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" |
fe4fa4b8 | 39 | |
a167b17e DC |
40 | #include <linux/kthread.h> |
41 | #include <linux/freezer.h> | |
42 | ||
78ae5256 DC |
43 | /* |
44 | * The inode lookup is done in batches to keep the amount of lock traffic and | |
45 | * radix tree lookups to a minimum. The batch size is a trade off between | |
46 | * lookup reduction and stack usage. This is in the reclaim path, so we can't | |
47 | * be too greedy. | |
48 | */ | |
49 | #define XFS_LOOKUP_BATCH 32 | |
50 | ||
e13de955 DC |
51 | STATIC int |
52 | xfs_inode_ag_walk_grab( | |
53 | struct xfs_inode *ip) | |
54 | { | |
55 | struct inode *inode = VFS_I(ip); | |
56 | ||
1a3e8f3d DC |
57 | ASSERT(rcu_read_lock_held()); |
58 | ||
59 | /* | |
60 | * check for stale RCU freed inode | |
61 | * | |
62 | * If the inode has been reallocated, it doesn't matter if it's not in | |
63 | * the AG we are walking - we are walking for writeback, so if it | |
64 | * passes all the "valid inode" checks and is dirty, then we'll write | |
65 | * it back anyway. If it has been reallocated and still being | |
66 | * initialised, the XFS_INEW check below will catch it. | |
67 | */ | |
68 | spin_lock(&ip->i_flags_lock); | |
69 | if (!ip->i_ino) | |
70 | goto out_unlock_noent; | |
71 | ||
72 | /* avoid new or reclaimable inodes. Leave for reclaim code to flush */ | |
73 | if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM)) | |
74 | goto out_unlock_noent; | |
75 | spin_unlock(&ip->i_flags_lock); | |
76 | ||
e13de955 DC |
77 | /* nothing to sync during shutdown */ |
78 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) | |
79 | return EFSCORRUPTED; | |
80 | ||
e13de955 DC |
81 | /* If we can't grab the inode, it must on it's way to reclaim. */ |
82 | if (!igrab(inode)) | |
83 | return ENOENT; | |
84 | ||
85 | if (is_bad_inode(inode)) { | |
86 | IRELE(ip); | |
87 | return ENOENT; | |
88 | } | |
89 | ||
90 | /* inode is valid */ | |
91 | return 0; | |
1a3e8f3d DC |
92 | |
93 | out_unlock_noent: | |
94 | spin_unlock(&ip->i_flags_lock); | |
95 | return ENOENT; | |
e13de955 DC |
96 | } |
97 | ||
75f3cb13 DC |
98 | STATIC int |
99 | xfs_inode_ag_walk( | |
100 | struct xfs_mount *mp, | |
5017e97d | 101 | struct xfs_perag *pag, |
75f3cb13 DC |
102 | int (*execute)(struct xfs_inode *ip, |
103 | struct xfs_perag *pag, int flags), | |
65d0f205 | 104 | int flags) |
75f3cb13 | 105 | { |
75f3cb13 DC |
106 | uint32_t first_index; |
107 | int last_error = 0; | |
108 | int skipped; | |
65d0f205 | 109 | int done; |
78ae5256 | 110 | int nr_found; |
75f3cb13 DC |
111 | |
112 | restart: | |
65d0f205 | 113 | done = 0; |
75f3cb13 DC |
114 | skipped = 0; |
115 | first_index = 0; | |
78ae5256 | 116 | nr_found = 0; |
75f3cb13 | 117 | do { |
78ae5256 | 118 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
75f3cb13 | 119 | int error = 0; |
78ae5256 | 120 | int i; |
75f3cb13 | 121 | |
1a3e8f3d | 122 | rcu_read_lock(); |
65d0f205 | 123 | nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, |
78ae5256 DC |
124 | (void **)batch, first_index, |
125 | XFS_LOOKUP_BATCH); | |
65d0f205 | 126 | if (!nr_found) { |
1a3e8f3d | 127 | rcu_read_unlock(); |
75f3cb13 | 128 | break; |
c8e20be0 | 129 | } |
75f3cb13 | 130 | |
65d0f205 | 131 | /* |
78ae5256 DC |
132 | * Grab the inodes before we drop the lock. if we found |
133 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 134 | */ |
78ae5256 DC |
135 | for (i = 0; i < nr_found; i++) { |
136 | struct xfs_inode *ip = batch[i]; | |
137 | ||
138 | if (done || xfs_inode_ag_walk_grab(ip)) | |
139 | batch[i] = NULL; | |
140 | ||
141 | /* | |
1a3e8f3d DC |
142 | * Update the index for the next lookup. Catch |
143 | * overflows into the next AG range which can occur if | |
144 | * we have inodes in the last block of the AG and we | |
145 | * are currently pointing to the last inode. | |
146 | * | |
147 | * Because we may see inodes that are from the wrong AG | |
148 | * due to RCU freeing and reallocation, only update the | |
149 | * index if it lies in this AG. It was a race that lead | |
150 | * us to see this inode, so another lookup from the | |
151 | * same index will not find it again. | |
78ae5256 | 152 | */ |
1a3e8f3d DC |
153 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno) |
154 | continue; | |
78ae5256 DC |
155 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
156 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
157 | done = 1; | |
e13de955 | 158 | } |
78ae5256 DC |
159 | |
160 | /* unlock now we've grabbed the inodes. */ | |
1a3e8f3d | 161 | rcu_read_unlock(); |
e13de955 | 162 | |
78ae5256 DC |
163 | for (i = 0; i < nr_found; i++) { |
164 | if (!batch[i]) | |
165 | continue; | |
166 | error = execute(batch[i], pag, flags); | |
167 | IRELE(batch[i]); | |
168 | if (error == EAGAIN) { | |
169 | skipped++; | |
170 | continue; | |
171 | } | |
172 | if (error && last_error != EFSCORRUPTED) | |
173 | last_error = error; | |
75f3cb13 | 174 | } |
c8e20be0 DC |
175 | |
176 | /* bail out if the filesystem is corrupted. */ | |
75f3cb13 DC |
177 | if (error == EFSCORRUPTED) |
178 | break; | |
179 | ||
8daaa831 DC |
180 | cond_resched(); |
181 | ||
78ae5256 | 182 | } while (nr_found && !done); |
75f3cb13 DC |
183 | |
184 | if (skipped) { | |
185 | delay(1); | |
186 | goto restart; | |
187 | } | |
75f3cb13 DC |
188 | return last_error; |
189 | } | |
190 | ||
fe588ed3 | 191 | int |
75f3cb13 DC |
192 | xfs_inode_ag_iterator( |
193 | struct xfs_mount *mp, | |
194 | int (*execute)(struct xfs_inode *ip, | |
195 | struct xfs_perag *pag, int flags), | |
65d0f205 | 196 | int flags) |
75f3cb13 | 197 | { |
16fd5367 | 198 | struct xfs_perag *pag; |
75f3cb13 DC |
199 | int error = 0; |
200 | int last_error = 0; | |
201 | xfs_agnumber_t ag; | |
202 | ||
16fd5367 | 203 | ag = 0; |
65d0f205 DC |
204 | while ((pag = xfs_perag_get(mp, ag))) { |
205 | ag = pag->pag_agno + 1; | |
206 | error = xfs_inode_ag_walk(mp, pag, execute, flags); | |
5017e97d | 207 | xfs_perag_put(pag); |
75f3cb13 DC |
208 | if (error) { |
209 | last_error = error; | |
210 | if (error == EFSCORRUPTED) | |
211 | break; | |
212 | } | |
213 | } | |
214 | return XFS_ERROR(last_error); | |
215 | } | |
216 | ||
a7b339f1 DC |
217 | /* |
218 | * Queue a new inode reclaim pass if there are reclaimable inodes and there | |
219 | * isn't a reclaim pass already in progress. By default it runs every 5s based | |
5889608d | 220 | * on the xfs periodic sync default of 30s. Perhaps this should have it's own |
a7b339f1 DC |
221 | * tunable, but that can be done if this method proves to be ineffective or too |
222 | * aggressive. | |
223 | */ | |
224 | static void | |
5889608d | 225 | xfs_reclaim_work_queue( |
a7b339f1 | 226 | struct xfs_mount *mp) |
a167b17e | 227 | { |
a167b17e | 228 | |
a7b339f1 DC |
229 | rcu_read_lock(); |
230 | if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) { | |
5889608d | 231 | queue_delayed_work(mp->m_reclaim_workqueue, &mp->m_reclaim_work, |
a7b339f1 | 232 | msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10)); |
a167b17e | 233 | } |
a7b339f1 DC |
234 | rcu_read_unlock(); |
235 | } | |
a167b17e | 236 | |
a7b339f1 DC |
237 | /* |
238 | * This is a fast pass over the inode cache to try to get reclaim moving on as | |
239 | * many inodes as possible in a short period of time. It kicks itself every few | |
240 | * seconds, as well as being kicked by the inode cache shrinker when memory | |
241 | * goes low. It scans as quickly as possible avoiding locked inodes or those | |
242 | * already being flushed, and once done schedules a future pass. | |
243 | */ | |
33c7a2bc | 244 | void |
a7b339f1 DC |
245 | xfs_reclaim_worker( |
246 | struct work_struct *work) | |
247 | { | |
248 | struct xfs_mount *mp = container_of(to_delayed_work(work), | |
249 | struct xfs_mount, m_reclaim_work); | |
250 | ||
251 | xfs_reclaim_inodes(mp, SYNC_TRYLOCK); | |
5889608d | 252 | xfs_reclaim_work_queue(mp); |
a7b339f1 DC |
253 | } |
254 | ||
bc990f5c CH |
255 | void |
256 | __xfs_inode_set_reclaim_tag( | |
257 | struct xfs_perag *pag, | |
258 | struct xfs_inode *ip) | |
259 | { | |
260 | radix_tree_tag_set(&pag->pag_ici_root, | |
261 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
262 | XFS_ICI_RECLAIM_TAG); | |
16fd5367 DC |
263 | |
264 | if (!pag->pag_ici_reclaimable) { | |
265 | /* propagate the reclaim tag up into the perag radix tree */ | |
266 | spin_lock(&ip->i_mount->m_perag_lock); | |
267 | radix_tree_tag_set(&ip->i_mount->m_perag_tree, | |
268 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
269 | XFS_ICI_RECLAIM_TAG); | |
270 | spin_unlock(&ip->i_mount->m_perag_lock); | |
a7b339f1 DC |
271 | |
272 | /* schedule periodic background inode reclaim */ | |
5889608d | 273 | xfs_reclaim_work_queue(ip->i_mount); |
a7b339f1 | 274 | |
16fd5367 DC |
275 | trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno, |
276 | -1, _RET_IP_); | |
277 | } | |
9bf729c0 | 278 | pag->pag_ici_reclaimable++; |
bc990f5c CH |
279 | } |
280 | ||
11654513 DC |
281 | /* |
282 | * We set the inode flag atomically with the radix tree tag. | |
283 | * Once we get tag lookups on the radix tree, this inode flag | |
284 | * can go away. | |
285 | */ | |
396beb85 DC |
286 | void |
287 | xfs_inode_set_reclaim_tag( | |
288 | xfs_inode_t *ip) | |
289 | { | |
5017e97d DC |
290 | struct xfs_mount *mp = ip->i_mount; |
291 | struct xfs_perag *pag; | |
396beb85 | 292 | |
5017e97d | 293 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); |
1a427ab0 | 294 | spin_lock(&pag->pag_ici_lock); |
396beb85 | 295 | spin_lock(&ip->i_flags_lock); |
bc990f5c | 296 | __xfs_inode_set_reclaim_tag(pag, ip); |
11654513 | 297 | __xfs_iflags_set(ip, XFS_IRECLAIMABLE); |
396beb85 | 298 | spin_unlock(&ip->i_flags_lock); |
1a427ab0 | 299 | spin_unlock(&pag->pag_ici_lock); |
5017e97d | 300 | xfs_perag_put(pag); |
396beb85 DC |
301 | } |
302 | ||
081003ff JW |
303 | STATIC void |
304 | __xfs_inode_clear_reclaim( | |
396beb85 DC |
305 | xfs_perag_t *pag, |
306 | xfs_inode_t *ip) | |
307 | { | |
9bf729c0 | 308 | pag->pag_ici_reclaimable--; |
16fd5367 DC |
309 | if (!pag->pag_ici_reclaimable) { |
310 | /* clear the reclaim tag from the perag radix tree */ | |
311 | spin_lock(&ip->i_mount->m_perag_lock); | |
312 | radix_tree_tag_clear(&ip->i_mount->m_perag_tree, | |
313 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
314 | XFS_ICI_RECLAIM_TAG); | |
315 | spin_unlock(&ip->i_mount->m_perag_lock); | |
316 | trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno, | |
317 | -1, _RET_IP_); | |
318 | } | |
396beb85 DC |
319 | } |
320 | ||
081003ff JW |
321 | void |
322 | __xfs_inode_clear_reclaim_tag( | |
323 | xfs_mount_t *mp, | |
324 | xfs_perag_t *pag, | |
325 | xfs_inode_t *ip) | |
326 | { | |
327 | radix_tree_tag_clear(&pag->pag_ici_root, | |
328 | XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG); | |
329 | __xfs_inode_clear_reclaim(pag, ip); | |
330 | } | |
331 | ||
e3a20c0b DC |
332 | /* |
333 | * Grab the inode for reclaim exclusively. | |
334 | * Return 0 if we grabbed it, non-zero otherwise. | |
335 | */ | |
336 | STATIC int | |
337 | xfs_reclaim_inode_grab( | |
338 | struct xfs_inode *ip, | |
339 | int flags) | |
340 | { | |
1a3e8f3d DC |
341 | ASSERT(rcu_read_lock_held()); |
342 | ||
343 | /* quick check for stale RCU freed inode */ | |
344 | if (!ip->i_ino) | |
345 | return 1; | |
e3a20c0b DC |
346 | |
347 | /* | |
474fce06 CH |
348 | * If we are asked for non-blocking operation, do unlocked checks to |
349 | * see if the inode already is being flushed or in reclaim to avoid | |
350 | * lock traffic. | |
e3a20c0b DC |
351 | */ |
352 | if ((flags & SYNC_TRYLOCK) && | |
474fce06 | 353 | __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM)) |
e3a20c0b | 354 | return 1; |
e3a20c0b DC |
355 | |
356 | /* | |
357 | * The radix tree lock here protects a thread in xfs_iget from racing | |
358 | * with us starting reclaim on the inode. Once we have the | |
359 | * XFS_IRECLAIM flag set it will not touch us. | |
1a3e8f3d DC |
360 | * |
361 | * Due to RCU lookup, we may find inodes that have been freed and only | |
362 | * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that | |
363 | * aren't candidates for reclaim at all, so we must check the | |
364 | * XFS_IRECLAIMABLE is set first before proceeding to reclaim. | |
e3a20c0b DC |
365 | */ |
366 | spin_lock(&ip->i_flags_lock); | |
1a3e8f3d DC |
367 | if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) || |
368 | __xfs_iflags_test(ip, XFS_IRECLAIM)) { | |
369 | /* not a reclaim candidate. */ | |
e3a20c0b DC |
370 | spin_unlock(&ip->i_flags_lock); |
371 | return 1; | |
372 | } | |
373 | __xfs_iflags_set(ip, XFS_IRECLAIM); | |
374 | spin_unlock(&ip->i_flags_lock); | |
375 | return 0; | |
376 | } | |
377 | ||
777df5af | 378 | /* |
8a48088f CH |
379 | * Inodes in different states need to be treated differently. The following |
380 | * table lists the inode states and the reclaim actions necessary: | |
777df5af DC |
381 | * |
382 | * inode state iflush ret required action | |
383 | * --------------- ---------- --------------- | |
384 | * bad - reclaim | |
385 | * shutdown EIO unpin and reclaim | |
386 | * clean, unpinned 0 reclaim | |
387 | * stale, unpinned 0 reclaim | |
c854363e DC |
388 | * clean, pinned(*) 0 requeue |
389 | * stale, pinned EAGAIN requeue | |
8a48088f CH |
390 | * dirty, async - requeue |
391 | * dirty, sync 0 reclaim | |
777df5af DC |
392 | * |
393 | * (*) dgc: I don't think the clean, pinned state is possible but it gets | |
394 | * handled anyway given the order of checks implemented. | |
395 | * | |
c854363e DC |
396 | * Also, because we get the flush lock first, we know that any inode that has |
397 | * been flushed delwri has had the flush completed by the time we check that | |
8a48088f | 398 | * the inode is clean. |
c854363e | 399 | * |
8a48088f CH |
400 | * Note that because the inode is flushed delayed write by AIL pushing, the |
401 | * flush lock may already be held here and waiting on it can result in very | |
402 | * long latencies. Hence for sync reclaims, where we wait on the flush lock, | |
403 | * the caller should push the AIL first before trying to reclaim inodes to | |
404 | * minimise the amount of time spent waiting. For background relaim, we only | |
405 | * bother to reclaim clean inodes anyway. | |
c854363e | 406 | * |
777df5af DC |
407 | * Hence the order of actions after gaining the locks should be: |
408 | * bad => reclaim | |
409 | * shutdown => unpin and reclaim | |
8a48088f | 410 | * pinned, async => requeue |
c854363e | 411 | * pinned, sync => unpin |
777df5af DC |
412 | * stale => reclaim |
413 | * clean => reclaim | |
8a48088f | 414 | * dirty, async => requeue |
c854363e | 415 | * dirty, sync => flush, wait and reclaim |
777df5af | 416 | */ |
75f3cb13 | 417 | STATIC int |
c8e20be0 | 418 | xfs_reclaim_inode( |
75f3cb13 DC |
419 | struct xfs_inode *ip, |
420 | struct xfs_perag *pag, | |
c8e20be0 | 421 | int sync_mode) |
fce08f2f | 422 | { |
4c46819a CH |
423 | struct xfs_buf *bp = NULL; |
424 | int error; | |
777df5af | 425 | |
1bfd8d04 DC |
426 | restart: |
427 | error = 0; | |
c8e20be0 | 428 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
c854363e DC |
429 | if (!xfs_iflock_nowait(ip)) { |
430 | if (!(sync_mode & SYNC_WAIT)) | |
431 | goto out; | |
432 | xfs_iflock(ip); | |
433 | } | |
7a3be02b | 434 | |
777df5af DC |
435 | if (is_bad_inode(VFS_I(ip))) |
436 | goto reclaim; | |
437 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) { | |
438 | xfs_iunpin_wait(ip); | |
04913fdd | 439 | xfs_iflush_abort(ip, false); |
777df5af DC |
440 | goto reclaim; |
441 | } | |
c854363e | 442 | if (xfs_ipincount(ip)) { |
8a48088f CH |
443 | if (!(sync_mode & SYNC_WAIT)) |
444 | goto out_ifunlock; | |
777df5af | 445 | xfs_iunpin_wait(ip); |
c854363e | 446 | } |
777df5af DC |
447 | if (xfs_iflags_test(ip, XFS_ISTALE)) |
448 | goto reclaim; | |
449 | if (xfs_inode_clean(ip)) | |
450 | goto reclaim; | |
451 | ||
8a48088f CH |
452 | /* |
453 | * Never flush out dirty data during non-blocking reclaim, as it would | |
454 | * just contend with AIL pushing trying to do the same job. | |
455 | */ | |
456 | if (!(sync_mode & SYNC_WAIT)) | |
457 | goto out_ifunlock; | |
458 | ||
1bfd8d04 DC |
459 | /* |
460 | * Now we have an inode that needs flushing. | |
461 | * | |
4c46819a | 462 | * Note that xfs_iflush will never block on the inode buffer lock, as |
1bfd8d04 | 463 | * xfs_ifree_cluster() can lock the inode buffer before it locks the |
4c46819a | 464 | * ip->i_lock, and we are doing the exact opposite here. As a result, |
475ee413 CH |
465 | * doing a blocking xfs_imap_to_bp() to get the cluster buffer would |
466 | * result in an ABBA deadlock with xfs_ifree_cluster(). | |
1bfd8d04 DC |
467 | * |
468 | * As xfs_ifree_cluser() must gather all inodes that are active in the | |
469 | * cache to mark them stale, if we hit this case we don't actually want | |
470 | * to do IO here - we want the inode marked stale so we can simply | |
4c46819a CH |
471 | * reclaim it. Hence if we get an EAGAIN error here, just unlock the |
472 | * inode, back off and try again. Hopefully the next pass through will | |
473 | * see the stale flag set on the inode. | |
1bfd8d04 | 474 | */ |
4c46819a | 475 | error = xfs_iflush(ip, &bp); |
8a48088f CH |
476 | if (error == EAGAIN) { |
477 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
478 | /* backoff longer than in xfs_ifree_cluster */ | |
479 | delay(2); | |
480 | goto restart; | |
c854363e | 481 | } |
c854363e | 482 | |
4c46819a CH |
483 | if (!error) { |
484 | error = xfs_bwrite(bp); | |
485 | xfs_buf_relse(bp); | |
486 | } | |
487 | ||
488 | xfs_iflock(ip); | |
777df5af DC |
489 | reclaim: |
490 | xfs_ifunlock(ip); | |
c8e20be0 | 491 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
492 | |
493 | XFS_STATS_INC(xs_ig_reclaims); | |
494 | /* | |
495 | * Remove the inode from the per-AG radix tree. | |
496 | * | |
497 | * Because radix_tree_delete won't complain even if the item was never | |
498 | * added to the tree assert that it's been there before to catch | |
499 | * problems with the inode life time early on. | |
500 | */ | |
1a427ab0 | 501 | spin_lock(&pag->pag_ici_lock); |
2f11feab DC |
502 | if (!radix_tree_delete(&pag->pag_ici_root, |
503 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino))) | |
504 | ASSERT(0); | |
081003ff | 505 | __xfs_inode_clear_reclaim(pag, ip); |
1a427ab0 | 506 | spin_unlock(&pag->pag_ici_lock); |
2f11feab DC |
507 | |
508 | /* | |
509 | * Here we do an (almost) spurious inode lock in order to coordinate | |
510 | * with inode cache radix tree lookups. This is because the lookup | |
511 | * can reference the inodes in the cache without taking references. | |
512 | * | |
513 | * We make that OK here by ensuring that we wait until the inode is | |
ad637a10 | 514 | * unlocked after the lookup before we go ahead and free it. |
2f11feab | 515 | */ |
ad637a10 | 516 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
2f11feab | 517 | xfs_qm_dqdetach(ip); |
ad637a10 | 518 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
519 | |
520 | xfs_inode_free(ip); | |
ad637a10 | 521 | return error; |
8a48088f CH |
522 | |
523 | out_ifunlock: | |
524 | xfs_ifunlock(ip); | |
525 | out: | |
526 | xfs_iflags_clear(ip, XFS_IRECLAIM); | |
527 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
528 | /* | |
529 | * We could return EAGAIN here to make reclaim rescan the inode tree in | |
530 | * a short while. However, this just burns CPU time scanning the tree | |
5889608d DC |
531 | * waiting for IO to complete and the reclaim work never goes back to |
532 | * the idle state. Instead, return 0 to let the next scheduled | |
533 | * background reclaim attempt to reclaim the inode again. | |
8a48088f CH |
534 | */ |
535 | return 0; | |
7a3be02b DC |
536 | } |
537 | ||
65d0f205 DC |
538 | /* |
539 | * Walk the AGs and reclaim the inodes in them. Even if the filesystem is | |
540 | * corrupted, we still want to try to reclaim all the inodes. If we don't, | |
541 | * then a shut down during filesystem unmount reclaim walk leak all the | |
542 | * unreclaimed inodes. | |
543 | */ | |
544 | int | |
545 | xfs_reclaim_inodes_ag( | |
546 | struct xfs_mount *mp, | |
547 | int flags, | |
548 | int *nr_to_scan) | |
549 | { | |
550 | struct xfs_perag *pag; | |
551 | int error = 0; | |
552 | int last_error = 0; | |
553 | xfs_agnumber_t ag; | |
69b491c2 DC |
554 | int trylock = flags & SYNC_TRYLOCK; |
555 | int skipped; | |
65d0f205 | 556 | |
69b491c2 | 557 | restart: |
65d0f205 | 558 | ag = 0; |
69b491c2 | 559 | skipped = 0; |
65d0f205 DC |
560 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
561 | unsigned long first_index = 0; | |
562 | int done = 0; | |
e3a20c0b | 563 | int nr_found = 0; |
65d0f205 DC |
564 | |
565 | ag = pag->pag_agno + 1; | |
566 | ||
69b491c2 DC |
567 | if (trylock) { |
568 | if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) { | |
569 | skipped++; | |
f83282a8 | 570 | xfs_perag_put(pag); |
69b491c2 DC |
571 | continue; |
572 | } | |
573 | first_index = pag->pag_ici_reclaim_cursor; | |
574 | } else | |
575 | mutex_lock(&pag->pag_ici_reclaim_lock); | |
576 | ||
65d0f205 | 577 | do { |
e3a20c0b DC |
578 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
579 | int i; | |
65d0f205 | 580 | |
1a3e8f3d | 581 | rcu_read_lock(); |
e3a20c0b DC |
582 | nr_found = radix_tree_gang_lookup_tag( |
583 | &pag->pag_ici_root, | |
584 | (void **)batch, first_index, | |
585 | XFS_LOOKUP_BATCH, | |
65d0f205 DC |
586 | XFS_ICI_RECLAIM_TAG); |
587 | if (!nr_found) { | |
b2232219 | 588 | done = 1; |
1a3e8f3d | 589 | rcu_read_unlock(); |
65d0f205 DC |
590 | break; |
591 | } | |
592 | ||
593 | /* | |
e3a20c0b DC |
594 | * Grab the inodes before we drop the lock. if we found |
595 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 596 | */ |
e3a20c0b DC |
597 | for (i = 0; i < nr_found; i++) { |
598 | struct xfs_inode *ip = batch[i]; | |
599 | ||
600 | if (done || xfs_reclaim_inode_grab(ip, flags)) | |
601 | batch[i] = NULL; | |
602 | ||
603 | /* | |
604 | * Update the index for the next lookup. Catch | |
605 | * overflows into the next AG range which can | |
606 | * occur if we have inodes in the last block of | |
607 | * the AG and we are currently pointing to the | |
608 | * last inode. | |
1a3e8f3d DC |
609 | * |
610 | * Because we may see inodes that are from the | |
611 | * wrong AG due to RCU freeing and | |
612 | * reallocation, only update the index if it | |
613 | * lies in this AG. It was a race that lead us | |
614 | * to see this inode, so another lookup from | |
615 | * the same index will not find it again. | |
e3a20c0b | 616 | */ |
1a3e8f3d DC |
617 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != |
618 | pag->pag_agno) | |
619 | continue; | |
e3a20c0b DC |
620 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
621 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
622 | done = 1; | |
623 | } | |
65d0f205 | 624 | |
e3a20c0b | 625 | /* unlock now we've grabbed the inodes. */ |
1a3e8f3d | 626 | rcu_read_unlock(); |
e3a20c0b DC |
627 | |
628 | for (i = 0; i < nr_found; i++) { | |
629 | if (!batch[i]) | |
630 | continue; | |
631 | error = xfs_reclaim_inode(batch[i], pag, flags); | |
632 | if (error && last_error != EFSCORRUPTED) | |
633 | last_error = error; | |
634 | } | |
635 | ||
636 | *nr_to_scan -= XFS_LOOKUP_BATCH; | |
65d0f205 | 637 | |
8daaa831 DC |
638 | cond_resched(); |
639 | ||
e3a20c0b | 640 | } while (nr_found && !done && *nr_to_scan > 0); |
65d0f205 | 641 | |
69b491c2 DC |
642 | if (trylock && !done) |
643 | pag->pag_ici_reclaim_cursor = first_index; | |
644 | else | |
645 | pag->pag_ici_reclaim_cursor = 0; | |
646 | mutex_unlock(&pag->pag_ici_reclaim_lock); | |
65d0f205 DC |
647 | xfs_perag_put(pag); |
648 | } | |
69b491c2 DC |
649 | |
650 | /* | |
651 | * if we skipped any AG, and we still have scan count remaining, do | |
652 | * another pass this time using blocking reclaim semantics (i.e | |
653 | * waiting on the reclaim locks and ignoring the reclaim cursors). This | |
654 | * ensure that when we get more reclaimers than AGs we block rather | |
655 | * than spin trying to execute reclaim. | |
656 | */ | |
8daaa831 | 657 | if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) { |
69b491c2 DC |
658 | trylock = 0; |
659 | goto restart; | |
660 | } | |
65d0f205 DC |
661 | return XFS_ERROR(last_error); |
662 | } | |
663 | ||
7a3be02b DC |
664 | int |
665 | xfs_reclaim_inodes( | |
666 | xfs_mount_t *mp, | |
7a3be02b DC |
667 | int mode) |
668 | { | |
65d0f205 DC |
669 | int nr_to_scan = INT_MAX; |
670 | ||
671 | return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan); | |
9bf729c0 DC |
672 | } |
673 | ||
674 | /* | |
8daaa831 | 675 | * Scan a certain number of inodes for reclaim. |
a7b339f1 DC |
676 | * |
677 | * When called we make sure that there is a background (fast) inode reclaim in | |
8daaa831 | 678 | * progress, while we will throttle the speed of reclaim via doing synchronous |
a7b339f1 DC |
679 | * reclaim of inodes. That means if we come across dirty inodes, we wait for |
680 | * them to be cleaned, which we hope will not be very long due to the | |
681 | * background walker having already kicked the IO off on those dirty inodes. | |
9bf729c0 | 682 | */ |
8daaa831 DC |
683 | void |
684 | xfs_reclaim_inodes_nr( | |
685 | struct xfs_mount *mp, | |
686 | int nr_to_scan) | |
9bf729c0 | 687 | { |
8daaa831 | 688 | /* kick background reclaimer and push the AIL */ |
5889608d | 689 | xfs_reclaim_work_queue(mp); |
8daaa831 | 690 | xfs_ail_push_all(mp->m_ail); |
a7b339f1 | 691 | |
8daaa831 DC |
692 | xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan); |
693 | } | |
9bf729c0 | 694 | |
8daaa831 DC |
695 | /* |
696 | * Return the number of reclaimable inodes in the filesystem for | |
697 | * the shrinker to determine how much to reclaim. | |
698 | */ | |
699 | int | |
700 | xfs_reclaim_inodes_count( | |
701 | struct xfs_mount *mp) | |
702 | { | |
703 | struct xfs_perag *pag; | |
704 | xfs_agnumber_t ag = 0; | |
705 | int reclaimable = 0; | |
9bf729c0 | 706 | |
65d0f205 DC |
707 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
708 | ag = pag->pag_agno + 1; | |
70e60ce7 DC |
709 | reclaimable += pag->pag_ici_reclaimable; |
710 | xfs_perag_put(pag); | |
9bf729c0 | 711 | } |
9bf729c0 DC |
712 | return reclaimable; |
713 | } | |
714 |