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" | |
6ca1c906 | 20 | #include "xfs_format.h" |
239880ef DC |
21 | #include "xfs_log_format.h" |
22 | #include "xfs_trans_resv.h" | |
fe4fa4b8 | 23 | #include "xfs_inum.h" |
fe4fa4b8 DC |
24 | #include "xfs_sb.h" |
25 | #include "xfs_ag.h" | |
fe4fa4b8 | 26 | #include "xfs_mount.h" |
fe4fa4b8 | 27 | #include "xfs_inode.h" |
fe4fa4b8 | 28 | #include "xfs_error.h" |
239880ef DC |
29 | #include "xfs_trans.h" |
30 | #include "xfs_trans_priv.h" | |
fe4fa4b8 | 31 | #include "xfs_inode_item.h" |
7d095257 | 32 | #include "xfs_quota.h" |
0b1b213f | 33 | #include "xfs_trace.h" |
6d8b79cf | 34 | #include "xfs_icache.h" |
c24b5dfa | 35 | #include "xfs_bmap_util.h" |
dc06f398 BF |
36 | #include "xfs_quota.h" |
37 | #include "xfs_dquot_item.h" | |
38 | #include "xfs_dquot.h" | |
fe4fa4b8 | 39 | |
a167b17e DC |
40 | #include <linux/kthread.h> |
41 | #include <linux/freezer.h> | |
42 | ||
33479e05 DC |
43 | STATIC void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp, |
44 | struct xfs_perag *pag, struct xfs_inode *ip); | |
45 | ||
46 | /* | |
47 | * Allocate and initialise an xfs_inode. | |
48 | */ | |
638f4416 | 49 | struct xfs_inode * |
33479e05 DC |
50 | xfs_inode_alloc( |
51 | struct xfs_mount *mp, | |
52 | xfs_ino_t ino) | |
53 | { | |
54 | struct xfs_inode *ip; | |
55 | ||
56 | /* | |
57 | * if this didn't occur in transactions, we could use | |
58 | * KM_MAYFAIL and return NULL here on ENOMEM. Set the | |
59 | * code up to do this anyway. | |
60 | */ | |
61 | ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP); | |
62 | if (!ip) | |
63 | return NULL; | |
64 | if (inode_init_always(mp->m_super, VFS_I(ip))) { | |
65 | kmem_zone_free(xfs_inode_zone, ip); | |
66 | return NULL; | |
67 | } | |
68 | ||
69 | ASSERT(atomic_read(&ip->i_pincount) == 0); | |
70 | ASSERT(!spin_is_locked(&ip->i_flags_lock)); | |
71 | ASSERT(!xfs_isiflocked(ip)); | |
72 | ASSERT(ip->i_ino == 0); | |
73 | ||
74 | mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino); | |
75 | ||
76 | /* initialise the xfs inode */ | |
77 | ip->i_ino = ino; | |
78 | ip->i_mount = mp; | |
79 | memset(&ip->i_imap, 0, sizeof(struct xfs_imap)); | |
80 | ip->i_afp = NULL; | |
81 | memset(&ip->i_df, 0, sizeof(xfs_ifork_t)); | |
82 | ip->i_flags = 0; | |
83 | ip->i_delayed_blks = 0; | |
84 | memset(&ip->i_d, 0, sizeof(xfs_icdinode_t)); | |
85 | ||
86 | return ip; | |
87 | } | |
88 | ||
89 | STATIC void | |
90 | xfs_inode_free_callback( | |
91 | struct rcu_head *head) | |
92 | { | |
93 | struct inode *inode = container_of(head, struct inode, i_rcu); | |
94 | struct xfs_inode *ip = XFS_I(inode); | |
95 | ||
96 | kmem_zone_free(xfs_inode_zone, ip); | |
97 | } | |
98 | ||
638f4416 | 99 | void |
33479e05 DC |
100 | xfs_inode_free( |
101 | struct xfs_inode *ip) | |
102 | { | |
103 | switch (ip->i_d.di_mode & S_IFMT) { | |
104 | case S_IFREG: | |
105 | case S_IFDIR: | |
106 | case S_IFLNK: | |
107 | xfs_idestroy_fork(ip, XFS_DATA_FORK); | |
108 | break; | |
109 | } | |
110 | ||
111 | if (ip->i_afp) | |
112 | xfs_idestroy_fork(ip, XFS_ATTR_FORK); | |
113 | ||
114 | if (ip->i_itemp) { | |
115 | ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL)); | |
116 | xfs_inode_item_destroy(ip); | |
117 | ip->i_itemp = NULL; | |
118 | } | |
119 | ||
33479e05 DC |
120 | /* |
121 | * Because we use RCU freeing we need to ensure the inode always | |
122 | * appears to be reclaimed with an invalid inode number when in the | |
123 | * free state. The ip->i_flags_lock provides the barrier against lookup | |
124 | * races. | |
125 | */ | |
126 | spin_lock(&ip->i_flags_lock); | |
127 | ip->i_flags = XFS_IRECLAIM; | |
128 | ip->i_ino = 0; | |
129 | spin_unlock(&ip->i_flags_lock); | |
130 | ||
b313a5f1 DC |
131 | /* asserts to verify all state is correct here */ |
132 | ASSERT(atomic_read(&ip->i_pincount) == 0); | |
133 | ASSERT(!xfs_isiflocked(ip)); | |
134 | ||
33479e05 DC |
135 | call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback); |
136 | } | |
137 | ||
138 | /* | |
139 | * Check the validity of the inode we just found it the cache | |
140 | */ | |
141 | static int | |
142 | xfs_iget_cache_hit( | |
143 | struct xfs_perag *pag, | |
144 | struct xfs_inode *ip, | |
145 | xfs_ino_t ino, | |
146 | int flags, | |
147 | int lock_flags) __releases(RCU) | |
148 | { | |
149 | struct inode *inode = VFS_I(ip); | |
150 | struct xfs_mount *mp = ip->i_mount; | |
151 | int error; | |
152 | ||
153 | /* | |
154 | * check for re-use of an inode within an RCU grace period due to the | |
155 | * radix tree nodes not being updated yet. We monitor for this by | |
156 | * setting the inode number to zero before freeing the inode structure. | |
157 | * If the inode has been reallocated and set up, then the inode number | |
158 | * will not match, so check for that, too. | |
159 | */ | |
160 | spin_lock(&ip->i_flags_lock); | |
161 | if (ip->i_ino != ino) { | |
162 | trace_xfs_iget_skip(ip); | |
163 | XFS_STATS_INC(xs_ig_frecycle); | |
2451337d | 164 | error = -EAGAIN; |
33479e05 DC |
165 | goto out_error; |
166 | } | |
167 | ||
168 | ||
169 | /* | |
170 | * If we are racing with another cache hit that is currently | |
171 | * instantiating this inode or currently recycling it out of | |
172 | * reclaimabe state, wait for the initialisation to complete | |
173 | * before continuing. | |
174 | * | |
175 | * XXX(hch): eventually we should do something equivalent to | |
176 | * wait_on_inode to wait for these flags to be cleared | |
177 | * instead of polling for it. | |
178 | */ | |
179 | if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) { | |
180 | trace_xfs_iget_skip(ip); | |
181 | XFS_STATS_INC(xs_ig_frecycle); | |
2451337d | 182 | error = -EAGAIN; |
33479e05 DC |
183 | goto out_error; |
184 | } | |
185 | ||
186 | /* | |
187 | * If lookup is racing with unlink return an error immediately. | |
188 | */ | |
189 | if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) { | |
2451337d | 190 | error = -ENOENT; |
33479e05 DC |
191 | goto out_error; |
192 | } | |
193 | ||
194 | /* | |
195 | * If IRECLAIMABLE is set, we've torn down the VFS inode already. | |
196 | * Need to carefully get it back into useable state. | |
197 | */ | |
198 | if (ip->i_flags & XFS_IRECLAIMABLE) { | |
199 | trace_xfs_iget_reclaim(ip); | |
200 | ||
201 | /* | |
202 | * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode | |
203 | * from stomping over us while we recycle the inode. We can't | |
204 | * clear the radix tree reclaimable tag yet as it requires | |
205 | * pag_ici_lock to be held exclusive. | |
206 | */ | |
207 | ip->i_flags |= XFS_IRECLAIM; | |
208 | ||
209 | spin_unlock(&ip->i_flags_lock); | |
210 | rcu_read_unlock(); | |
211 | ||
2451337d | 212 | error = inode_init_always(mp->m_super, inode); |
33479e05 DC |
213 | if (error) { |
214 | /* | |
215 | * Re-initializing the inode failed, and we are in deep | |
216 | * trouble. Try to re-add it to the reclaim list. | |
217 | */ | |
218 | rcu_read_lock(); | |
219 | spin_lock(&ip->i_flags_lock); | |
220 | ||
221 | ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM); | |
222 | ASSERT(ip->i_flags & XFS_IRECLAIMABLE); | |
223 | trace_xfs_iget_reclaim_fail(ip); | |
224 | goto out_error; | |
225 | } | |
226 | ||
227 | spin_lock(&pag->pag_ici_lock); | |
228 | spin_lock(&ip->i_flags_lock); | |
229 | ||
230 | /* | |
231 | * Clear the per-lifetime state in the inode as we are now | |
232 | * effectively a new inode and need to return to the initial | |
233 | * state before reuse occurs. | |
234 | */ | |
235 | ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS; | |
236 | ip->i_flags |= XFS_INEW; | |
237 | __xfs_inode_clear_reclaim_tag(mp, pag, ip); | |
238 | inode->i_state = I_NEW; | |
239 | ||
240 | ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock)); | |
241 | mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino); | |
242 | ||
243 | spin_unlock(&ip->i_flags_lock); | |
244 | spin_unlock(&pag->pag_ici_lock); | |
245 | } else { | |
246 | /* If the VFS inode is being torn down, pause and try again. */ | |
247 | if (!igrab(inode)) { | |
248 | trace_xfs_iget_skip(ip); | |
2451337d | 249 | error = -EAGAIN; |
33479e05 DC |
250 | goto out_error; |
251 | } | |
252 | ||
253 | /* We've got a live one. */ | |
254 | spin_unlock(&ip->i_flags_lock); | |
255 | rcu_read_unlock(); | |
256 | trace_xfs_iget_hit(ip); | |
257 | } | |
258 | ||
259 | if (lock_flags != 0) | |
260 | xfs_ilock(ip, lock_flags); | |
261 | ||
262 | xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE); | |
263 | XFS_STATS_INC(xs_ig_found); | |
264 | ||
265 | return 0; | |
266 | ||
267 | out_error: | |
268 | spin_unlock(&ip->i_flags_lock); | |
269 | rcu_read_unlock(); | |
270 | return error; | |
271 | } | |
272 | ||
273 | ||
274 | static int | |
275 | xfs_iget_cache_miss( | |
276 | struct xfs_mount *mp, | |
277 | struct xfs_perag *pag, | |
278 | xfs_trans_t *tp, | |
279 | xfs_ino_t ino, | |
280 | struct xfs_inode **ipp, | |
281 | int flags, | |
282 | int lock_flags) | |
283 | { | |
284 | struct xfs_inode *ip; | |
285 | int error; | |
286 | xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino); | |
287 | int iflags; | |
288 | ||
289 | ip = xfs_inode_alloc(mp, ino); | |
290 | if (!ip) | |
2451337d | 291 | return -ENOMEM; |
33479e05 DC |
292 | |
293 | error = xfs_iread(mp, tp, ip, flags); | |
294 | if (error) | |
295 | goto out_destroy; | |
296 | ||
297 | trace_xfs_iget_miss(ip); | |
298 | ||
299 | if ((ip->i_d.di_mode == 0) && !(flags & XFS_IGET_CREATE)) { | |
2451337d | 300 | error = -ENOENT; |
33479e05 DC |
301 | goto out_destroy; |
302 | } | |
303 | ||
304 | /* | |
305 | * Preload the radix tree so we can insert safely under the | |
306 | * write spinlock. Note that we cannot sleep inside the preload | |
307 | * region. Since we can be called from transaction context, don't | |
308 | * recurse into the file system. | |
309 | */ | |
310 | if (radix_tree_preload(GFP_NOFS)) { | |
2451337d | 311 | error = -EAGAIN; |
33479e05 DC |
312 | goto out_destroy; |
313 | } | |
314 | ||
315 | /* | |
316 | * Because the inode hasn't been added to the radix-tree yet it can't | |
317 | * be found by another thread, so we can do the non-sleeping lock here. | |
318 | */ | |
319 | if (lock_flags) { | |
320 | if (!xfs_ilock_nowait(ip, lock_flags)) | |
321 | BUG(); | |
322 | } | |
323 | ||
324 | /* | |
325 | * These values must be set before inserting the inode into the radix | |
326 | * tree as the moment it is inserted a concurrent lookup (allowed by the | |
327 | * RCU locking mechanism) can find it and that lookup must see that this | |
328 | * is an inode currently under construction (i.e. that XFS_INEW is set). | |
329 | * The ip->i_flags_lock that protects the XFS_INEW flag forms the | |
330 | * memory barrier that ensures this detection works correctly at lookup | |
331 | * time. | |
332 | */ | |
333 | iflags = XFS_INEW; | |
334 | if (flags & XFS_IGET_DONTCACHE) | |
335 | iflags |= XFS_IDONTCACHE; | |
113a5683 CS |
336 | ip->i_udquot = NULL; |
337 | ip->i_gdquot = NULL; | |
92f8ff73 | 338 | ip->i_pdquot = NULL; |
33479e05 DC |
339 | xfs_iflags_set(ip, iflags); |
340 | ||
341 | /* insert the new inode */ | |
342 | spin_lock(&pag->pag_ici_lock); | |
343 | error = radix_tree_insert(&pag->pag_ici_root, agino, ip); | |
344 | if (unlikely(error)) { | |
345 | WARN_ON(error != -EEXIST); | |
346 | XFS_STATS_INC(xs_ig_dup); | |
2451337d | 347 | error = -EAGAIN; |
33479e05 DC |
348 | goto out_preload_end; |
349 | } | |
350 | spin_unlock(&pag->pag_ici_lock); | |
351 | radix_tree_preload_end(); | |
352 | ||
353 | *ipp = ip; | |
354 | return 0; | |
355 | ||
356 | out_preload_end: | |
357 | spin_unlock(&pag->pag_ici_lock); | |
358 | radix_tree_preload_end(); | |
359 | if (lock_flags) | |
360 | xfs_iunlock(ip, lock_flags); | |
361 | out_destroy: | |
362 | __destroy_inode(VFS_I(ip)); | |
363 | xfs_inode_free(ip); | |
364 | return error; | |
365 | } | |
366 | ||
367 | /* | |
368 | * Look up an inode by number in the given file system. | |
369 | * The inode is looked up in the cache held in each AG. | |
370 | * If the inode is found in the cache, initialise the vfs inode | |
371 | * if necessary. | |
372 | * | |
373 | * If it is not in core, read it in from the file system's device, | |
374 | * add it to the cache and initialise the vfs inode. | |
375 | * | |
376 | * The inode is locked according to the value of the lock_flags parameter. | |
377 | * This flag parameter indicates how and if the inode's IO lock and inode lock | |
378 | * should be taken. | |
379 | * | |
380 | * mp -- the mount point structure for the current file system. It points | |
381 | * to the inode hash table. | |
382 | * tp -- a pointer to the current transaction if there is one. This is | |
383 | * simply passed through to the xfs_iread() call. | |
384 | * ino -- the number of the inode desired. This is the unique identifier | |
385 | * within the file system for the inode being requested. | |
386 | * lock_flags -- flags indicating how to lock the inode. See the comment | |
387 | * for xfs_ilock() for a list of valid values. | |
388 | */ | |
389 | int | |
390 | xfs_iget( | |
391 | xfs_mount_t *mp, | |
392 | xfs_trans_t *tp, | |
393 | xfs_ino_t ino, | |
394 | uint flags, | |
395 | uint lock_flags, | |
396 | xfs_inode_t **ipp) | |
397 | { | |
398 | xfs_inode_t *ip; | |
399 | int error; | |
400 | xfs_perag_t *pag; | |
401 | xfs_agino_t agino; | |
402 | ||
403 | /* | |
404 | * xfs_reclaim_inode() uses the ILOCK to ensure an inode | |
405 | * doesn't get freed while it's being referenced during a | |
406 | * radix tree traversal here. It assumes this function | |
407 | * aqcuires only the ILOCK (and therefore it has no need to | |
408 | * involve the IOLOCK in this synchronization). | |
409 | */ | |
410 | ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0); | |
411 | ||
412 | /* reject inode numbers outside existing AGs */ | |
413 | if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount) | |
2451337d | 414 | return -EINVAL; |
33479e05 DC |
415 | |
416 | /* get the perag structure and ensure that it's inode capable */ | |
417 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino)); | |
418 | agino = XFS_INO_TO_AGINO(mp, ino); | |
419 | ||
420 | again: | |
421 | error = 0; | |
422 | rcu_read_lock(); | |
423 | ip = radix_tree_lookup(&pag->pag_ici_root, agino); | |
424 | ||
425 | if (ip) { | |
426 | error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags); | |
427 | if (error) | |
428 | goto out_error_or_again; | |
429 | } else { | |
430 | rcu_read_unlock(); | |
431 | XFS_STATS_INC(xs_ig_missed); | |
432 | ||
433 | error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip, | |
434 | flags, lock_flags); | |
435 | if (error) | |
436 | goto out_error_or_again; | |
437 | } | |
438 | xfs_perag_put(pag); | |
439 | ||
440 | *ipp = ip; | |
441 | ||
442 | /* | |
443 | * If we have a real type for an on-disk inode, we can set ops(&unlock) | |
444 | * now. If it's a new inode being created, xfs_ialloc will handle it. | |
445 | */ | |
446 | if (xfs_iflags_test(ip, XFS_INEW) && ip->i_d.di_mode != 0) | |
447 | xfs_setup_inode(ip); | |
448 | return 0; | |
449 | ||
450 | out_error_or_again: | |
2451337d | 451 | if (error == -EAGAIN) { |
33479e05 DC |
452 | delay(1); |
453 | goto again; | |
454 | } | |
455 | xfs_perag_put(pag); | |
456 | return error; | |
457 | } | |
458 | ||
78ae5256 DC |
459 | /* |
460 | * The inode lookup is done in batches to keep the amount of lock traffic and | |
461 | * radix tree lookups to a minimum. The batch size is a trade off between | |
462 | * lookup reduction and stack usage. This is in the reclaim path, so we can't | |
463 | * be too greedy. | |
464 | */ | |
465 | #define XFS_LOOKUP_BATCH 32 | |
466 | ||
e13de955 DC |
467 | STATIC int |
468 | xfs_inode_ag_walk_grab( | |
469 | struct xfs_inode *ip) | |
470 | { | |
471 | struct inode *inode = VFS_I(ip); | |
472 | ||
1a3e8f3d DC |
473 | ASSERT(rcu_read_lock_held()); |
474 | ||
475 | /* | |
476 | * check for stale RCU freed inode | |
477 | * | |
478 | * If the inode has been reallocated, it doesn't matter if it's not in | |
479 | * the AG we are walking - we are walking for writeback, so if it | |
480 | * passes all the "valid inode" checks and is dirty, then we'll write | |
481 | * it back anyway. If it has been reallocated and still being | |
482 | * initialised, the XFS_INEW check below will catch it. | |
483 | */ | |
484 | spin_lock(&ip->i_flags_lock); | |
485 | if (!ip->i_ino) | |
486 | goto out_unlock_noent; | |
487 | ||
488 | /* avoid new or reclaimable inodes. Leave for reclaim code to flush */ | |
489 | if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM)) | |
490 | goto out_unlock_noent; | |
491 | spin_unlock(&ip->i_flags_lock); | |
492 | ||
e13de955 DC |
493 | /* nothing to sync during shutdown */ |
494 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) | |
2451337d | 495 | return -EFSCORRUPTED; |
e13de955 | 496 | |
e13de955 DC |
497 | /* If we can't grab the inode, it must on it's way to reclaim. */ |
498 | if (!igrab(inode)) | |
2451337d | 499 | return -ENOENT; |
e13de955 | 500 | |
e13de955 DC |
501 | /* inode is valid */ |
502 | return 0; | |
1a3e8f3d DC |
503 | |
504 | out_unlock_noent: | |
505 | spin_unlock(&ip->i_flags_lock); | |
2451337d | 506 | return -ENOENT; |
e13de955 DC |
507 | } |
508 | ||
75f3cb13 DC |
509 | STATIC int |
510 | xfs_inode_ag_walk( | |
511 | struct xfs_mount *mp, | |
5017e97d | 512 | struct xfs_perag *pag, |
e0094008 | 513 | int (*execute)(struct xfs_inode *ip, int flags, |
a454f742 BF |
514 | void *args), |
515 | int flags, | |
516 | void *args, | |
517 | int tag) | |
75f3cb13 | 518 | { |
75f3cb13 DC |
519 | uint32_t first_index; |
520 | int last_error = 0; | |
521 | int skipped; | |
65d0f205 | 522 | int done; |
78ae5256 | 523 | int nr_found; |
75f3cb13 DC |
524 | |
525 | restart: | |
65d0f205 | 526 | done = 0; |
75f3cb13 DC |
527 | skipped = 0; |
528 | first_index = 0; | |
78ae5256 | 529 | nr_found = 0; |
75f3cb13 | 530 | do { |
78ae5256 | 531 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
75f3cb13 | 532 | int error = 0; |
78ae5256 | 533 | int i; |
75f3cb13 | 534 | |
1a3e8f3d | 535 | rcu_read_lock(); |
a454f742 BF |
536 | |
537 | if (tag == -1) | |
538 | nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, | |
78ae5256 DC |
539 | (void **)batch, first_index, |
540 | XFS_LOOKUP_BATCH); | |
a454f742 BF |
541 | else |
542 | nr_found = radix_tree_gang_lookup_tag( | |
543 | &pag->pag_ici_root, | |
544 | (void **) batch, first_index, | |
545 | XFS_LOOKUP_BATCH, tag); | |
546 | ||
65d0f205 | 547 | if (!nr_found) { |
1a3e8f3d | 548 | rcu_read_unlock(); |
75f3cb13 | 549 | break; |
c8e20be0 | 550 | } |
75f3cb13 | 551 | |
65d0f205 | 552 | /* |
78ae5256 DC |
553 | * Grab the inodes before we drop the lock. if we found |
554 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 555 | */ |
78ae5256 DC |
556 | for (i = 0; i < nr_found; i++) { |
557 | struct xfs_inode *ip = batch[i]; | |
558 | ||
559 | if (done || xfs_inode_ag_walk_grab(ip)) | |
560 | batch[i] = NULL; | |
561 | ||
562 | /* | |
1a3e8f3d DC |
563 | * Update the index for the next lookup. Catch |
564 | * overflows into the next AG range which can occur if | |
565 | * we have inodes in the last block of the AG and we | |
566 | * are currently pointing to the last inode. | |
567 | * | |
568 | * Because we may see inodes that are from the wrong AG | |
569 | * due to RCU freeing and reallocation, only update the | |
570 | * index if it lies in this AG. It was a race that lead | |
571 | * us to see this inode, so another lookup from the | |
572 | * same index will not find it again. | |
78ae5256 | 573 | */ |
1a3e8f3d DC |
574 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno) |
575 | continue; | |
78ae5256 DC |
576 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
577 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
578 | done = 1; | |
e13de955 | 579 | } |
78ae5256 DC |
580 | |
581 | /* unlock now we've grabbed the inodes. */ | |
1a3e8f3d | 582 | rcu_read_unlock(); |
e13de955 | 583 | |
78ae5256 DC |
584 | for (i = 0; i < nr_found; i++) { |
585 | if (!batch[i]) | |
586 | continue; | |
e0094008 | 587 | error = execute(batch[i], flags, args); |
78ae5256 | 588 | IRELE(batch[i]); |
2451337d | 589 | if (error == -EAGAIN) { |
78ae5256 DC |
590 | skipped++; |
591 | continue; | |
592 | } | |
2451337d | 593 | if (error && last_error != -EFSCORRUPTED) |
78ae5256 | 594 | last_error = error; |
75f3cb13 | 595 | } |
c8e20be0 DC |
596 | |
597 | /* bail out if the filesystem is corrupted. */ | |
2451337d | 598 | if (error == -EFSCORRUPTED) |
75f3cb13 DC |
599 | break; |
600 | ||
8daaa831 DC |
601 | cond_resched(); |
602 | ||
78ae5256 | 603 | } while (nr_found && !done); |
75f3cb13 DC |
604 | |
605 | if (skipped) { | |
606 | delay(1); | |
607 | goto restart; | |
608 | } | |
75f3cb13 DC |
609 | return last_error; |
610 | } | |
611 | ||
579b62fa BF |
612 | /* |
613 | * Background scanning to trim post-EOF preallocated space. This is queued | |
b9fe5052 | 614 | * based on the 'speculative_prealloc_lifetime' tunable (5m by default). |
579b62fa BF |
615 | */ |
616 | STATIC void | |
617 | xfs_queue_eofblocks( | |
618 | struct xfs_mount *mp) | |
619 | { | |
620 | rcu_read_lock(); | |
621 | if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_EOFBLOCKS_TAG)) | |
622 | queue_delayed_work(mp->m_eofblocks_workqueue, | |
623 | &mp->m_eofblocks_work, | |
624 | msecs_to_jiffies(xfs_eofb_secs * 1000)); | |
625 | rcu_read_unlock(); | |
626 | } | |
627 | ||
628 | void | |
629 | xfs_eofblocks_worker( | |
630 | struct work_struct *work) | |
631 | { | |
632 | struct xfs_mount *mp = container_of(to_delayed_work(work), | |
633 | struct xfs_mount, m_eofblocks_work); | |
634 | xfs_icache_free_eofblocks(mp, NULL); | |
635 | xfs_queue_eofblocks(mp); | |
636 | } | |
637 | ||
fe588ed3 | 638 | int |
75f3cb13 DC |
639 | xfs_inode_ag_iterator( |
640 | struct xfs_mount *mp, | |
e0094008 | 641 | int (*execute)(struct xfs_inode *ip, int flags, |
a454f742 BF |
642 | void *args), |
643 | int flags, | |
644 | void *args) | |
75f3cb13 | 645 | { |
16fd5367 | 646 | struct xfs_perag *pag; |
75f3cb13 DC |
647 | int error = 0; |
648 | int last_error = 0; | |
649 | xfs_agnumber_t ag; | |
650 | ||
16fd5367 | 651 | ag = 0; |
65d0f205 DC |
652 | while ((pag = xfs_perag_get(mp, ag))) { |
653 | ag = pag->pag_agno + 1; | |
a454f742 BF |
654 | error = xfs_inode_ag_walk(mp, pag, execute, flags, args, -1); |
655 | xfs_perag_put(pag); | |
656 | if (error) { | |
657 | last_error = error; | |
2451337d | 658 | if (error == -EFSCORRUPTED) |
a454f742 BF |
659 | break; |
660 | } | |
661 | } | |
b474c7ae | 662 | return last_error; |
a454f742 BF |
663 | } |
664 | ||
665 | int | |
666 | xfs_inode_ag_iterator_tag( | |
667 | struct xfs_mount *mp, | |
e0094008 | 668 | int (*execute)(struct xfs_inode *ip, int flags, |
a454f742 BF |
669 | void *args), |
670 | int flags, | |
671 | void *args, | |
672 | int tag) | |
673 | { | |
674 | struct xfs_perag *pag; | |
675 | int error = 0; | |
676 | int last_error = 0; | |
677 | xfs_agnumber_t ag; | |
678 | ||
679 | ag = 0; | |
680 | while ((pag = xfs_perag_get_tag(mp, ag, tag))) { | |
681 | ag = pag->pag_agno + 1; | |
682 | error = xfs_inode_ag_walk(mp, pag, execute, flags, args, tag); | |
5017e97d | 683 | xfs_perag_put(pag); |
75f3cb13 DC |
684 | if (error) { |
685 | last_error = error; | |
2451337d | 686 | if (error == -EFSCORRUPTED) |
75f3cb13 DC |
687 | break; |
688 | } | |
689 | } | |
b474c7ae | 690 | return last_error; |
75f3cb13 DC |
691 | } |
692 | ||
a7b339f1 DC |
693 | /* |
694 | * Queue a new inode reclaim pass if there are reclaimable inodes and there | |
695 | * isn't a reclaim pass already in progress. By default it runs every 5s based | |
5889608d | 696 | * on the xfs periodic sync default of 30s. Perhaps this should have it's own |
a7b339f1 DC |
697 | * tunable, but that can be done if this method proves to be ineffective or too |
698 | * aggressive. | |
699 | */ | |
700 | static void | |
5889608d | 701 | xfs_reclaim_work_queue( |
a7b339f1 | 702 | struct xfs_mount *mp) |
a167b17e | 703 | { |
a167b17e | 704 | |
a7b339f1 DC |
705 | rcu_read_lock(); |
706 | if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) { | |
5889608d | 707 | queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work, |
a7b339f1 | 708 | msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10)); |
a167b17e | 709 | } |
a7b339f1 DC |
710 | rcu_read_unlock(); |
711 | } | |
a167b17e | 712 | |
a7b339f1 DC |
713 | /* |
714 | * This is a fast pass over the inode cache to try to get reclaim moving on as | |
715 | * many inodes as possible in a short period of time. It kicks itself every few | |
716 | * seconds, as well as being kicked by the inode cache shrinker when memory | |
717 | * goes low. It scans as quickly as possible avoiding locked inodes or those | |
718 | * already being flushed, and once done schedules a future pass. | |
719 | */ | |
33c7a2bc | 720 | void |
a7b339f1 DC |
721 | xfs_reclaim_worker( |
722 | struct work_struct *work) | |
723 | { | |
724 | struct xfs_mount *mp = container_of(to_delayed_work(work), | |
725 | struct xfs_mount, m_reclaim_work); | |
726 | ||
727 | xfs_reclaim_inodes(mp, SYNC_TRYLOCK); | |
5889608d | 728 | xfs_reclaim_work_queue(mp); |
a7b339f1 DC |
729 | } |
730 | ||
33479e05 | 731 | static void |
bc990f5c CH |
732 | __xfs_inode_set_reclaim_tag( |
733 | struct xfs_perag *pag, | |
734 | struct xfs_inode *ip) | |
735 | { | |
736 | radix_tree_tag_set(&pag->pag_ici_root, | |
737 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
738 | XFS_ICI_RECLAIM_TAG); | |
16fd5367 DC |
739 | |
740 | if (!pag->pag_ici_reclaimable) { | |
741 | /* propagate the reclaim tag up into the perag radix tree */ | |
742 | spin_lock(&ip->i_mount->m_perag_lock); | |
743 | radix_tree_tag_set(&ip->i_mount->m_perag_tree, | |
744 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
745 | XFS_ICI_RECLAIM_TAG); | |
746 | spin_unlock(&ip->i_mount->m_perag_lock); | |
a7b339f1 DC |
747 | |
748 | /* schedule periodic background inode reclaim */ | |
5889608d | 749 | xfs_reclaim_work_queue(ip->i_mount); |
a7b339f1 | 750 | |
16fd5367 DC |
751 | trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno, |
752 | -1, _RET_IP_); | |
753 | } | |
9bf729c0 | 754 | pag->pag_ici_reclaimable++; |
bc990f5c CH |
755 | } |
756 | ||
11654513 DC |
757 | /* |
758 | * We set the inode flag atomically with the radix tree tag. | |
759 | * Once we get tag lookups on the radix tree, this inode flag | |
760 | * can go away. | |
761 | */ | |
396beb85 DC |
762 | void |
763 | xfs_inode_set_reclaim_tag( | |
764 | xfs_inode_t *ip) | |
765 | { | |
5017e97d DC |
766 | struct xfs_mount *mp = ip->i_mount; |
767 | struct xfs_perag *pag; | |
396beb85 | 768 | |
5017e97d | 769 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); |
1a427ab0 | 770 | spin_lock(&pag->pag_ici_lock); |
396beb85 | 771 | spin_lock(&ip->i_flags_lock); |
bc990f5c | 772 | __xfs_inode_set_reclaim_tag(pag, ip); |
11654513 | 773 | __xfs_iflags_set(ip, XFS_IRECLAIMABLE); |
396beb85 | 774 | spin_unlock(&ip->i_flags_lock); |
1a427ab0 | 775 | spin_unlock(&pag->pag_ici_lock); |
5017e97d | 776 | xfs_perag_put(pag); |
396beb85 DC |
777 | } |
778 | ||
081003ff JW |
779 | STATIC void |
780 | __xfs_inode_clear_reclaim( | |
396beb85 DC |
781 | xfs_perag_t *pag, |
782 | xfs_inode_t *ip) | |
783 | { | |
9bf729c0 | 784 | pag->pag_ici_reclaimable--; |
16fd5367 DC |
785 | if (!pag->pag_ici_reclaimable) { |
786 | /* clear the reclaim tag from the perag radix tree */ | |
787 | spin_lock(&ip->i_mount->m_perag_lock); | |
788 | radix_tree_tag_clear(&ip->i_mount->m_perag_tree, | |
789 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
790 | XFS_ICI_RECLAIM_TAG); | |
791 | spin_unlock(&ip->i_mount->m_perag_lock); | |
792 | trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno, | |
793 | -1, _RET_IP_); | |
794 | } | |
396beb85 DC |
795 | } |
796 | ||
33479e05 | 797 | STATIC void |
081003ff JW |
798 | __xfs_inode_clear_reclaim_tag( |
799 | xfs_mount_t *mp, | |
800 | xfs_perag_t *pag, | |
801 | xfs_inode_t *ip) | |
802 | { | |
803 | radix_tree_tag_clear(&pag->pag_ici_root, | |
804 | XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG); | |
805 | __xfs_inode_clear_reclaim(pag, ip); | |
806 | } | |
807 | ||
e3a20c0b DC |
808 | /* |
809 | * Grab the inode for reclaim exclusively. | |
810 | * Return 0 if we grabbed it, non-zero otherwise. | |
811 | */ | |
812 | STATIC int | |
813 | xfs_reclaim_inode_grab( | |
814 | struct xfs_inode *ip, | |
815 | int flags) | |
816 | { | |
1a3e8f3d DC |
817 | ASSERT(rcu_read_lock_held()); |
818 | ||
819 | /* quick check for stale RCU freed inode */ | |
820 | if (!ip->i_ino) | |
821 | return 1; | |
e3a20c0b DC |
822 | |
823 | /* | |
474fce06 CH |
824 | * If we are asked for non-blocking operation, do unlocked checks to |
825 | * see if the inode already is being flushed or in reclaim to avoid | |
826 | * lock traffic. | |
e3a20c0b DC |
827 | */ |
828 | if ((flags & SYNC_TRYLOCK) && | |
474fce06 | 829 | __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM)) |
e3a20c0b | 830 | return 1; |
e3a20c0b DC |
831 | |
832 | /* | |
833 | * The radix tree lock here protects a thread in xfs_iget from racing | |
834 | * with us starting reclaim on the inode. Once we have the | |
835 | * XFS_IRECLAIM flag set it will not touch us. | |
1a3e8f3d DC |
836 | * |
837 | * Due to RCU lookup, we may find inodes that have been freed and only | |
838 | * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that | |
839 | * aren't candidates for reclaim at all, so we must check the | |
840 | * XFS_IRECLAIMABLE is set first before proceeding to reclaim. | |
e3a20c0b DC |
841 | */ |
842 | spin_lock(&ip->i_flags_lock); | |
1a3e8f3d DC |
843 | if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) || |
844 | __xfs_iflags_test(ip, XFS_IRECLAIM)) { | |
845 | /* not a reclaim candidate. */ | |
e3a20c0b DC |
846 | spin_unlock(&ip->i_flags_lock); |
847 | return 1; | |
848 | } | |
849 | __xfs_iflags_set(ip, XFS_IRECLAIM); | |
850 | spin_unlock(&ip->i_flags_lock); | |
851 | return 0; | |
852 | } | |
853 | ||
777df5af | 854 | /* |
8a48088f CH |
855 | * Inodes in different states need to be treated differently. The following |
856 | * table lists the inode states and the reclaim actions necessary: | |
777df5af DC |
857 | * |
858 | * inode state iflush ret required action | |
859 | * --------------- ---------- --------------- | |
860 | * bad - reclaim | |
861 | * shutdown EIO unpin and reclaim | |
862 | * clean, unpinned 0 reclaim | |
863 | * stale, unpinned 0 reclaim | |
c854363e DC |
864 | * clean, pinned(*) 0 requeue |
865 | * stale, pinned EAGAIN requeue | |
8a48088f CH |
866 | * dirty, async - requeue |
867 | * dirty, sync 0 reclaim | |
777df5af DC |
868 | * |
869 | * (*) dgc: I don't think the clean, pinned state is possible but it gets | |
870 | * handled anyway given the order of checks implemented. | |
871 | * | |
c854363e DC |
872 | * Also, because we get the flush lock first, we know that any inode that has |
873 | * been flushed delwri has had the flush completed by the time we check that | |
8a48088f | 874 | * the inode is clean. |
c854363e | 875 | * |
8a48088f CH |
876 | * Note that because the inode is flushed delayed write by AIL pushing, the |
877 | * flush lock may already be held here and waiting on it can result in very | |
878 | * long latencies. Hence for sync reclaims, where we wait on the flush lock, | |
879 | * the caller should push the AIL first before trying to reclaim inodes to | |
880 | * minimise the amount of time spent waiting. For background relaim, we only | |
881 | * bother to reclaim clean inodes anyway. | |
c854363e | 882 | * |
777df5af DC |
883 | * Hence the order of actions after gaining the locks should be: |
884 | * bad => reclaim | |
885 | * shutdown => unpin and reclaim | |
8a48088f | 886 | * pinned, async => requeue |
c854363e | 887 | * pinned, sync => unpin |
777df5af DC |
888 | * stale => reclaim |
889 | * clean => reclaim | |
8a48088f | 890 | * dirty, async => requeue |
c854363e | 891 | * dirty, sync => flush, wait and reclaim |
777df5af | 892 | */ |
75f3cb13 | 893 | STATIC int |
c8e20be0 | 894 | xfs_reclaim_inode( |
75f3cb13 DC |
895 | struct xfs_inode *ip, |
896 | struct xfs_perag *pag, | |
c8e20be0 | 897 | int sync_mode) |
fce08f2f | 898 | { |
4c46819a CH |
899 | struct xfs_buf *bp = NULL; |
900 | int error; | |
777df5af | 901 | |
1bfd8d04 DC |
902 | restart: |
903 | error = 0; | |
c8e20be0 | 904 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
c854363e DC |
905 | if (!xfs_iflock_nowait(ip)) { |
906 | if (!(sync_mode & SYNC_WAIT)) | |
907 | goto out; | |
908 | xfs_iflock(ip); | |
909 | } | |
7a3be02b | 910 | |
777df5af DC |
911 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) { |
912 | xfs_iunpin_wait(ip); | |
04913fdd | 913 | xfs_iflush_abort(ip, false); |
777df5af DC |
914 | goto reclaim; |
915 | } | |
c854363e | 916 | if (xfs_ipincount(ip)) { |
8a48088f CH |
917 | if (!(sync_mode & SYNC_WAIT)) |
918 | goto out_ifunlock; | |
777df5af | 919 | xfs_iunpin_wait(ip); |
c854363e | 920 | } |
777df5af DC |
921 | if (xfs_iflags_test(ip, XFS_ISTALE)) |
922 | goto reclaim; | |
923 | if (xfs_inode_clean(ip)) | |
924 | goto reclaim; | |
925 | ||
8a48088f CH |
926 | /* |
927 | * Never flush out dirty data during non-blocking reclaim, as it would | |
928 | * just contend with AIL pushing trying to do the same job. | |
929 | */ | |
930 | if (!(sync_mode & SYNC_WAIT)) | |
931 | goto out_ifunlock; | |
932 | ||
1bfd8d04 DC |
933 | /* |
934 | * Now we have an inode that needs flushing. | |
935 | * | |
4c46819a | 936 | * Note that xfs_iflush will never block on the inode buffer lock, as |
1bfd8d04 | 937 | * xfs_ifree_cluster() can lock the inode buffer before it locks the |
4c46819a | 938 | * ip->i_lock, and we are doing the exact opposite here. As a result, |
475ee413 CH |
939 | * doing a blocking xfs_imap_to_bp() to get the cluster buffer would |
940 | * result in an ABBA deadlock with xfs_ifree_cluster(). | |
1bfd8d04 DC |
941 | * |
942 | * As xfs_ifree_cluser() must gather all inodes that are active in the | |
943 | * cache to mark them stale, if we hit this case we don't actually want | |
944 | * to do IO here - we want the inode marked stale so we can simply | |
4c46819a CH |
945 | * reclaim it. Hence if we get an EAGAIN error here, just unlock the |
946 | * inode, back off and try again. Hopefully the next pass through will | |
947 | * see the stale flag set on the inode. | |
1bfd8d04 | 948 | */ |
4c46819a | 949 | error = xfs_iflush(ip, &bp); |
2451337d | 950 | if (error == -EAGAIN) { |
8a48088f CH |
951 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
952 | /* backoff longer than in xfs_ifree_cluster */ | |
953 | delay(2); | |
954 | goto restart; | |
c854363e | 955 | } |
c854363e | 956 | |
4c46819a CH |
957 | if (!error) { |
958 | error = xfs_bwrite(bp); | |
959 | xfs_buf_relse(bp); | |
960 | } | |
961 | ||
962 | xfs_iflock(ip); | |
777df5af DC |
963 | reclaim: |
964 | xfs_ifunlock(ip); | |
c8e20be0 | 965 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
966 | |
967 | XFS_STATS_INC(xs_ig_reclaims); | |
968 | /* | |
969 | * Remove the inode from the per-AG radix tree. | |
970 | * | |
971 | * Because radix_tree_delete won't complain even if the item was never | |
972 | * added to the tree assert that it's been there before to catch | |
973 | * problems with the inode life time early on. | |
974 | */ | |
1a427ab0 | 975 | spin_lock(&pag->pag_ici_lock); |
2f11feab DC |
976 | if (!radix_tree_delete(&pag->pag_ici_root, |
977 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino))) | |
978 | ASSERT(0); | |
081003ff | 979 | __xfs_inode_clear_reclaim(pag, ip); |
1a427ab0 | 980 | spin_unlock(&pag->pag_ici_lock); |
2f11feab DC |
981 | |
982 | /* | |
983 | * Here we do an (almost) spurious inode lock in order to coordinate | |
984 | * with inode cache radix tree lookups. This is because the lookup | |
985 | * can reference the inodes in the cache without taking references. | |
986 | * | |
987 | * We make that OK here by ensuring that we wait until the inode is | |
ad637a10 | 988 | * unlocked after the lookup before we go ahead and free it. |
2f11feab | 989 | */ |
ad637a10 | 990 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
2f11feab | 991 | xfs_qm_dqdetach(ip); |
ad637a10 | 992 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
993 | |
994 | xfs_inode_free(ip); | |
ad637a10 | 995 | return error; |
8a48088f CH |
996 | |
997 | out_ifunlock: | |
998 | xfs_ifunlock(ip); | |
999 | out: | |
1000 | xfs_iflags_clear(ip, XFS_IRECLAIM); | |
1001 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
1002 | /* | |
2451337d | 1003 | * We could return -EAGAIN here to make reclaim rescan the inode tree in |
8a48088f | 1004 | * a short while. However, this just burns CPU time scanning the tree |
5889608d DC |
1005 | * waiting for IO to complete and the reclaim work never goes back to |
1006 | * the idle state. Instead, return 0 to let the next scheduled | |
1007 | * background reclaim attempt to reclaim the inode again. | |
8a48088f CH |
1008 | */ |
1009 | return 0; | |
7a3be02b DC |
1010 | } |
1011 | ||
65d0f205 DC |
1012 | /* |
1013 | * Walk the AGs and reclaim the inodes in them. Even if the filesystem is | |
1014 | * corrupted, we still want to try to reclaim all the inodes. If we don't, | |
1015 | * then a shut down during filesystem unmount reclaim walk leak all the | |
1016 | * unreclaimed inodes. | |
1017 | */ | |
33479e05 | 1018 | STATIC int |
65d0f205 DC |
1019 | xfs_reclaim_inodes_ag( |
1020 | struct xfs_mount *mp, | |
1021 | int flags, | |
1022 | int *nr_to_scan) | |
1023 | { | |
1024 | struct xfs_perag *pag; | |
1025 | int error = 0; | |
1026 | int last_error = 0; | |
1027 | xfs_agnumber_t ag; | |
69b491c2 DC |
1028 | int trylock = flags & SYNC_TRYLOCK; |
1029 | int skipped; | |
65d0f205 | 1030 | |
69b491c2 | 1031 | restart: |
65d0f205 | 1032 | ag = 0; |
69b491c2 | 1033 | skipped = 0; |
65d0f205 DC |
1034 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
1035 | unsigned long first_index = 0; | |
1036 | int done = 0; | |
e3a20c0b | 1037 | int nr_found = 0; |
65d0f205 DC |
1038 | |
1039 | ag = pag->pag_agno + 1; | |
1040 | ||
69b491c2 DC |
1041 | if (trylock) { |
1042 | if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) { | |
1043 | skipped++; | |
f83282a8 | 1044 | xfs_perag_put(pag); |
69b491c2 DC |
1045 | continue; |
1046 | } | |
1047 | first_index = pag->pag_ici_reclaim_cursor; | |
1048 | } else | |
1049 | mutex_lock(&pag->pag_ici_reclaim_lock); | |
1050 | ||
65d0f205 | 1051 | do { |
e3a20c0b DC |
1052 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
1053 | int i; | |
65d0f205 | 1054 | |
1a3e8f3d | 1055 | rcu_read_lock(); |
e3a20c0b DC |
1056 | nr_found = radix_tree_gang_lookup_tag( |
1057 | &pag->pag_ici_root, | |
1058 | (void **)batch, first_index, | |
1059 | XFS_LOOKUP_BATCH, | |
65d0f205 DC |
1060 | XFS_ICI_RECLAIM_TAG); |
1061 | if (!nr_found) { | |
b2232219 | 1062 | done = 1; |
1a3e8f3d | 1063 | rcu_read_unlock(); |
65d0f205 DC |
1064 | break; |
1065 | } | |
1066 | ||
1067 | /* | |
e3a20c0b DC |
1068 | * Grab the inodes before we drop the lock. if we found |
1069 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 1070 | */ |
e3a20c0b DC |
1071 | for (i = 0; i < nr_found; i++) { |
1072 | struct xfs_inode *ip = batch[i]; | |
1073 | ||
1074 | if (done || xfs_reclaim_inode_grab(ip, flags)) | |
1075 | batch[i] = NULL; | |
1076 | ||
1077 | /* | |
1078 | * Update the index for the next lookup. Catch | |
1079 | * overflows into the next AG range which can | |
1080 | * occur if we have inodes in the last block of | |
1081 | * the AG and we are currently pointing to the | |
1082 | * last inode. | |
1a3e8f3d DC |
1083 | * |
1084 | * Because we may see inodes that are from the | |
1085 | * wrong AG due to RCU freeing and | |
1086 | * reallocation, only update the index if it | |
1087 | * lies in this AG. It was a race that lead us | |
1088 | * to see this inode, so another lookup from | |
1089 | * the same index will not find it again. | |
e3a20c0b | 1090 | */ |
1a3e8f3d DC |
1091 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != |
1092 | pag->pag_agno) | |
1093 | continue; | |
e3a20c0b DC |
1094 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
1095 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
1096 | done = 1; | |
1097 | } | |
65d0f205 | 1098 | |
e3a20c0b | 1099 | /* unlock now we've grabbed the inodes. */ |
1a3e8f3d | 1100 | rcu_read_unlock(); |
e3a20c0b DC |
1101 | |
1102 | for (i = 0; i < nr_found; i++) { | |
1103 | if (!batch[i]) | |
1104 | continue; | |
1105 | error = xfs_reclaim_inode(batch[i], pag, flags); | |
2451337d | 1106 | if (error && last_error != -EFSCORRUPTED) |
e3a20c0b DC |
1107 | last_error = error; |
1108 | } | |
1109 | ||
1110 | *nr_to_scan -= XFS_LOOKUP_BATCH; | |
65d0f205 | 1111 | |
8daaa831 DC |
1112 | cond_resched(); |
1113 | ||
e3a20c0b | 1114 | } while (nr_found && !done && *nr_to_scan > 0); |
65d0f205 | 1115 | |
69b491c2 DC |
1116 | if (trylock && !done) |
1117 | pag->pag_ici_reclaim_cursor = first_index; | |
1118 | else | |
1119 | pag->pag_ici_reclaim_cursor = 0; | |
1120 | mutex_unlock(&pag->pag_ici_reclaim_lock); | |
65d0f205 DC |
1121 | xfs_perag_put(pag); |
1122 | } | |
69b491c2 DC |
1123 | |
1124 | /* | |
1125 | * if we skipped any AG, and we still have scan count remaining, do | |
1126 | * another pass this time using blocking reclaim semantics (i.e | |
1127 | * waiting on the reclaim locks and ignoring the reclaim cursors). This | |
1128 | * ensure that when we get more reclaimers than AGs we block rather | |
1129 | * than spin trying to execute reclaim. | |
1130 | */ | |
8daaa831 | 1131 | if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) { |
69b491c2 DC |
1132 | trylock = 0; |
1133 | goto restart; | |
1134 | } | |
b474c7ae | 1135 | return last_error; |
65d0f205 DC |
1136 | } |
1137 | ||
7a3be02b DC |
1138 | int |
1139 | xfs_reclaim_inodes( | |
1140 | xfs_mount_t *mp, | |
7a3be02b DC |
1141 | int mode) |
1142 | { | |
65d0f205 DC |
1143 | int nr_to_scan = INT_MAX; |
1144 | ||
1145 | return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan); | |
9bf729c0 DC |
1146 | } |
1147 | ||
1148 | /* | |
8daaa831 | 1149 | * Scan a certain number of inodes for reclaim. |
a7b339f1 DC |
1150 | * |
1151 | * When called we make sure that there is a background (fast) inode reclaim in | |
8daaa831 | 1152 | * progress, while we will throttle the speed of reclaim via doing synchronous |
a7b339f1 DC |
1153 | * reclaim of inodes. That means if we come across dirty inodes, we wait for |
1154 | * them to be cleaned, which we hope will not be very long due to the | |
1155 | * background walker having already kicked the IO off on those dirty inodes. | |
9bf729c0 | 1156 | */ |
0a234c6d | 1157 | long |
8daaa831 DC |
1158 | xfs_reclaim_inodes_nr( |
1159 | struct xfs_mount *mp, | |
1160 | int nr_to_scan) | |
9bf729c0 | 1161 | { |
8daaa831 | 1162 | /* kick background reclaimer and push the AIL */ |
5889608d | 1163 | xfs_reclaim_work_queue(mp); |
8daaa831 | 1164 | xfs_ail_push_all(mp->m_ail); |
a7b339f1 | 1165 | |
0a234c6d | 1166 | return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan); |
8daaa831 | 1167 | } |
9bf729c0 | 1168 | |
8daaa831 DC |
1169 | /* |
1170 | * Return the number of reclaimable inodes in the filesystem for | |
1171 | * the shrinker to determine how much to reclaim. | |
1172 | */ | |
1173 | int | |
1174 | xfs_reclaim_inodes_count( | |
1175 | struct xfs_mount *mp) | |
1176 | { | |
1177 | struct xfs_perag *pag; | |
1178 | xfs_agnumber_t ag = 0; | |
1179 | int reclaimable = 0; | |
9bf729c0 | 1180 | |
65d0f205 DC |
1181 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
1182 | ag = pag->pag_agno + 1; | |
70e60ce7 DC |
1183 | reclaimable += pag->pag_ici_reclaimable; |
1184 | xfs_perag_put(pag); | |
9bf729c0 | 1185 | } |
9bf729c0 DC |
1186 | return reclaimable; |
1187 | } | |
1188 | ||
3e3f9f58 BF |
1189 | STATIC int |
1190 | xfs_inode_match_id( | |
1191 | struct xfs_inode *ip, | |
1192 | struct xfs_eofblocks *eofb) | |
1193 | { | |
b9fe5052 DE |
1194 | if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) && |
1195 | !uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid)) | |
1b556048 | 1196 | return 0; |
3e3f9f58 | 1197 | |
b9fe5052 DE |
1198 | if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) && |
1199 | !gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid)) | |
1b556048 BF |
1200 | return 0; |
1201 | ||
b9fe5052 | 1202 | if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) && |
1b556048 BF |
1203 | xfs_get_projid(ip) != eofb->eof_prid) |
1204 | return 0; | |
1205 | ||
1206 | return 1; | |
3e3f9f58 BF |
1207 | } |
1208 | ||
f4526397 BF |
1209 | /* |
1210 | * A union-based inode filtering algorithm. Process the inode if any of the | |
1211 | * criteria match. This is for global/internal scans only. | |
1212 | */ | |
1213 | STATIC int | |
1214 | xfs_inode_match_id_union( | |
1215 | struct xfs_inode *ip, | |
1216 | struct xfs_eofblocks *eofb) | |
1217 | { | |
1218 | if ((eofb->eof_flags & XFS_EOF_FLAGS_UID) && | |
1219 | uid_eq(VFS_I(ip)->i_uid, eofb->eof_uid)) | |
1220 | return 1; | |
1221 | ||
1222 | if ((eofb->eof_flags & XFS_EOF_FLAGS_GID) && | |
1223 | gid_eq(VFS_I(ip)->i_gid, eofb->eof_gid)) | |
1224 | return 1; | |
1225 | ||
1226 | if ((eofb->eof_flags & XFS_EOF_FLAGS_PRID) && | |
1227 | xfs_get_projid(ip) == eofb->eof_prid) | |
1228 | return 1; | |
1229 | ||
1230 | return 0; | |
1231 | } | |
1232 | ||
41176a68 BF |
1233 | STATIC int |
1234 | xfs_inode_free_eofblocks( | |
1235 | struct xfs_inode *ip, | |
41176a68 BF |
1236 | int flags, |
1237 | void *args) | |
1238 | { | |
1239 | int ret; | |
3e3f9f58 | 1240 | struct xfs_eofblocks *eofb = args; |
5400da7d | 1241 | bool need_iolock = true; |
f4526397 | 1242 | int match; |
5400da7d BF |
1243 | |
1244 | ASSERT(!eofb || (eofb && eofb->eof_scan_owner != 0)); | |
41176a68 BF |
1245 | |
1246 | if (!xfs_can_free_eofblocks(ip, false)) { | |
1247 | /* inode could be preallocated or append-only */ | |
1248 | trace_xfs_inode_free_eofblocks_invalid(ip); | |
1249 | xfs_inode_clear_eofblocks_tag(ip); | |
1250 | return 0; | |
1251 | } | |
1252 | ||
1253 | /* | |
1254 | * If the mapping is dirty the operation can block and wait for some | |
1255 | * time. Unless we are waiting, skip it. | |
1256 | */ | |
1257 | if (!(flags & SYNC_WAIT) && | |
1258 | mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY)) | |
1259 | return 0; | |
1260 | ||
00ca79a0 | 1261 | if (eofb) { |
f4526397 BF |
1262 | if (eofb->eof_flags & XFS_EOF_FLAGS_UNION) |
1263 | match = xfs_inode_match_id_union(ip, eofb); | |
1264 | else | |
1265 | match = xfs_inode_match_id(ip, eofb); | |
1266 | if (!match) | |
00ca79a0 BF |
1267 | return 0; |
1268 | ||
1269 | /* skip the inode if the file size is too small */ | |
1270 | if (eofb->eof_flags & XFS_EOF_FLAGS_MINFILESIZE && | |
1271 | XFS_ISIZE(ip) < eofb->eof_min_file_size) | |
1272 | return 0; | |
5400da7d BF |
1273 | |
1274 | /* | |
1275 | * A scan owner implies we already hold the iolock. Skip it in | |
1276 | * xfs_free_eofblocks() to avoid deadlock. This also eliminates | |
1277 | * the possibility of EAGAIN being returned. | |
1278 | */ | |
1279 | if (eofb->eof_scan_owner == ip->i_ino) | |
1280 | need_iolock = false; | |
00ca79a0 | 1281 | } |
3e3f9f58 | 1282 | |
5400da7d | 1283 | ret = xfs_free_eofblocks(ip->i_mount, ip, need_iolock); |
41176a68 BF |
1284 | |
1285 | /* don't revisit the inode if we're not waiting */ | |
2451337d | 1286 | if (ret == -EAGAIN && !(flags & SYNC_WAIT)) |
41176a68 BF |
1287 | ret = 0; |
1288 | ||
1289 | return ret; | |
1290 | } | |
1291 | ||
1292 | int | |
1293 | xfs_icache_free_eofblocks( | |
1294 | struct xfs_mount *mp, | |
8ca149de | 1295 | struct xfs_eofblocks *eofb) |
41176a68 | 1296 | { |
8ca149de BF |
1297 | int flags = SYNC_TRYLOCK; |
1298 | ||
1299 | if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC)) | |
1300 | flags = SYNC_WAIT; | |
1301 | ||
41176a68 | 1302 | return xfs_inode_ag_iterator_tag(mp, xfs_inode_free_eofblocks, flags, |
8ca149de | 1303 | eofb, XFS_ICI_EOFBLOCKS_TAG); |
41176a68 BF |
1304 | } |
1305 | ||
dc06f398 BF |
1306 | /* |
1307 | * Run eofblocks scans on the quotas applicable to the inode. For inodes with | |
1308 | * multiple quotas, we don't know exactly which quota caused an allocation | |
1309 | * failure. We make a best effort by including each quota under low free space | |
1310 | * conditions (less than 1% free space) in the scan. | |
1311 | */ | |
1312 | int | |
1313 | xfs_inode_free_quota_eofblocks( | |
1314 | struct xfs_inode *ip) | |
1315 | { | |
1316 | int scan = 0; | |
1317 | struct xfs_eofblocks eofb = {0}; | |
1318 | struct xfs_dquot *dq; | |
1319 | ||
1320 | ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL)); | |
1321 | ||
1322 | /* | |
1323 | * Set the scan owner to avoid a potential livelock. Otherwise, the scan | |
1324 | * can repeatedly trylock on the inode we're currently processing. We | |
1325 | * run a sync scan to increase effectiveness and use the union filter to | |
1326 | * cover all applicable quotas in a single scan. | |
1327 | */ | |
1328 | eofb.eof_scan_owner = ip->i_ino; | |
1329 | eofb.eof_flags = XFS_EOF_FLAGS_UNION|XFS_EOF_FLAGS_SYNC; | |
1330 | ||
1331 | if (XFS_IS_UQUOTA_ENFORCED(ip->i_mount)) { | |
1332 | dq = xfs_inode_dquot(ip, XFS_DQ_USER); | |
1333 | if (dq && xfs_dquot_lowsp(dq)) { | |
1334 | eofb.eof_uid = VFS_I(ip)->i_uid; | |
1335 | eofb.eof_flags |= XFS_EOF_FLAGS_UID; | |
1336 | scan = 1; | |
1337 | } | |
1338 | } | |
1339 | ||
1340 | if (XFS_IS_GQUOTA_ENFORCED(ip->i_mount)) { | |
1341 | dq = xfs_inode_dquot(ip, XFS_DQ_GROUP); | |
1342 | if (dq && xfs_dquot_lowsp(dq)) { | |
1343 | eofb.eof_gid = VFS_I(ip)->i_gid; | |
1344 | eofb.eof_flags |= XFS_EOF_FLAGS_GID; | |
1345 | scan = 1; | |
1346 | } | |
1347 | } | |
1348 | ||
1349 | if (scan) | |
1350 | xfs_icache_free_eofblocks(ip->i_mount, &eofb); | |
1351 | ||
1352 | return scan; | |
1353 | } | |
1354 | ||
27b52867 BF |
1355 | void |
1356 | xfs_inode_set_eofblocks_tag( | |
1357 | xfs_inode_t *ip) | |
1358 | { | |
1359 | struct xfs_mount *mp = ip->i_mount; | |
1360 | struct xfs_perag *pag; | |
1361 | int tagged; | |
1362 | ||
1363 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); | |
1364 | spin_lock(&pag->pag_ici_lock); | |
1365 | trace_xfs_inode_set_eofblocks_tag(ip); | |
1366 | ||
1367 | tagged = radix_tree_tagged(&pag->pag_ici_root, | |
1368 | XFS_ICI_EOFBLOCKS_TAG); | |
1369 | radix_tree_tag_set(&pag->pag_ici_root, | |
1370 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
1371 | XFS_ICI_EOFBLOCKS_TAG); | |
1372 | if (!tagged) { | |
1373 | /* propagate the eofblocks tag up into the perag radix tree */ | |
1374 | spin_lock(&ip->i_mount->m_perag_lock); | |
1375 | radix_tree_tag_set(&ip->i_mount->m_perag_tree, | |
1376 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
1377 | XFS_ICI_EOFBLOCKS_TAG); | |
1378 | spin_unlock(&ip->i_mount->m_perag_lock); | |
579b62fa BF |
1379 | |
1380 | /* kick off background trimming */ | |
1381 | xfs_queue_eofblocks(ip->i_mount); | |
27b52867 BF |
1382 | |
1383 | trace_xfs_perag_set_eofblocks(ip->i_mount, pag->pag_agno, | |
1384 | -1, _RET_IP_); | |
1385 | } | |
1386 | ||
1387 | spin_unlock(&pag->pag_ici_lock); | |
1388 | xfs_perag_put(pag); | |
1389 | } | |
1390 | ||
1391 | void | |
1392 | xfs_inode_clear_eofblocks_tag( | |
1393 | xfs_inode_t *ip) | |
1394 | { | |
1395 | struct xfs_mount *mp = ip->i_mount; | |
1396 | struct xfs_perag *pag; | |
1397 | ||
1398 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); | |
1399 | spin_lock(&pag->pag_ici_lock); | |
1400 | trace_xfs_inode_clear_eofblocks_tag(ip); | |
1401 | ||
1402 | radix_tree_tag_clear(&pag->pag_ici_root, | |
1403 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
1404 | XFS_ICI_EOFBLOCKS_TAG); | |
1405 | if (!radix_tree_tagged(&pag->pag_ici_root, XFS_ICI_EOFBLOCKS_TAG)) { | |
1406 | /* clear the eofblocks tag from the perag radix tree */ | |
1407 | spin_lock(&ip->i_mount->m_perag_lock); | |
1408 | radix_tree_tag_clear(&ip->i_mount->m_perag_tree, | |
1409 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
1410 | XFS_ICI_EOFBLOCKS_TAG); | |
1411 | spin_unlock(&ip->i_mount->m_perag_lock); | |
1412 | trace_xfs_perag_clear_eofblocks(ip->i_mount, pag->pag_agno, | |
1413 | -1, _RET_IP_); | |
1414 | } | |
1415 | ||
1416 | spin_unlock(&pag->pag_ici_lock); | |
1417 | xfs_perag_put(pag); | |
1418 | } | |
1419 |