gc->items, 0);
1966
1967 if (!node)
1968 return;
1969
1970 /*
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.
1974 */
1975 nofs_flag = memalloc_nofs_save();
1976
1977 ip = llist_entry(node, struct xfs_inode, i_gclist);
1978 trace_xfs_inodegc_worker(mp, READ_ONCE(gc->shrinker_hits));
1979
1980 WRITE_ONCE(gc->shrinker_hits, 0);
1981 llist_for_each_entry_safe(ip, n, node, i_gclist) {
1982 int error;
1983
1984 xfs_iflags_set(ip, XFS_INACTIVATING);
1985 error = xfs_inodegc_inactivate(ip);
1986 if (error && !gc->error)
1987 gc->error = error;
1988 }
1989
1990 memalloc_nofs_restore(nofs_flag);
1991}
1992
1993/*
1994 * Expedite all pending inodegc work to run immediately. This does not wait for
1995 * completion of the work.
1996 */
1997void
1998xfs_inodegc_push(
1999 struct xfs_mount *mp)
2000{
2001 if (!xfs_is_inodegc_enabled(mp))
2002 return;
2003 trace_xfs_inodegc_push(mp, __return_address);
2004 xfs_inodegc_queue_all(mp);
2005}
2006
2007/*
2008 * Force all currently queued inode inactivation work to run immediately and
2009 * wait for the work to finish.
2010 */
2011int
2012xfs_inodegc_flush(
2013 struct xfs_mount *mp)
2014{
2015 xfs_inodegc_push(mp);
2016 trace_xfs_inodegc_flush(mp, __return_address);
2017 return xfs_inodegc_wait_all(mp);
2018}
2019
2020/*
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.
2024 */
2025void
2026xfs_inodegc_stop(
2027 struct xfs_mount *mp)
2028{
2029 bool rerun;
2030
2031 if (!xfs_clear_inodegc_enabled(mp))
2032 return;
2033
2034 /*
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.
2043 */
2044 xfs_inodegc_queue_all(mp);
2045 do {
2046 flush_workqueue(mp->m_inodegc_wq);
2047 rerun = xfs_inodegc_queue_all(mp);
2048 } while (rerun);
2049
2050 trace_xfs_inodegc_stop(mp, __return_address);
2051}
2052
2053/*
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.
2057 */
2058void
2059xfs_inodegc_start(
2060 struct xfs_mount *mp)
2061{
2062 if (xfs_set_inodegc_enabled(mp))
2063 return;
2064
2065 trace_xfs_inodegc_start(mp, __return_address);
2066 xfs_inodegc_queue_all(mp);
2067}
2068
2069#ifdef CONFIG_XFS_RT
2070static inline bool
2071xfs_inodegc_want_queue_rt_file(
2072 struct xfs_inode *ip)
2073{
2074 struct xfs_mount *mp = ip->i_mount;
2075
2076 if (!XFS_IS_REALTIME_INODE(ip))
2077 return false;
2078
2079 if (__percpu_counter_compare(&mp->m_frextents,
2080 mp->m_low_rtexts[XFS_LOWSP_5_PCNT],
2081 XFS_FDBLOCKS_BATCH) < 0)
2082 return true;
2083
2084 return false;
2085}
2086#else
2087# define xfs_inodegc_want_queue_rt_file(ip) (false)
2088#endif /* CONFIG_XFS_RT */
2089
2090/*
2091 * Schedule the inactivation worker when:
2092 *
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.
2096 */
2097static inline bool
2098xfs_inodegc_want_queue_work(
2099 struct xfs_inode *ip,
2100 unsigned int items)
2101{
2102 struct xfs_mount *mp = ip->i_mount;
2103
2104 if (items > mp->m_ino_geo.inodes_per_cluster)
2105 return true;
2106
2107 if (__percpu_counter_compare(&mp->m_fdblocks,
2108 mp->m_low_space[XFS_LOWSP_5_PCNT],
2109 XFS_FDBLOCKS_BATCH) < 0)
2110 return true;
2111
2112 if (xfs_inodegc_want_queue_rt_file(ip))
2113 return true;
2114
2115 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
2116 return true;
2117
2118 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
2119 return true;
2120
2121 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
2122 return true;
2123
2124 return false;
2125}
2126
2127/*
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.
2130 */
2131#define XFS_INODEGC_MAX_BACKLOG (4 * XFS_INODES_PER_CHUNK)
2132
2133/*
2134 * Make the frontend wait for inactivations when:
2135 *
2136 * - Memory shrinkers queued the inactivation worker and it hasn't finished.
2137 * - The queue depth exceeds the maximum allowable percpu backlog.
2138 *
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.
2143 */
2144static inline bool
2145xfs_inodegc_want_flush_work(
2146 struct xfs_inode *ip,
2147 unsigned int items,
2148 unsigned int shrinker_hits)
2149{
2150 if (current->flags & PF_MEMALLOC_NOFS)
2151 return false;
2152
2153 if (shrinker_hits > 0)
2154 return true;
2155
2156 if (items > XFS_INODEGC_MAX_BACKLOG)
2157 return true;
2158
2159 return false;
2160}
2161
2162/*
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
2165 * workers.
2166 */
2167static void
2168xfs_inodegc_queue(
2169 struct xfs_inode *ip)
2170{
2171 struct xfs_mount *mp = ip->i_mount;
2172 struct xfs_inodegc *gc;
2173 int items;
2174 unsigned int shrinker_hits;
2175 unsigned int cpu_nr;
2176 unsigned long queue_delay = 1;
2177
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);
2182
2183 cpu_nr = get_cpu();
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);
2189
2190 /*
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.
2194 */
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);
2198
2199 /*
2200 * We queue the work while holding the current CPU so that the work
2201 * is scheduled to run on this CPU.
2202 */
2203 if (!xfs_is_inodegc_enabled(mp)) {
2204 put_cpu();
2205 return;
2206 }
2207
2208 if (xfs_inodegc_want_queue_work(ip, items))
2209 queue_delay = 0;
2210
2211 trace_xfs_inodegc_queue(mp, __return_address);
2212 mod_delayed_work_on(current_cpu(), mp->m_inodegc_wq, &gc->work,
2213 queue_delay);
2214 put_cpu();
2215
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);
2219 }
2220}
2221
2222/*
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.
2225 *
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
2230 * all inodes.
2231 */
2232void
2233xfs_inode_mark_reclaimable(
2234 struct xfs_inode *ip)
2235{
2236 struct xfs_mount *mp = ip->i_mount;
2237 bool need_inactive;
2238
2239 XFS_STATS_INC(mp, vn_reclaim);
2240
2241 /*
2242 * We should never get here with any of the reclaim flags already set.
2243 */
2244 ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2245
2246 need_inactive = xfs_inode_needs_inactive(ip);
2247 if (need_inactive) {
2248 xfs_inodegc_queue(ip);
2249 return;
2250 }
2251
2252 /* Going straight to reclaim, so drop the dquots. */
2253 xfs_qm_dqdetach(ip);
2254 xfs_inodegc_set_reclaimable(ip);
2255}
2256
2257/*
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.
2261 *
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.
2266 */
2267#define XFS_INODEGC_SHRINKER_COUNT (1UL << DEF_PRIORITY)
2268#define XFS_INODEGC_SHRINKER_BATCH ((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2269
2270static unsigned long
2271xfs_inodegc_shrinker_count(
2272 struct shrinker *shrink,
2273 struct shrink_control *sc)
2274{
2275 struct xfs_mount *mp = shrink->private_data;
2276 struct xfs_inodegc *gc;
2277 int cpu;
2278
2279 if (!xfs_is_inodegc_enabled(mp))
2280 return 0;
2281
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;
2286 }
2287
2288 return 0;
2289}
2290
2291static unsigned long
2292xfs_inodegc_shrinker_scan(
2293 struct shrinker *shrink,
2294 struct shrink_control *sc)
2295{
2296 struct xfs_mount *mp = shrink->private_data;
2297 struct xfs_inodegc *gc;
2298 int cpu;
2299 bool no_items = true;
2300
2301 if (!xfs_is_inodegc_enabled(mp))
2302 return SHRINK_STOP;
2303
2304 trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2305
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);
2310
2311 WRITE_ONCE(gc->shrinker_hits, h + 1);
2312 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
2313 no_items = false;
2314 }
2315 }
2316
2317 /*
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.
2320 */
2321 if (no_items)
2322 return LONG_MAX;
2323
2324 return SHRINK_STOP;
2325}
2326
2327/* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2328int
2329xfs_inodegc_register_shrinker(
2330 struct xfs_mount *mp)
2331{
2332 mp->m_inodegc_shrinker = shrinker_alloc(SHRINKER_NONSLAB,
2333 "xfs-inodegc:%s",
2334 mp->m_super->s_id);
2335 if (!mp->m_inodegc_shrinker)
2336 return -ENOMEM;
2337
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;
2343
2344 shrinker_register(mp->m_inodegc_shrinker);
2345
2346 return 0;
2347}