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
29#include <linux/iversion.h>
30
31/* Radix tree tags for incore inode tree. */
32
33/* inode is to be reclaimed */
34#define XFS_ICI_RECLAIM_TAG 0
35/* Inode has speculative preallocations (posteof or cow) to clean. */
36#define XFS_ICI_BLOCKGC_TAG 1
37
38/*
39 * The goal for walking incore inodes. These can correspond with incore inode
40 * radix tree tags when convenient. Avoid existing XFS_IWALK namespace.
41 */
42enum xfs_icwalk_goal {
43 /* Goals directly associated with tagged inodes. */
44 XFS_ICWALK_BLOCKGC = XFS_ICI_BLOCKGC_TAG,
45 XFS_ICWALK_RECLAIM = XFS_ICI_RECLAIM_TAG,
46};
47
48static int xfs_icwalk(struct xfs_mount *mp,
49 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
50static int xfs_icwalk_ag(struct xfs_perag *pag,
51 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
52
53/*
54 * Private inode cache walk flags for struct xfs_icwalk. Must not
55 * coincide with XFS_ICWALK_FLAGS_VALID.
56 */
57
58/* Stop scanning after icw_scan_limit inodes. */
59#define XFS_ICWALK_FLAG_SCAN_LIMIT (1U << 28)
60
61#define XFS_ICWALK_FLAG_RECLAIM_SICK (1U << 27)
62#define XFS_ICWALK_FLAG_UNION (1U << 26) /* union filter algorithm */
63
64#define XFS_ICWALK_PRIVATE_FLAGS (XFS_ICWALK_FLAG_SCAN_LIMIT | \
65 XFS_ICWALK_FLAG_RECLAIM_SICK | \
66 XFS_ICWALK_FLAG_UNION)
67
68/*
69 * Allocate and initialise an xfs_inode.
70 */
71struct xfs_inode *
72xfs_inode_alloc(
73 struct xfs_mount *mp,
74 xfs_ino_t ino)
75{
76 struct xfs_inode *ip;
77
78 /*
79 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL
80 * and return NULL here on ENOMEM.
81 */
82 ip = alloc_inode_sb(mp->m_super, xfs_inode_cache, GFP_KERNEL | __GFP_NOFAIL);
83
84 if (inode_init_always(mp->m_super, VFS_I(ip))) {
85 kmem_cache_free(xfs_inode_cache, ip);
86 return NULL;
87 }
88
89 /* VFS doesn't initialise i_mode or i_state! */
90 VFS_I(ip)->i_mode = 0;
91 VFS_I(ip)->i_state = 0;
92 mapping_set_large_folios(VFS_I(ip)->i_mapping);
93
94 XFS_STATS_INC(mp, vn_active);
95 ASSERT(atomic_read(&ip->i_pincount) == 0);
96 ASSERT(ip->i_ino == 0);
97
98 /* initialise the xfs inode */
99 ip->i_ino = ino;
100 ip->i_mount = mp;
101 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
102 ip->i_cowfp = NULL;
103 memset(&ip->i_af, 0, sizeof(ip->i_af));
104 ip->i_af.if_format = XFS_DINODE_FMT_EXTENTS;
105 memset(&ip->i_df, 0, sizeof(ip->i_df));
106 ip->i_flags = 0;
107 ip->i_delayed_blks = 0;
108 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
109 ip->i_nblocks = 0;
110 ip->i_forkoff = 0;
111 ip->i_sick = 0;
112 ip->i_checked = 0;
113 INIT_WORK(&ip->i_ioend_work, xfs_end_io);
114 INIT_LIST_HEAD(&ip->i_ioend_list);
115 spin_lock_init(&ip->i_ioend_lock);
116 ip->i_next_unlinked = NULLAGINO;
117 ip->i_prev_unlinked = 0;
118
119 return ip;
120}
121
122STATIC void
123xfs_inode_free_callback(
124 struct rcu_head *head)
125{
126 struct inode *inode = container_of(head, struct inode, i_rcu);
127 struct xfs_inode *ip = XFS_I(inode);
128
129 switch (VFS_I(ip)->i_mode & S_IFMT) {
130 case S_IFREG:
131 case S_IFDIR:
132 case S_IFLNK:
133 xfs_idestroy_fork(&ip->i_df);
134 break;
135 }
136
137 xfs_ifork_zap_attr(ip);
138
139 if (ip->i_cowfp) {
140 xfs_idestroy_fork(ip->i_cowfp);
141 kmem_cache_free(xfs_ifork_cache, ip->i_cowfp);
142 }
143 if (ip->i_itemp) {
144 ASSERT(!test_bit(XFS_LI_IN_AIL,
145 &ip->i_itemp->ili_item.li_flags));
146 xfs_inode_item_destroy(ip);
147 ip->i_itemp = NULL;
148 }
149
150 kmem_cache_free(xfs_inode_cache, ip);
151}
152
153static void
154__xfs_inode_free(
155 struct xfs_inode *ip)
156{
157 /* asserts to verify all state is correct here */
158 ASSERT(atomic_read(&ip->i_pincount) == 0);
159 ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
160 XFS_STATS_DEC(ip->i_mount, vn_active);
161
162 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
163}
164
165void
166xfs_inode_free(
167 struct xfs_inode *ip)
168{
169 ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
170
171 /*
172 * Because we use RCU freeing we need to ensure the inode always
173 * appears to be reclaimed with an invalid inode number when in the
174 * free state. The ip->i_flags_lock provides the barrier against lookup
175 * races.
176 */
177 spin_lock(&ip->i_flags_lock);
178 ip->i_flags = XFS_IRECLAIM;
179 ip->i_ino = 0;
180 spin_unlock(&ip->i_flags_lock);
181
182 __xfs_inode_free(ip);
183}
184
185/*
186 * Queue background inode reclaim work if there are reclaimable inodes and there
187 * isn't reclaim work already scheduled or in progress.
188 */
189static void
190xfs_reclaim_work_queue(
191 struct xfs_mount *mp)
192{
193
194 rcu_read_lock();
195 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
196 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
197 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
198 }
199 rcu_read_unlock();
200}
201
202/*
203 * Background scanning to trim preallocated space. This is queued based on the
204 * 'speculative_prealloc_lifetime' tunable (5m by default).
205 */
206static inline void
207xfs_blockgc_queue(
208 struct xfs_perag *pag)
209{
210 struct xfs_mount *mp = pag->pag_mount;
211
212 if (!xfs_is_blockgc_enabled(mp))
213 return;
214
215 rcu_read_lock();
216 if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
217 queue_delayed_work(pag->pag_mount->m_blockgc_wq,
218 &pag->pag_blockgc_work,
219 msecs_to_jiffies(xfs_blockgc_secs * 1000));
220 rcu_read_unlock();
221}
222
223/* Set a tag on both the AG incore inode tree and the AG radix tree. */
224static void
225xfs_perag_set_inode_tag(
226 struct xfs_perag *pag,
227 xfs_agino_t agino,
228 unsigned int tag)
229{
230 struct xfs_mount *mp = pag->pag_mount;
231 bool was_tagged;
232
233 lockdep_assert_held(&pag->pag_ici_lock);
234
235 was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
236 radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
237
238 if (tag == XFS_ICI_RECLAIM_TAG)
239 pag->pag_ici_reclaimable++;
240
241 if (was_tagged)
242 return;
243
244 /* propagate the tag up into the perag radix tree */
245 spin_lock(&mp->m_perag_lock);
246 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno, tag);
247 spin_unlock(&mp->m_perag_lock);
248
249 /* start background work */
250 switch (tag) {
251 case XFS_ICI_RECLAIM_TAG:
252 xfs_reclaim_work_queue(mp);
253 break;
254 case XFS_ICI_BLOCKGC_TAG:
255 xfs_blockgc_queue(pag);
256 break;
257 }
258
259 trace_xfs_perag_set_inode_tag(pag, _RET_IP_);
260}
261
262/* Clear a tag on both the AG incore inode tree and the AG radix tree. */
263static void
264xfs_perag_clear_inode_tag(
265 struct xfs_perag *pag,
266 xfs_agino_t agino,
267 unsigned int tag)
268{
269 struct xfs_mount *mp = pag->pag_mount;
270
271 lockdep_assert_held(&pag->pag_ici_lock);
272
273 /*
274 * Reclaim can signal (with a null agino) that it cleared its own tag
275 * by removing the inode from the radix tree.
276 */
277 if (agino != NULLAGINO)
278 radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
279 else
280 ASSERT(tag == XFS_ICI_RECLAIM_TAG);
281
282 if (tag == XFS_ICI_RECLAIM_TAG)
283 pag->pag_ici_reclaimable--;
284
285 if (radix_tree_tagged(&pag->pag_ici_root, tag))
286 return;
287
288 /* clear the tag from the perag radix tree */
289 spin_lock(&mp->m_perag_lock);
290 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno, tag);
291 spin_unlock(&mp->m_perag_lock);
292
293 trace_xfs_perag_clear_inode_tag(pag, _RET_IP_);
294}
295
296/*
297 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
298 * part of the structure. This is made more complex by the fact we store
299 * information about the on-disk values in the VFS inode and so we can't just
300 * overwrite the values unconditionally. Hence we save the parameters we
301 * need to retain across reinitialisation, and rewrite them into the VFS inode
302 * after reinitialisation even if it fails.
303 */
304static int
305xfs_reinit_inode(
306 struct xfs_mount *mp,
307 struct inode *inode)
308{
309 int error;
310 uint32_t nlink = inode->i_nlink;
311 uint32_t generation = inode->i_generation;
312 uint64_t version = inode_peek_iversion(inode);
313 umode_t mode = inode->i_mode;
314 dev_t dev = inode->i_rdev;
315 kuid_t uid = inode->i_uid;
316 kgid_t gid = inode->i_gid;
317
318 error = inode_init_always(mp->m_super, inode);
319
320 set_nlink(inode, nlink);
321 inode->i_generation = generation;
322 inode_set_iversion_queried(inode, version);
323 inode->i_mode = mode;
324 inode->i_rdev = dev;
325 inode->i_uid = uid;
326 inode->i_gid = gid;
327 mapping_set_large_folios(inode->i_mapping);
328 return error;
329}
330
331/*
332 * Carefully nudge an inode whose VFS state has been torn down back into a
333 * usable state. Drops the i_flags_lock and the rcu read lock.
334 */
335static int
336xfs_iget_recycle(
337 struct xfs_perag *pag,
338 struct xfs_inode *ip) __releases(&ip->i_flags_lock)
339{
340 struct xfs_mount *mp = ip->i_mount;
341 struct inode *inode = VFS_I(ip);
342 int error;
343
344 trace_xfs_iget_recycle(ip);
345
346 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
347 return -EAGAIN;
348
349 /*
350 * We need to make it look like the inode is being reclaimed to prevent
351 * the actual reclaim workers from stomping over us while we recycle
352 * the inode. We can't clear the radix tree tag yet as it requires
353 * pag_ici_lock to be held exclusive.
354 */
355 ip->i_flags |= XFS_IRECLAIM;
356
357 spin_unlock(&ip->i_flags_lock);
358 rcu_read_unlock();
359
360 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
361 error = xfs_reinit_inode(mp, inode);
362 xfs_iunlock(ip, XFS_ILOCK_EXCL);
363 if (error) {
364 /*
365 * Re-initializing the inode failed, and we are in deep
366 * trouble. Try to re-add it to the reclaim list.
367 */
368 rcu_read_lock();
369 spin_lock(&ip->i_flags_lock);
370 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
371 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
372 spin_unlock(&ip->i_flags_lock);
373 rcu_read_unlock();
374
375 trace_xfs_iget_recycle_fail(ip);
376 return error;
377 }
378
379 spin_lock(&pag->pag_ici_lock);
380 spin_lock(&ip->i_flags_lock);
381
382 /*
383 * Clear the per-lifetime state in the inode as we are now effectively
384 * a new inode and need to return to the initial state before reuse
385 * occurs.
386 */
387 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
388 ip->i_flags |= XFS_INEW;
389 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
390 XFS_ICI_RECLAIM_TAG);
391 inode->i_state = I_NEW;
392 spin_unlock(&ip->i_flags_lock);
393 spin_unlock(&pag->pag_ici_lock);
394
395 return 0;
396}
397
398/*
399 * If we are allocating a new inode, then check what was returned is
400 * actually a free, empty inode. If we are not allocating an inode,
401 * then check we didn't find a free inode.
402 *
403 * Returns:
404 * 0 if the inode free state matches the lookup context
405 * -ENOENT if the inode is free and we are not allocating
406 * -EFSCORRUPTED if there is any state mismatch at all
407 */
408static int
409xfs_iget_check_free_state(
410 struct xfs_inode *ip,
411 int flags)
412{
413 if (flags & XFS_IGET_CREATE) {
414 /* should be a free inode */
415 if (VFS_I(ip)->i_mode != 0) {
416 xfs_warn(ip->i_mount,
417"Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
418 ip->i_ino, VFS_I(ip)->i_mode);
419 xfs_agno_mark_sick(ip->i_mount,
420 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
421 XFS_SICK_AG_INOBT);
422 return -EFSCORRUPTED;
423 }
424
425 if (ip->i_nblocks != 0) {
426 xfs_warn(ip->i_mount,
427"Corruption detected! Free inode 0x%llx has blocks allocated!",
428 ip->i_ino);
429 xfs_agno_mark_sick(ip->i_mount,
430 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
431 XFS_SICK_AG_INOBT);
432 return -EFSCORRUPTED;
433 }
434 return 0;
435 }
436
437 /* should be an allocated inode */
438 if (VFS_I(ip)->i_mode == 0)
439 return -ENOENT;
440
441 return 0;
442}
443
444/* Make all pending inactivation work start immediately. */
445static bool
446xfs_inodegc_queue_all(
447 struct xfs_mount *mp)
448{
449 struct xfs_inodegc *gc;
450 int cpu;
451 bool ret = false;
452
453 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
454 gc = per_cpu_ptr(mp->m_inodegc, cpu);
455 if (!llist_empty(&gc->list)) {
456 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
457 ret = true;
458 }
459 }
460
461 return ret;
462}
463
464/* Wait for all queued work and collect errors */
465static int
466xfs_inodegc_wait_all(
467 struct xfs_mount *mp)
468{
469 int cpu;
470 int error = 0;
471
472 flush_workqueue(mp->m_inodegc_wq);
473 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
474 struct xfs_inodegc *gc;
475
476 gc = per_cpu_ptr(mp->m_inodegc, cpu);
477 if (gc->error && !error)
478 error = gc->error;
479 gc->error = 0;
480 }
481
482 return error;
483}
484
485/*
486 * Check the validity of the inode we just found it the cache
487 */
488static int
489xfs_iget_cache_hit(
490 struct xfs_perag *pag,
491 struct xfs_inode *ip,
492 xfs_ino_t ino,
493 int flags,
494 int lock_flags) __releases(RCU)
495{
496 struct inode *inode = VFS_I(ip);
497 struct xfs_mount *mp = ip->i_mount;
498 int error;
499
500 /*
501 * check for re-use of an inode within an RCU grace period due to the
502 * radix tree nodes not being updated yet. We monitor for this by
503 * setting the inode number to zero before freeing the inode structure.
504 * If the inode has been reallocated and set up, then the inode number
505 * will not match, so check for that, too.
506 */
507 spin_lock(&ip->i_flags_lock);
508 if (ip->i_ino != ino)
509 goto out_skip;
510
511 /*
512 * If we are racing with another cache hit that is currently
513 * instantiating this inode or currently recycling it out of
514 * reclaimable state, wait for the initialisation to complete
515 * before continuing.
516 *
517 * If we're racing with the inactivation worker we also want to wait.
518 * If we're creating a new file, it's possible that the worker
519 * previously marked the inode as free on disk but hasn't finished
520 * updating the incore state yet. The AGI buffer will be dirty and
521 * locked to the icreate transaction, so a synchronous push of the
522 * inodegc workers would result in deadlock. For a regular iget, the
523 * worker is running already, so we might as well wait.
524 *
525 * XXX(hch): eventually we should do something equivalent to
526 * wait_on_inode to wait for these flags to be cleared
527 * instead of polling for it.
528 */
529 if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM | XFS_INACTIVATING))
530 goto out_skip;
531
532 if (ip->i_flags & XFS_NEED_INACTIVE) {
533 /* Unlinked inodes cannot be re-grabbed. */
534 if (VFS_I(ip)->i_nlink == 0) {
535 error = -ENOENT;
536 goto out_error;
537 }
538 goto out_inodegc_flush;
539 }
540
541 /*
542 * Check the inode free state is valid. This also detects lookup
543 * racing with unlinks.
544 */
545 error = xfs_iget_check_free_state(ip, flags);
546 if (error)
547 goto out_error;
548
549 /* Skip inodes that have no vfs state. */
550 if ((flags & XFS_IGET_INCORE) &&
551 (ip->i_flags & XFS_IRECLAIMABLE))
552 goto out_skip;
553
554 /* The inode fits the selection criteria; process it. */
555 if (ip->i_flags & XFS_IRECLAIMABLE) {
556 /* Drops i_flags_lock and RCU read lock. */
557 error = xfs_iget_recycle(pag, ip);
558 if (error == -EAGAIN)
559 goto out_skip;
560 if (error)
561 return error;
562 } else {
563 /* If the VFS inode is being torn down, pause and try again. */
564 if (!igrab(inode))
565 goto out_skip;
566
567 /* We've got a live one. */
568 spin_unlock(&ip->i_flags_lock);
569 rcu_read_unlock();
570 trace_xfs_iget_hit(ip);
571 }
572
573 if (lock_flags != 0)
574 xfs_ilock(ip, lock_flags);
575
576 if (!(flags & XFS_IGET_INCORE))
577 xfs_iflags_clear(ip, XFS_ISTALE);
578 XFS_STATS_INC(mp, xs_ig_found);
579
580 return 0;
581
582out_skip:
583 trace_xfs_iget_skip(ip);
584 XFS_STATS_INC(mp, xs_ig_frecycle);
585 error = -EAGAIN;
586out_error:
587 spin_unlock(&ip->i_flags_lock);
588 rcu_read_unlock();
589 return error;
590
591out_inodegc_flush:
592 spin_unlock(&ip->i_flags_lock);
593 rcu_read_unlock();
594 /*
595 * Do not wait for the workers, because the caller could hold an AGI
596 * buffer lock. We're just going to sleep in a loop anyway.
597 */
598 if (xfs_is_inodegc_enabled(mp))
599 xfs_inodegc_queue_all(mp);
600 return -EAGAIN;
601}
602
603static int
604xfs_iget_cache_miss(
605 struct xfs_mount *mp,
606 struct xfs_perag *pag,
607 xfs_trans_t *tp,
608 xfs_ino_t ino,
609 struct xfs_inode **ipp,
610 int flags,
611 int lock_flags)
612{
613 struct xfs_inode *ip;
614 int error;
615 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
616 int iflags;
617
618 ip = xfs_inode_alloc(mp, ino);
619 if (!ip)
620 return -ENOMEM;
621
622 error = xfs_imap(pag, tp, ip->i_ino, &ip->i_imap, flags);
623 if (error)
624 goto out_destroy;
625
626 /*
627 * For version 5 superblocks, if we are initialising a new inode and we
628 * are not utilising the XFS_FEAT_IKEEP inode cluster mode, we can
629 * simply build the new inode core with a random generation number.
630 *
631 * For version 4 (and older) superblocks, log recovery is dependent on
632 * the i_flushiter field being initialised from the current on-disk
633 * value and hence we must also read the inode off disk even when
634 * initializing new inodes.
635 */
636 if (xfs_has_v3inodes(mp) &&
637 (flags & XFS_IGET_CREATE) && !xfs_has_ikeep(mp)) {
638 VFS_I(ip)->i_generation = get_random_u32();
639 } else {
640 struct xfs_buf *bp;
641
642 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
643 if (error)
644 goto out_destroy;
645
646 error = xfs_inode_from_disk(ip,
647 xfs_buf_offset(bp, ip->i_imap.im_boffset));
648 if (!error)
649 xfs_buf_set_ref(bp, XFS_INO_REF);
650 else
651 xfs_inode_mark_sick(ip, XFS_SICK_INO_CORE);
652 xfs_trans_brelse(tp, bp);
653
654 if (error)
655 goto out_destroy;
656 }
657
658 trace_xfs_iget_miss(ip);
659
660 /*
661 * Check the inode free state is valid. This also detects lookup
662 * racing with unlinks.
663 */
664 error = xfs_iget_check_free_state(ip, flags);
665 if (error)
666 goto out_destroy;
667
668 /*
669 * Preload the radix tree so we can insert safely under the
670 * write spinlock. Note that we cannot sleep inside the preload
671 * region.
672 */
673 if (radix_tree_preload(GFP_KERNEL | __GFP_NOLOCKDEP)) {
674 error = -EAGAIN;
675 goto out_destroy;
676 }
677
678 /*
679 * Because the inode hasn't been added to the radix-tree yet it can't
680 * be found by another thread, so we can do the non-sleeping lock here.
681 */
682 if (lock_flags) {
683 if (!xfs_ilock_nowait(ip, lock_flags))
684 BUG();
685 }
686
687 /*
688 * These values must be set before inserting the inode into the radix
689 * tree as the moment it is inserted a concurrent lookup (allowed by the
690 * RCU locking mechanism) can find it and that lookup must see that this
691 * is an inode currently under construction (i.e. that XFS_INEW is set).
692 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
693 * memory barrier that ensures this detection works correctly at lookup
694 * time.
695 */
696 iflags = XFS_INEW;
697 if (flags & XFS_IGET_DONTCACHE)
698 d_mark_dontcache(VFS_I(ip));
699 ip->i_udquot = NULL;
700 ip->i_gdquot = NULL;
701 ip->i_pdquot = NULL;
702 xfs_iflags_set(ip, iflags);
703
704 /* insert the new inode */
705 spin_lock(&pag->pag_ici_lock);
706 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
707 if (unlikely(error)) {
708 WARN_ON(error != -EEXIST);
709 XFS_STATS_INC(mp, xs_ig_dup);
710 error = -EAGAIN;
711 goto out_preload_end;
712 }
713 spin_unlock(&pag->pag_ici_lock);
714 radix_tree_preload_end();
715
716 *ipp = ip;
717 return 0;
718
719out_preload_end:
720 spin_unlock(&pag->pag_ici_lock);
721 radix_tree_preload_end();
722 if (lock_flags)
723 xfs_iunlock(ip, lock_flags);
724out_destroy:
725 __destroy_inode(VFS_I(ip));
726 xfs_inode_free(ip);
727 return error;
728}
729
730/*
731 * Look up an inode by number in the given file system. The inode is looked up
732 * in the cache held in each AG. If the inode is found in the cache, initialise
733 * the vfs inode if necessary.
734 *
735 * If it is not in core, read it in from the file system's device, add it to the
736 * cache and initialise the vfs inode.
737 *
738 * The inode is locked according to the value of the lock_flags parameter.
739 * Inode lookup is only done during metadata operations and not as part of the
740 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
741 */
742int
743xfs_iget(
744 struct xfs_mount *mp,
745 struct xfs_trans *tp,
746 xfs_ino_t ino,
747 uint flags,
748 uint lock_flags,
749 struct xfs_inode **ipp)
750{
751 struct xfs_inode *ip;
752 struct xfs_perag *pag;
753 xfs_agino_t agino;
754 int error;
755
756 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
757
758 /* reject inode numbers outside existing AGs */
759 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
760 return -EINVAL;
761
762 XFS_STATS_INC(mp, xs_ig_attempts);
763
764 /* get the perag structure and ensure that it's inode capable */
765 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
766 agino = XFS_INO_TO_AGINO(mp, ino);
767
768again:
769 error = 0;
770 rcu_read_lock();
771 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
772
773 if (ip) {
774 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
775 if (error)
776 goto out_error_or_again;
777 } else {
778 rcu_read_unlock();
779 if (flags & XFS_IGET_INCORE) {
780 error = -ENODATA;
781 goto out_error_or_again;
782 }
783 XFS_STATS_INC(mp, xs_ig_missed);
784
785 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
786 flags, lock_flags);
787 if (error)
788 goto out_error_or_again;
789 }
790 xfs_perag_put(pag);
791
792 *ipp = ip;
793
794 /*
795 * If we have a real type for an on-disk inode, we can setup the inode
796 * now. If it's a new inode being created, xfs_init_new_inode will
797 * handle it.
798 */
799 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
800 xfs_setup_existing_inode(ip);
801 return 0;
802
803out_error_or_again:
804 if (!(flags & (XFS_IGET_INCORE | XFS_IGET_NORETRY)) &&
805 error == -EAGAIN) {
806 delay(1);
807 goto again;
808 }
809 xfs_perag_put(pag);
810 return error;
811}
812
813/*
814 * Grab the inode for reclaim exclusively.
815 *
816 * We have found this inode via a lookup under RCU, so the inode may have
817 * already been freed, or it may be in the process of being recycled by
818 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
819 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
820 * will not be set. Hence we need to check for both these flag conditions to
821 * avoid inodes that are no longer reclaim candidates.
822 *
823 * Note: checking for other state flags here, under the i_flags_lock or not, is
824 * racy and should be avoided. Those races should be resolved only after we have
825 * ensured that we are able to reclaim this inode and the world can see that we
826 * are going to reclaim it.
827 *
828 * Return true if we grabbed it, false otherwise.
829 */
830static bool
831xfs_reclaim_igrab(
832 struct xfs_inode *ip,
833 struct xfs_icwalk *icw)
834{
835 ASSERT(rcu_read_lock_held());
836
837 spin_lock(&ip->i_flags_lock);
838 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
839 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
840 /* not a reclaim candidate. */
841 spin_unlock(&ip->i_flags_lock);
842 return false;
843 }
844
845 /* Don't reclaim a sick inode unless the caller asked for it. */
846 if (ip->i_sick &&
847 (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
848 spin_unlock(&ip->i_flags_lock);
849 return false;
850 }
851
852 __xfs_iflags_set(ip, XFS_IRECLAIM);
853 spin_unlock(&ip->i_flags_lock);
854 return true;
855}
856
857/*
858 * Inode reclaim is non-blocking, so the default action if progress cannot be
859 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
860 * XFS_IRECLAIM flag. If we are in a shutdown state, we don't care about
861 * blocking anymore and hence we can wait for the inode to be able to reclaim
862 * it.
863 *
864 * We do no IO here - if callers require inodes to be cleaned they must push the
865 * AIL first to trigger writeback of dirty inodes. This enables writeback to be
866 * done in the background in a non-blocking manner, and enables memory reclaim
867 * to make progress without blocking.
868 */
869static void
870xfs_reclaim_inode(
871 struct xfs_inode *ip,
872 struct xfs_perag *pag)
873{
874 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
875
876 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
877 goto out;
878 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
879 goto out_iunlock;
880
881 /*
882 * Check for log shutdown because aborting the inode can move the log
883 * tail and corrupt in memory state. This is fine if the log is shut
884 * down, but if the log is still active and only the mount is shut down
885 * then the in-memory log tail movement caused by the abort can be
886 * incorrectly propagated to disk.
887 */
888 if (xlog_is_shutdown(ip->i_mount->m_log)) {
889 xfs_iunpin_wait(ip);
890 xfs_iflush_shutdown_abort(ip);
891 goto reclaim;
892 }
893 if (xfs_ipincount(ip))
894 goto out_clear_flush;
895 if (!xfs_inode_clean(ip))
896 goto out_clear_flush;
897
898 xfs_iflags_clear(ip, XFS_IFLUSHING);
899reclaim:
900 trace_xfs_inode_reclaiming(ip);
901
902 /*
903 * Because we use RCU freeing we need to ensure the inode always appears
904 * to be reclaimed with an invalid inode number when in the free state.
905 * We do this as early as possible under the ILOCK so that
906 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
907 * detect races with us here. By doing this, we guarantee that once
908 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
909 * it will see either a valid inode that will serialise correctly, or it
910 * will see an invalid inode that it can skip.
911 */
912 spin_lock(&ip->i_flags_lock);
913 ip->i_flags = XFS_IRECLAIM;
914 ip->i_ino = 0;
915 ip->i_sick = 0;
916 ip->i_checked = 0;
917 spin_unlock(&ip->i_flags_lock);
918
919 ASSERT(!ip->i_itemp || ip->i_itemp->ili_item.li_buf == NULL);
920 xfs_iunlock(ip, XFS_ILOCK_EXCL);
921
922 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
923 /*
924 * Remove the inode from the per-AG radix tree.
925 *
926 * Because radix_tree_delete won't complain even if the item was never
927 * added to the tree assert that it's been there before to catch
928 * problems with the inode life time early on.
929 */
930 spin_lock(&pag->pag_ici_lock);
931 if (!radix_tree_delete(&pag->pag_ici_root,
932 XFS_INO_TO_AGINO(ip->i_mount, ino)))
933 ASSERT(0);
934 xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
935 spin_unlock(&pag->pag_ici_lock);
936
937 /*
938 * Here we do an (almost) spurious inode lock in order to coordinate
939 * with inode cache radix tree lookups. This is because the lookup
940 * can reference the inodes in the cache without taking references.
941 *
942 * We make that OK here by ensuring that we wait until the inode is
943 * unlocked after the lookup before we go ahead and free it.
944 */
945 xfs_ilock(ip, XFS_ILOCK_EXCL);
946 ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
947 xfs_iunlock(ip, XFS_ILOCK_EXCL);
948 ASSERT(xfs_inode_clean(ip));
949
950 __xfs_inode_free(ip);
951 return;
952
953out_clear_flush:
954 xfs_iflags_clear(ip, XFS_IFLUSHING);
955out_iunlock:
956 xfs_iunlock(ip, XFS_ILOCK_EXCL);
957out:
958 xfs_iflags_clear(ip, XFS_IRECLAIM);
959}
960
961/* Reclaim sick inodes if we're unmounting or the fs went down. */
962static inline bool
963xfs_want_reclaim_sick(
964 struct xfs_mount *mp)
965{
966 return xfs_is_unmounting(mp) || xfs_has_norecovery(mp) ||
967 xfs_is_shutdown(mp);
968}
969
970void
971xfs_reclaim_inodes(
972 struct xfs_mount *mp)
973{
974 struct xfs_icwalk icw = {
975 .icw_flags = 0,
976 };
977
978 if (xfs_want_reclaim_sick(mp))
979 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
980
981 while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
982 xfs_ail_push_all_sync(mp->m_ail);
983 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
984 }
985}
986
987/*
988 * The shrinker infrastructure determines how many inodes we should scan for
989 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
990 * push the AIL here. We also want to proactively free up memory if we can to
991 * minimise the amount of work memory reclaim has to do so we kick the
992 * background reclaim if it isn't already scheduled.
993 */
994long
995xfs_reclaim_inodes_nr(
996 struct xfs_mount *mp,
997 unsigned long nr_to_scan)
998{
999 struct xfs_icwalk icw = {
1000 .icw_flags = XFS_ICWALK_FLAG_SCAN_LIMIT,
1001 .icw_scan_limit = min_t(unsigned long, LONG_MAX, nr_to_scan),
1002 };
1003
1004 if (xfs_want_reclaim_sick(mp))
1005 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1006
1007 /* kick background reclaimer and push the AIL */
1008 xfs_reclaim_work_queue(mp);
1009 xfs_ail_push_all(mp->m_ail);
1010
1011 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1012 return 0;
1013}
1014
1015/*
1016 * Return the number of reclaimable inodes in the filesystem for
1017 * the shrinker to determine how much to reclaim.
1018 */
1019long
1020xfs_reclaim_inodes_count(
1021 struct xfs_mount *mp)
1022{
1023 struct xfs_perag *pag;
1024 xfs_agnumber_t ag = 0;
1025 long reclaimable = 0;
1026
1027 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1028 ag = pag->pag_agno + 1;
1029 reclaimable += pag->pag_ici_reclaimable;
1030 xfs_perag_put(pag);
1031 }
1032 return reclaimable;
1033}
1034
1035STATIC bool
1036xfs_icwalk_match_id(
1037 struct xfs_inode *ip,
1038 struct xfs_icwalk *icw)
1039{
1040 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1041 !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1042 return false;
1043
1044 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1045 !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1046 return false;
1047
1048 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1049 ip->i_projid != icw->icw_prid)
1050 return false;
1051
1052 return true;
1053}
1054
1055/*
1056 * A union-based inode filtering algorithm. Process the inode if any of the
1057 * criteria match. This is for global/internal scans only.
1058 */
1059STATIC bool
1060xfs_icwalk_match_id_union(
1061 struct xfs_inode *ip,
1062 struct xfs_icwalk *icw)
1063{
1064 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1065 uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1066 return true;
1067
1068 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1069 gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1070 return true;
1071
1072 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1073 ip->i_projid == icw->icw_prid)
1074 return true;
1075
1076 return false;
1077}
1078
1079/*
1080 * Is this inode @ip eligible for eof/cow block reclamation, given some
1081 * filtering parameters @icw? The inode is eligible if @icw is null or
1082 * if the predicate functions match.
1083 */
1084static bool
1085xfs_icwalk_match(
1086 struct xfs_inode *ip,
1087 struct xfs_icwalk *icw)
1088{
1089 bool match;
1090
1091 if (!icw)
1092 return true;
1093
1094 if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1095 match = xfs_icwalk_match_id_union(ip, icw);
1096 else
1097 match = xfs_icwalk_match_id(ip, icw);
1098 if (!match)
1099 return false;
1100
1101 /* skip the inode if the file size is too small */
1102 if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1103 XFS_ISIZE(ip) < icw->icw_min_file_size)
1104 return false;
1105
1106 return true;
1107}
1108
1109/*
1110 * This is a fast pass over the inode cache to try to get reclaim moving on as
1111 * many inodes as possible in a short period of time. It kicks itself every few
1112 * seconds, as well as being kicked by the inode cache shrinker when memory
1113 * goes low.
1114 */
1115void
1116xfs_reclaim_worker(
1117 struct work_struct *work)
1118{
1119 struct xfs_mount *mp = container_of(to_delayed_work(work),
1120 struct xfs_mount, m_reclaim_work);
1121
1122 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1123 xfs_reclaim_work_queue(mp);
1124}
1125
1126STATIC int
1127xfs_inode_free_eofblocks(
1128 struct xfs_inode *ip,
1129 struct xfs_icwalk *icw,
1130 unsigned int *lockflags)
1131{
1132 bool wait;
1133
1134 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1135
1136 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1137 return 0;
1138
1139 /*
1140 * If the mapping is dirty the operation can block and wait for some
1141 * time. Unless we are waiting, skip it.
1142 */
1143 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1144 return 0;
1145
1146 if (!xfs_icwalk_match(ip, icw))
1147 return 0;
1148
1149 /*
1150 * If the caller is waiting, return -EAGAIN to keep the background
1151 * scanner moving and revisit the inode in a subsequent pass.
1152 */
1153 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1154 if (wait)
1155 return -EAGAIN;
1156 return 0;
1157 }
1158 *lockflags |= XFS_IOLOCK_EXCL;
1159
1160 if (xfs_can_free_eofblocks(ip, false))
1161 return xfs_free_eofblocks(ip);
1162
1163 /* inode could be preallocated or append-only */
1164 trace_xfs_inode_free_eofblocks_invalid(ip);
1165 xfs_inode_clear_eofblocks_tag(ip);
1166 return 0;
1167}
1168
1169static void
1170xfs_blockgc_set_iflag(
1171 struct xfs_inode *ip,
1172 unsigned long iflag)
1173{
1174 struct xfs_mount *mp = ip->i_mount;
1175 struct xfs_perag *pag;
1176
1177 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1178
1179 /*
1180 * Don't bother locking the AG and looking up in the radix trees
1181 * if we already know that we have the tag set.
1182 */
1183 if (ip->i_flags & iflag)
1184 return;
1185 spin_lock(&ip->i_flags_lock);
1186 ip->i_flags |= iflag;
1187 spin_unlock(&ip->i_flags_lock);
1188
1189 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1190 spin_lock(&pag->pag_ici_lock);
1191
1192 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1193 XFS_ICI_BLOCKGC_TAG);
1194
1195 spin_unlock(&pag->pag_ici_lock);
1196 xfs_perag_put(pag);
1197}
1198
1199void
1200xfs_inode_set_eofblocks_tag(
1201 xfs_inode_t *ip)
1202{
1203 trace_xfs_inode_set_eofblocks_tag(ip);
1204 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1205}
1206
1207static void
1208xfs_blockgc_clear_iflag(
1209 struct xfs_inode *ip,
1210 unsigned long iflag)
1211{
1212 struct xfs_mount *mp = ip->i_mount;
1213 struct xfs_perag *pag;
1214 bool clear_tag;
1215
1216 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1217
1218 spin_lock(&ip->i_flags_lock);
1219 ip->i_flags &= ~iflag;
1220 clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1221 spin_unlock(&ip->i_flags_lock);
1222
1223 if (!clear_tag)
1224 return;
1225
1226 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1227 spin_lock(&pag->pag_ici_lock);
1228
1229 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1230 XFS_ICI_BLOCKGC_TAG);
1231
1232 spin_unlock(&pag->pag_ici_lock);
1233 xfs_perag_put(pag);
1234}
1235
1236void
1237xfs_inode_clear_eofblocks_tag(
1238 xfs_inode_t *ip)
1239{
1240 trace_xfs_inode_clear_eofblocks_tag(ip);
1241 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1242}
1243
1244/*
1245 * Set ourselves up to free CoW blocks from this file. If it's already clean
1246 * then we can bail out quickly, but otherwise we must back off if the file
1247 * is undergoing some kind of write.
1248 */
1249static bool
1250xfs_prep_free_cowblocks(
1251 struct xfs_inode *ip)
1252{
1253 /*
1254 * Just clear the tag if we have an empty cow fork or none at all. It's
1255 * possible the inode was fully unshared since it was originally tagged.
1256 */
1257 if (!xfs_inode_has_cow_data(ip)) {
1258 trace_xfs_inode_free_cowblocks_invalid(ip);
1259 xfs_inode_clear_cowblocks_tag(ip);
1260 return false;
1261 }
1262
1263 /*
1264 * If the mapping is dirty or under writeback we cannot touch the
1265 * CoW fork. Leave it alone if we're in the midst of a directio.
1266 */
1267 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1268 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1269 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1270 atomic_read(&VFS_I(ip)->i_dio_count))
1271 return false;
1272
1273 return true;
1274}
1275
1276/*
1277 * Automatic CoW Reservation Freeing
1278 *
1279 * These functions automatically garbage collect leftover CoW reservations
1280 * that were made on behalf of a cowextsize hint when we start to run out
1281 * of quota or when the reservations sit around for too long. If the file
1282 * has dirty pages or is undergoing writeback, its CoW reservations will
1283 * be retained.
1284 *
1285 * The actual garbage collection piggybacks off the same code that runs
1286 * the speculative EOF preallocation garbage collector.
1287 */
1288STATIC int
1289xfs_inode_free_cowblocks(
1290 struct xfs_inode *ip,
1291 struct xfs_icwalk *icw,
1292 unsigned int *lockflags)
1293{
1294 bool wait;
1295 int ret = 0;
1296
1297 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1298
1299 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1300 return 0;
1301
1302 if (!xfs_prep_free_cowblocks(ip))
1303 return 0;
1304
1305 if (!xfs_icwalk_match(ip, icw))
1306 return 0;
1307
1308 /*
1309 * If the caller is waiting, return -EAGAIN to keep the background
1310 * scanner moving and revisit the inode in a subsequent pass.
1311 */
1312 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1313 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1314 if (wait)
1315 return -EAGAIN;
1316 return 0;
1317 }
1318 *lockflags |= XFS_IOLOCK_EXCL;
1319
1320 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1321 if (wait)
1322 return -EAGAIN;
1323 return 0;
1324 }
1325 *lockflags |= XFS_MMAPLOCK_EXCL;
1326
1327 /*
1328 * Check again, nobody else should be able to dirty blocks or change
1329 * the reflink iflag now that we have the first two locks held.
1330 */
1331 if (xfs_prep_free_cowblocks(ip))
1332 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1333 return ret;
1334}
1335
1336void
1337xfs_inode_set_cowblocks_tag(
1338 xfs_inode_t *ip)
1339{
1340 trace_xfs_inode_set_cowblocks_tag(ip);
1341 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1342}
1343
1344void
1345xfs_inode_clear_cowblocks_tag(
1346 xfs_inode_t *ip)
1347{
1348 trace_xfs_inode_clear_cowblocks_tag(ip);
1349 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1350}
1351
1352/* Disable post-EOF and CoW block auto-reclamation. */
1353void
1354xfs_blockgc_stop(
1355 struct xfs_mount *mp)
1356{
1357 struct xfs_perag *pag;
1358 xfs_agnumber_t agno;
1359
1360 if (!xfs_clear_blockgc_enabled(mp))
1361 return;
1362
1363 for_each_perag(mp, agno, pag)
1364 cancel_delayed_work_sync(&pag->pag_blockgc_work);
1365 trace_xfs_blockgc_stop(mp, __return_address);
1366}
1367
1368/* Enable post-EOF and CoW block auto-reclamation. */
1369void
1370xfs_blockgc_start(
1371 struct xfs_mount *mp)
1372{
1373 struct xfs_perag *pag;
1374 xfs_agnumber_t agno;
1375
1376 if (xfs_set_blockgc_enabled(mp))
1377 return;
1378
1379 trace_xfs_blockgc_start(mp, __return_address);
1380 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1381 xfs_blockgc_queue(pag);
1382}
1383
1384/* Don't try to run block gc on an inode that's in any of these states. */
1385#define XFS_BLOCKGC_NOGRAB_IFLAGS (XFS_INEW | \
1386 XFS_NEED_INACTIVE | \
1387 XFS_INACTIVATING | \
1388 XFS_IRECLAIMABLE | \
1389 XFS_IRECLAIM)
1390/*
1391 * Decide if the given @ip is eligible for garbage collection of speculative
1392 * preallocations, and grab it if so. Returns true if it's ready to go or
1393 * false if we should just ignore it.
1394 */
1395static bool
1396xfs_blockgc_igrab(
1397 struct xfs_inode *ip)
1398{
1399 struct inode *inode = VFS_I(ip);
1400
1401 ASSERT(rcu_read_lock_held());
1402
1403 /* Check for stale RCU freed inode */
1404 spin_lock(&ip->i_flags_lock);
1405 if (!ip->i_ino)
1406 goto out_unlock_noent;
1407
1408 if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1409 goto out_unlock_noent;
1410 spin_unlock(&ip->i_flags_lock);
1411
1412 /* nothing to sync during shutdown */
1413 if (xfs_is_shutdown(ip->i_mount))
1414 return false;
1415
1416 /* If we can't grab the inode, it must on it's way to reclaim. */
1417 if (!igrab(inode))
1418 return false;
1419
1420 /* inode is valid */
1421 return true;
1422
1423out_unlock_noent:
1424 spin_unlock(&ip->i_flags_lock);
1425 return false;
1426}
1427
1428/* Scan one incore inode for block preallocations that we can remove. */
1429static int
1430xfs_blockgc_scan_inode(
1431 struct xfs_inode *ip,
1432 struct xfs_icwalk *icw)
1433{
1434 unsigned int lockflags = 0;
1435 int error;
1436
1437 error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1438 if (error)
1439 goto unlock;
1440
1441 error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1442unlock:
1443 if (lockflags)
1444 xfs_iunlock(ip, lockflags);
1445 xfs_irele(ip);
1446 return error;
1447}
1448
1449/* Background worker that trims preallocated space. */
1450void
1451xfs_blockgc_worker(
1452 struct work_struct *work)
1453{
1454 struct xfs_perag *pag = container_of(to_delayed_work(work),
1455 struct xfs_perag, pag_blockgc_work);
1456 struct xfs_mount *mp = pag->pag_mount;
1457 int error;
1458
1459 trace_xfs_blockgc_worker(mp, __return_address);
1460
1461 error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1462 if (error)
1463 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1464 pag->pag_agno, error);
1465 xfs_blockgc_queue(pag);
1466}
1467
1468/*
1469 * Try to free space in the filesystem by purging inactive inodes, eofblocks
1470 * and cowblocks.
1471 */
1472int
1473xfs_blockgc_free_space(
1474 struct xfs_mount *mp,
1475 struct xfs_icwalk *icw)
1476{
1477 int error;
1478
1479 trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1480
1481 error = xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1482 if (error)
1483 return error;
1484
1485 return xfs_inodegc_flush(mp);
1486}
1487
1488/*
1489 * Reclaim all the free space that we can by scheduling the background blockgc
1490 * and inodegc workers immediately and waiting for them all to clear.
1491 */
1492int
1493xfs_blockgc_flush_all(
1494 struct xfs_mount *mp)
1495{
1496 struct xfs_perag *pag;
1497 xfs_agnumber_t agno;
1498
1499 trace_xfs_blockgc_flush_all(mp, __return_address);
1500
1501 /*
1502 * For each blockgc worker, move its queue time up to now. If it
1503 * wasn't queued, it will not be requeued. Then flush whatever's
1504 * left.
1505 */
1506 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1507 mod_delayed_work(pag->pag_mount->m_blockgc_wq,
1508 &pag->pag_blockgc_work, 0);
1509
1510 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1511 flush_delayed_work(&pag->pag_blockgc_work);
1512
1513 return xfs_inodegc_flush(mp);
1514}
1515
1516/*
1517 * Run cow/eofblocks scans on the supplied dquots. We don't know exactly which
1518 * quota caused an allocation failure, so we make a best effort by including
1519 * each quota under low free space conditions (less than 1% free space) in the
1520 * scan.
1521 *
1522 * Callers must not hold any inode's ILOCK. If requesting a synchronous scan
1523 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1524 * MMAPLOCK.
1525 */
1526int
1527xfs_blockgc_free_dquots(
1528 struct xfs_mount *mp,
1529 struct xfs_dquot *udqp,
1530 struct xfs_dquot *gdqp,
1531 struct xfs_dquot *pdqp,
1532 unsigned int iwalk_flags)
1533{
1534 struct xfs_icwalk icw = {0};
1535 bool do_work = false;
1536
1537 if (!udqp && !gdqp && !pdqp)
1538 return 0;
1539
1540 /*
1541 * Run a scan to free blocks using the union filter to cover all
1542 * applicable quotas in a single scan.
1543 */
1544 icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1545
1546 if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1547 icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1548 icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1549 do_work = true;
1550 }
1551
1552 if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1553 icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1554 icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1555 do_work = true;
1556 }
1557
1558 if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1559 icw.icw_prid = pdqp->q_id;
1560 icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1561 do_work = true;
1562 }
1563
1564 if (!do_work)
1565 return 0;
1566
1567 return xfs_blockgc_free_space(mp, &icw);
1568}
1569
1570/* Run cow/eofblocks scans on the quotas attached to the inode. */
1571int
1572xfs_blockgc_free_quota(
1573 struct xfs_inode *ip,
1574 unsigned int iwalk_flags)
1575{
1576 return xfs_blockgc_free_dquots(ip->i_mount,
1577 xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1578 xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1579 xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1580}
1581
1582/* XFS Inode Cache Walking Code */
1583
1584/*
1585 * The inode lookup is done in batches to keep the amount of lock traffic and
1586 * radix tree lookups to a minimum. The batch size is a trade off between
1587 * lookup reduction and stack usage. This is in the reclaim path, so we can't
1588 * be too greedy.
1589 */
1590#define XFS_LOOKUP_BATCH 32
1591
1592
1593/*
1594 * Decide if we want to grab this inode in anticipation of doing work towards
1595 * the goal.
1596 */
1597static inline bool
1598xfs_icwalk_igrab(
1599 enum xfs_icwalk_goal goal,
1600 struct xfs_inode *ip,
1601 struct xfs_icwalk *icw)
1602{
1603 switch (goal) {
1604 case XFS_ICWALK_BLOCKGC:
1605 return xfs_blockgc_igrab(ip);
1606 case XFS_ICWALK_RECLAIM:
1607 return xfs_reclaim_igrab(ip, icw);
1608 default:
1609 return false;
1610 }
1611}
1612
1613/*
1614 * Process an inode. Each processing function must handle any state changes
1615 * made by the icwalk igrab function. Return -EAGAIN to skip an inode.
1616 */
1617static inline int
1618xfs_icwalk_process_inode(
1619 enum xfs_icwalk_goal goal,
1620 struct xfs_inode *ip,
1621 struct xfs_perag *pag,
1622 struct xfs_icwalk *icw)
1623{
1624 int error = 0;
1625
1626 switch (goal) {
1627 case XFS_ICWALK_BLOCKGC:
1628 error = xfs_blockgc_scan_inode(ip, icw);
1629 break;
1630 case XFS_ICWALK_RECLAIM:
1631 xfs_reclaim_inode(ip, pag);
1632 break;
1633 }
1634 return error;
1635}
1636
1637/*
1638 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1639 * process them in some manner.
1640 */
1641static int
1642xfs_icwalk_ag(
1643 struct xfs_perag *pag,
1644 enum xfs_icwalk_goal goal,
1645 struct xfs_icwalk *icw)
1646{
1647 struct xfs_mount *mp = pag->pag_mount;
1648 uint32_t first_index;
1649 int last_error = 0;
1650 int skipped;
1651 bool done;
1652 int nr_found;
1653
1654restart:
1655 done = false;
1656 skipped = 0;
1657 if (goal == XFS_ICWALK_RECLAIM)
1658 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1659 else
1660 first_index = 0;
1661 nr_found = 0;
1662 do {
1663 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1664 int error = 0;
1665 int i;
1666
1667 rcu_read_lock();
1668
1669 nr_found = radix_tree_gang_lookup_tag(&pag->pag_ici_root,
1670 (void **) batch, first_index,
1671 XFS_LOOKUP_BATCH, goal);
1672 if (!nr_found) {
1673 done = true;
1674 rcu_read_unlock();
1675 break;
1676 }
1677
1678 /*
1679 * Grab the inodes before we drop the lock. if we found
1680 * nothing, nr == 0 and the loop will be skipped.
1681 */
1682 for (i = 0; i < nr_found; i++) {
1683 struct xfs_inode *ip = batch[i];
1684
1685 if (done || !xfs_icwalk_igrab(goal, ip, icw))
1686 batch[i] = NULL;
1687
1688 /*
1689 * Update the index for the next lookup. Catch
1690 * overflows into the next AG range which can occur if
1691 * we have inodes in the last block of the AG and we
1692 * are currently pointing to the last inode.
1693 *
1694 * Because we may see inodes that are from the wrong AG
1695 * due to RCU freeing and reallocation, only update the
1696 * index if it lies in this AG. It was a race that lead
1697 * us to see this inode, so another lookup from the
1698 * same index will not find it again.
1699 */
1700 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
1701 continue;
1702 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1703 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1704 done = true;
1705 }
1706
1707 /* unlock now we've grabbed the inodes. */
1708 rcu_read_unlock();
1709
1710 for (i = 0; i < nr_found; i++) {
1711 if (!batch[i])
1712 continue;
1713 error = xfs_icwalk_process_inode(goal, batch[i], pag,
1714 icw);
1715 if (error == -EAGAIN) {
1716 skipped++;
1717 continue;
1718 }
1719 if (error && last_error != -EFSCORRUPTED)
1720 last_error = error;
1721 }
1722
1723 /* bail out if the filesystem is corrupted. */
1724 if (error == -EFSCORRUPTED)
1725 break;
1726
1727 cond_resched();
1728
1729 if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1730 icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1731 if (icw->icw_scan_limit <= 0)
1732 break;
1733 }
1734 } while (nr_found && !done);
1735
1736 if (goal == XFS_ICWALK_RECLAIM) {
1737 if (done)
1738 first_index = 0;
1739 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1740 }
1741
1742 if (skipped) {
1743 delay(1);
1744 goto restart;
1745 }
1746 return last_error;
1747}
1748
1749/* Walk all incore inodes to achieve a given goal. */
1750static int
1751xfs_icwalk(
1752 struct xfs_mount *mp,
1753 enum xfs_icwalk_goal goal,
1754 struct xfs_icwalk *icw)
1755{
1756 struct xfs_perag *pag;
1757 int error = 0;
1758 int last_error = 0;
1759 xfs_agnumber_t agno;
1760
1761 for_each_perag_tag(mp, agno, pag, goal) {
1762 error = xfs_icwalk_ag(pag, goal, icw);
1763 if (error) {
1764 last_error = error;
1765 if (error == -EFSCORRUPTED) {
1766 xfs_perag_rele(pag);
1767 break;
1768 }
1769 }
1770 }
1771 return last_error;
1772 BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1773}
1774
1775#ifdef DEBUG
1776static void
1777xfs_check_delalloc(
1778 struct xfs_inode *ip,
1779 int whichfork)
1780{
1781 struct xfs_ifork *ifp = xfs_ifork_ptr(ip, whichfork);
1782 struct xfs_bmbt_irec got;
1783 struct xfs_iext_cursor icur;
1784
1785 if (!ifp || !xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got))
1786 return;
1787 do {
1788 if (isnullstartblock(got.br_startblock)) {
1789 xfs_warn(ip->i_mount,
1790 "ino %llx %s fork has delalloc extent at [0x%llx:0x%llx]",
1791 ip->i_ino,
1792 whichfork == XFS_DATA_FORK ? "data" : "cow",
1793 got.br_startoff, got.br_blockcount);
1794 }
1795 } while (xfs_iext_next_extent(ifp, &icur, &got));
1796}
1797#else
1798#define xfs_check_delalloc(ip, whichfork) do { } while (0)
1799#endif
1800
1801/* Schedule the inode for reclaim. */
1802static void
1803xfs_inodegc_set_reclaimable(
1804 struct xfs_inode *ip)
1805{
1806 struct xfs_mount *mp = ip->i_mount;
1807 struct xfs_perag *pag;
1808
1809 if (!xfs_is_shutdown(mp) && ip->i_delayed_blks) {
1810 xfs_check_delalloc(ip, XFS_DATA_FORK);
1811 xfs_check_delalloc(ip, XFS_COW_FORK);
1812 ASSERT(0);
1813 }
1814
1815 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1816 spin_lock(&pag->pag_ici_lock);
1817 spin_lock(&ip->i_flags_lock);
1818
1819 trace_xfs_inode_set_reclaimable(ip);
1820 ip->i_flags &= ~(XFS_NEED_INACTIVE | XFS_INACTIVATING);
1821 ip->i_flags |= XFS_IRECLAIMABLE;
1822 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1823 XFS_ICI_RECLAIM_TAG);
1824
1825 spin_unlock(&ip->i_flags_lock);
1826 spin_unlock(&pag->pag_ici_lock);
1827 xfs_perag_put(pag);
1828}
1829
1830/*
1831 * Free all speculative preallocations and possibly even the inode itself.
1832 * This is the last chance to make changes to an otherwise unreferenced file
1833 * before incore reclamation happens.
1834 */
1835static int
1836xfs_inodegc_inactivate(
1837 struct xfs_inode *ip)
1838{
1839 int error;
1840
1841 trace_xfs_inode_inactivating(ip);
1842 error = xfs_inactive(ip);
1843 xfs_inodegc_set_reclaimable(ip);
1844 return error;
1845
1846}
1847
1848void
1849xfs_inodegc_worker(
1850 struct work_struct *work)
1851{
1852 struct xfs_inodegc *gc = container_of(to_delayed_work(work),
1853 struct xfs_inodegc, work);
1854 struct llist_node *node = llist_del_all(&gc->list);
1855 struct xfs_inode *ip, *n;
1856 struct xfs_mount *mp = gc->mp;
1857 unsigned int nofs_flag;
1858
1859 /*
1860 * Clear the cpu mask bit and ensure that we have seen the latest
1861 * update of the gc structure associated with this CPU. This matches
1862 * with the release semantics used when setting the cpumask bit in
1863 * xfs_inodegc_queue.
1864 */
1865 cpumask_clear_cpu(gc->cpu, &mp->m_inodegc_cpumask);
1866 smp_mb__after_atomic();
1867
1868 WRITE_ONCE(gc->items, 0);
1869
1870 if (!node)
1871 return;
1872
1873 /*
1874 * We can allocate memory here while doing writeback on behalf of
1875 * memory reclaim. To avoid memory allocation deadlocks set the
1876 * task-wide nofs context for the following operations.
1877 */
1878 nofs_flag = memalloc_nofs_save();
1879
1880 ip = llist_entry(node, struct xfs_inode, i_gclist);
1881 trace_xfs_inodegc_worker(mp, READ_ONCE(gc->shrinker_hits));
1882
1883 WRITE_ONCE(gc->shrinker_hits, 0);
1884 llist_for_each_entry_safe(ip, n, node, i_gclist) {
1885 int error;
1886
1887 xfs_iflags_set(ip, XFS_INACTIVATING);
1888 error = xfs_inodegc_inactivate(ip);
1889 if (error && !gc->error)
1890 gc->error = error;
1891 }
1892
1893 memalloc_nofs_restore(nofs_flag);
1894}
1895
1896/*
1897 * Expedite all pending inodegc work to run immediately. This does not wait for
1898 * completion of the work.
1899 */
1900void
1901xfs_inodegc_push(
1902 struct xfs_mount *mp)
1903{
1904 if (!xfs_is_inodegc_enabled(mp))
1905 return;
1906 trace_xfs_inodegc_push(mp, __return_address);
1907 xfs_inodegc_queue_all(mp);
1908}
1909
1910/*
1911 * Force all currently queued inode inactivation work to run immediately and
1912 * wait for the work to finish.
1913 */
1914int
1915xfs_inodegc_flush(
1916 struct xfs_mount *mp)
1917{
1918 xfs_inodegc_push(mp);
1919 trace_xfs_inodegc_flush(mp, __return_address);
1920 return xfs_inodegc_wait_all(mp);
1921}
1922
1923/*
1924 * Flush all the pending work and then disable the inode inactivation background
1925 * workers and wait for them to stop. Caller must hold sb->s_umount to
1926 * coordinate changes in the inodegc_enabled state.
1927 */
1928void
1929xfs_inodegc_stop(
1930 struct xfs_mount *mp)
1931{
1932 bool rerun;
1933
1934 if (!xfs_clear_inodegc_enabled(mp))
1935 return;
1936
1937 /*
1938 * Drain all pending inodegc work, including inodes that could be
1939 * queued by racing xfs_inodegc_queue or xfs_inodegc_shrinker_scan
1940 * threads that sample the inodegc state just prior to us clearing it.
1941 * The inodegc flag state prevents new threads from queuing more
1942 * inodes, so we queue pending work items and flush the workqueue until
1943 * all inodegc lists are empty. IOWs, we cannot use drain_workqueue
1944 * here because it does not allow other unserialized mechanisms to
1945 * reschedule inodegc work while this draining is in progress.
1946 */
1947 xfs_inodegc_queue_all(mp);
1948 do {
1949 flush_workqueue(mp->m_inodegc_wq);
1950 rerun = xfs_inodegc_queue_all(mp);
1951 } while (rerun);
1952
1953 trace_xfs_inodegc_stop(mp, __return_address);
1954}
1955
1956/*
1957 * Enable the inode inactivation background workers and schedule deferred inode
1958 * inactivation work if there is any. Caller must hold sb->s_umount to
1959 * coordinate changes in the inodegc_enabled state.
1960 */
1961void
1962xfs_inodegc_start(
1963 struct xfs_mount *mp)
1964{
1965 if (xfs_set_inodegc_enabled(mp))
1966 return;
1967
1968 trace_xfs_inodegc_start(mp, __return_address);
1969 xfs_inodegc_queue_all(mp);
1970}
1971
1972#ifdef CONFIG_XFS_RT
1973static inline bool
1974xfs_inodegc_want_queue_rt_file(
1975 struct xfs_inode *ip)
1976{
1977 struct xfs_mount *mp = ip->i_mount;
1978
1979 if (!XFS_IS_REALTIME_INODE(ip))
1980 return false;
1981
1982 if (__percpu_counter_compare(&mp->m_frextents,
1983 mp->m_low_rtexts[XFS_LOWSP_5_PCNT],
1984 XFS_FDBLOCKS_BATCH) < 0)
1985 return true;
1986
1987 return false;
1988}
1989#else
1990# define xfs_inodegc_want_queue_rt_file(ip) (false)
1991#endif /* CONFIG_XFS_RT */
1992
1993/*
1994 * Schedule the inactivation worker when:
1995 *
1996 * - We've accumulated more than one inode cluster buffer's worth of inodes.
1997 * - There is less than 5% free space left.
1998 * - Any of the quotas for this inode are near an enforcement limit.
1999 */
2000static inline bool
2001xfs_inodegc_want_queue_work(
2002 struct xfs_inode *ip,
2003 unsigned int items)
2004{
2005 struct xfs_mount *mp = ip->i_mount;
2006
2007 if (items > mp->m_ino_geo.inodes_per_cluster)
2008 return true;
2009
2010 if (__percpu_counter_compare(&mp->m_fdblocks,
2011 mp->m_low_space[XFS_LOWSP_5_PCNT],
2012 XFS_FDBLOCKS_BATCH) < 0)
2013 return true;
2014
2015 if (xfs_inodegc_want_queue_rt_file(ip))
2016 return true;
2017
2018 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_USER))
2019 return true;
2020
2021 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_GROUP))
2022 return true;
2023
2024 if (xfs_inode_near_dquot_enforcement(ip, XFS_DQTYPE_PROJ))
2025 return true;
2026
2027 return false;
2028}
2029
2030/*
2031 * Upper bound on the number of inodes in each AG that can be queued for
2032 * inactivation at any given time, to avoid monopolizing the workqueue.
2033 */
2034#define XFS_INODEGC_MAX_BACKLOG (4 * XFS_INODES_PER_CHUNK)
2035
2036/*
2037 * Make the frontend wait for inactivations when:
2038 *
2039 * - Memory shrinkers queued the inactivation worker and it hasn't finished.
2040 * - The queue depth exceeds the maximum allowable percpu backlog.
2041 *
2042 * Note: If we are in a NOFS context here (e.g. current thread is running a
2043 * transaction) the we don't want to block here as inodegc progress may require
2044 * filesystem resources we hold to make progress and that could result in a
2045 * deadlock. Hence we skip out of here if we are in a scoped NOFS context.
2046 */
2047static inline bool
2048xfs_inodegc_want_flush_work(
2049 struct xfs_inode *ip,
2050 unsigned int items,
2051 unsigned int shrinker_hits)
2052{
2053 if (current->flags & PF_MEMALLOC_NOFS)
2054 return false;
2055
2056 if (shrinker_hits > 0)
2057 return true;
2058
2059 if (items > XFS_INODEGC_MAX_BACKLOG)
2060 return true;
2061
2062 return false;
2063}
2064
2065/*
2066 * Queue a background inactivation worker if there are inodes that need to be
2067 * inactivated and higher level xfs code hasn't disabled the background
2068 * workers.
2069 */
2070static void
2071xfs_inodegc_queue(
2072 struct xfs_inode *ip)
2073{
2074 struct xfs_mount *mp = ip->i_mount;
2075 struct xfs_inodegc *gc;
2076 int items;
2077 unsigned int shrinker_hits;
2078 unsigned int cpu_nr;
2079 unsigned long queue_delay = 1;
2080
2081 trace_xfs_inode_set_need_inactive(ip);
2082 spin_lock(&ip->i_flags_lock);
2083 ip->i_flags |= XFS_NEED_INACTIVE;
2084 spin_unlock(&ip->i_flags_lock);
2085
2086 cpu_nr = get_cpu();
2087 gc = this_cpu_ptr(mp->m_inodegc);
2088 llist_add(&ip->i_gclist, &gc->list);
2089 items = READ_ONCE(gc->items);
2090 WRITE_ONCE(gc->items, items + 1);
2091 shrinker_hits = READ_ONCE(gc->shrinker_hits);
2092
2093 /*
2094 * Ensure the list add is always seen by anyone who finds the cpumask
2095 * bit set. This effectively gives the cpumask bit set operation
2096 * release ordering semantics.
2097 */
2098 smp_mb__before_atomic();
2099 if (!cpumask_test_cpu(cpu_nr, &mp->m_inodegc_cpumask))
2100 cpumask_test_and_set_cpu(cpu_nr, &mp->m_inodegc_cpumask);
2101
2102 /*
2103 * We queue the work while holding the current CPU so that the work
2104 * is scheduled to run on this CPU.
2105 */
2106 if (!xfs_is_inodegc_enabled(mp)) {
2107 put_cpu();
2108 return;
2109 }
2110
2111 if (xfs_inodegc_want_queue_work(ip, items))
2112 queue_delay = 0;
2113
2114 trace_xfs_inodegc_queue(mp, __return_address);
2115 mod_delayed_work_on(current_cpu(), mp->m_inodegc_wq, &gc->work,
2116 queue_delay);
2117 put_cpu();
2118
2119 if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
2120 trace_xfs_inodegc_throttle(mp, __return_address);
2121 flush_delayed_work(&gc->work);
2122 }
2123}
2124
2125/*
2126 * We set the inode flag atomically with the radix tree tag. Once we get tag
2127 * lookups on the radix tree, this inode flag can go away.
2128 *
2129 * We always use background reclaim here because even if the inode is clean, it
2130 * still may be under IO and hence we have wait for IO completion to occur
2131 * before we can reclaim the inode. The background reclaim path handles this
2132 * more efficiently than we can here, so simply let background reclaim tear down
2133 * all inodes.
2134 */
2135void
2136xfs_inode_mark_reclaimable(
2137 struct xfs_inode *ip)
2138{
2139 struct xfs_mount *mp = ip->i_mount;
2140 bool need_inactive;
2141
2142 XFS_STATS_INC(mp, vn_reclaim);
2143
2144 /*
2145 * We should never get here with any of the reclaim flags already set.
2146 */
2147 ASSERT_ALWAYS(!xfs_iflags_test(ip, XFS_ALL_IRECLAIM_FLAGS));
2148
2149 need_inactive = xfs_inode_needs_inactive(ip);
2150 if (need_inactive) {
2151 xfs_inodegc_queue(ip);
2152 return;
2153 }
2154
2155 /* Going straight to reclaim, so drop the dquots. */
2156 xfs_qm_dqdetach(ip);
2157 xfs_inodegc_set_reclaimable(ip);
2158}
2159
2160/*
2161 * Register a phony shrinker so that we can run background inodegc sooner when
2162 * there's memory pressure. Inactivation does not itself free any memory but
2163 * it does make inodes reclaimable, which eventually frees memory.
2164 *
2165 * The count function, seek value, and batch value are crafted to trigger the
2166 * scan function during the second round of scanning. Hopefully this means
2167 * that we reclaimed enough memory that initiating metadata transactions won't
2168 * make things worse.
2169 */
2170#define XFS_INODEGC_SHRINKER_COUNT (1UL << DEF_PRIORITY)
2171#define XFS_INODEGC_SHRINKER_BATCH ((XFS_INODEGC_SHRINKER_COUNT / 2) + 1)
2172
2173static unsigned long
2174xfs_inodegc_shrinker_count(
2175 struct shrinker *shrink,
2176 struct shrink_control *sc)
2177{
2178 struct xfs_mount *mp = shrink->private_data;
2179 struct xfs_inodegc *gc;
2180 int cpu;
2181
2182 if (!xfs_is_inodegc_enabled(mp))
2183 return 0;
2184
2185 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
2186 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2187 if (!llist_empty(&gc->list))
2188 return XFS_INODEGC_SHRINKER_COUNT;
2189 }
2190
2191 return 0;
2192}
2193
2194static unsigned long
2195xfs_inodegc_shrinker_scan(
2196 struct shrinker *shrink,
2197 struct shrink_control *sc)
2198{
2199 struct xfs_mount *mp = shrink->private_data;
2200 struct xfs_inodegc *gc;
2201 int cpu;
2202 bool no_items = true;
2203
2204 if (!xfs_is_inodegc_enabled(mp))
2205 return SHRINK_STOP;
2206
2207 trace_xfs_inodegc_shrinker_scan(mp, sc, __return_address);
2208
2209 for_each_cpu(cpu, &mp->m_inodegc_cpumask) {
2210 gc = per_cpu_ptr(mp->m_inodegc, cpu);
2211 if (!llist_empty(&gc->list)) {
2212 unsigned int h = READ_ONCE(gc->shrinker_hits);
2213
2214 WRITE_ONCE(gc->shrinker_hits, h + 1);
2215 mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
2216 no_items = false;
2217 }
2218 }
2219
2220 /*
2221 * If there are no inodes to inactivate, we don't want the shrinker
2222 * to think there's deferred work to call us back about.
2223 */
2224 if (no_items)
2225 return LONG_MAX;
2226
2227 return SHRINK_STOP;
2228}
2229
2230/* Register a shrinker so we can accelerate inodegc and throttle queuing. */
2231int
2232xfs_inodegc_register_shrinker(
2233 struct xfs_mount *mp)
2234{
2235 mp->m_inodegc_shrinker = shrinker_alloc(SHRINKER_NONSLAB,
2236 "xfs-inodegc:%s",
2237 mp->m_super->s_id);
2238 if (!mp->m_inodegc_shrinker)
2239 return -ENOMEM;
2240
2241 mp->m_inodegc_shrinker->count_objects = xfs_inodegc_shrinker_count;
2242 mp->m_inodegc_shrinker->scan_objects = xfs_inodegc_shrinker_scan;
2243 mp->m_inodegc_shrinker->seeks = 0;
2244 mp->m_inodegc_shrinker->batch = XFS_INODEGC_SHRINKER_BATCH;
2245 mp->m_inodegc_shrinker->private_data = mp;
2246
2247 shrinker_register(mp->m_inodegc_shrinker);
2248
2249 return 0;
2250}