Loading...
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
27#include <linux/iversion.h>
28
29/* Radix tree tags for incore inode tree. */
30
31/* inode is to be reclaimed */
32#define XFS_ICI_RECLAIM_TAG 0
33/* Inode has speculative preallocations (posteof or cow) to clean. */
34#define XFS_ICI_BLOCKGC_TAG 1
35
36/*
37 * The goal for walking incore inodes. These can correspond with incore inode
38 * radix tree tags when convenient. Avoid existing XFS_IWALK namespace.
39 */
40enum xfs_icwalk_goal {
41 /* Goals that are not related to tags; these must be < 0. */
42 XFS_ICWALK_DQRELE = -1,
43
44 /* Goals directly associated with tagged inodes. */
45 XFS_ICWALK_BLOCKGC = XFS_ICI_BLOCKGC_TAG,
46 XFS_ICWALK_RECLAIM = XFS_ICI_RECLAIM_TAG,
47};
48
49#define XFS_ICWALK_NULL_TAG (-1U)
50
51/* Compute the inode radix tree tag for this goal. */
52static inline unsigned int
53xfs_icwalk_tag(enum xfs_icwalk_goal goal)
54{
55 return goal < 0 ? XFS_ICWALK_NULL_TAG : goal;
56}
57
58static int xfs_icwalk(struct xfs_mount *mp,
59 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
60static int xfs_icwalk_ag(struct xfs_perag *pag,
61 enum xfs_icwalk_goal goal, struct xfs_icwalk *icw);
62
63/*
64 * Private inode cache walk flags for struct xfs_icwalk. Must not
65 * coincide with XFS_ICWALK_FLAGS_VALID.
66 */
67#define XFS_ICWALK_FLAG_DROP_UDQUOT (1U << 31)
68#define XFS_ICWALK_FLAG_DROP_GDQUOT (1U << 30)
69#define XFS_ICWALK_FLAG_DROP_PDQUOT (1U << 29)
70
71/* Stop scanning after icw_scan_limit inodes. */
72#define XFS_ICWALK_FLAG_SCAN_LIMIT (1U << 28)
73
74#define XFS_ICWALK_FLAG_RECLAIM_SICK (1U << 27)
75#define XFS_ICWALK_FLAG_UNION (1U << 26) /* union filter algorithm */
76
77#define XFS_ICWALK_PRIVATE_FLAGS (XFS_ICWALK_FLAG_DROP_UDQUOT | \
78 XFS_ICWALK_FLAG_DROP_GDQUOT | \
79 XFS_ICWALK_FLAG_DROP_PDQUOT | \
80 XFS_ICWALK_FLAG_SCAN_LIMIT | \
81 XFS_ICWALK_FLAG_RECLAIM_SICK | \
82 XFS_ICWALK_FLAG_UNION)
83
84/*
85 * Allocate and initialise an xfs_inode.
86 */
87struct xfs_inode *
88xfs_inode_alloc(
89 struct xfs_mount *mp,
90 xfs_ino_t ino)
91{
92 struct xfs_inode *ip;
93
94 /*
95 * XXX: If this didn't occur in transactions, we could drop GFP_NOFAIL
96 * and return NULL here on ENOMEM.
97 */
98 ip = kmem_cache_alloc(xfs_inode_zone, GFP_KERNEL | __GFP_NOFAIL);
99
100 if (inode_init_always(mp->m_super, VFS_I(ip))) {
101 kmem_cache_free(xfs_inode_zone, ip);
102 return NULL;
103 }
104
105 /* VFS doesn't initialise i_mode! */
106 VFS_I(ip)->i_mode = 0;
107
108 XFS_STATS_INC(mp, vn_active);
109 ASSERT(atomic_read(&ip->i_pincount) == 0);
110 ASSERT(ip->i_ino == 0);
111
112 /* initialise the xfs inode */
113 ip->i_ino = ino;
114 ip->i_mount = mp;
115 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
116 ip->i_afp = NULL;
117 ip->i_cowfp = NULL;
118 memset(&ip->i_df, 0, sizeof(ip->i_df));
119 ip->i_flags = 0;
120 ip->i_delayed_blks = 0;
121 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
122 ip->i_nblocks = 0;
123 ip->i_forkoff = 0;
124 ip->i_sick = 0;
125 ip->i_checked = 0;
126 INIT_WORK(&ip->i_ioend_work, xfs_end_io);
127 INIT_LIST_HEAD(&ip->i_ioend_list);
128 spin_lock_init(&ip->i_ioend_lock);
129
130 return ip;
131}
132
133STATIC void
134xfs_inode_free_callback(
135 struct rcu_head *head)
136{
137 struct inode *inode = container_of(head, struct inode, i_rcu);
138 struct xfs_inode *ip = XFS_I(inode);
139
140 switch (VFS_I(ip)->i_mode & S_IFMT) {
141 case S_IFREG:
142 case S_IFDIR:
143 case S_IFLNK:
144 xfs_idestroy_fork(&ip->i_df);
145 break;
146 }
147
148 if (ip->i_afp) {
149 xfs_idestroy_fork(ip->i_afp);
150 kmem_cache_free(xfs_ifork_zone, ip->i_afp);
151 }
152 if (ip->i_cowfp) {
153 xfs_idestroy_fork(ip->i_cowfp);
154 kmem_cache_free(xfs_ifork_zone, ip->i_cowfp);
155 }
156 if (ip->i_itemp) {
157 ASSERT(!test_bit(XFS_LI_IN_AIL,
158 &ip->i_itemp->ili_item.li_flags));
159 xfs_inode_item_destroy(ip);
160 ip->i_itemp = NULL;
161 }
162
163 kmem_cache_free(xfs_inode_zone, ip);
164}
165
166static void
167__xfs_inode_free(
168 struct xfs_inode *ip)
169{
170 /* asserts to verify all state is correct here */
171 ASSERT(atomic_read(&ip->i_pincount) == 0);
172 ASSERT(!ip->i_itemp || list_empty(&ip->i_itemp->ili_item.li_bio_list));
173 XFS_STATS_DEC(ip->i_mount, vn_active);
174
175 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
176}
177
178void
179xfs_inode_free(
180 struct xfs_inode *ip)
181{
182 ASSERT(!xfs_iflags_test(ip, XFS_IFLUSHING));
183
184 /*
185 * Because we use RCU freeing we need to ensure the inode always
186 * appears to be reclaimed with an invalid inode number when in the
187 * free state. The ip->i_flags_lock provides the barrier against lookup
188 * races.
189 */
190 spin_lock(&ip->i_flags_lock);
191 ip->i_flags = XFS_IRECLAIM;
192 ip->i_ino = 0;
193 spin_unlock(&ip->i_flags_lock);
194
195 __xfs_inode_free(ip);
196}
197
198/*
199 * Queue background inode reclaim work if there are reclaimable inodes and there
200 * isn't reclaim work already scheduled or in progress.
201 */
202static void
203xfs_reclaim_work_queue(
204 struct xfs_mount *mp)
205{
206
207 rcu_read_lock();
208 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
209 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
210 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
211 }
212 rcu_read_unlock();
213}
214
215/*
216 * Background scanning to trim preallocated space. This is queued based on the
217 * 'speculative_prealloc_lifetime' tunable (5m by default).
218 */
219static inline void
220xfs_blockgc_queue(
221 struct xfs_perag *pag)
222{
223 rcu_read_lock();
224 if (radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_BLOCKGC_TAG))
225 queue_delayed_work(pag->pag_mount->m_gc_workqueue,
226 &pag->pag_blockgc_work,
227 msecs_to_jiffies(xfs_blockgc_secs * 1000));
228 rcu_read_unlock();
229}
230
231/* Set a tag on both the AG incore inode tree and the AG radix tree. */
232static void
233xfs_perag_set_inode_tag(
234 struct xfs_perag *pag,
235 xfs_agino_t agino,
236 unsigned int tag)
237{
238 struct xfs_mount *mp = pag->pag_mount;
239 bool was_tagged;
240
241 lockdep_assert_held(&pag->pag_ici_lock);
242
243 was_tagged = radix_tree_tagged(&pag->pag_ici_root, tag);
244 radix_tree_tag_set(&pag->pag_ici_root, agino, tag);
245
246 if (tag == XFS_ICI_RECLAIM_TAG)
247 pag->pag_ici_reclaimable++;
248
249 if (was_tagged)
250 return;
251
252 /* propagate the tag up into the perag radix tree */
253 spin_lock(&mp->m_perag_lock);
254 radix_tree_tag_set(&mp->m_perag_tree, pag->pag_agno, tag);
255 spin_unlock(&mp->m_perag_lock);
256
257 /* start background work */
258 switch (tag) {
259 case XFS_ICI_RECLAIM_TAG:
260 xfs_reclaim_work_queue(mp);
261 break;
262 case XFS_ICI_BLOCKGC_TAG:
263 xfs_blockgc_queue(pag);
264 break;
265 }
266
267 trace_xfs_perag_set_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
268}
269
270/* Clear a tag on both the AG incore inode tree and the AG radix tree. */
271static void
272xfs_perag_clear_inode_tag(
273 struct xfs_perag *pag,
274 xfs_agino_t agino,
275 unsigned int tag)
276{
277 struct xfs_mount *mp = pag->pag_mount;
278
279 lockdep_assert_held(&pag->pag_ici_lock);
280
281 /*
282 * Reclaim can signal (with a null agino) that it cleared its own tag
283 * by removing the inode from the radix tree.
284 */
285 if (agino != NULLAGINO)
286 radix_tree_tag_clear(&pag->pag_ici_root, agino, tag);
287 else
288 ASSERT(tag == XFS_ICI_RECLAIM_TAG);
289
290 if (tag == XFS_ICI_RECLAIM_TAG)
291 pag->pag_ici_reclaimable--;
292
293 if (radix_tree_tagged(&pag->pag_ici_root, tag))
294 return;
295
296 /* clear the tag from the perag radix tree */
297 spin_lock(&mp->m_perag_lock);
298 radix_tree_tag_clear(&mp->m_perag_tree, pag->pag_agno, tag);
299 spin_unlock(&mp->m_perag_lock);
300
301 trace_xfs_perag_clear_inode_tag(mp, pag->pag_agno, tag, _RET_IP_);
302}
303
304/*
305 * We set the inode flag atomically with the radix tree tag.
306 * Once we get tag lookups on the radix tree, this inode flag
307 * can go away.
308 */
309void
310xfs_inode_mark_reclaimable(
311 struct xfs_inode *ip)
312{
313 struct xfs_mount *mp = ip->i_mount;
314 struct xfs_perag *pag;
315
316 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
317 spin_lock(&pag->pag_ici_lock);
318 spin_lock(&ip->i_flags_lock);
319
320 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
321 XFS_ICI_RECLAIM_TAG);
322 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
323
324 spin_unlock(&ip->i_flags_lock);
325 spin_unlock(&pag->pag_ici_lock);
326 xfs_perag_put(pag);
327}
328
329static inline void
330xfs_inew_wait(
331 struct xfs_inode *ip)
332{
333 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_INEW_BIT);
334 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_INEW_BIT);
335
336 do {
337 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
338 if (!xfs_iflags_test(ip, XFS_INEW))
339 break;
340 schedule();
341 } while (true);
342 finish_wait(wq, &wait.wq_entry);
343}
344
345/*
346 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
347 * part of the structure. This is made more complex by the fact we store
348 * information about the on-disk values in the VFS inode and so we can't just
349 * overwrite the values unconditionally. Hence we save the parameters we
350 * need to retain across reinitialisation, and rewrite them into the VFS inode
351 * after reinitialisation even if it fails.
352 */
353static int
354xfs_reinit_inode(
355 struct xfs_mount *mp,
356 struct inode *inode)
357{
358 int error;
359 uint32_t nlink = inode->i_nlink;
360 uint32_t generation = inode->i_generation;
361 uint64_t version = inode_peek_iversion(inode);
362 umode_t mode = inode->i_mode;
363 dev_t dev = inode->i_rdev;
364 kuid_t uid = inode->i_uid;
365 kgid_t gid = inode->i_gid;
366
367 error = inode_init_always(mp->m_super, inode);
368
369 set_nlink(inode, nlink);
370 inode->i_generation = generation;
371 inode_set_iversion_queried(inode, version);
372 inode->i_mode = mode;
373 inode->i_rdev = dev;
374 inode->i_uid = uid;
375 inode->i_gid = gid;
376 return error;
377}
378
379/*
380 * Carefully nudge an inode whose VFS state has been torn down back into a
381 * usable state. Drops the i_flags_lock and the rcu read lock.
382 */
383static int
384xfs_iget_recycle(
385 struct xfs_perag *pag,
386 struct xfs_inode *ip) __releases(&ip->i_flags_lock)
387{
388 struct xfs_mount *mp = ip->i_mount;
389 struct inode *inode = VFS_I(ip);
390 int error;
391
392 trace_xfs_iget_recycle(ip);
393
394 /*
395 * We need to make it look like the inode is being reclaimed to prevent
396 * the actual reclaim workers from stomping over us while we recycle
397 * the inode. We can't clear the radix tree tag yet as it requires
398 * pag_ici_lock to be held exclusive.
399 */
400 ip->i_flags |= XFS_IRECLAIM;
401
402 spin_unlock(&ip->i_flags_lock);
403 rcu_read_unlock();
404
405 ASSERT(!rwsem_is_locked(&inode->i_rwsem));
406 error = xfs_reinit_inode(mp, inode);
407 if (error) {
408 bool wake;
409
410 /*
411 * Re-initializing the inode failed, and we are in deep
412 * trouble. Try to re-add it to the reclaim list.
413 */
414 rcu_read_lock();
415 spin_lock(&ip->i_flags_lock);
416 wake = !!__xfs_iflags_test(ip, XFS_INEW);
417 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
418 if (wake)
419 wake_up_bit(&ip->i_flags, __XFS_INEW_BIT);
420 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
421 spin_unlock(&ip->i_flags_lock);
422 rcu_read_unlock();
423
424 trace_xfs_iget_recycle_fail(ip);
425 return error;
426 }
427
428 spin_lock(&pag->pag_ici_lock);
429 spin_lock(&ip->i_flags_lock);
430
431 /*
432 * Clear the per-lifetime state in the inode as we are now effectively
433 * a new inode and need to return to the initial state before reuse
434 * occurs.
435 */
436 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
437 ip->i_flags |= XFS_INEW;
438 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
439 XFS_ICI_RECLAIM_TAG);
440 inode->i_state = I_NEW;
441 spin_unlock(&ip->i_flags_lock);
442 spin_unlock(&pag->pag_ici_lock);
443
444 return 0;
445}
446
447/*
448 * If we are allocating a new inode, then check what was returned is
449 * actually a free, empty inode. If we are not allocating an inode,
450 * then check we didn't find a free inode.
451 *
452 * Returns:
453 * 0 if the inode free state matches the lookup context
454 * -ENOENT if the inode is free and we are not allocating
455 * -EFSCORRUPTED if there is any state mismatch at all
456 */
457static int
458xfs_iget_check_free_state(
459 struct xfs_inode *ip,
460 int flags)
461{
462 if (flags & XFS_IGET_CREATE) {
463 /* should be a free inode */
464 if (VFS_I(ip)->i_mode != 0) {
465 xfs_warn(ip->i_mount,
466"Corruption detected! Free inode 0x%llx not marked free! (mode 0x%x)",
467 ip->i_ino, VFS_I(ip)->i_mode);
468 return -EFSCORRUPTED;
469 }
470
471 if (ip->i_nblocks != 0) {
472 xfs_warn(ip->i_mount,
473"Corruption detected! Free inode 0x%llx has blocks allocated!",
474 ip->i_ino);
475 return -EFSCORRUPTED;
476 }
477 return 0;
478 }
479
480 /* should be an allocated inode */
481 if (VFS_I(ip)->i_mode == 0)
482 return -ENOENT;
483
484 return 0;
485}
486
487/*
488 * Check the validity of the inode we just found it the cache
489 */
490static int
491xfs_iget_cache_hit(
492 struct xfs_perag *pag,
493 struct xfs_inode *ip,
494 xfs_ino_t ino,
495 int flags,
496 int lock_flags) __releases(RCU)
497{
498 struct inode *inode = VFS_I(ip);
499 struct xfs_mount *mp = ip->i_mount;
500 int error;
501
502 /*
503 * check for re-use of an inode within an RCU grace period due to the
504 * radix tree nodes not being updated yet. We monitor for this by
505 * setting the inode number to zero before freeing the inode structure.
506 * If the inode has been reallocated and set up, then the inode number
507 * will not match, so check for that, too.
508 */
509 spin_lock(&ip->i_flags_lock);
510 if (ip->i_ino != ino)
511 goto out_skip;
512
513 /*
514 * If we are racing with another cache hit that is currently
515 * instantiating this inode or currently recycling it out of
516 * reclaimable state, wait for the initialisation to complete
517 * before continuing.
518 *
519 * XXX(hch): eventually we should do something equivalent to
520 * wait_on_inode to wait for these flags to be cleared
521 * instead of polling for it.
522 */
523 if (ip->i_flags & (XFS_INEW | XFS_IRECLAIM))
524 goto out_skip;
525
526 /*
527 * Check the inode free state is valid. This also detects lookup
528 * racing with unlinks.
529 */
530 error = xfs_iget_check_free_state(ip, flags);
531 if (error)
532 goto out_error;
533
534 /* Skip inodes that have no vfs state. */
535 if ((flags & XFS_IGET_INCORE) &&
536 (ip->i_flags & XFS_IRECLAIMABLE))
537 goto out_skip;
538
539 /* The inode fits the selection criteria; process it. */
540 if (ip->i_flags & XFS_IRECLAIMABLE) {
541 /* Drops i_flags_lock and RCU read lock. */
542 error = xfs_iget_recycle(pag, ip);
543 if (error)
544 return error;
545 } else {
546 /* If the VFS inode is being torn down, pause and try again. */
547 if (!igrab(inode))
548 goto out_skip;
549
550 /* We've got a live one. */
551 spin_unlock(&ip->i_flags_lock);
552 rcu_read_unlock();
553 trace_xfs_iget_hit(ip);
554 }
555
556 if (lock_flags != 0)
557 xfs_ilock(ip, lock_flags);
558
559 if (!(flags & XFS_IGET_INCORE))
560 xfs_iflags_clear(ip, XFS_ISTALE);
561 XFS_STATS_INC(mp, xs_ig_found);
562
563 return 0;
564
565out_skip:
566 trace_xfs_iget_skip(ip);
567 XFS_STATS_INC(mp, xs_ig_frecycle);
568 error = -EAGAIN;
569out_error:
570 spin_unlock(&ip->i_flags_lock);
571 rcu_read_unlock();
572 return error;
573}
574
575static int
576xfs_iget_cache_miss(
577 struct xfs_mount *mp,
578 struct xfs_perag *pag,
579 xfs_trans_t *tp,
580 xfs_ino_t ino,
581 struct xfs_inode **ipp,
582 int flags,
583 int lock_flags)
584{
585 struct xfs_inode *ip;
586 int error;
587 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
588 int iflags;
589
590 ip = xfs_inode_alloc(mp, ino);
591 if (!ip)
592 return -ENOMEM;
593
594 error = xfs_imap(mp, tp, ip->i_ino, &ip->i_imap, flags);
595 if (error)
596 goto out_destroy;
597
598 /*
599 * For version 5 superblocks, if we are initialising a new inode and we
600 * are not utilising the XFS_MOUNT_IKEEP inode cluster mode, we can
601 * simply build the new inode core with a random generation number.
602 *
603 * For version 4 (and older) superblocks, log recovery is dependent on
604 * the i_flushiter field being initialised from the current on-disk
605 * value and hence we must also read the inode off disk even when
606 * initializing new inodes.
607 */
608 if (xfs_sb_version_has_v3inode(&mp->m_sb) &&
609 (flags & XFS_IGET_CREATE) && !(mp->m_flags & XFS_MOUNT_IKEEP)) {
610 VFS_I(ip)->i_generation = prandom_u32();
611 } else {
612 struct xfs_buf *bp;
613
614 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &bp);
615 if (error)
616 goto out_destroy;
617
618 error = xfs_inode_from_disk(ip,
619 xfs_buf_offset(bp, ip->i_imap.im_boffset));
620 if (!error)
621 xfs_buf_set_ref(bp, XFS_INO_REF);
622 xfs_trans_brelse(tp, bp);
623
624 if (error)
625 goto out_destroy;
626 }
627
628 trace_xfs_iget_miss(ip);
629
630 /*
631 * Check the inode free state is valid. This also detects lookup
632 * racing with unlinks.
633 */
634 error = xfs_iget_check_free_state(ip, flags);
635 if (error)
636 goto out_destroy;
637
638 /*
639 * Preload the radix tree so we can insert safely under the
640 * write spinlock. Note that we cannot sleep inside the preload
641 * region. Since we can be called from transaction context, don't
642 * recurse into the file system.
643 */
644 if (radix_tree_preload(GFP_NOFS)) {
645 error = -EAGAIN;
646 goto out_destroy;
647 }
648
649 /*
650 * Because the inode hasn't been added to the radix-tree yet it can't
651 * be found by another thread, so we can do the non-sleeping lock here.
652 */
653 if (lock_flags) {
654 if (!xfs_ilock_nowait(ip, lock_flags))
655 BUG();
656 }
657
658 /*
659 * These values must be set before inserting the inode into the radix
660 * tree as the moment it is inserted a concurrent lookup (allowed by the
661 * RCU locking mechanism) can find it and that lookup must see that this
662 * is an inode currently under construction (i.e. that XFS_INEW is set).
663 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
664 * memory barrier that ensures this detection works correctly at lookup
665 * time.
666 */
667 iflags = XFS_INEW;
668 if (flags & XFS_IGET_DONTCACHE)
669 d_mark_dontcache(VFS_I(ip));
670 ip->i_udquot = NULL;
671 ip->i_gdquot = NULL;
672 ip->i_pdquot = NULL;
673 xfs_iflags_set(ip, iflags);
674
675 /* insert the new inode */
676 spin_lock(&pag->pag_ici_lock);
677 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
678 if (unlikely(error)) {
679 WARN_ON(error != -EEXIST);
680 XFS_STATS_INC(mp, xs_ig_dup);
681 error = -EAGAIN;
682 goto out_preload_end;
683 }
684 spin_unlock(&pag->pag_ici_lock);
685 radix_tree_preload_end();
686
687 *ipp = ip;
688 return 0;
689
690out_preload_end:
691 spin_unlock(&pag->pag_ici_lock);
692 radix_tree_preload_end();
693 if (lock_flags)
694 xfs_iunlock(ip, lock_flags);
695out_destroy:
696 __destroy_inode(VFS_I(ip));
697 xfs_inode_free(ip);
698 return error;
699}
700
701/*
702 * Look up an inode by number in the given file system. The inode is looked up
703 * in the cache held in each AG. If the inode is found in the cache, initialise
704 * the vfs inode if necessary.
705 *
706 * If it is not in core, read it in from the file system's device, add it to the
707 * cache and initialise the vfs inode.
708 *
709 * The inode is locked according to the value of the lock_flags parameter.
710 * Inode lookup is only done during metadata operations and not as part of the
711 * data IO path. Hence we only allow locking of the XFS_ILOCK during lookup.
712 */
713int
714xfs_iget(
715 struct xfs_mount *mp,
716 struct xfs_trans *tp,
717 xfs_ino_t ino,
718 uint flags,
719 uint lock_flags,
720 struct xfs_inode **ipp)
721{
722 struct xfs_inode *ip;
723 struct xfs_perag *pag;
724 xfs_agino_t agino;
725 int error;
726
727 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
728
729 /* reject inode numbers outside existing AGs */
730 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
731 return -EINVAL;
732
733 XFS_STATS_INC(mp, xs_ig_attempts);
734
735 /* get the perag structure and ensure that it's inode capable */
736 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
737 agino = XFS_INO_TO_AGINO(mp, ino);
738
739again:
740 error = 0;
741 rcu_read_lock();
742 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
743
744 if (ip) {
745 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
746 if (error)
747 goto out_error_or_again;
748 } else {
749 rcu_read_unlock();
750 if (flags & XFS_IGET_INCORE) {
751 error = -ENODATA;
752 goto out_error_or_again;
753 }
754 XFS_STATS_INC(mp, xs_ig_missed);
755
756 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
757 flags, lock_flags);
758 if (error)
759 goto out_error_or_again;
760 }
761 xfs_perag_put(pag);
762
763 *ipp = ip;
764
765 /*
766 * If we have a real type for an on-disk inode, we can setup the inode
767 * now. If it's a new inode being created, xfs_ialloc will handle it.
768 */
769 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
770 xfs_setup_existing_inode(ip);
771 return 0;
772
773out_error_or_again:
774 if (!(flags & XFS_IGET_INCORE) && error == -EAGAIN) {
775 delay(1);
776 goto again;
777 }
778 xfs_perag_put(pag);
779 return error;
780}
781
782/*
783 * "Is this a cached inode that's also allocated?"
784 *
785 * Look up an inode by number in the given file system. If the inode is
786 * in cache and isn't in purgatory, return 1 if the inode is allocated
787 * and 0 if it is not. For all other cases (not in cache, being torn
788 * down, etc.), return a negative error code.
789 *
790 * The caller has to prevent inode allocation and freeing activity,
791 * presumably by locking the AGI buffer. This is to ensure that an
792 * inode cannot transition from allocated to freed until the caller is
793 * ready to allow that. If the inode is in an intermediate state (new,
794 * reclaimable, or being reclaimed), -EAGAIN will be returned; if the
795 * inode is not in the cache, -ENOENT will be returned. The caller must
796 * deal with these scenarios appropriately.
797 *
798 * This is a specialized use case for the online scrubber; if you're
799 * reading this, you probably want xfs_iget.
800 */
801int
802xfs_icache_inode_is_allocated(
803 struct xfs_mount *mp,
804 struct xfs_trans *tp,
805 xfs_ino_t ino,
806 bool *inuse)
807{
808 struct xfs_inode *ip;
809 int error;
810
811 error = xfs_iget(mp, tp, ino, XFS_IGET_INCORE, 0, &ip);
812 if (error)
813 return error;
814
815 *inuse = !!(VFS_I(ip)->i_mode);
816 xfs_irele(ip);
817 return 0;
818}
819
820#ifdef CONFIG_XFS_QUOTA
821/* Decide if we want to grab this inode to drop its dquots. */
822static bool
823xfs_dqrele_igrab(
824 struct xfs_inode *ip)
825{
826 bool ret = false;
827
828 ASSERT(rcu_read_lock_held());
829
830 /* Check for stale RCU freed inode */
831 spin_lock(&ip->i_flags_lock);
832 if (!ip->i_ino)
833 goto out_unlock;
834
835 /*
836 * Skip inodes that are anywhere in the reclaim machinery because we
837 * drop dquots before tagging an inode for reclamation.
838 */
839 if (ip->i_flags & (XFS_IRECLAIM | XFS_IRECLAIMABLE))
840 goto out_unlock;
841
842 /*
843 * The inode looks alive; try to grab a VFS reference so that it won't
844 * get destroyed. If we got the reference, return true to say that
845 * we grabbed the inode.
846 *
847 * If we can't get the reference, then we know the inode had its VFS
848 * state torn down and hasn't yet entered the reclaim machinery. Since
849 * we also know that dquots are detached from an inode before it enters
850 * reclaim, we can skip the inode.
851 */
852 ret = igrab(VFS_I(ip)) != NULL;
853
854out_unlock:
855 spin_unlock(&ip->i_flags_lock);
856 return ret;
857}
858
859/* Drop this inode's dquots. */
860static void
861xfs_dqrele_inode(
862 struct xfs_inode *ip,
863 struct xfs_icwalk *icw)
864{
865 if (xfs_iflags_test(ip, XFS_INEW))
866 xfs_inew_wait(ip);
867
868 xfs_ilock(ip, XFS_ILOCK_EXCL);
869 if (icw->icw_flags & XFS_ICWALK_FLAG_DROP_UDQUOT) {
870 xfs_qm_dqrele(ip->i_udquot);
871 ip->i_udquot = NULL;
872 }
873 if (icw->icw_flags & XFS_ICWALK_FLAG_DROP_GDQUOT) {
874 xfs_qm_dqrele(ip->i_gdquot);
875 ip->i_gdquot = NULL;
876 }
877 if (icw->icw_flags & XFS_ICWALK_FLAG_DROP_PDQUOT) {
878 xfs_qm_dqrele(ip->i_pdquot);
879 ip->i_pdquot = NULL;
880 }
881 xfs_iunlock(ip, XFS_ILOCK_EXCL);
882 xfs_irele(ip);
883}
884
885/*
886 * Detach all dquots from incore inodes if we can. The caller must already
887 * have dropped the relevant XFS_[UGP]QUOTA_ACTIVE flags so that dquots will
888 * not get reattached.
889 */
890int
891xfs_dqrele_all_inodes(
892 struct xfs_mount *mp,
893 unsigned int qflags)
894{
895 struct xfs_icwalk icw = { .icw_flags = 0 };
896
897 if (qflags & XFS_UQUOTA_ACCT)
898 icw.icw_flags |= XFS_ICWALK_FLAG_DROP_UDQUOT;
899 if (qflags & XFS_GQUOTA_ACCT)
900 icw.icw_flags |= XFS_ICWALK_FLAG_DROP_GDQUOT;
901 if (qflags & XFS_PQUOTA_ACCT)
902 icw.icw_flags |= XFS_ICWALK_FLAG_DROP_PDQUOT;
903
904 return xfs_icwalk(mp, XFS_ICWALK_DQRELE, &icw);
905}
906#else
907# define xfs_dqrele_igrab(ip) (false)
908# define xfs_dqrele_inode(ip, priv) ((void)0)
909#endif /* CONFIG_XFS_QUOTA */
910
911/*
912 * Grab the inode for reclaim exclusively.
913 *
914 * We have found this inode via a lookup under RCU, so the inode may have
915 * already been freed, or it may be in the process of being recycled by
916 * xfs_iget(). In both cases, the inode will have XFS_IRECLAIM set. If the inode
917 * has been fully recycled by the time we get the i_flags_lock, XFS_IRECLAIMABLE
918 * will not be set. Hence we need to check for both these flag conditions to
919 * avoid inodes that are no longer reclaim candidates.
920 *
921 * Note: checking for other state flags here, under the i_flags_lock or not, is
922 * racy and should be avoided. Those races should be resolved only after we have
923 * ensured that we are able to reclaim this inode and the world can see that we
924 * are going to reclaim it.
925 *
926 * Return true if we grabbed it, false otherwise.
927 */
928static bool
929xfs_reclaim_igrab(
930 struct xfs_inode *ip,
931 struct xfs_icwalk *icw)
932{
933 ASSERT(rcu_read_lock_held());
934
935 spin_lock(&ip->i_flags_lock);
936 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
937 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
938 /* not a reclaim candidate. */
939 spin_unlock(&ip->i_flags_lock);
940 return false;
941 }
942
943 /* Don't reclaim a sick inode unless the caller asked for it. */
944 if (ip->i_sick &&
945 (!icw || !(icw->icw_flags & XFS_ICWALK_FLAG_RECLAIM_SICK))) {
946 spin_unlock(&ip->i_flags_lock);
947 return false;
948 }
949
950 __xfs_iflags_set(ip, XFS_IRECLAIM);
951 spin_unlock(&ip->i_flags_lock);
952 return true;
953}
954
955/*
956 * Inode reclaim is non-blocking, so the default action if progress cannot be
957 * made is to "requeue" the inode for reclaim by unlocking it and clearing the
958 * XFS_IRECLAIM flag. If we are in a shutdown state, we don't care about
959 * blocking anymore and hence we can wait for the inode to be able to reclaim
960 * it.
961 *
962 * We do no IO here - if callers require inodes to be cleaned they must push the
963 * AIL first to trigger writeback of dirty inodes. This enables writeback to be
964 * done in the background in a non-blocking manner, and enables memory reclaim
965 * to make progress without blocking.
966 */
967static void
968xfs_reclaim_inode(
969 struct xfs_inode *ip,
970 struct xfs_perag *pag)
971{
972 xfs_ino_t ino = ip->i_ino; /* for radix_tree_delete */
973
974 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
975 goto out;
976 if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
977 goto out_iunlock;
978
979 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
980 xfs_iunpin_wait(ip);
981 xfs_iflush_abort(ip);
982 goto reclaim;
983 }
984 if (xfs_ipincount(ip))
985 goto out_clear_flush;
986 if (!xfs_inode_clean(ip))
987 goto out_clear_flush;
988
989 xfs_iflags_clear(ip, XFS_IFLUSHING);
990reclaim:
991
992 /*
993 * Because we use RCU freeing we need to ensure the inode always appears
994 * to be reclaimed with an invalid inode number when in the free state.
995 * We do this as early as possible under the ILOCK so that
996 * xfs_iflush_cluster() and xfs_ifree_cluster() can be guaranteed to
997 * detect races with us here. By doing this, we guarantee that once
998 * xfs_iflush_cluster() or xfs_ifree_cluster() has locked XFS_ILOCK that
999 * it will see either a valid inode that will serialise correctly, or it
1000 * will see an invalid inode that it can skip.
1001 */
1002 spin_lock(&ip->i_flags_lock);
1003 ip->i_flags = XFS_IRECLAIM;
1004 ip->i_ino = 0;
1005 ip->i_sick = 0;
1006 ip->i_checked = 0;
1007 spin_unlock(&ip->i_flags_lock);
1008
1009 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1010
1011 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1012 /*
1013 * Remove the inode from the per-AG radix tree.
1014 *
1015 * Because radix_tree_delete won't complain even if the item was never
1016 * added to the tree assert that it's been there before to catch
1017 * problems with the inode life time early on.
1018 */
1019 spin_lock(&pag->pag_ici_lock);
1020 if (!radix_tree_delete(&pag->pag_ici_root,
1021 XFS_INO_TO_AGINO(ip->i_mount, ino)))
1022 ASSERT(0);
1023 xfs_perag_clear_inode_tag(pag, NULLAGINO, XFS_ICI_RECLAIM_TAG);
1024 spin_unlock(&pag->pag_ici_lock);
1025
1026 /*
1027 * Here we do an (almost) spurious inode lock in order to coordinate
1028 * with inode cache radix tree lookups. This is because the lookup
1029 * can reference the inodes in the cache without taking references.
1030 *
1031 * We make that OK here by ensuring that we wait until the inode is
1032 * unlocked after the lookup before we go ahead and free it.
1033 */
1034 xfs_ilock(ip, XFS_ILOCK_EXCL);
1035 ASSERT(!ip->i_udquot && !ip->i_gdquot && !ip->i_pdquot);
1036 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1037 ASSERT(xfs_inode_clean(ip));
1038
1039 __xfs_inode_free(ip);
1040 return;
1041
1042out_clear_flush:
1043 xfs_iflags_clear(ip, XFS_IFLUSHING);
1044out_iunlock:
1045 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1046out:
1047 xfs_iflags_clear(ip, XFS_IRECLAIM);
1048}
1049
1050/* Reclaim sick inodes if we're unmounting or the fs went down. */
1051static inline bool
1052xfs_want_reclaim_sick(
1053 struct xfs_mount *mp)
1054{
1055 return (mp->m_flags & XFS_MOUNT_UNMOUNTING) ||
1056 (mp->m_flags & XFS_MOUNT_NORECOVERY) ||
1057 XFS_FORCED_SHUTDOWN(mp);
1058}
1059
1060void
1061xfs_reclaim_inodes(
1062 struct xfs_mount *mp)
1063{
1064 struct xfs_icwalk icw = {
1065 .icw_flags = 0,
1066 };
1067
1068 if (xfs_want_reclaim_sick(mp))
1069 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1070
1071 while (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
1072 xfs_ail_push_all_sync(mp->m_ail);
1073 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1074 }
1075}
1076
1077/*
1078 * The shrinker infrastructure determines how many inodes we should scan for
1079 * reclaim. We want as many clean inodes ready to reclaim as possible, so we
1080 * push the AIL here. We also want to proactively free up memory if we can to
1081 * minimise the amount of work memory reclaim has to do so we kick the
1082 * background reclaim if it isn't already scheduled.
1083 */
1084long
1085xfs_reclaim_inodes_nr(
1086 struct xfs_mount *mp,
1087 unsigned long nr_to_scan)
1088{
1089 struct xfs_icwalk icw = {
1090 .icw_flags = XFS_ICWALK_FLAG_SCAN_LIMIT,
1091 .icw_scan_limit = min_t(unsigned long, LONG_MAX, nr_to_scan),
1092 };
1093
1094 if (xfs_want_reclaim_sick(mp))
1095 icw.icw_flags |= XFS_ICWALK_FLAG_RECLAIM_SICK;
1096
1097 /* kick background reclaimer and push the AIL */
1098 xfs_reclaim_work_queue(mp);
1099 xfs_ail_push_all(mp->m_ail);
1100
1101 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, &icw);
1102 return 0;
1103}
1104
1105/*
1106 * Return the number of reclaimable inodes in the filesystem for
1107 * the shrinker to determine how much to reclaim.
1108 */
1109long
1110xfs_reclaim_inodes_count(
1111 struct xfs_mount *mp)
1112{
1113 struct xfs_perag *pag;
1114 xfs_agnumber_t ag = 0;
1115 long reclaimable = 0;
1116
1117 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1118 ag = pag->pag_agno + 1;
1119 reclaimable += pag->pag_ici_reclaimable;
1120 xfs_perag_put(pag);
1121 }
1122 return reclaimable;
1123}
1124
1125STATIC bool
1126xfs_icwalk_match_id(
1127 struct xfs_inode *ip,
1128 struct xfs_icwalk *icw)
1129{
1130 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1131 !uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1132 return false;
1133
1134 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1135 !gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1136 return false;
1137
1138 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1139 ip->i_projid != icw->icw_prid)
1140 return false;
1141
1142 return true;
1143}
1144
1145/*
1146 * A union-based inode filtering algorithm. Process the inode if any of the
1147 * criteria match. This is for global/internal scans only.
1148 */
1149STATIC bool
1150xfs_icwalk_match_id_union(
1151 struct xfs_inode *ip,
1152 struct xfs_icwalk *icw)
1153{
1154 if ((icw->icw_flags & XFS_ICWALK_FLAG_UID) &&
1155 uid_eq(VFS_I(ip)->i_uid, icw->icw_uid))
1156 return true;
1157
1158 if ((icw->icw_flags & XFS_ICWALK_FLAG_GID) &&
1159 gid_eq(VFS_I(ip)->i_gid, icw->icw_gid))
1160 return true;
1161
1162 if ((icw->icw_flags & XFS_ICWALK_FLAG_PRID) &&
1163 ip->i_projid == icw->icw_prid)
1164 return true;
1165
1166 return false;
1167}
1168
1169/*
1170 * Is this inode @ip eligible for eof/cow block reclamation, given some
1171 * filtering parameters @icw? The inode is eligible if @icw is null or
1172 * if the predicate functions match.
1173 */
1174static bool
1175xfs_icwalk_match(
1176 struct xfs_inode *ip,
1177 struct xfs_icwalk *icw)
1178{
1179 bool match;
1180
1181 if (!icw)
1182 return true;
1183
1184 if (icw->icw_flags & XFS_ICWALK_FLAG_UNION)
1185 match = xfs_icwalk_match_id_union(ip, icw);
1186 else
1187 match = xfs_icwalk_match_id(ip, icw);
1188 if (!match)
1189 return false;
1190
1191 /* skip the inode if the file size is too small */
1192 if ((icw->icw_flags & XFS_ICWALK_FLAG_MINFILESIZE) &&
1193 XFS_ISIZE(ip) < icw->icw_min_file_size)
1194 return false;
1195
1196 return true;
1197}
1198
1199/*
1200 * This is a fast pass over the inode cache to try to get reclaim moving on as
1201 * many inodes as possible in a short period of time. It kicks itself every few
1202 * seconds, as well as being kicked by the inode cache shrinker when memory
1203 * goes low.
1204 */
1205void
1206xfs_reclaim_worker(
1207 struct work_struct *work)
1208{
1209 struct xfs_mount *mp = container_of(to_delayed_work(work),
1210 struct xfs_mount, m_reclaim_work);
1211
1212 xfs_icwalk(mp, XFS_ICWALK_RECLAIM, NULL);
1213 xfs_reclaim_work_queue(mp);
1214}
1215
1216STATIC int
1217xfs_inode_free_eofblocks(
1218 struct xfs_inode *ip,
1219 struct xfs_icwalk *icw,
1220 unsigned int *lockflags)
1221{
1222 bool wait;
1223
1224 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1225
1226 if (!xfs_iflags_test(ip, XFS_IEOFBLOCKS))
1227 return 0;
1228
1229 /*
1230 * If the mapping is dirty the operation can block and wait for some
1231 * time. Unless we are waiting, skip it.
1232 */
1233 if (!wait && mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1234 return 0;
1235
1236 if (!xfs_icwalk_match(ip, icw))
1237 return 0;
1238
1239 /*
1240 * If the caller is waiting, return -EAGAIN to keep the background
1241 * scanner moving and revisit the inode in a subsequent pass.
1242 */
1243 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1244 if (wait)
1245 return -EAGAIN;
1246 return 0;
1247 }
1248 *lockflags |= XFS_IOLOCK_EXCL;
1249
1250 if (xfs_can_free_eofblocks(ip, false))
1251 return xfs_free_eofblocks(ip);
1252
1253 /* inode could be preallocated or append-only */
1254 trace_xfs_inode_free_eofblocks_invalid(ip);
1255 xfs_inode_clear_eofblocks_tag(ip);
1256 return 0;
1257}
1258
1259static void
1260xfs_blockgc_set_iflag(
1261 struct xfs_inode *ip,
1262 unsigned long iflag)
1263{
1264 struct xfs_mount *mp = ip->i_mount;
1265 struct xfs_perag *pag;
1266
1267 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1268
1269 /*
1270 * Don't bother locking the AG and looking up in the radix trees
1271 * if we already know that we have the tag set.
1272 */
1273 if (ip->i_flags & iflag)
1274 return;
1275 spin_lock(&ip->i_flags_lock);
1276 ip->i_flags |= iflag;
1277 spin_unlock(&ip->i_flags_lock);
1278
1279 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1280 spin_lock(&pag->pag_ici_lock);
1281
1282 xfs_perag_set_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1283 XFS_ICI_BLOCKGC_TAG);
1284
1285 spin_unlock(&pag->pag_ici_lock);
1286 xfs_perag_put(pag);
1287}
1288
1289void
1290xfs_inode_set_eofblocks_tag(
1291 xfs_inode_t *ip)
1292{
1293 trace_xfs_inode_set_eofblocks_tag(ip);
1294 return xfs_blockgc_set_iflag(ip, XFS_IEOFBLOCKS);
1295}
1296
1297static void
1298xfs_blockgc_clear_iflag(
1299 struct xfs_inode *ip,
1300 unsigned long iflag)
1301{
1302 struct xfs_mount *mp = ip->i_mount;
1303 struct xfs_perag *pag;
1304 bool clear_tag;
1305
1306 ASSERT((iflag & ~(XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0);
1307
1308 spin_lock(&ip->i_flags_lock);
1309 ip->i_flags &= ~iflag;
1310 clear_tag = (ip->i_flags & (XFS_IEOFBLOCKS | XFS_ICOWBLOCKS)) == 0;
1311 spin_unlock(&ip->i_flags_lock);
1312
1313 if (!clear_tag)
1314 return;
1315
1316 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1317 spin_lock(&pag->pag_ici_lock);
1318
1319 xfs_perag_clear_inode_tag(pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
1320 XFS_ICI_BLOCKGC_TAG);
1321
1322 spin_unlock(&pag->pag_ici_lock);
1323 xfs_perag_put(pag);
1324}
1325
1326void
1327xfs_inode_clear_eofblocks_tag(
1328 xfs_inode_t *ip)
1329{
1330 trace_xfs_inode_clear_eofblocks_tag(ip);
1331 return xfs_blockgc_clear_iflag(ip, XFS_IEOFBLOCKS);
1332}
1333
1334/*
1335 * Set ourselves up to free CoW blocks from this file. If it's already clean
1336 * then we can bail out quickly, but otherwise we must back off if the file
1337 * is undergoing some kind of write.
1338 */
1339static bool
1340xfs_prep_free_cowblocks(
1341 struct xfs_inode *ip)
1342{
1343 /*
1344 * Just clear the tag if we have an empty cow fork or none at all. It's
1345 * possible the inode was fully unshared since it was originally tagged.
1346 */
1347 if (!xfs_inode_has_cow_data(ip)) {
1348 trace_xfs_inode_free_cowblocks_invalid(ip);
1349 xfs_inode_clear_cowblocks_tag(ip);
1350 return false;
1351 }
1352
1353 /*
1354 * If the mapping is dirty or under writeback we cannot touch the
1355 * CoW fork. Leave it alone if we're in the midst of a directio.
1356 */
1357 if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
1358 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
1359 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
1360 atomic_read(&VFS_I(ip)->i_dio_count))
1361 return false;
1362
1363 return true;
1364}
1365
1366/*
1367 * Automatic CoW Reservation Freeing
1368 *
1369 * These functions automatically garbage collect leftover CoW reservations
1370 * that were made on behalf of a cowextsize hint when we start to run out
1371 * of quota or when the reservations sit around for too long. If the file
1372 * has dirty pages or is undergoing writeback, its CoW reservations will
1373 * be retained.
1374 *
1375 * The actual garbage collection piggybacks off the same code that runs
1376 * the speculative EOF preallocation garbage collector.
1377 */
1378STATIC int
1379xfs_inode_free_cowblocks(
1380 struct xfs_inode *ip,
1381 struct xfs_icwalk *icw,
1382 unsigned int *lockflags)
1383{
1384 bool wait;
1385 int ret = 0;
1386
1387 wait = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
1388
1389 if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
1390 return 0;
1391
1392 if (!xfs_prep_free_cowblocks(ip))
1393 return 0;
1394
1395 if (!xfs_icwalk_match(ip, icw))
1396 return 0;
1397
1398 /*
1399 * If the caller is waiting, return -EAGAIN to keep the background
1400 * scanner moving and revisit the inode in a subsequent pass.
1401 */
1402 if (!(*lockflags & XFS_IOLOCK_EXCL) &&
1403 !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1404 if (wait)
1405 return -EAGAIN;
1406 return 0;
1407 }
1408 *lockflags |= XFS_IOLOCK_EXCL;
1409
1410 if (!xfs_ilock_nowait(ip, XFS_MMAPLOCK_EXCL)) {
1411 if (wait)
1412 return -EAGAIN;
1413 return 0;
1414 }
1415 *lockflags |= XFS_MMAPLOCK_EXCL;
1416
1417 /*
1418 * Check again, nobody else should be able to dirty blocks or change
1419 * the reflink iflag now that we have the first two locks held.
1420 */
1421 if (xfs_prep_free_cowblocks(ip))
1422 ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
1423 return ret;
1424}
1425
1426void
1427xfs_inode_set_cowblocks_tag(
1428 xfs_inode_t *ip)
1429{
1430 trace_xfs_inode_set_cowblocks_tag(ip);
1431 return xfs_blockgc_set_iflag(ip, XFS_ICOWBLOCKS);
1432}
1433
1434void
1435xfs_inode_clear_cowblocks_tag(
1436 xfs_inode_t *ip)
1437{
1438 trace_xfs_inode_clear_cowblocks_tag(ip);
1439 return xfs_blockgc_clear_iflag(ip, XFS_ICOWBLOCKS);
1440}
1441
1442/* Disable post-EOF and CoW block auto-reclamation. */
1443void
1444xfs_blockgc_stop(
1445 struct xfs_mount *mp)
1446{
1447 struct xfs_perag *pag;
1448 xfs_agnumber_t agno;
1449
1450 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1451 cancel_delayed_work_sync(&pag->pag_blockgc_work);
1452}
1453
1454/* Enable post-EOF and CoW block auto-reclamation. */
1455void
1456xfs_blockgc_start(
1457 struct xfs_mount *mp)
1458{
1459 struct xfs_perag *pag;
1460 xfs_agnumber_t agno;
1461
1462 for_each_perag_tag(mp, agno, pag, XFS_ICI_BLOCKGC_TAG)
1463 xfs_blockgc_queue(pag);
1464}
1465
1466/* Don't try to run block gc on an inode that's in any of these states. */
1467#define XFS_BLOCKGC_NOGRAB_IFLAGS (XFS_INEW | \
1468 XFS_IRECLAIMABLE | \
1469 XFS_IRECLAIM)
1470/*
1471 * Decide if the given @ip is eligible for garbage collection of speculative
1472 * preallocations, and grab it if so. Returns true if it's ready to go or
1473 * false if we should just ignore it.
1474 */
1475static bool
1476xfs_blockgc_igrab(
1477 struct xfs_inode *ip)
1478{
1479 struct inode *inode = VFS_I(ip);
1480
1481 ASSERT(rcu_read_lock_held());
1482
1483 /* Check for stale RCU freed inode */
1484 spin_lock(&ip->i_flags_lock);
1485 if (!ip->i_ino)
1486 goto out_unlock_noent;
1487
1488 if (ip->i_flags & XFS_BLOCKGC_NOGRAB_IFLAGS)
1489 goto out_unlock_noent;
1490 spin_unlock(&ip->i_flags_lock);
1491
1492 /* nothing to sync during shutdown */
1493 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
1494 return false;
1495
1496 /* If we can't grab the inode, it must on it's way to reclaim. */
1497 if (!igrab(inode))
1498 return false;
1499
1500 /* inode is valid */
1501 return true;
1502
1503out_unlock_noent:
1504 spin_unlock(&ip->i_flags_lock);
1505 return false;
1506}
1507
1508/* Scan one incore inode for block preallocations that we can remove. */
1509static int
1510xfs_blockgc_scan_inode(
1511 struct xfs_inode *ip,
1512 struct xfs_icwalk *icw)
1513{
1514 unsigned int lockflags = 0;
1515 int error;
1516
1517 error = xfs_inode_free_eofblocks(ip, icw, &lockflags);
1518 if (error)
1519 goto unlock;
1520
1521 error = xfs_inode_free_cowblocks(ip, icw, &lockflags);
1522unlock:
1523 if (lockflags)
1524 xfs_iunlock(ip, lockflags);
1525 xfs_irele(ip);
1526 return error;
1527}
1528
1529/* Background worker that trims preallocated space. */
1530void
1531xfs_blockgc_worker(
1532 struct work_struct *work)
1533{
1534 struct xfs_perag *pag = container_of(to_delayed_work(work),
1535 struct xfs_perag, pag_blockgc_work);
1536 struct xfs_mount *mp = pag->pag_mount;
1537 int error;
1538
1539 if (!sb_start_write_trylock(mp->m_super))
1540 return;
1541 error = xfs_icwalk_ag(pag, XFS_ICWALK_BLOCKGC, NULL);
1542 if (error)
1543 xfs_info(mp, "AG %u preallocation gc worker failed, err=%d",
1544 pag->pag_agno, error);
1545 sb_end_write(mp->m_super);
1546 xfs_blockgc_queue(pag);
1547}
1548
1549/*
1550 * Try to free space in the filesystem by purging eofblocks and cowblocks.
1551 */
1552int
1553xfs_blockgc_free_space(
1554 struct xfs_mount *mp,
1555 struct xfs_icwalk *icw)
1556{
1557 trace_xfs_blockgc_free_space(mp, icw, _RET_IP_);
1558
1559 return xfs_icwalk(mp, XFS_ICWALK_BLOCKGC, icw);
1560}
1561
1562/*
1563 * Run cow/eofblocks scans on the supplied dquots. We don't know exactly which
1564 * quota caused an allocation failure, so we make a best effort by including
1565 * each quota under low free space conditions (less than 1% free space) in the
1566 * scan.
1567 *
1568 * Callers must not hold any inode's ILOCK. If requesting a synchronous scan
1569 * (XFS_ICWALK_FLAG_SYNC), the caller also must not hold any inode's IOLOCK or
1570 * MMAPLOCK.
1571 */
1572int
1573xfs_blockgc_free_dquots(
1574 struct xfs_mount *mp,
1575 struct xfs_dquot *udqp,
1576 struct xfs_dquot *gdqp,
1577 struct xfs_dquot *pdqp,
1578 unsigned int iwalk_flags)
1579{
1580 struct xfs_icwalk icw = {0};
1581 bool do_work = false;
1582
1583 if (!udqp && !gdqp && !pdqp)
1584 return 0;
1585
1586 /*
1587 * Run a scan to free blocks using the union filter to cover all
1588 * applicable quotas in a single scan.
1589 */
1590 icw.icw_flags = XFS_ICWALK_FLAG_UNION | iwalk_flags;
1591
1592 if (XFS_IS_UQUOTA_ENFORCED(mp) && udqp && xfs_dquot_lowsp(udqp)) {
1593 icw.icw_uid = make_kuid(mp->m_super->s_user_ns, udqp->q_id);
1594 icw.icw_flags |= XFS_ICWALK_FLAG_UID;
1595 do_work = true;
1596 }
1597
1598 if (XFS_IS_UQUOTA_ENFORCED(mp) && gdqp && xfs_dquot_lowsp(gdqp)) {
1599 icw.icw_gid = make_kgid(mp->m_super->s_user_ns, gdqp->q_id);
1600 icw.icw_flags |= XFS_ICWALK_FLAG_GID;
1601 do_work = true;
1602 }
1603
1604 if (XFS_IS_PQUOTA_ENFORCED(mp) && pdqp && xfs_dquot_lowsp(pdqp)) {
1605 icw.icw_prid = pdqp->q_id;
1606 icw.icw_flags |= XFS_ICWALK_FLAG_PRID;
1607 do_work = true;
1608 }
1609
1610 if (!do_work)
1611 return 0;
1612
1613 return xfs_blockgc_free_space(mp, &icw);
1614}
1615
1616/* Run cow/eofblocks scans on the quotas attached to the inode. */
1617int
1618xfs_blockgc_free_quota(
1619 struct xfs_inode *ip,
1620 unsigned int iwalk_flags)
1621{
1622 return xfs_blockgc_free_dquots(ip->i_mount,
1623 xfs_inode_dquot(ip, XFS_DQTYPE_USER),
1624 xfs_inode_dquot(ip, XFS_DQTYPE_GROUP),
1625 xfs_inode_dquot(ip, XFS_DQTYPE_PROJ), iwalk_flags);
1626}
1627
1628/* XFS Inode Cache Walking Code */
1629
1630/*
1631 * The inode lookup is done in batches to keep the amount of lock traffic and
1632 * radix tree lookups to a minimum. The batch size is a trade off between
1633 * lookup reduction and stack usage. This is in the reclaim path, so we can't
1634 * be too greedy.
1635 */
1636#define XFS_LOOKUP_BATCH 32
1637
1638
1639/*
1640 * Decide if we want to grab this inode in anticipation of doing work towards
1641 * the goal.
1642 */
1643static inline bool
1644xfs_icwalk_igrab(
1645 enum xfs_icwalk_goal goal,
1646 struct xfs_inode *ip,
1647 struct xfs_icwalk *icw)
1648{
1649 switch (goal) {
1650 case XFS_ICWALK_DQRELE:
1651 return xfs_dqrele_igrab(ip);
1652 case XFS_ICWALK_BLOCKGC:
1653 return xfs_blockgc_igrab(ip);
1654 case XFS_ICWALK_RECLAIM:
1655 return xfs_reclaim_igrab(ip, icw);
1656 default:
1657 return false;
1658 }
1659}
1660
1661/*
1662 * Process an inode. Each processing function must handle any state changes
1663 * made by the icwalk igrab function. Return -EAGAIN to skip an inode.
1664 */
1665static inline int
1666xfs_icwalk_process_inode(
1667 enum xfs_icwalk_goal goal,
1668 struct xfs_inode *ip,
1669 struct xfs_perag *pag,
1670 struct xfs_icwalk *icw)
1671{
1672 int error = 0;
1673
1674 switch (goal) {
1675 case XFS_ICWALK_DQRELE:
1676 xfs_dqrele_inode(ip, icw);
1677 break;
1678 case XFS_ICWALK_BLOCKGC:
1679 error = xfs_blockgc_scan_inode(ip, icw);
1680 break;
1681 case XFS_ICWALK_RECLAIM:
1682 xfs_reclaim_inode(ip, pag);
1683 break;
1684 }
1685 return error;
1686}
1687
1688/*
1689 * For a given per-AG structure @pag and a goal, grab qualifying inodes and
1690 * process them in some manner.
1691 */
1692static int
1693xfs_icwalk_ag(
1694 struct xfs_perag *pag,
1695 enum xfs_icwalk_goal goal,
1696 struct xfs_icwalk *icw)
1697{
1698 struct xfs_mount *mp = pag->pag_mount;
1699 uint32_t first_index;
1700 int last_error = 0;
1701 int skipped;
1702 bool done;
1703 int nr_found;
1704
1705restart:
1706 done = false;
1707 skipped = 0;
1708 if (goal == XFS_ICWALK_RECLAIM)
1709 first_index = READ_ONCE(pag->pag_ici_reclaim_cursor);
1710 else
1711 first_index = 0;
1712 nr_found = 0;
1713 do {
1714 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1715 unsigned int tag = xfs_icwalk_tag(goal);
1716 int error = 0;
1717 int i;
1718
1719 rcu_read_lock();
1720
1721 if (tag == XFS_ICWALK_NULL_TAG)
1722 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
1723 (void **)batch, first_index,
1724 XFS_LOOKUP_BATCH);
1725 else
1726 nr_found = radix_tree_gang_lookup_tag(
1727 &pag->pag_ici_root,
1728 (void **) batch, first_index,
1729 XFS_LOOKUP_BATCH, tag);
1730
1731 if (!nr_found) {
1732 done = true;
1733 rcu_read_unlock();
1734 break;
1735 }
1736
1737 /*
1738 * Grab the inodes before we drop the lock. if we found
1739 * nothing, nr == 0 and the loop will be skipped.
1740 */
1741 for (i = 0; i < nr_found; i++) {
1742 struct xfs_inode *ip = batch[i];
1743
1744 if (done || !xfs_icwalk_igrab(goal, ip, icw))
1745 batch[i] = NULL;
1746
1747 /*
1748 * Update the index for the next lookup. Catch
1749 * overflows into the next AG range which can occur if
1750 * we have inodes in the last block of the AG and we
1751 * are currently pointing to the last inode.
1752 *
1753 * Because we may see inodes that are from the wrong AG
1754 * due to RCU freeing and reallocation, only update the
1755 * index if it lies in this AG. It was a race that lead
1756 * us to see this inode, so another lookup from the
1757 * same index will not find it again.
1758 */
1759 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
1760 continue;
1761 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1762 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1763 done = true;
1764 }
1765
1766 /* unlock now we've grabbed the inodes. */
1767 rcu_read_unlock();
1768
1769 for (i = 0; i < nr_found; i++) {
1770 if (!batch[i])
1771 continue;
1772 error = xfs_icwalk_process_inode(goal, batch[i], pag,
1773 icw);
1774 if (error == -EAGAIN) {
1775 skipped++;
1776 continue;
1777 }
1778 if (error && last_error != -EFSCORRUPTED)
1779 last_error = error;
1780 }
1781
1782 /* bail out if the filesystem is corrupted. */
1783 if (error == -EFSCORRUPTED)
1784 break;
1785
1786 cond_resched();
1787
1788 if (icw && (icw->icw_flags & XFS_ICWALK_FLAG_SCAN_LIMIT)) {
1789 icw->icw_scan_limit -= XFS_LOOKUP_BATCH;
1790 if (icw->icw_scan_limit <= 0)
1791 break;
1792 }
1793 } while (nr_found && !done);
1794
1795 if (goal == XFS_ICWALK_RECLAIM) {
1796 if (done)
1797 first_index = 0;
1798 WRITE_ONCE(pag->pag_ici_reclaim_cursor, first_index);
1799 }
1800
1801 if (skipped) {
1802 delay(1);
1803 goto restart;
1804 }
1805 return last_error;
1806}
1807
1808/* Fetch the next (possibly tagged) per-AG structure. */
1809static inline struct xfs_perag *
1810xfs_icwalk_get_perag(
1811 struct xfs_mount *mp,
1812 xfs_agnumber_t agno,
1813 enum xfs_icwalk_goal goal)
1814{
1815 unsigned int tag = xfs_icwalk_tag(goal);
1816
1817 if (tag == XFS_ICWALK_NULL_TAG)
1818 return xfs_perag_get(mp, agno);
1819 return xfs_perag_get_tag(mp, agno, tag);
1820}
1821
1822/* Walk all incore inodes to achieve a given goal. */
1823static int
1824xfs_icwalk(
1825 struct xfs_mount *mp,
1826 enum xfs_icwalk_goal goal,
1827 struct xfs_icwalk *icw)
1828{
1829 struct xfs_perag *pag;
1830 int error = 0;
1831 int last_error = 0;
1832 xfs_agnumber_t agno = 0;
1833
1834 while ((pag = xfs_icwalk_get_perag(mp, agno, goal))) {
1835 agno = pag->pag_agno + 1;
1836 error = xfs_icwalk_ag(pag, goal, icw);
1837 xfs_perag_put(pag);
1838 if (error) {
1839 last_error = error;
1840 if (error == -EFSCORRUPTED)
1841 break;
1842 }
1843 }
1844 return last_error;
1845 BUILD_BUG_ON(XFS_ICWALK_PRIVATE_FLAGS & XFS_ICWALK_FLAGS_VALID);
1846}
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
37#include <linux/kthread.h>
38#include <linux/freezer.h>
39
40STATIC void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp,
41 struct xfs_perag *pag, struct xfs_inode *ip);
42
43/*
44 * Allocate and initialise an xfs_inode.
45 */
46struct xfs_inode *
47xfs_inode_alloc(
48 struct xfs_mount *mp,
49 xfs_ino_t ino)
50{
51 struct xfs_inode *ip;
52
53 /*
54 * if this didn't occur in transactions, we could use
55 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
56 * code up to do this anyway.
57 */
58 ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
59 if (!ip)
60 return NULL;
61 if (inode_init_always(mp->m_super, VFS_I(ip))) {
62 kmem_zone_free(xfs_inode_zone, ip);
63 return NULL;
64 }
65
66 /* VFS doesn't initialise i_mode! */
67 VFS_I(ip)->i_mode = 0;
68
69 XFS_STATS_INC(mp, vn_active);
70 ASSERT(atomic_read(&ip->i_pincount) == 0);
71 ASSERT(!spin_is_locked(&ip->i_flags_lock));
72 ASSERT(!xfs_isiflocked(ip));
73 ASSERT(ip->i_ino == 0);
74
75 mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
76
77 /* initialise the xfs inode */
78 ip->i_ino = ino;
79 ip->i_mount = mp;
80 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
81 ip->i_afp = NULL;
82 memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
83 ip->i_flags = 0;
84 ip->i_delayed_blks = 0;
85 memset(&ip->i_d, 0, sizeof(ip->i_d));
86
87 return ip;
88}
89
90STATIC void
91xfs_inode_free_callback(
92 struct rcu_head *head)
93{
94 struct inode *inode = container_of(head, struct inode, i_rcu);
95 struct xfs_inode *ip = XFS_I(inode);
96
97 kmem_zone_free(xfs_inode_zone, ip);
98}
99
100void
101xfs_inode_free(
102 struct xfs_inode *ip)
103{
104 switch (VFS_I(ip)->i_mode & S_IFMT) {
105 case S_IFREG:
106 case S_IFDIR:
107 case S_IFLNK:
108 xfs_idestroy_fork(ip, XFS_DATA_FORK);
109 break;
110 }
111
112 if (ip->i_afp)
113 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
114
115 if (ip->i_itemp) {
116 ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
117 xfs_inode_item_destroy(ip);
118 ip->i_itemp = NULL;
119 }
120
121 /*
122 * Because we use RCU freeing we need to ensure the inode always
123 * appears to be reclaimed with an invalid inode number when in the
124 * free state. The ip->i_flags_lock provides the barrier against lookup
125 * races.
126 */
127 spin_lock(&ip->i_flags_lock);
128 ip->i_flags = XFS_IRECLAIM;
129 ip->i_ino = 0;
130 spin_unlock(&ip->i_flags_lock);
131
132 /* asserts to verify all state is correct here */
133 ASSERT(atomic_read(&ip->i_pincount) == 0);
134 ASSERT(!xfs_isiflocked(ip));
135 XFS_STATS_DEC(ip->i_mount, vn_active);
136
137 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
138}
139
140/*
141 * When we recycle a reclaimable inode, we need to re-initialise the VFS inode
142 * part of the structure. This is made more complex by the fact we store
143 * information about the on-disk values in the VFS inode and so we can't just
144 * overwrite the values unconditionally. Hence we save the parameters we
145 * need to retain across reinitialisation, and rewrite them into the VFS inode
146 * after reinitialisation even if it fails.
147 */
148static int
149xfs_reinit_inode(
150 struct xfs_mount *mp,
151 struct inode *inode)
152{
153 int error;
154 uint32_t nlink = inode->i_nlink;
155 uint32_t generation = inode->i_generation;
156 uint64_t version = inode->i_version;
157 umode_t mode = inode->i_mode;
158
159 error = inode_init_always(mp->m_super, inode);
160
161 set_nlink(inode, nlink);
162 inode->i_generation = generation;
163 inode->i_version = version;
164 inode->i_mode = mode;
165 return error;
166}
167
168/*
169 * Check the validity of the inode we just found it the cache
170 */
171static int
172xfs_iget_cache_hit(
173 struct xfs_perag *pag,
174 struct xfs_inode *ip,
175 xfs_ino_t ino,
176 int flags,
177 int lock_flags) __releases(RCU)
178{
179 struct inode *inode = VFS_I(ip);
180 struct xfs_mount *mp = ip->i_mount;
181 int error;
182
183 /*
184 * check for re-use of an inode within an RCU grace period due to the
185 * radix tree nodes not being updated yet. We monitor for this by
186 * setting the inode number to zero before freeing the inode structure.
187 * If the inode has been reallocated and set up, then the inode number
188 * will not match, so check for that, too.
189 */
190 spin_lock(&ip->i_flags_lock);
191 if (ip->i_ino != ino) {
192 trace_xfs_iget_skip(ip);
193 XFS_STATS_INC(mp, xs_ig_frecycle);
194 error = -EAGAIN;
195 goto out_error;
196 }
197
198
199 /*
200 * If we are racing with another cache hit that is currently
201 * instantiating this inode or currently recycling it out of
202 * reclaimabe state, wait for the initialisation to complete
203 * before continuing.
204 *
205 * XXX(hch): eventually we should do something equivalent to
206 * wait_on_inode to wait for these flags to be cleared
207 * instead of polling for it.
208 */
209 if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
210 trace_xfs_iget_skip(ip);
211 XFS_STATS_INC(mp, xs_ig_frecycle);
212 error = -EAGAIN;
213 goto out_error;
214 }
215
216 /*
217 * If lookup is racing with unlink return an error immediately.
218 */
219 if (VFS_I(ip)->i_mode == 0 && !(flags & XFS_IGET_CREATE)) {
220 error = -ENOENT;
221 goto out_error;
222 }
223
224 /*
225 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
226 * Need to carefully get it back into useable state.
227 */
228 if (ip->i_flags & XFS_IRECLAIMABLE) {
229 trace_xfs_iget_reclaim(ip);
230
231 /*
232 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
233 * from stomping over us while we recycle the inode. We can't
234 * clear the radix tree reclaimable tag yet as it requires
235 * pag_ici_lock to be held exclusive.
236 */
237 ip->i_flags |= XFS_IRECLAIM;
238
239 spin_unlock(&ip->i_flags_lock);
240 rcu_read_unlock();
241
242 error = xfs_reinit_inode(mp, inode);
243 if (error) {
244 /*
245 * Re-initializing the inode failed, and we are in deep
246 * trouble. Try to re-add it to the reclaim list.
247 */
248 rcu_read_lock();
249 spin_lock(&ip->i_flags_lock);
250
251 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
252 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
253 trace_xfs_iget_reclaim_fail(ip);
254 goto out_error;
255 }
256
257 spin_lock(&pag->pag_ici_lock);
258 spin_lock(&ip->i_flags_lock);
259
260 /*
261 * Clear the per-lifetime state in the inode as we are now
262 * effectively a new inode and need to return to the initial
263 * state before reuse occurs.
264 */
265 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
266 ip->i_flags |= XFS_INEW;
267 __xfs_inode_clear_reclaim_tag(mp, pag, ip);
268 inode->i_state = I_NEW;
269
270 ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock));
271 mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
272
273 spin_unlock(&ip->i_flags_lock);
274 spin_unlock(&pag->pag_ici_lock);
275 } else {
276 /* If the VFS inode is being torn down, pause and try again. */
277 if (!igrab(inode)) {
278 trace_xfs_iget_skip(ip);
279 error = -EAGAIN;
280 goto out_error;
281 }
282
283 /* We've got a live one. */
284 spin_unlock(&ip->i_flags_lock);
285 rcu_read_unlock();
286 trace_xfs_iget_hit(ip);
287 }
288
289 if (lock_flags != 0)
290 xfs_ilock(ip, lock_flags);
291
292 xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
293 XFS_STATS_INC(mp, xs_ig_found);
294
295 return 0;
296
297out_error:
298 spin_unlock(&ip->i_flags_lock);
299 rcu_read_unlock();
300 return error;
301}
302
303
304static int
305xfs_iget_cache_miss(
306 struct xfs_mount *mp,
307 struct xfs_perag *pag,
308 xfs_trans_t *tp,
309 xfs_ino_t ino,
310 struct xfs_inode **ipp,
311 int flags,
312 int lock_flags)
313{
314 struct xfs_inode *ip;
315 int error;
316 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
317 int iflags;
318
319 ip = xfs_inode_alloc(mp, ino);
320 if (!ip)
321 return -ENOMEM;
322
323 error = xfs_iread(mp, tp, ip, flags);
324 if (error)
325 goto out_destroy;
326
327 trace_xfs_iget_miss(ip);
328
329 if ((VFS_I(ip)->i_mode == 0) && !(flags & XFS_IGET_CREATE)) {
330 error = -ENOENT;
331 goto out_destroy;
332 }
333
334 /*
335 * Preload the radix tree so we can insert safely under the
336 * write spinlock. Note that we cannot sleep inside the preload
337 * region. Since we can be called from transaction context, don't
338 * recurse into the file system.
339 */
340 if (radix_tree_preload(GFP_NOFS)) {
341 error = -EAGAIN;
342 goto out_destroy;
343 }
344
345 /*
346 * Because the inode hasn't been added to the radix-tree yet it can't
347 * be found by another thread, so we can do the non-sleeping lock here.
348 */
349 if (lock_flags) {
350 if (!xfs_ilock_nowait(ip, lock_flags))
351 BUG();
352 }
353
354 /*
355 * These values must be set before inserting the inode into the radix
356 * tree as the moment it is inserted a concurrent lookup (allowed by the
357 * RCU locking mechanism) can find it and that lookup must see that this
358 * is an inode currently under construction (i.e. that XFS_INEW is set).
359 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
360 * memory barrier that ensures this detection works correctly at lookup
361 * time.
362 */
363 iflags = XFS_INEW;
364 if (flags & XFS_IGET_DONTCACHE)
365 iflags |= XFS_IDONTCACHE;
366 ip->i_udquot = NULL;
367 ip->i_gdquot = NULL;
368 ip->i_pdquot = NULL;
369 xfs_iflags_set(ip, iflags);
370
371 /* insert the new inode */
372 spin_lock(&pag->pag_ici_lock);
373 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
374 if (unlikely(error)) {
375 WARN_ON(error != -EEXIST);
376 XFS_STATS_INC(mp, xs_ig_dup);
377 error = -EAGAIN;
378 goto out_preload_end;
379 }
380 spin_unlock(&pag->pag_ici_lock);
381 radix_tree_preload_end();
382
383 *ipp = ip;
384 return 0;
385
386out_preload_end:
387 spin_unlock(&pag->pag_ici_lock);
388 radix_tree_preload_end();
389 if (lock_flags)
390 xfs_iunlock(ip, lock_flags);
391out_destroy:
392 __destroy_inode(VFS_I(ip));
393 xfs_inode_free(ip);
394 return error;
395}
396
397/*
398 * Look up an inode by number in the given file system.
399 * The inode is looked up in the cache held in each AG.
400 * If the inode is found in the cache, initialise the vfs inode
401 * if necessary.
402 *
403 * If it is not in core, read it in from the file system's device,
404 * add it to the cache and initialise the vfs inode.
405 *
406 * The inode is locked according to the value of the lock_flags parameter.
407 * This flag parameter indicates how and if the inode's IO lock and inode lock
408 * should be taken.
409 *
410 * mp -- the mount point structure for the current file system. It points
411 * to the inode hash table.
412 * tp -- a pointer to the current transaction if there is one. This is
413 * simply passed through to the xfs_iread() call.
414 * ino -- the number of the inode desired. This is the unique identifier
415 * within the file system for the inode being requested.
416 * lock_flags -- flags indicating how to lock the inode. See the comment
417 * for xfs_ilock() for a list of valid values.
418 */
419int
420xfs_iget(
421 xfs_mount_t *mp,
422 xfs_trans_t *tp,
423 xfs_ino_t ino,
424 uint flags,
425 uint lock_flags,
426 xfs_inode_t **ipp)
427{
428 xfs_inode_t *ip;
429 int error;
430 xfs_perag_t *pag;
431 xfs_agino_t agino;
432
433 /*
434 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
435 * doesn't get freed while it's being referenced during a
436 * radix tree traversal here. It assumes this function
437 * aqcuires only the ILOCK (and therefore it has no need to
438 * involve the IOLOCK in this synchronization).
439 */
440 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
441
442 /* reject inode numbers outside existing AGs */
443 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
444 return -EINVAL;
445
446 XFS_STATS_INC(mp, xs_ig_attempts);
447
448 /* get the perag structure and ensure that it's inode capable */
449 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
450 agino = XFS_INO_TO_AGINO(mp, ino);
451
452again:
453 error = 0;
454 rcu_read_lock();
455 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
456
457 if (ip) {
458 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
459 if (error)
460 goto out_error_or_again;
461 } else {
462 rcu_read_unlock();
463 XFS_STATS_INC(mp, xs_ig_missed);
464
465 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
466 flags, lock_flags);
467 if (error)
468 goto out_error_or_again;
469 }
470 xfs_perag_put(pag);
471
472 *ipp = ip;
473
474 /*
475 * If we have a real type for an on-disk inode, we can setup the inode
476 * now. If it's a new inode being created, xfs_ialloc will handle it.
477 */
478 if (xfs_iflags_test(ip, XFS_INEW) && VFS_I(ip)->i_mode != 0)
479 xfs_setup_existing_inode(ip);
480 return 0;
481
482out_error_or_again:
483 if (error == -EAGAIN) {
484 delay(1);
485 goto again;
486 }
487 xfs_perag_put(pag);
488 return error;
489}
490
491/*
492 * The inode lookup is done in batches to keep the amount of lock traffic and
493 * radix tree lookups to a minimum. The batch size is a trade off between
494 * lookup reduction and stack usage. This is in the reclaim path, so we can't
495 * be too greedy.
496 */
497#define XFS_LOOKUP_BATCH 32
498
499STATIC int
500xfs_inode_ag_walk_grab(
501 struct xfs_inode *ip)
502{
503 struct inode *inode = VFS_I(ip);
504
505 ASSERT(rcu_read_lock_held());
506
507 /*
508 * check for stale RCU freed inode
509 *
510 * If the inode has been reallocated, it doesn't matter if it's not in
511 * the AG we are walking - we are walking for writeback, so if it
512 * passes all the "valid inode" checks and is dirty, then we'll write
513 * it back anyway. If it has been reallocated and still being
514 * initialised, the XFS_INEW check below will catch it.
515 */
516 spin_lock(&ip->i_flags_lock);
517 if (!ip->i_ino)
518 goto out_unlock_noent;
519
520 /* avoid new or reclaimable inodes. Leave for reclaim code to flush */
521 if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM))
522 goto out_unlock_noent;
523 spin_unlock(&ip->i_flags_lock);
524
525 /* nothing to sync during shutdown */
526 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
527 return -EFSCORRUPTED;
528
529 /* If we can't grab the inode, it must on it's way to reclaim. */
530 if (!igrab(inode))
531 return -ENOENT;
532
533 /* inode is valid */
534 return 0;
535
536out_unlock_noent:
537 spin_unlock(&ip->i_flags_lock);
538 return -ENOENT;
539}
540
541STATIC int
542xfs_inode_ag_walk(
543 struct xfs_mount *mp,
544 struct xfs_perag *pag,
545 int (*execute)(struct xfs_inode *ip, int flags,
546 void *args),
547 int flags,
548 void *args,
549 int tag)
550{
551 uint32_t first_index;
552 int last_error = 0;
553 int skipped;
554 int done;
555 int nr_found;
556
557restart:
558 done = 0;
559 skipped = 0;
560 first_index = 0;
561 nr_found = 0;
562 do {
563 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
564 int error = 0;
565 int i;
566
567 rcu_read_lock();
568
569 if (tag == -1)
570 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
571 (void **)batch, first_index,
572 XFS_LOOKUP_BATCH);
573 else
574 nr_found = radix_tree_gang_lookup_tag(
575 &pag->pag_ici_root,
576 (void **) batch, first_index,
577 XFS_LOOKUP_BATCH, tag);
578
579 if (!nr_found) {
580 rcu_read_unlock();
581 break;
582 }
583
584 /*
585 * Grab the inodes before we drop the lock. if we found
586 * nothing, nr == 0 and the loop will be skipped.
587 */
588 for (i = 0; i < nr_found; i++) {
589 struct xfs_inode *ip = batch[i];
590
591 if (done || xfs_inode_ag_walk_grab(ip))
592 batch[i] = NULL;
593
594 /*
595 * Update the index for the next lookup. Catch
596 * overflows into the next AG range which can occur if
597 * we have inodes in the last block of the AG and we
598 * are currently pointing to the last inode.
599 *
600 * Because we may see inodes that are from the wrong AG
601 * due to RCU freeing and reallocation, only update the
602 * index if it lies in this AG. It was a race that lead
603 * us to see this inode, so another lookup from the
604 * same index will not find it again.
605 */
606 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
607 continue;
608 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
609 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
610 done = 1;
611 }
612
613 /* unlock now we've grabbed the inodes. */
614 rcu_read_unlock();
615
616 for (i = 0; i < nr_found; i++) {
617 if (!batch[i])
618 continue;
619 error = execute(batch[i], flags, args);
620 IRELE(batch[i]);
621 if (error == -EAGAIN) {
622 skipped++;
623 continue;
624 }
625 if (error && last_error != -EFSCORRUPTED)
626 last_error = error;
627 }
628
629 /* bail out if the filesystem is corrupted. */
630 if (error == -EFSCORRUPTED)
631 break;
632
633 cond_resched();
634
635 } while (nr_found && !done);
636
637 if (skipped) {
638 delay(1);
639 goto restart;
640 }
641 return last_error;
642}
643
644/*
645 * Background scanning to trim post-EOF preallocated space. This is queued
646 * based on the 'speculative_prealloc_lifetime' tunable (5m by default).
647 */
648STATIC void
649xfs_queue_eofblocks(
650 struct xfs_mount *mp)
651{
652 rcu_read_lock();
653 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG))
654 queue_delayed_work(mp->m_eofblocks_workqueue,
655 &mp->m_eofblocks_work,
656 msecs_to_jiffies(xfs_eofb_secs * 1000));
657 rcu_read_unlock();
658}
659
660void
661xfs_eofblocks_worker(
662 struct work_struct *work)
663{
664 struct xfs_mount *mp = container_of(to_delayed_work(work),
665 struct xfs_mount, m_eofblocks_work);
666 xfs_icache_free_eofblocks(mp, NULL);
667 xfs_queue_eofblocks(mp);
668}
669
670int
671xfs_inode_ag_iterator(
672 struct xfs_mount *mp,
673 int (*execute)(struct xfs_inode *ip, int flags,
674 void *args),
675 int flags,
676 void *args)
677{
678 struct xfs_perag *pag;
679 int error = 0;
680 int last_error = 0;
681 xfs_agnumber_t ag;
682
683 ag = 0;
684 while ((pag = xfs_perag_get(mp, ag))) {
685 ag = pag->pag_agno + 1;
686 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1);
687 xfs_perag_put(pag);
688 if (error) {
689 last_error = error;
690 if (error == -EFSCORRUPTED)
691 break;
692 }
693 }
694 return last_error;
695}
696
697int
698xfs_inode_ag_iterator_tag(
699 struct xfs_mount *mp,
700 int (*execute)(struct xfs_inode *ip, int flags,
701 void *args),
702 int flags,
703 void *args,
704 int tag)
705{
706 struct xfs_perag *pag;
707 int error = 0;
708 int last_error = 0;
709 xfs_agnumber_t ag;
710
711 ag = 0;
712 while ((pag = xfs_perag_get_tag(mp, ag, tag))) {
713 ag = pag->pag_agno + 1;
714 error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag);
715 xfs_perag_put(pag);
716 if (error) {
717 last_error = error;
718 if (error == -EFSCORRUPTED)
719 break;
720 }
721 }
722 return last_error;
723}
724
725/*
726 * Queue a new inode reclaim pass if there are reclaimable inodes and there
727 * isn't a reclaim pass already in progress. By default it runs every 5s based
728 * on the xfs periodic sync default of 30s. Perhaps this should have it's own
729 * tunable, but that can be done if this method proves to be ineffective or too
730 * aggressive.
731 */
732static void
733xfs_reclaim_work_queue(
734 struct xfs_mount *mp)
735{
736
737 rcu_read_lock();
738 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
739 queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work,
740 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
741 }
742 rcu_read_unlock();
743}
744
745/*
746 * This is a fast pass over the inode cache to try to get reclaim moving on as
747 * many inodes as possible in a short period of time. It kicks itself every few
748 * seconds, as well as being kicked by the inode cache shrinker when memory
749 * goes low. It scans as quickly as possible avoiding locked inodes or those
750 * already being flushed, and once done schedules a future pass.
751 */
752void
753xfs_reclaim_worker(
754 struct work_struct *work)
755{
756 struct xfs_mount *mp = container_of(to_delayed_work(work),
757 struct xfs_mount, m_reclaim_work);
758
759 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
760 xfs_reclaim_work_queue(mp);
761}
762
763static void
764__xfs_inode_set_reclaim_tag(
765 struct xfs_perag *pag,
766 struct xfs_inode *ip)
767{
768 radix_tree_tag_set(&pag->pag_ici_root,
769 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
770 XFS_ICI_RECLAIM_TAG);
771
772 if (!pag->pag_ici_reclaimable) {
773 /* propagate the reclaim tag up into the perag radix tree */
774 spin_lock(&ip->i_mount->m_perag_lock);
775 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
776 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
777 XFS_ICI_RECLAIM_TAG);
778 spin_unlock(&ip->i_mount->m_perag_lock);
779
780 /* schedule periodic background inode reclaim */
781 xfs_reclaim_work_queue(ip->i_mount);
782
783 trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno,
784 -1, _RET_IP_);
785 }
786 pag->pag_ici_reclaimable++;
787}
788
789/*
790 * We set the inode flag atomically with the radix tree tag.
791 * Once we get tag lookups on the radix tree, this inode flag
792 * can go away.
793 */
794void
795xfs_inode_set_reclaim_tag(
796 xfs_inode_t *ip)
797{
798 struct xfs_mount *mp = ip->i_mount;
799 struct xfs_perag *pag;
800
801 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
802 spin_lock(&pag->pag_ici_lock);
803 spin_lock(&ip->i_flags_lock);
804 __xfs_inode_set_reclaim_tag(pag, ip);
805 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
806 spin_unlock(&ip->i_flags_lock);
807 spin_unlock(&pag->pag_ici_lock);
808 xfs_perag_put(pag);
809}
810
811STATIC void
812__xfs_inode_clear_reclaim(
813 xfs_perag_t *pag,
814 xfs_inode_t *ip)
815{
816 pag->pag_ici_reclaimable--;
817 if (!pag->pag_ici_reclaimable) {
818 /* clear the reclaim tag from the perag radix tree */
819 spin_lock(&ip->i_mount->m_perag_lock);
820 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
821 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
822 XFS_ICI_RECLAIM_TAG);
823 spin_unlock(&ip->i_mount->m_perag_lock);
824 trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno,
825 -1, _RET_IP_);
826 }
827}
828
829STATIC void
830__xfs_inode_clear_reclaim_tag(
831 xfs_mount_t *mp,
832 xfs_perag_t *pag,
833 xfs_inode_t *ip)
834{
835 radix_tree_tag_clear(&pag->pag_ici_root,
836 XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
837 __xfs_inode_clear_reclaim(pag, ip);
838}
839
840/*
841 * Grab the inode for reclaim exclusively.
842 * Return 0 if we grabbed it, non-zero otherwise.
843 */
844STATIC int
845xfs_reclaim_inode_grab(
846 struct xfs_inode *ip,
847 int flags)
848{
849 ASSERT(rcu_read_lock_held());
850
851 /* quick check for stale RCU freed inode */
852 if (!ip->i_ino)
853 return 1;
854
855 /*
856 * If we are asked for non-blocking operation, do unlocked checks to
857 * see if the inode already is being flushed or in reclaim to avoid
858 * lock traffic.
859 */
860 if ((flags & SYNC_TRYLOCK) &&
861 __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
862 return 1;
863
864 /*
865 * The radix tree lock here protects a thread in xfs_iget from racing
866 * with us starting reclaim on the inode. Once we have the
867 * XFS_IRECLAIM flag set it will not touch us.
868 *
869 * Due to RCU lookup, we may find inodes that have been freed and only
870 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
871 * aren't candidates for reclaim at all, so we must check the
872 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
873 */
874 spin_lock(&ip->i_flags_lock);
875 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
876 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
877 /* not a reclaim candidate. */
878 spin_unlock(&ip->i_flags_lock);
879 return 1;
880 }
881 __xfs_iflags_set(ip, XFS_IRECLAIM);
882 spin_unlock(&ip->i_flags_lock);
883 return 0;
884}
885
886/*
887 * Inodes in different states need to be treated differently. The following
888 * table lists the inode states and the reclaim actions necessary:
889 *
890 * inode state iflush ret required action
891 * --------------- ---------- ---------------
892 * bad - reclaim
893 * shutdown EIO unpin and reclaim
894 * clean, unpinned 0 reclaim
895 * stale, unpinned 0 reclaim
896 * clean, pinned(*) 0 requeue
897 * stale, pinned EAGAIN requeue
898 * dirty, async - requeue
899 * dirty, sync 0 reclaim
900 *
901 * (*) dgc: I don't think the clean, pinned state is possible but it gets
902 * handled anyway given the order of checks implemented.
903 *
904 * Also, because we get the flush lock first, we know that any inode that has
905 * been flushed delwri has had the flush completed by the time we check that
906 * the inode is clean.
907 *
908 * Note that because the inode is flushed delayed write by AIL pushing, the
909 * flush lock may already be held here and waiting on it can result in very
910 * long latencies. Hence for sync reclaims, where we wait on the flush lock,
911 * the caller should push the AIL first before trying to reclaim inodes to
912 * minimise the amount of time spent waiting. For background relaim, we only
913 * bother to reclaim clean inodes anyway.
914 *
915 * Hence the order of actions after gaining the locks should be:
916 * bad => reclaim
917 * shutdown => unpin and reclaim
918 * pinned, async => requeue
919 * pinned, sync => unpin
920 * stale => reclaim
921 * clean => reclaim
922 * dirty, async => requeue
923 * dirty, sync => flush, wait and reclaim
924 */
925STATIC int
926xfs_reclaim_inode(
927 struct xfs_inode *ip,
928 struct xfs_perag *pag,
929 int sync_mode)
930{
931 struct xfs_buf *bp = NULL;
932 int error;
933
934restart:
935 error = 0;
936 xfs_ilock(ip, XFS_ILOCK_EXCL);
937 if (!xfs_iflock_nowait(ip)) {
938 if (!(sync_mode & SYNC_WAIT))
939 goto out;
940 xfs_iflock(ip);
941 }
942
943 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
944 xfs_iunpin_wait(ip);
945 xfs_iflush_abort(ip, false);
946 goto reclaim;
947 }
948 if (xfs_ipincount(ip)) {
949 if (!(sync_mode & SYNC_WAIT))
950 goto out_ifunlock;
951 xfs_iunpin_wait(ip);
952 }
953 if (xfs_iflags_test(ip, XFS_ISTALE))
954 goto reclaim;
955 if (xfs_inode_clean(ip))
956 goto reclaim;
957
958 /*
959 * Never flush out dirty data during non-blocking reclaim, as it would
960 * just contend with AIL pushing trying to do the same job.
961 */
962 if (!(sync_mode & SYNC_WAIT))
963 goto out_ifunlock;
964
965 /*
966 * Now we have an inode that needs flushing.
967 *
968 * Note that xfs_iflush will never block on the inode buffer lock, as
969 * xfs_ifree_cluster() can lock the inode buffer before it locks the
970 * ip->i_lock, and we are doing the exact opposite here. As a result,
971 * doing a blocking xfs_imap_to_bp() to get the cluster buffer would
972 * result in an ABBA deadlock with xfs_ifree_cluster().
973 *
974 * As xfs_ifree_cluser() must gather all inodes that are active in the
975 * cache to mark them stale, if we hit this case we don't actually want
976 * to do IO here - we want the inode marked stale so we can simply
977 * reclaim it. Hence if we get an EAGAIN error here, just unlock the
978 * inode, back off and try again. Hopefully the next pass through will
979 * see the stale flag set on the inode.
980 */
981 error = xfs_iflush(ip, &bp);
982 if (error == -EAGAIN) {
983 xfs_iunlock(ip, XFS_ILOCK_EXCL);
984 /* backoff longer than in xfs_ifree_cluster */
985 delay(2);
986 goto restart;
987 }
988
989 if (!error) {
990 error = xfs_bwrite(bp);
991 xfs_buf_relse(bp);
992 }
993
994 xfs_iflock(ip);
995reclaim:
996 xfs_ifunlock(ip);
997 xfs_iunlock(ip, XFS_ILOCK_EXCL);
998
999 XFS_STATS_INC(ip->i_mount, xs_ig_reclaims);
1000 /*
1001 * Remove the inode from the per-AG radix tree.
1002 *
1003 * Because radix_tree_delete won't complain even if the item was never
1004 * added to the tree assert that it's been there before to catch
1005 * problems with the inode life time early on.
1006 */
1007 spin_lock(&pag->pag_ici_lock);
1008 if (!radix_tree_delete(&pag->pag_ici_root,
1009 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino)))
1010 ASSERT(0);
1011 __xfs_inode_clear_reclaim(pag, ip);
1012 spin_unlock(&pag->pag_ici_lock);
1013
1014 /*
1015 * Here we do an (almost) spurious inode lock in order to coordinate
1016 * with inode cache radix tree lookups. This is because the lookup
1017 * can reference the inodes in the cache without taking references.
1018 *
1019 * We make that OK here by ensuring that we wait until the inode is
1020 * unlocked after the lookup before we go ahead and free it.
1021 */
1022 xfs_ilock(ip, XFS_ILOCK_EXCL);
1023 xfs_qm_dqdetach(ip);
1024 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1025
1026 xfs_inode_free(ip);
1027 return error;
1028
1029out_ifunlock:
1030 xfs_ifunlock(ip);
1031out:
1032 xfs_iflags_clear(ip, XFS_IRECLAIM);
1033 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1034 /*
1035 * We could return -EAGAIN here to make reclaim rescan the inode tree in
1036 * a short while. However, this just burns CPU time scanning the tree
1037 * waiting for IO to complete and the reclaim work never goes back to
1038 * the idle state. Instead, return 0 to let the next scheduled
1039 * background reclaim attempt to reclaim the inode again.
1040 */
1041 return 0;
1042}
1043
1044/*
1045 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
1046 * corrupted, we still want to try to reclaim all the inodes. If we don't,
1047 * then a shut down during filesystem unmount reclaim walk leak all the
1048 * unreclaimed inodes.
1049 */
1050STATIC int
1051xfs_reclaim_inodes_ag(
1052 struct xfs_mount *mp,
1053 int flags,
1054 int *nr_to_scan)
1055{
1056 struct xfs_perag *pag;
1057 int error = 0;
1058 int last_error = 0;
1059 xfs_agnumber_t ag;
1060 int trylock = flags & SYNC_TRYLOCK;
1061 int skipped;
1062
1063restart:
1064 ag = 0;
1065 skipped = 0;
1066 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1067 unsigned long first_index = 0;
1068 int done = 0;
1069 int nr_found = 0;
1070
1071 ag = pag->pag_agno + 1;
1072
1073 if (trylock) {
1074 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
1075 skipped++;
1076 xfs_perag_put(pag);
1077 continue;
1078 }
1079 first_index = pag->pag_ici_reclaim_cursor;
1080 } else
1081 mutex_lock(&pag->pag_ici_reclaim_lock);
1082
1083 do {
1084 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
1085 int i;
1086
1087 rcu_read_lock();
1088 nr_found = radix_tree_gang_lookup_tag(
1089 &pag->pag_ici_root,
1090 (void **)batch, first_index,
1091 XFS_LOOKUP_BATCH,
1092 XFS_ICI_RECLAIM_TAG);
1093 if (!nr_found) {
1094 done = 1;
1095 rcu_read_unlock();
1096 break;
1097 }
1098
1099 /*
1100 * Grab the inodes before we drop the lock. if we found
1101 * nothing, nr == 0 and the loop will be skipped.
1102 */
1103 for (i = 0; i < nr_found; i++) {
1104 struct xfs_inode *ip = batch[i];
1105
1106 if (done || xfs_reclaim_inode_grab(ip, flags))
1107 batch[i] = NULL;
1108
1109 /*
1110 * Update the index for the next lookup. Catch
1111 * overflows into the next AG range which can
1112 * occur if we have inodes in the last block of
1113 * the AG and we are currently pointing to the
1114 * last inode.
1115 *
1116 * Because we may see inodes that are from the
1117 * wrong AG due to RCU freeing and
1118 * reallocation, only update the index if it
1119 * lies in this AG. It was a race that lead us
1120 * to see this inode, so another lookup from
1121 * the same index will not find it again.
1122 */
1123 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
1124 pag->pag_agno)
1125 continue;
1126 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
1127 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
1128 done = 1;
1129 }
1130
1131 /* unlock now we've grabbed the inodes. */
1132 rcu_read_unlock();
1133
1134 for (i = 0; i < nr_found; i++) {
1135 if (!batch[i])
1136 continue;
1137 error = xfs_reclaim_inode(batch[i], pag, flags);
1138 if (error && last_error != -EFSCORRUPTED)
1139 last_error = error;
1140 }
1141
1142 *nr_to_scan -= XFS_LOOKUP_BATCH;
1143
1144 cond_resched();
1145
1146 } while (nr_found && !done && *nr_to_scan > 0);
1147
1148 if (trylock && !done)
1149 pag->pag_ici_reclaim_cursor = first_index;
1150 else
1151 pag->pag_ici_reclaim_cursor = 0;
1152 mutex_unlock(&pag->pag_ici_reclaim_lock);
1153 xfs_perag_put(pag);
1154 }
1155
1156 /*
1157 * if we skipped any AG, and we still have scan count remaining, do
1158 * another pass this time using blocking reclaim semantics (i.e
1159 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1160 * ensure that when we get more reclaimers than AGs we block rather
1161 * than spin trying to execute reclaim.
1162 */
1163 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
1164 trylock = 0;
1165 goto restart;
1166 }
1167 return last_error;
1168}
1169
1170int
1171xfs_reclaim_inodes(
1172 xfs_mount_t *mp,
1173 int mode)
1174{
1175 int nr_to_scan = INT_MAX;
1176
1177 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
1178}
1179
1180/*
1181 * Scan a certain number of inodes for reclaim.
1182 *
1183 * When called we make sure that there is a background (fast) inode reclaim in
1184 * progress, while we will throttle the speed of reclaim via doing synchronous
1185 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1186 * them to be cleaned, which we hope will not be very long due to the
1187 * background walker having already kicked the IO off on those dirty inodes.
1188 */
1189long
1190xfs_reclaim_inodes_nr(
1191 struct xfs_mount *mp,
1192 int nr_to_scan)
1193{
1194 /* kick background reclaimer and push the AIL */
1195 xfs_reclaim_work_queue(mp);
1196 xfs_ail_push_all(mp->m_ail);
1197
1198 return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1199}
1200
1201/*
1202 * Return the number of reclaimable inodes in the filesystem for
1203 * the shrinker to determine how much to reclaim.
1204 */
1205int
1206xfs_reclaim_inodes_count(
1207 struct xfs_mount *mp)
1208{
1209 struct xfs_perag *pag;
1210 xfs_agnumber_t ag = 0;
1211 int reclaimable = 0;
1212
1213 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1214 ag = pag->pag_agno + 1;
1215 reclaimable += pag->pag_ici_reclaimable;
1216 xfs_perag_put(pag);
1217 }
1218 return reclaimable;
1219}
1220
1221STATIC int
1222xfs_inode_match_id(
1223 struct xfs_inode *ip,
1224 struct xfs_eofblocks *eofb)
1225{
1226 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1227 !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1228 return 0;
1229
1230 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1231 !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1232 return 0;
1233
1234 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1235 xfs_get_projid(ip) != eofb->eof_prid)
1236 return 0;
1237
1238 return 1;
1239}
1240
1241/*
1242 * A union-based inode filtering algorithm. Process the inode if any of the
1243 * criteria match. This is for global/internal scans only.
1244 */
1245STATIC int
1246xfs_inode_match_id_union(
1247 struct xfs_inode *ip,
1248 struct xfs_eofblocks *eofb)
1249{
1250 if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) &&
1251 uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid))
1252 return 1;
1253
1254 if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) &&
1255 gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid))
1256 return 1;
1257
1258 if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) &&
1259 xfs_get_projid(ip) == eofb->eof_prid)
1260 return 1;
1261
1262 return 0;
1263}
1264
1265STATIC int
1266xfs_inode_free_eofblocks(
1267 struct xfs_inode *ip,
1268 int flags,
1269 void *args)
1270{
1271 int ret;
1272 struct xfs_eofblocks *eofb = args;
1273 bool need_iolock = true;
1274 int match;
1275
1276 ASSERT(!eofb || (eofb && eofb->eof_scan_owner != 0));
1277
1278 if (!xfs_can_free_eofblocks(ip, false)) {
1279 /* inode could be preallocated or append-only */
1280 trace_xfs_inode_free_eofblocks_invalid(ip);
1281 xfs_inode_clear_eofblocks_tag(ip);
1282 return 0;
1283 }
1284
1285 /*
1286 * If the mapping is dirty the operation can block and wait for some
1287 * time. Unless we are waiting, skip it.
1288 */
1289 if (!(flags & SYNC_WAIT) &&
1290 mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY))
1291 return 0;
1292
1293 if (eofb) {
1294 if (eofb->eof_flags & XFS_EOF_FLAGS_UNION)
1295 match = xfs_inode_match_id_union(ip, eofb);
1296 else
1297 match = xfs_inode_match_id(ip, eofb);
1298 if (!match)
1299 return 0;
1300
1301 /* skip the inode if the file size is too small */
1302 if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE &&
1303 XFS_ISIZE(ip) < eofb->eof_min_file_size)
1304 return 0;
1305
1306 /*
1307 * A scan owner implies we already hold the iolock. Skip it in
1308 * xfs_free_eofblocks() to avoid deadlock. This also eliminates
1309 * the possibility of EAGAIN being returned.
1310 */
1311 if (eofb->eof_scan_owner == ip->i_ino)
1312 need_iolock = false;
1313 }
1314
1315 ret = xfs_free_eofblocks(ip->i_mount, ip, need_iolock);
1316
1317 /* don't revisit the inode if we're not waiting */
1318 if (ret == -EAGAIN && !(flags & SYNC_WAIT))
1319 ret = 0;
1320
1321 return ret;
1322}
1323
1324int
1325xfs_icache_free_eofblocks(
1326 struct xfs_mount *mp,
1327 struct xfs_eofblocks *eofb)
1328{
1329 int flags = SYNC_TRYLOCK;
1330
1331 if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
1332 flags = SYNC_WAIT;
1333
1334 return xfs_inode_ag_iterator_tag(mp, xfs_inode_free_eofblocks, flags,
1335 eofb, XFS_ICI_EOFBLOCKS_TAG);
1336}
1337
1338/*
1339 * Run eofblocks scans on the quotas applicable to the inode. For inodes with
1340 * multiple quotas, we don't know exactly which quota caused an allocation
1341 * failure. We make a best effort by including each quota under low free space
1342 * conditions (less than 1% free space) in the scan.
1343 */
1344int
1345xfs_inode_free_quota_eofblocks(
1346 struct xfs_inode *ip)
1347{
1348 int scan = 0;
1349 struct xfs_eofblocks eofb = {0};
1350 struct xfs_dquot *dq;
1351
1352 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1353
1354 /*
1355 * Set the scan owner to avoid a potential livelock. Otherwise, the scan
1356 * can repeatedly trylock on the inode we're currently processing. We
1357 * run a sync scan to increase effectiveness and use the union filter to
1358 * cover all applicable quotas in a single scan.
1359 */
1360 eofb.eof_scan_owner = ip->i_ino;
1361 eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC;
1362
1363 if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) {
1364 dq = xfs_inode_dquot(ip, XFS_DQ_USER);
1365 if (dq && xfs_dquot_lowsp(dq)) {
1366 eofb.eof_uid = VFS_I(ip)->i_uid;
1367 eofb.eof_flags |= XFS_EOF_FLAGS_UID;
1368 scan = 1;
1369 }
1370 }
1371
1372 if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) {
1373 dq = xfs_inode_dquot(ip, XFS_DQ_GROUP);
1374 if (dq && xfs_dquot_lowsp(dq)) {
1375 eofb.eof_gid = VFS_I(ip)->i_gid;
1376 eofb.eof_flags |= XFS_EOF_FLAGS_GID;
1377 scan = 1;
1378 }
1379 }
1380
1381 if (scan)
1382 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
1383
1384 return scan;
1385}
1386
1387void
1388xfs_inode_set_eofblocks_tag(
1389 xfs_inode_t *ip)
1390{
1391 struct xfs_mount *mp = ip->i_mount;
1392 struct xfs_perag *pag;
1393 int tagged;
1394
1395 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1396 spin_lock(&pag->pag_ici_lock);
1397 trace_xfs_inode_set_eofblocks_tag(ip);
1398
1399 tagged = radix_tree_tagged(&pag->pag_ici_root,
1400 XFS_ICI_EOFBLOCKS_TAG);
1401 radix_tree_tag_set(&pag->pag_ici_root,
1402 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1403 XFS_ICI_EOFBLOCKS_TAG);
1404 if (!tagged) {
1405 /* propagate the eofblocks tag up into the perag radix tree */
1406 spin_lock(&ip->i_mount->m_perag_lock);
1407 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
1408 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1409 XFS_ICI_EOFBLOCKS_TAG);
1410 spin_unlock(&ip->i_mount->m_perag_lock);
1411
1412 /* kick off background trimming */
1413 xfs_queue_eofblocks(ip->i_mount);
1414
1415 trace_xfs_perag_set_eofblocks(ip->i_mount, pag->pag_agno,
1416 -1, _RET_IP_);
1417 }
1418
1419 spin_unlock(&pag->pag_ici_lock);
1420 xfs_perag_put(pag);
1421}
1422
1423void
1424xfs_inode_clear_eofblocks_tag(
1425 xfs_inode_t *ip)
1426{
1427 struct xfs_mount *mp = ip->i_mount;
1428 struct xfs_perag *pag;
1429
1430 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1431 spin_lock(&pag->pag_ici_lock);
1432 trace_xfs_inode_clear_eofblocks_tag(ip);
1433
1434 radix_tree_tag_clear(&pag->pag_ici_root,
1435 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
1436 XFS_ICI_EOFBLOCKS_TAG);
1437 if (!radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_EOFBLOCKS_TAG)) {
1438 /* clear the eofblocks tag from the perag radix tree */
1439 spin_lock(&ip->i_mount->m_perag_lock);
1440 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
1441 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
1442 XFS_ICI_EOFBLOCKS_TAG);
1443 spin_unlock(&ip->i_mount->m_perag_lock);
1444 trace_xfs_perag_clear_eofblocks(ip->i_mount, pag->pag_agno,
1445 -1, _RET_IP_);
1446 }
1447
1448 spin_unlock(&pag->pag_ici_lock);
1449 xfs_perag_put(pag);
1450}
1451