Commit | Line | Data |
---|---|---|
fe4fa4b8 DC |
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_types.h" | |
fe4fa4b8 | 21 | #include "xfs_log.h" |
f661f1e0 | 22 | #include "xfs_log_priv.h" |
fe4fa4b8 DC |
23 | #include "xfs_inum.h" |
24 | #include "xfs_trans.h" | |
fd074841 | 25 | #include "xfs_trans_priv.h" |
fe4fa4b8 DC |
26 | #include "xfs_sb.h" |
27 | #include "xfs_ag.h" | |
fe4fa4b8 DC |
28 | #include "xfs_mount.h" |
29 | #include "xfs_bmap_btree.h" | |
fe4fa4b8 DC |
30 | #include "xfs_inode.h" |
31 | #include "xfs_dinode.h" | |
32 | #include "xfs_error.h" | |
fe4fa4b8 DC |
33 | #include "xfs_filestream.h" |
34 | #include "xfs_vnodeops.h" | |
fe4fa4b8 | 35 | #include "xfs_inode_item.h" |
7d095257 | 36 | #include "xfs_quota.h" |
0b1b213f | 37 | #include "xfs_trace.h" |
1a387d3b | 38 | #include "xfs_fsops.h" |
6d8b79cf | 39 | #include "xfs_icache.h" |
fe4fa4b8 | 40 | |
a167b17e DC |
41 | #include <linux/kthread.h> |
42 | #include <linux/freezer.h> | |
43 | ||
33479e05 DC |
44 | STATIC void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp, |
45 | struct xfs_perag *pag, struct xfs_inode *ip); | |
46 | ||
47 | /* | |
48 | * Allocate and initialise an xfs_inode. | |
49 | */ | |
50 | STATIC struct xfs_inode * | |
51 | xfs_inode_alloc( | |
52 | struct xfs_mount *mp, | |
53 | xfs_ino_t ino) | |
54 | { | |
55 | struct xfs_inode *ip; | |
56 | ||
57 | /* | |
58 | * if this didn't occur in transactions, we could use | |
59 | * KM_MAYFAIL and return NULL here on ENOMEM. Set the | |
60 | * code up to do this anyway. | |
61 | */ | |
62 | ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP); | |
63 | if (!ip) | |
64 | return NULL; | |
65 | if (inode_init_always(mp->m_super, VFS_I(ip))) { | |
66 | kmem_zone_free(xfs_inode_zone, ip); | |
67 | return NULL; | |
68 | } | |
69 | ||
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(xfs_icdinode_t)); | |
86 | ||
87 | return ip; | |
88 | } | |
89 | ||
90 | STATIC void | |
91 | xfs_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 | ||
100 | STATIC void | |
101 | xfs_inode_free( | |
102 | struct xfs_inode *ip) | |
103 | { | |
104 | switch (ip->i_d.di_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 | /* asserts to verify all state is correct here */ | |
122 | ASSERT(atomic_read(&ip->i_pincount) == 0); | |
123 | ASSERT(!spin_is_locked(&ip->i_flags_lock)); | |
124 | ASSERT(!xfs_isiflocked(ip)); | |
125 | ||
126 | /* | |
127 | * Because we use RCU freeing we need to ensure the inode always | |
128 | * appears to be reclaimed with an invalid inode number when in the | |
129 | * free state. The ip->i_flags_lock provides the barrier against lookup | |
130 | * races. | |
131 | */ | |
132 | spin_lock(&ip->i_flags_lock); | |
133 | ip->i_flags = XFS_IRECLAIM; | |
134 | ip->i_ino = 0; | |
135 | spin_unlock(&ip->i_flags_lock); | |
136 | ||
137 | call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback); | |
138 | } | |
139 | ||
140 | /* | |
141 | * Check the validity of the inode we just found it the cache | |
142 | */ | |
143 | static int | |
144 | xfs_iget_cache_hit( | |
145 | struct xfs_perag *pag, | |
146 | struct xfs_inode *ip, | |
147 | xfs_ino_t ino, | |
148 | int flags, | |
149 | int lock_flags) __releases(RCU) | |
150 | { | |
151 | struct inode *inode = VFS_I(ip); | |
152 | struct xfs_mount *mp = ip->i_mount; | |
153 | int error; | |
154 | ||
155 | /* | |
156 | * check for re-use of an inode within an RCU grace period due to the | |
157 | * radix tree nodes not being updated yet. We monitor for this by | |
158 | * setting the inode number to zero before freeing the inode structure. | |
159 | * If the inode has been reallocated and set up, then the inode number | |
160 | * will not match, so check for that, too. | |
161 | */ | |
162 | spin_lock(&ip->i_flags_lock); | |
163 | if (ip->i_ino != ino) { | |
164 | trace_xfs_iget_skip(ip); | |
165 | XFS_STATS_INC(xs_ig_frecycle); | |
166 | error = EAGAIN; | |
167 | goto out_error; | |
168 | } | |
169 | ||
170 | ||
171 | /* | |
172 | * If we are racing with another cache hit that is currently | |
173 | * instantiating this inode or currently recycling it out of | |
174 | * reclaimabe state, wait for the initialisation to complete | |
175 | * before continuing. | |
176 | * | |
177 | * XXX(hch): eventually we should do something equivalent to | |
178 | * wait_on_inode to wait for these flags to be cleared | |
179 | * instead of polling for it. | |
180 | */ | |
181 | if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) { | |
182 | trace_xfs_iget_skip(ip); | |
183 | XFS_STATS_INC(xs_ig_frecycle); | |
184 | error = EAGAIN; | |
185 | goto out_error; | |
186 | } | |
187 | ||
188 | /* | |
189 | * If lookup is racing with unlink return an error immediately. | |
190 | */ | |
191 | if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) { | |
192 | error = ENOENT; | |
193 | goto out_error; | |
194 | } | |
195 | ||
196 | /* | |
197 | * If IRECLAIMABLE is set, we've torn down the VFS inode already. | |
198 | * Need to carefully get it back into useable state. | |
199 | */ | |
200 | if (ip->i_flags & XFS_IRECLAIMABLE) { | |
201 | trace_xfs_iget_reclaim(ip); | |
202 | ||
203 | /* | |
204 | * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode | |
205 | * from stomping over us while we recycle the inode. We can't | |
206 | * clear the radix tree reclaimable tag yet as it requires | |
207 | * pag_ici_lock to be held exclusive. | |
208 | */ | |
209 | ip->i_flags |= XFS_IRECLAIM; | |
210 | ||
211 | spin_unlock(&ip->i_flags_lock); | |
212 | rcu_read_unlock(); | |
213 | ||
214 | error = -inode_init_always(mp->m_super, inode); | |
215 | if (error) { | |
216 | /* | |
217 | * Re-initializing the inode failed, and we are in deep | |
218 | * trouble. Try to re-add it to the reclaim list. | |
219 | */ | |
220 | rcu_read_lock(); | |
221 | spin_lock(&ip->i_flags_lock); | |
222 | ||
223 | ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM); | |
224 | ASSERT(ip->i_flags & XFS_IRECLAIMABLE); | |
225 | trace_xfs_iget_reclaim_fail(ip); | |
226 | goto out_error; | |
227 | } | |
228 | ||
229 | spin_lock(&pag->pag_ici_lock); | |
230 | spin_lock(&ip->i_flags_lock); | |
231 | ||
232 | /* | |
233 | * Clear the per-lifetime state in the inode as we are now | |
234 | * effectively a new inode and need to return to the initial | |
235 | * state before reuse occurs. | |
236 | */ | |
237 | ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS; | |
238 | ip->i_flags |= XFS_INEW; | |
239 | __xfs_inode_clear_reclaim_tag(mp, pag, ip); | |
240 | inode->i_state = I_NEW; | |
241 | ||
242 | ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock)); | |
243 | mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino); | |
244 | ||
245 | spin_unlock(&ip->i_flags_lock); | |
246 | spin_unlock(&pag->pag_ici_lock); | |
247 | } else { | |
248 | /* If the VFS inode is being torn down, pause and try again. */ | |
249 | if (!igrab(inode)) { | |
250 | trace_xfs_iget_skip(ip); | |
251 | error = EAGAIN; | |
252 | goto out_error; | |
253 | } | |
254 | ||
255 | /* We've got a live one. */ | |
256 | spin_unlock(&ip->i_flags_lock); | |
257 | rcu_read_unlock(); | |
258 | trace_xfs_iget_hit(ip); | |
259 | } | |
260 | ||
261 | if (lock_flags != 0) | |
262 | xfs_ilock(ip, lock_flags); | |
263 | ||
264 | xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE); | |
265 | XFS_STATS_INC(xs_ig_found); | |
266 | ||
267 | return 0; | |
268 | ||
269 | out_error: | |
270 | spin_unlock(&ip->i_flags_lock); | |
271 | rcu_read_unlock(); | |
272 | return error; | |
273 | } | |
274 | ||
275 | ||
276 | static int | |
277 | xfs_iget_cache_miss( | |
278 | struct xfs_mount *mp, | |
279 | struct xfs_perag *pag, | |
280 | xfs_trans_t *tp, | |
281 | xfs_ino_t ino, | |
282 | struct xfs_inode **ipp, | |
283 | int flags, | |
284 | int lock_flags) | |
285 | { | |
286 | struct xfs_inode *ip; | |
287 | int error; | |
288 | xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino); | |
289 | int iflags; | |
290 | ||
291 | ip = xfs_inode_alloc(mp, ino); | |
292 | if (!ip) | |
293 | return ENOMEM; | |
294 | ||
295 | error = xfs_iread(mp, tp, ip, flags); | |
296 | if (error) | |
297 | goto out_destroy; | |
298 | ||
299 | trace_xfs_iget_miss(ip); | |
300 | ||
301 | if ((ip->i_d.di_mode == 0) && !(flags & XFS_IGET_CREATE)) { | |
302 | error = ENOENT; | |
303 | goto out_destroy; | |
304 | } | |
305 | ||
306 | /* | |
307 | * Preload the radix tree so we can insert safely under the | |
308 | * write spinlock. Note that we cannot sleep inside the preload | |
309 | * region. Since we can be called from transaction context, don't | |
310 | * recurse into the file system. | |
311 | */ | |
312 | if (radix_tree_preload(GFP_NOFS)) { | |
313 | error = EAGAIN; | |
314 | goto out_destroy; | |
315 | } | |
316 | ||
317 | /* | |
318 | * Because the inode hasn't been added to the radix-tree yet it can't | |
319 | * be found by another thread, so we can do the non-sleeping lock here. | |
320 | */ | |
321 | if (lock_flags) { | |
322 | if (!xfs_ilock_nowait(ip, lock_flags)) | |
323 | BUG(); | |
324 | } | |
325 | ||
326 | /* | |
327 | * These values must be set before inserting the inode into the radix | |
328 | * tree as the moment it is inserted a concurrent lookup (allowed by the | |
329 | * RCU locking mechanism) can find it and that lookup must see that this | |
330 | * is an inode currently under construction (i.e. that XFS_INEW is set). | |
331 | * The ip->i_flags_lock that protects the XFS_INEW flag forms the | |
332 | * memory barrier that ensures this detection works correctly at lookup | |
333 | * time. | |
334 | */ | |
335 | iflags = XFS_INEW; | |
336 | if (flags & XFS_IGET_DONTCACHE) | |
337 | iflags |= XFS_IDONTCACHE; | |
113a5683 CS |
338 | ip->i_udquot = NULL; |
339 | ip->i_gdquot = NULL; | |
92f8ff73 | 340 | ip->i_pdquot = NULL; |
33479e05 DC |
341 | xfs_iflags_set(ip, iflags); |
342 | ||
343 | /* insert the new inode */ | |
344 | spin_lock(&pag->pag_ici_lock); | |
345 | error = radix_tree_insert(&pag->pag_ici_root, agino, ip); | |
346 | if (unlikely(error)) { | |
347 | WARN_ON(error != -EEXIST); | |
348 | XFS_STATS_INC(xs_ig_dup); | |
349 | error = EAGAIN; | |
350 | goto out_preload_end; | |
351 | } | |
352 | spin_unlock(&pag->pag_ici_lock); | |
353 | radix_tree_preload_end(); | |
354 | ||
355 | *ipp = ip; | |
356 | return 0; | |
357 | ||
358 | out_preload_end: | |
359 | spin_unlock(&pag->pag_ici_lock); | |
360 | radix_tree_preload_end(); | |
361 | if (lock_flags) | |
362 | xfs_iunlock(ip, lock_flags); | |
363 | out_destroy: | |
364 | __destroy_inode(VFS_I(ip)); | |
365 | xfs_inode_free(ip); | |
366 | return error; | |
367 | } | |
368 | ||
369 | /* | |
370 | * Look up an inode by number in the given file system. | |
371 | * The inode is looked up in the cache held in each AG. | |
372 | * If the inode is found in the cache, initialise the vfs inode | |
373 | * if necessary. | |
374 | * | |
375 | * If it is not in core, read it in from the file system's device, | |
376 | * add it to the cache and initialise the vfs inode. | |
377 | * | |
378 | * The inode is locked according to the value of the lock_flags parameter. | |
379 | * This flag parameter indicates how and if the inode's IO lock and inode lock | |
380 | * should be taken. | |
381 | * | |
382 | * mp -- the mount point structure for the current file system. It points | |
383 | * to the inode hash table. | |
384 | * tp -- a pointer to the current transaction if there is one. This is | |
385 | * simply passed through to the xfs_iread() call. | |
386 | * ino -- the number of the inode desired. This is the unique identifier | |
387 | * within the file system for the inode being requested. | |
388 | * lock_flags -- flags indicating how to lock the inode. See the comment | |
389 | * for xfs_ilock() for a list of valid values. | |
390 | */ | |
391 | int | |
392 | xfs_iget( | |
393 | xfs_mount_t *mp, | |
394 | xfs_trans_t *tp, | |
395 | xfs_ino_t ino, | |
396 | uint flags, | |
397 | uint lock_flags, | |
398 | xfs_inode_t **ipp) | |
399 | { | |
400 | xfs_inode_t *ip; | |
401 | int error; | |
402 | xfs_perag_t *pag; | |
403 | xfs_agino_t agino; | |
404 | ||
405 | /* | |
406 | * xfs_reclaim_inode() uses the ILOCK to ensure an inode | |
407 | * doesn't get freed while it's being referenced during a | |
408 | * radix tree traversal here. It assumes this function | |
409 | * aqcuires only the ILOCK (and therefore it has no need to | |
410 | * involve the IOLOCK in this synchronization). | |
411 | */ | |
412 | ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0); | |
413 | ||
414 | /* reject inode numbers outside existing AGs */ | |
415 | if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount) | |
416 | return EINVAL; | |
417 | ||
418 | /* get the perag structure and ensure that it's inode capable */ | |
419 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino)); | |
420 | agino = XFS_INO_TO_AGINO(mp, ino); | |
421 | ||
422 | again: | |
423 | error = 0; | |
424 | rcu_read_lock(); | |
425 | ip = radix_tree_lookup(&pag->pag_ici_root, agino); | |
426 | ||
427 | if (ip) { | |
428 | error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags); | |
429 | if (error) | |
430 | goto out_error_or_again; | |
431 | } else { | |
432 | rcu_read_unlock(); | |
433 | XFS_STATS_INC(xs_ig_missed); | |
434 | ||
435 | error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip, | |
436 | flags, lock_flags); | |
437 | if (error) | |
438 | goto out_error_or_again; | |
439 | } | |
440 | xfs_perag_put(pag); | |
441 | ||
442 | *ipp = ip; | |
443 | ||
444 | /* | |
445 | * If we have a real type for an on-disk inode, we can set ops(&unlock) | |
446 | * now. If it's a new inode being created, xfs_ialloc will handle it. | |
447 | */ | |
448 | if (xfs_iflags_test(ip, XFS_INEW) && ip->i_d.di_mode != 0) | |
449 | xfs_setup_inode(ip); | |
450 | return 0; | |
451 | ||
452 | out_error_or_again: | |
453 | if (error == EAGAIN) { | |
454 | delay(1); | |
455 | goto again; | |
456 | } | |
457 | xfs_perag_put(pag); | |
458 | return error; | |
459 | } | |
460 | ||
78ae5256 DC |
461 | /* |
462 | * The inode lookup is done in batches to keep the amount of lock traffic and | |
463 | * radix tree lookups to a minimum. The batch size is a trade off between | |
464 | * lookup reduction and stack usage. This is in the reclaim path, so we can't | |
465 | * be too greedy. | |
466 | */ | |
467 | #define XFS_LOOKUP_BATCH 32 | |
468 | ||
e13de955 DC |
469 | STATIC int |
470 | xfs_inode_ag_walk_grab( | |
471 | struct xfs_inode *ip) | |
472 | { | |
473 | struct inode *inode = VFS_I(ip); | |
474 | ||
1a3e8f3d DC |
475 | ASSERT(rcu_read_lock_held()); |
476 | ||
477 | /* | |
478 | * check for stale RCU freed inode | |
479 | * | |
480 | * If the inode has been reallocated, it doesn't matter if it's not in | |
481 | * the AG we are walking - we are walking for writeback, so if it | |
482 | * passes all the "valid inode" checks and is dirty, then we'll write | |
483 | * it back anyway. If it has been reallocated and still being | |
484 | * initialised, the XFS_INEW check below will catch it. | |
485 | */ | |
486 | spin_lock(&ip->i_flags_lock); | |
487 | if (!ip->i_ino) | |
488 | goto out_unlock_noent; | |
489 | ||
490 | /* avoid new or reclaimable inodes. Leave for reclaim code to flush */ | |
491 | if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM)) | |
492 | goto out_unlock_noent; | |
493 | spin_unlock(&ip->i_flags_lock); | |
494 | ||
e13de955 DC |
495 | /* nothing to sync during shutdown */ |
496 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) | |
497 | return EFSCORRUPTED; | |
498 | ||
e13de955 DC |
499 | /* If we can't grab the inode, it must on it's way to reclaim. */ |
500 | if (!igrab(inode)) | |
501 | return ENOENT; | |
502 | ||
503 | if (is_bad_inode(inode)) { | |
504 | IRELE(ip); | |
505 | return ENOENT; | |
506 | } | |
507 | ||
508 | /* inode is valid */ | |
509 | return 0; | |
1a3e8f3d DC |
510 | |
511 | out_unlock_noent: | |
512 | spin_unlock(&ip->i_flags_lock); | |
513 | return ENOENT; | |
e13de955 DC |
514 | } |
515 | ||
75f3cb13 DC |
516 | STATIC int |
517 | xfs_inode_ag_walk( | |
518 | struct xfs_mount *mp, | |
5017e97d | 519 | struct xfs_perag *pag, |
75f3cb13 | 520 | int (*execute)(struct xfs_inode *ip, |
a454f742 BF |
521 | struct xfs_perag *pag, int flags, |
522 | void *args), | |
523 | int flags, | |
524 | void *args, | |
525 | int tag) | |
75f3cb13 | 526 | { |
75f3cb13 DC |
527 | uint32_t first_index; |
528 | int last_error = 0; | |
529 | int skipped; | |
65d0f205 | 530 | int done; |
78ae5256 | 531 | int nr_found; |
75f3cb13 DC |
532 | |
533 | restart: | |
65d0f205 | 534 | done = 0; |
75f3cb13 DC |
535 | skipped = 0; |
536 | first_index = 0; | |
78ae5256 | 537 | nr_found = 0; |
75f3cb13 | 538 | do { |
78ae5256 | 539 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
75f3cb13 | 540 | int error = 0; |
78ae5256 | 541 | int i; |
75f3cb13 | 542 | |
1a3e8f3d | 543 | rcu_read_lock(); |
a454f742 BF |
544 | |
545 | if (tag == -1) | |
546 | nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, | |
78ae5256 DC |
547 | (void **)batch, first_index, |
548 | XFS_LOOKUP_BATCH); | |
a454f742 BF |
549 | else |
550 | nr_found = radix_tree_gang_lookup_tag( | |
551 | &pag->pag_ici_root, | |
552 | (void **) batch, first_index, | |
553 | XFS_LOOKUP_BATCH, tag); | |
554 | ||
65d0f205 | 555 | if (!nr_found) { |
1a3e8f3d | 556 | rcu_read_unlock(); |
75f3cb13 | 557 | break; |
c8e20be0 | 558 | } |
75f3cb13 | 559 | |
65d0f205 | 560 | /* |
78ae5256 DC |
561 | * Grab the inodes before we drop the lock. if we found |
562 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 563 | */ |
78ae5256 DC |
564 | for (i = 0; i < nr_found; i++) { |
565 | struct xfs_inode *ip = batch[i]; | |
566 | ||
567 | if (done || xfs_inode_ag_walk_grab(ip)) | |
568 | batch[i] = NULL; | |
569 | ||
570 | /* | |
1a3e8f3d DC |
571 | * Update the index for the next lookup. Catch |
572 | * overflows into the next AG range which can occur if | |
573 | * we have inodes in the last block of the AG and we | |
574 | * are currently pointing to the last inode. | |
575 | * | |
576 | * Because we may see inodes that are from the wrong AG | |
577 | * due to RCU freeing and reallocation, only update the | |
578 | * index if it lies in this AG. It was a race that lead | |
579 | * us to see this inode, so another lookup from the | |
580 | * same index will not find it again. | |
78ae5256 | 581 | */ |
1a3e8f3d DC |
582 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno) |
583 | continue; | |
78ae5256 DC |
584 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
585 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
586 | done = 1; | |
e13de955 | 587 | } |
78ae5256 DC |
588 | |
589 | /* unlock now we've grabbed the inodes. */ | |
1a3e8f3d | 590 | rcu_read_unlock(); |
e13de955 | 591 | |
78ae5256 DC |
592 | for (i = 0; i < nr_found; i++) { |
593 | if (!batch[i]) | |
594 | continue; | |
a454f742 | 595 | error = execute(batch[i], pag, flags, args); |
78ae5256 DC |
596 | IRELE(batch[i]); |
597 | if (error == EAGAIN) { | |
598 | skipped++; | |
599 | continue; | |
600 | } | |
601 | if (error && last_error != EFSCORRUPTED) | |
602 | last_error = error; | |
75f3cb13 | 603 | } |
c8e20be0 DC |
604 | |
605 | /* bail out if the filesystem is corrupted. */ | |
75f3cb13 DC |
606 | if (error == EFSCORRUPTED) |
607 | break; | |
608 | ||
8daaa831 DC |
609 | cond_resched(); |
610 | ||
78ae5256 | 611 | } while (nr_found && !done); |
75f3cb13 DC |
612 | |
613 | if (skipped) { | |
614 | delay(1); | |
615 | goto restart; | |
616 | } | |
75f3cb13 DC |
617 | return last_error; |
618 | } | |
619 | ||
579b62fa BF |
620 | /* |
621 | * Background scanning to trim post-EOF preallocated space. This is queued | |
622 | * based on the 'background_prealloc_discard_period' tunable (5m by default). | |
623 | */ | |
624 | STATIC void | |
625 | xfs_queue_eofblocks( | |
626 | struct xfs_mount *mp) | |
627 | { | |
628 | rcu_read_lock(); | |
629 | if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG)) | |
630 | queue_delayed_work(mp->m_eofblocks_workqueue, | |
631 | &mp->m_eofblocks_work, | |
632 | msecs_to_jiffies(xfs_eofb_secs * 1000)); | |
633 | rcu_read_unlock(); | |
634 | } | |
635 | ||
636 | void | |
637 | xfs_eofblocks_worker( | |
638 | struct work_struct *work) | |
639 | { | |
640 | struct xfs_mount *mp = container_of(to_delayed_work(work), | |
641 | struct xfs_mount, m_eofblocks_work); | |
642 | xfs_icache_free_eofblocks(mp, NULL); | |
643 | xfs_queue_eofblocks(mp); | |
644 | } | |
645 | ||
fe588ed3 | 646 | int |
75f3cb13 DC |
647 | xfs_inode_ag_iterator( |
648 | struct xfs_mount *mp, | |
649 | int (*execute)(struct xfs_inode *ip, | |
a454f742 BF |
650 | struct xfs_perag *pag, int flags, |
651 | void *args), | |
652 | int flags, | |
653 | void *args) | |
75f3cb13 | 654 | { |
16fd5367 | 655 | struct xfs_perag *pag; |
75f3cb13 DC |
656 | int error = 0; |
657 | int last_error = 0; | |
658 | xfs_agnumber_t ag; | |
659 | ||
16fd5367 | 660 | ag = 0; |
65d0f205 DC |
661 | while ((pag = xfs_perag_get(mp, ag))) { |
662 | ag = pag->pag_agno + 1; | |
a454f742 BF |
663 | error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1); |
664 | xfs_perag_put(pag); | |
665 | if (error) { | |
666 | last_error = error; | |
667 | if (error == EFSCORRUPTED) | |
668 | break; | |
669 | } | |
670 | } | |
671 | return XFS_ERROR(last_error); | |
672 | } | |
673 | ||
674 | int | |
675 | xfs_inode_ag_iterator_tag( | |
676 | struct xfs_mount *mp, | |
677 | int (*execute)(struct xfs_inode *ip, | |
678 | struct xfs_perag *pag, int flags, | |
679 | void *args), | |
680 | int flags, | |
681 | void *args, | |
682 | int tag) | |
683 | { | |
684 | struct xfs_perag *pag; | |
685 | int error = 0; | |
686 | int last_error = 0; | |
687 | xfs_agnumber_t ag; | |
688 | ||
689 | ag = 0; | |
690 | while ((pag = xfs_perag_get_tag(mp, ag, tag))) { | |
691 | ag = pag->pag_agno + 1; | |
692 | error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag); | |
5017e97d | 693 | xfs_perag_put(pag); |
75f3cb13 DC |
694 | if (error) { |
695 | last_error = error; | |
696 | if (error == EFSCORRUPTED) | |
697 | break; | |
698 | } | |
699 | } | |
700 | return XFS_ERROR(last_error); | |
701 | } | |
702 | ||
a7b339f1 DC |
703 | /* |
704 | * Queue a new inode reclaim pass if there are reclaimable inodes and there | |
705 | * isn't a reclaim pass already in progress. By default it runs every 5s based | |
5889608d | 706 | * on the xfs periodic sync default of 30s. Perhaps this should have it's own |
a7b339f1 DC |
707 | * tunable, but that can be done if this method proves to be ineffective or too |
708 | * aggressive. | |
709 | */ | |
710 | static void | |
5889608d | 711 | xfs_reclaim_work_queue( |
a7b339f1 | 712 | struct xfs_mount *mp) |
a167b17e | 713 | { |
a167b17e | 714 | |
a7b339f1 DC |
715 | rcu_read_lock(); |
716 | if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) { | |
5889608d | 717 | queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work, |
a7b339f1 | 718 | msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10)); |
a167b17e | 719 | } |
a7b339f1 DC |
720 | rcu_read_unlock(); |
721 | } | |
a167b17e | 722 | |
a7b339f1 DC |
723 | /* |
724 | * This is a fast pass over the inode cache to try to get reclaim moving on as | |
725 | * many inodes as possible in a short period of time. It kicks itself every few | |
726 | * seconds, as well as being kicked by the inode cache shrinker when memory | |
727 | * goes low. It scans as quickly as possible avoiding locked inodes or those | |
728 | * already being flushed, and once done schedules a future pass. | |
729 | */ | |
33c7a2bc | 730 | void |
a7b339f1 DC |
731 | xfs_reclaim_worker( |
732 | struct work_struct *work) | |
733 | { | |
734 | struct xfs_mount *mp = container_of(to_delayed_work(work), | |
735 | struct xfs_mount, m_reclaim_work); | |
736 | ||
737 | xfs_reclaim_inodes(mp, SYNC_TRYLOCK); | |
5889608d | 738 | xfs_reclaim_work_queue(mp); |
a7b339f1 DC |
739 | } |
740 | ||
33479e05 | 741 | static void |
bc990f5c CH |
742 | __xfs_inode_set_reclaim_tag( |
743 | struct xfs_perag *pag, | |
744 | struct xfs_inode *ip) | |
745 | { | |
746 | radix_tree_tag_set(&pag->pag_ici_root, | |
747 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
748 | XFS_ICI_RECLAIM_TAG); | |
16fd5367 DC |
749 | |
750 | if (!pag->pag_ici_reclaimable) { | |
751 | /* propagate the reclaim tag up into the perag radix tree */ | |
752 | spin_lock(&ip->i_mount->m_perag_lock); | |
753 | radix_tree_tag_set(&ip->i_mount->m_perag_tree, | |
754 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
755 | XFS_ICI_RECLAIM_TAG); | |
756 | spin_unlock(&ip->i_mount->m_perag_lock); | |
a7b339f1 DC |
757 | |
758 | /* schedule periodic background inode reclaim */ | |
5889608d | 759 | xfs_reclaim_work_queue(ip->i_mount); |
a7b339f1 | 760 | |
16fd5367 DC |
761 | trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno, |
762 | -1, _RET_IP_); | |
763 | } | |
9bf729c0 | 764 | pag->pag_ici_reclaimable++; |
bc990f5c CH |
765 | } |
766 | ||
11654513 DC |
767 | /* |
768 | * We set the inode flag atomically with the radix tree tag. | |
769 | * Once we get tag lookups on the radix tree, this inode flag | |
770 | * can go away. | |
771 | */ | |
396beb85 DC |
772 | void |
773 | xfs_inode_set_reclaim_tag( | |
774 | xfs_inode_t *ip) | |
775 | { | |
5017e97d DC |
776 | struct xfs_mount *mp = ip->i_mount; |
777 | struct xfs_perag *pag; | |
396beb85 | 778 | |
5017e97d | 779 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); |
1a427ab0 | 780 | spin_lock(&pag->pag_ici_lock); |
396beb85 | 781 | spin_lock(&ip->i_flags_lock); |
bc990f5c | 782 | __xfs_inode_set_reclaim_tag(pag, ip); |
11654513 | 783 | __xfs_iflags_set(ip, XFS_IRECLAIMABLE); |
396beb85 | 784 | spin_unlock(&ip->i_flags_lock); |
1a427ab0 | 785 | spin_unlock(&pag->pag_ici_lock); |
5017e97d | 786 | xfs_perag_put(pag); |
396beb85 DC |
787 | } |
788 | ||
081003ff JW |
789 | STATIC void |
790 | __xfs_inode_clear_reclaim( | |
396beb85 DC |
791 | xfs_perag_t *pag, |
792 | xfs_inode_t *ip) | |
793 | { | |
9bf729c0 | 794 | pag->pag_ici_reclaimable--; |
16fd5367 DC |
795 | if (!pag->pag_ici_reclaimable) { |
796 | /* clear the reclaim tag from the perag radix tree */ | |
797 | spin_lock(&ip->i_mount->m_perag_lock); | |
798 | radix_tree_tag_clear(&ip->i_mount->m_perag_tree, | |
799 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
800 | XFS_ICI_RECLAIM_TAG); | |
801 | spin_unlock(&ip->i_mount->m_perag_lock); | |
802 | trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno, | |
803 | -1, _RET_IP_); | |
804 | } | |
396beb85 DC |
805 | } |
806 | ||
33479e05 | 807 | STATIC void |
081003ff JW |
808 | __xfs_inode_clear_reclaim_tag( |
809 | xfs_mount_t *mp, | |
810 | xfs_perag_t *pag, | |
811 | xfs_inode_t *ip) | |
812 | { | |
813 | radix_tree_tag_clear(&pag->pag_ici_root, | |
814 | XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG); | |
815 | __xfs_inode_clear_reclaim(pag, ip); | |
816 | } | |
817 | ||
e3a20c0b DC |
818 | /* |
819 | * Grab the inode for reclaim exclusively. | |
820 | * Return 0 if we grabbed it, non-zero otherwise. | |
821 | */ | |
822 | STATIC int | |
823 | xfs_reclaim_inode_grab( | |
824 | struct xfs_inode *ip, | |
825 | int flags) | |
826 | { | |
1a3e8f3d DC |
827 | ASSERT(rcu_read_lock_held()); |
828 | ||
829 | /* quick check for stale RCU freed inode */ | |
830 | if (!ip->i_ino) | |
831 | return 1; | |
e3a20c0b DC |
832 | |
833 | /* | |
474fce06 CH |
834 | * If we are asked for non-blocking operation, do unlocked checks to |
835 | * see if the inode already is being flushed or in reclaim to avoid | |
836 | * lock traffic. | |
e3a20c0b DC |
837 | */ |
838 | if ((flags & SYNC_TRYLOCK) && | |
474fce06 | 839 | __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM)) |
e3a20c0b | 840 | return 1; |
e3a20c0b DC |
841 | |
842 | /* | |
843 | * The radix tree lock here protects a thread in xfs_iget from racing | |
844 | * with us starting reclaim on the inode. Once we have the | |
845 | * XFS_IRECLAIM flag set it will not touch us. | |
1a3e8f3d DC |
846 | * |
847 | * Due to RCU lookup, we may find inodes that have been freed and only | |
848 | * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that | |
849 | * aren't candidates for reclaim at all, so we must check the | |
850 | * XFS_IRECLAIMABLE is set first before proceeding to reclaim. | |
e3a20c0b DC |
851 | */ |
852 | spin_lock(&ip->i_flags_lock); | |
1a3e8f3d DC |
853 | if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) || |
854 | __xfs_iflags_test(ip, XFS_IRECLAIM)) { | |
855 | /* not a reclaim candidate. */ | |
e3a20c0b DC |
856 | spin_unlock(&ip->i_flags_lock); |
857 | return 1; | |
858 | } | |
859 | __xfs_iflags_set(ip, XFS_IRECLAIM); | |
860 | spin_unlock(&ip->i_flags_lock); | |
861 | return 0; | |
862 | } | |
863 | ||
777df5af | 864 | /* |
8a48088f CH |
865 | * Inodes in different states need to be treated differently. The following |
866 | * table lists the inode states and the reclaim actions necessary: | |
777df5af DC |
867 | * |
868 | * inode state iflush ret required action | |
869 | * --------------- ---------- --------------- | |
870 | * bad - reclaim | |
871 | * shutdown EIO unpin and reclaim | |
872 | * clean, unpinned 0 reclaim | |
873 | * stale, unpinned 0 reclaim | |
c854363e DC |
874 | * clean, pinned(*) 0 requeue |
875 | * stale, pinned EAGAIN requeue | |
8a48088f CH |
876 | * dirty, async - requeue |
877 | * dirty, sync 0 reclaim | |
777df5af DC |
878 | * |
879 | * (*) dgc: I don't think the clean, pinned state is possible but it gets | |
880 | * handled anyway given the order of checks implemented. | |
881 | * | |
c854363e DC |
882 | * Also, because we get the flush lock first, we know that any inode that has |
883 | * been flushed delwri has had the flush completed by the time we check that | |
8a48088f | 884 | * the inode is clean. |
c854363e | 885 | * |
8a48088f CH |
886 | * Note that because the inode is flushed delayed write by AIL pushing, the |
887 | * flush lock may already be held here and waiting on it can result in very | |
888 | * long latencies. Hence for sync reclaims, where we wait on the flush lock, | |
889 | * the caller should push the AIL first before trying to reclaim inodes to | |
890 | * minimise the amount of time spent waiting. For background relaim, we only | |
891 | * bother to reclaim clean inodes anyway. | |
c854363e | 892 | * |
777df5af DC |
893 | * Hence the order of actions after gaining the locks should be: |
894 | * bad => reclaim | |
895 | * shutdown => unpin and reclaim | |
8a48088f | 896 | * pinned, async => requeue |
c854363e | 897 | * pinned, sync => unpin |
777df5af DC |
898 | * stale => reclaim |
899 | * clean => reclaim | |
8a48088f | 900 | * dirty, async => requeue |
c854363e | 901 | * dirty, sync => flush, wait and reclaim |
777df5af | 902 | */ |
75f3cb13 | 903 | STATIC int |
c8e20be0 | 904 | xfs_reclaim_inode( |
75f3cb13 DC |
905 | struct xfs_inode *ip, |
906 | struct xfs_perag *pag, | |
c8e20be0 | 907 | int sync_mode) |
fce08f2f | 908 | { |
4c46819a CH |
909 | struct xfs_buf *bp = NULL; |
910 | int error; | |
777df5af | 911 | |
1bfd8d04 DC |
912 | restart: |
913 | error = 0; | |
c8e20be0 | 914 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
c854363e DC |
915 | if (!xfs_iflock_nowait(ip)) { |
916 | if (!(sync_mode & SYNC_WAIT)) | |
917 | goto out; | |
918 | xfs_iflock(ip); | |
919 | } | |
7a3be02b | 920 | |
777df5af DC |
921 | if (is_bad_inode(VFS_I(ip))) |
922 | goto reclaim; | |
923 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) { | |
924 | xfs_iunpin_wait(ip); | |
04913fdd | 925 | xfs_iflush_abort(ip, false); |
777df5af DC |
926 | goto reclaim; |
927 | } | |
c854363e | 928 | if (xfs_ipincount(ip)) { |
8a48088f CH |
929 | if (!(sync_mode & SYNC_WAIT)) |
930 | goto out_ifunlock; | |
777df5af | 931 | xfs_iunpin_wait(ip); |
c854363e | 932 | } |
777df5af DC |
933 | if (xfs_iflags_test(ip, XFS_ISTALE)) |
934 | goto reclaim; | |
935 | if (xfs_inode_clean(ip)) | |
936 | goto reclaim; | |
937 | ||
8a48088f CH |
938 | /* |
939 | * Never flush out dirty data during non-blocking reclaim, as it would | |
940 | * just contend with AIL pushing trying to do the same job. | |
941 | */ | |
942 | if (!(sync_mode & SYNC_WAIT)) | |
943 | goto out_ifunlock; | |
944 | ||
1bfd8d04 DC |
945 | /* |
946 | * Now we have an inode that needs flushing. | |
947 | * | |
4c46819a | 948 | * Note that xfs_iflush will never block on the inode buffer lock, as |
1bfd8d04 | 949 | * xfs_ifree_cluster() can lock the inode buffer before it locks the |
4c46819a | 950 | * ip->i_lock, and we are doing the exact opposite here. As a result, |
475ee413 CH |
951 | * doing a blocking xfs_imap_to_bp() to get the cluster buffer would |
952 | * result in an ABBA deadlock with xfs_ifree_cluster(). | |
1bfd8d04 DC |
953 | * |
954 | * As xfs_ifree_cluser() must gather all inodes that are active in the | |
955 | * cache to mark them stale, if we hit this case we don't actually want | |
956 | * to do IO here - we want the inode marked stale so we can simply | |
4c46819a CH |
957 | * reclaim it. Hence if we get an EAGAIN error here, just unlock the |
958 | * inode, back off and try again. Hopefully the next pass through will | |
959 | * see the stale flag set on the inode. | |
1bfd8d04 | 960 | */ |
4c46819a | 961 | error = xfs_iflush(ip, &bp); |
8a48088f CH |
962 | if (error == EAGAIN) { |
963 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
964 | /* backoff longer than in xfs_ifree_cluster */ | |
965 | delay(2); | |
966 | goto restart; | |
c854363e | 967 | } |
c854363e | 968 | |
4c46819a CH |
969 | if (!error) { |
970 | error = xfs_bwrite(bp); | |
971 | xfs_buf_relse(bp); | |
972 | } | |
973 | ||
974 | xfs_iflock(ip); | |
777df5af DC |
975 | reclaim: |
976 | xfs_ifunlock(ip); | |
c8e20be0 | 977 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
978 | |
979 | XFS_STATS_INC(xs_ig_reclaims); | |
980 | /* | |
981 | * Remove the inode from the per-AG radix tree. | |
982 | * | |
983 | * Because radix_tree_delete won't complain even if the item was never | |
984 | * added to the tree assert that it's been there before to catch | |
985 | * problems with the inode life time early on. | |
986 | */ | |
1a427ab0 | 987 | spin_lock(&pag->pag_ici_lock); |
2f11feab DC |
988 | if (!radix_tree_delete(&pag->pag_ici_root, |
989 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino))) | |
990 | ASSERT(0); | |
081003ff | 991 | __xfs_inode_clear_reclaim(pag, ip); |
1a427ab0 | 992 | spin_unlock(&pag->pag_ici_lock); |
2f11feab DC |
993 | |
994 | /* | |
995 | * Here we do an (almost) spurious inode lock in order to coordinate | |
996 | * with inode cache radix tree lookups. This is because the lookup | |
997 | * can reference the inodes in the cache without taking references. | |
998 | * | |
999 | * We make that OK here by ensuring that we wait until the inode is | |
ad637a10 | 1000 | * unlocked after the lookup before we go ahead and free it. |
2f11feab | 1001 | */ |
ad637a10 | 1002 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
2f11feab | 1003 | xfs_qm_dqdetach(ip); |
ad637a10 | 1004 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
1005 | |
1006 | xfs_inode_free(ip); | |
ad637a10 | 1007 | return error; |
8a48088f CH |
1008 | |
1009 | out_ifunlock: | |
1010 | xfs_ifunlock(ip); | |
1011 | out: | |
1012 | xfs_iflags_clear(ip, XFS_IRECLAIM); | |
1013 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
1014 | /* | |
1015 | * We could return EAGAIN here to make reclaim rescan the inode tree in | |
1016 | * a short while. However, this just burns CPU time scanning the tree | |
5889608d DC |
1017 | * waiting for IO to complete and the reclaim work never goes back to |
1018 | * the idle state. Instead, return 0 to let the next scheduled | |
1019 | * background reclaim attempt to reclaim the inode again. | |
8a48088f CH |
1020 | */ |
1021 | return 0; | |
7a3be02b DC |
1022 | } |
1023 | ||
65d0f205 DC |
1024 | /* |
1025 | * Walk the AGs and reclaim the inodes in them. Even if the filesystem is | |
1026 | * corrupted, we still want to try to reclaim all the inodes. If we don't, | |
1027 | * then a shut down during filesystem unmount reclaim walk leak all the | |
1028 | * unreclaimed inodes. | |
1029 | */ | |
33479e05 | 1030 | STATIC int |
65d0f205 DC |
1031 | xfs_reclaim_inodes_ag( |
1032 | struct xfs_mount *mp, | |
1033 | int flags, | |
1034 | int *nr_to_scan) | |
1035 | { | |
1036 | struct xfs_perag *pag; | |
1037 | int error = 0; | |
1038 | int last_error = 0; | |
1039 | xfs_agnumber_t ag; | |
69b491c2 DC |
1040 | int trylock = flags & SYNC_TRYLOCK; |
1041 | int skipped; | |
65d0f205 | 1042 | |
69b491c2 | 1043 | restart: |
65d0f205 | 1044 | ag = 0; |
69b491c2 | 1045 | skipped = 0; |
65d0f205 DC |
1046 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
1047 | unsigned long first_index = 0; | |
1048 | int done = 0; | |
e3a20c0b | 1049 | int nr_found = 0; |
65d0f205 DC |
1050 | |
1051 | ag = pag->pag_agno + 1; | |
1052 | ||
69b491c2 DC |
1053 | if (trylock) { |
1054 | if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) { | |
1055 | skipped++; | |
f83282a8 | 1056 | xfs_perag_put(pag); |
69b491c2 DC |
1057 | continue; |
1058 | } | |
1059 | first_index = pag->pag_ici_reclaim_cursor; | |
1060 | } else | |
1061 | mutex_lock(&pag->pag_ici_reclaim_lock); | |
1062 | ||
65d0f205 | 1063 | do { |
e3a20c0b DC |
1064 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
1065 | int i; | |
65d0f205 | 1066 | |
1a3e8f3d | 1067 | rcu_read_lock(); |
e3a20c0b DC |
1068 | nr_found = radix_tree_gang_lookup_tag( |
1069 | &pag->pag_ici_root, | |
1070 | (void **)batch, first_index, | |
1071 | XFS_LOOKUP_BATCH, | |
65d0f205 DC |
1072 | XFS_ICI_RECLAIM_TAG); |
1073 | if (!nr_found) { | |
b2232219 | 1074 | done = 1; |
1a3e8f3d | 1075 | rcu_read_unlock(); |
65d0f205 DC |
1076 | break; |
1077 | } | |
1078 | ||
1079 | /* | |
e3a20c0b DC |
1080 | * Grab the inodes before we drop the lock. if we found |
1081 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 1082 | */ |
e3a20c0b DC |
1083 | for (i = 0; i < nr_found; i++) { |
1084 | struct xfs_inode *ip = batch[i]; | |
1085 | ||
1086 | if (done || xfs_reclaim_inode_grab(ip, flags)) | |
1087 | batch[i] = NULL; | |
1088 | ||
1089 | /* | |
1090 | * Update the index for the next lookup. Catch | |
1091 | * overflows into the next AG range which can | |
1092 | * occur if we have inodes in the last block of | |
1093 | * the AG and we are currently pointing to the | |
1094 | * last inode. | |
1a3e8f3d DC |
1095 | * |
1096 | * Because we may see inodes that are from the | |
1097 | * wrong AG due to RCU freeing and | |
1098 | * reallocation, only update the index if it | |
1099 | * lies in this AG. It was a race that lead us | |
1100 | * to see this inode, so another lookup from | |
1101 | * the same index will not find it again. | |
e3a20c0b | 1102 | */ |
1a3e8f3d DC |
1103 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != |
1104 | pag->pag_agno) | |
1105 | continue; | |
e3a20c0b DC |
1106 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
1107 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
1108 | done = 1; | |
1109 | } | |
65d0f205 | 1110 | |
e3a20c0b | 1111 | /* unlock now we've grabbed the inodes. */ |
1a3e8f3d | 1112 | rcu_read_unlock(); |
e3a20c0b DC |
1113 | |
1114 | for (i = 0; i < nr_found; i++) { | |
1115 | if (!batch[i]) | |
1116 | continue; | |
1117 | error = xfs_reclaim_inode(batch[i], pag, flags); | |
1118 | if (error && last_error != EFSCORRUPTED) | |
1119 | last_error = error; | |
1120 | } | |
1121 | ||
1122 | *nr_to_scan -= XFS_LOOKUP_BATCH; | |
65d0f205 | 1123 | |
8daaa831 DC |
1124 | cond_resched(); |
1125 | ||
e3a20c0b | 1126 | } while (nr_found && !done && *nr_to_scan > 0); |
65d0f205 | 1127 | |
69b491c2 DC |
1128 | if (trylock && !done) |
1129 | pag->pag_ici_reclaim_cursor = first_index; | |
1130 | else | |
1131 | pag->pag_ici_reclaim_cursor = 0; | |
1132 | mutex_unlock(&pag->pag_ici_reclaim_lock); | |
65d0f205 DC |
1133 | xfs_perag_put(pag); |
1134 | } | |
69b491c2 DC |
1135 | |
1136 | /* | |
1137 | * if we skipped any AG, and we still have scan count remaining, do | |
1138 | * another pass this time using blocking reclaim semantics (i.e | |
1139 | * waiting on the reclaim locks and ignoring the reclaim cursors). This | |
1140 | * ensure that when we get more reclaimers than AGs we block rather | |
1141 | * than spin trying to execute reclaim. | |
1142 | */ | |
8daaa831 | 1143 | if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) { |
69b491c2 DC |
1144 | trylock = 0; |
1145 | goto restart; | |
1146 | } | |
65d0f205 DC |
1147 | return XFS_ERROR(last_error); |
1148 | } | |
1149 | ||
7a3be02b DC |
1150 | int |
1151 | xfs_reclaim_inodes( | |
1152 | xfs_mount_t *mp, | |
7a3be02b DC |
1153 | int mode) |
1154 | { | |
65d0f205 DC |
1155 | int nr_to_scan = INT_MAX; |
1156 | ||
1157 | return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan); | |
9bf729c0 DC |
1158 | } |
1159 | ||
1160 | /* | |
8daaa831 | 1161 | * Scan a certain number of inodes for reclaim. |
a7b339f1 DC |
1162 | * |
1163 | * When called we make sure that there is a background (fast) inode reclaim in | |
8daaa831 | 1164 | * progress, while we will throttle the speed of reclaim via doing synchronous |
a7b339f1 DC |
1165 | * reclaim of inodes. That means if we come across dirty inodes, we wait for |
1166 | * them to be cleaned, which we hope will not be very long due to the | |
1167 | * background walker having already kicked the IO off on those dirty inodes. | |
9bf729c0 | 1168 | */ |
8daaa831 DC |
1169 | void |
1170 | xfs_reclaim_inodes_nr( | |
1171 | struct xfs_mount *mp, | |
1172 | int nr_to_scan) | |
9bf729c0 | 1173 | { |
8daaa831 | 1174 | /* kick background reclaimer and push the AIL */ |
5889608d | 1175 | xfs_reclaim_work_queue(mp); |
8daaa831 | 1176 | xfs_ail_push_all(mp->m_ail); |
a7b339f1 | 1177 | |
8daaa831 DC |
1178 | xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan); |
1179 | } | |
9bf729c0 | 1180 | |
8daaa831 DC |
1181 | /* |
1182 | * Return the number of reclaimable inodes in the filesystem for | |
1183 | * the shrinker to determine how much to reclaim. | |
1184 | */ | |
1185 | int | |
1186 | xfs_reclaim_inodes_count( | |
1187 | struct xfs_mount *mp) | |
1188 | { | |
1189 | struct xfs_perag *pag; | |
1190 | xfs_agnumber_t ag = 0; | |
1191 | int reclaimable = 0; | |
9bf729c0 | 1192 | |
65d0f205 DC |
1193 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
1194 | ag = pag->pag_agno + 1; | |
70e60ce7 DC |
1195 | reclaimable += pag->pag_ici_reclaimable; |
1196 | xfs_perag_put(pag); | |
9bf729c0 | 1197 | } |
9bf729c0 DC |
1198 | return reclaimable; |
1199 | } | |
1200 | ||
3e3f9f58 BF |
1201 | STATIC int |
1202 | xfs_inode_match_id( | |
1203 | struct xfs_inode *ip, | |
1204 | struct xfs_eofblocks *eofb) | |
1205 | { | |
1b556048 BF |
1206 | if (eofb->eof_flags & XFS_EOF_FLAGS_UID && |
1207 | ip->i_d.di_uid != eofb->eof_uid) | |
1208 | return 0; | |
3e3f9f58 | 1209 | |
1b556048 BF |
1210 | if (eofb->eof_flags & XFS_EOF_FLAGS_GID && |
1211 | ip->i_d.di_gid != eofb->eof_gid) | |
1212 | return 0; | |
1213 | ||
1214 | if (eofb->eof_flags & XFS_EOF_FLAGS_PRID && | |
1215 | xfs_get_projid(ip) != eofb->eof_prid) | |
1216 | return 0; | |
1217 | ||
1218 | return 1; | |
3e3f9f58 BF |
1219 | } |
1220 | ||
41176a68 BF |
1221 | STATIC int |
1222 | xfs_inode_free_eofblocks( | |
1223 | struct xfs_inode *ip, | |
1224 | struct xfs_perag *pag, | |
1225 | int flags, | |
1226 | void *args) | |
1227 | { | |
1228 | int ret; | |
3e3f9f58 | 1229 | struct xfs_eofblocks *eofb = args; |
41176a68 BF |
1230 | |
1231 | if (!xfs_can_free_eofblocks(ip, false)) { | |
1232 | /* inode could be preallocated or append-only */ | |
1233 | trace_xfs_inode_free_eofblocks_invalid(ip); | |
1234 | xfs_inode_clear_eofblocks_tag(ip); | |
1235 | return 0; | |
1236 | } | |
1237 | ||
1238 | /* | |
1239 | * If the mapping is dirty the operation can block and wait for some | |
1240 | * time. Unless we are waiting, skip it. | |
1241 | */ | |
1242 | if (!(flags & SYNC_WAIT) && | |
1243 | mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY)) | |
1244 | return 0; | |
1245 | ||
00ca79a0 BF |
1246 | if (eofb) { |
1247 | if (!xfs_inode_match_id(ip, eofb)) | |
1248 | return 0; | |
1249 | ||
1250 | /* skip the inode if the file size is too small */ | |
1251 | if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE && | |
1252 | XFS_ISIZE(ip) < eofb->eof_min_file_size) | |
1253 | return 0; | |
1254 | } | |
3e3f9f58 | 1255 | |
41176a68 BF |
1256 | ret = xfs_free_eofblocks(ip->i_mount, ip, true); |
1257 | ||
1258 | /* don't revisit the inode if we're not waiting */ | |
1259 | if (ret == EAGAIN && !(flags & SYNC_WAIT)) | |
1260 | ret = 0; | |
1261 | ||
1262 | return ret; | |
1263 | } | |
1264 | ||
1265 | int | |
1266 | xfs_icache_free_eofblocks( | |
1267 | struct xfs_mount *mp, | |
8ca149de | 1268 | struct xfs_eofblocks *eofb) |
41176a68 | 1269 | { |
8ca149de BF |
1270 | int flags = SYNC_TRYLOCK; |
1271 | ||
1272 | if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC)) | |
1273 | flags = SYNC_WAIT; | |
1274 | ||
41176a68 | 1275 | return xfs_inode_ag_iterator_tag(mp, xfs_inode_free_eofblocks, flags, |
8ca149de | 1276 | eofb, XFS_ICI_EOFBLOCKS_TAG); |
41176a68 BF |
1277 | } |
1278 | ||
27b52867 BF |
1279 | void |
1280 | xfs_inode_set_eofblocks_tag( | |
1281 | xfs_inode_t *ip) | |
1282 | { | |
1283 | struct xfs_mount *mp = ip->i_mount; | |
1284 | struct xfs_perag *pag; | |
1285 | int tagged; | |
1286 | ||
1287 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); | |
1288 | spin_lock(&pag->pag_ici_lock); | |
1289 | trace_xfs_inode_set_eofblocks_tag(ip); | |
1290 | ||
1291 | tagged = radix_tree_tagged(&pag->pag_ici_root, | |
1292 | XFS_ICI_EOFBLOCKS_TAG); | |
1293 | radix_tree_tag_set(&pag->pag_ici_root, | |
1294 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
1295 | XFS_ICI_EOFBLOCKS_TAG); | |
1296 | if (!tagged) { | |
1297 | /* propagate the eofblocks tag up into the perag radix tree */ | |
1298 | spin_lock(&ip->i_mount->m_perag_lock); | |
1299 | radix_tree_tag_set(&ip->i_mount->m_perag_tree, | |
1300 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
1301 | XFS_ICI_EOFBLOCKS_TAG); | |
1302 | spin_unlock(&ip->i_mount->m_perag_lock); | |
579b62fa BF |
1303 | |
1304 | /* kick off background trimming */ | |
1305 | xfs_queue_eofblocks(ip->i_mount); | |
27b52867 BF |
1306 | |
1307 | trace_xfs_perag_set_eofblocks(ip->i_mount, pag->pag_agno, | |
1308 | -1, _RET_IP_); | |
1309 | } | |
1310 | ||
1311 | spin_unlock(&pag->pag_ici_lock); | |
1312 | xfs_perag_put(pag); | |
1313 | } | |
1314 | ||
1315 | void | |
1316 | xfs_inode_clear_eofblocks_tag( | |
1317 | xfs_inode_t *ip) | |
1318 | { | |
1319 | struct xfs_mount *mp = ip->i_mount; | |
1320 | struct xfs_perag *pag; | |
1321 | ||
1322 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); | |
1323 | spin_lock(&pag->pag_ici_lock); | |
1324 | trace_xfs_inode_clear_eofblocks_tag(ip); | |
1325 | ||
1326 | radix_tree_tag_clear(&pag->pag_ici_root, | |
1327 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
1328 | XFS_ICI_EOFBLOCKS_TAG); | |
1329 | if (!radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_EOFBLOCKS_TAG)) { | |
1330 | /* clear the eofblocks tag from the perag radix tree */ | |
1331 | spin_lock(&ip->i_mount->m_perag_lock); | |
1332 | radix_tree_tag_clear(&ip->i_mount->m_perag_tree, | |
1333 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
1334 | XFS_ICI_EOFBLOCKS_TAG); | |
1335 | spin_unlock(&ip->i_mount->m_perag_lock); | |
1336 | trace_xfs_perag_clear_eofblocks(ip->i_mount, pag->pag_agno, | |
1337 | -1, _RET_IP_); | |
1338 | } | |
1339 | ||
1340 | spin_unlock(&pag->pag_ici_lock); | |
1341 | xfs_perag_put(pag); | |
1342 | } | |
1343 |