]>
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 | ||
c6d09b66 DC |
43 | struct workqueue_struct *xfs_syncd_wq; /* sync workqueue */ |
44 | ||
78ae5256 DC |
45 | /* |
46 | * The inode lookup is done in batches to keep the amount of lock traffic and | |
47 | * radix tree lookups to a minimum. The batch size is a trade off between | |
48 | * lookup reduction and stack usage. This is in the reclaim path, so we can't | |
49 | * be too greedy. | |
50 | */ | |
51 | #define XFS_LOOKUP_BATCH 32 | |
52 | ||
e13de955 DC |
53 | STATIC int |
54 | xfs_inode_ag_walk_grab( | |
55 | struct xfs_inode *ip) | |
56 | { | |
57 | struct inode *inode = VFS_I(ip); | |
58 | ||
1a3e8f3d DC |
59 | ASSERT(rcu_read_lock_held()); |
60 | ||
61 | /* | |
62 | * check for stale RCU freed inode | |
63 | * | |
64 | * If the inode has been reallocated, it doesn't matter if it's not in | |
65 | * the AG we are walking - we are walking for writeback, so if it | |
66 | * passes all the "valid inode" checks and is dirty, then we'll write | |
67 | * it back anyway. If it has been reallocated and still being | |
68 | * initialised, the XFS_INEW check below will catch it. | |
69 | */ | |
70 | spin_lock(&ip->i_flags_lock); | |
71 | if (!ip->i_ino) | |
72 | goto out_unlock_noent; | |
73 | ||
74 | /* avoid new or reclaimable inodes. Leave for reclaim code to flush */ | |
75 | if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM)) | |
76 | goto out_unlock_noent; | |
77 | spin_unlock(&ip->i_flags_lock); | |
78 | ||
e13de955 DC |
79 | /* nothing to sync during shutdown */ |
80 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) | |
81 | return EFSCORRUPTED; | |
82 | ||
e13de955 DC |
83 | /* If we can't grab the inode, it must on it's way to reclaim. */ |
84 | if (!igrab(inode)) | |
85 | return ENOENT; | |
86 | ||
87 | if (is_bad_inode(inode)) { | |
88 | IRELE(ip); | |
89 | return ENOENT; | |
90 | } | |
91 | ||
92 | /* inode is valid */ | |
93 | return 0; | |
1a3e8f3d DC |
94 | |
95 | out_unlock_noent: | |
96 | spin_unlock(&ip->i_flags_lock); | |
97 | return ENOENT; | |
e13de955 DC |
98 | } |
99 | ||
75f3cb13 DC |
100 | STATIC int |
101 | xfs_inode_ag_walk( | |
102 | struct xfs_mount *mp, | |
5017e97d | 103 | struct xfs_perag *pag, |
75f3cb13 DC |
104 | int (*execute)(struct xfs_inode *ip, |
105 | struct xfs_perag *pag, int flags), | |
65d0f205 | 106 | int flags) |
75f3cb13 | 107 | { |
75f3cb13 DC |
108 | uint32_t first_index; |
109 | int last_error = 0; | |
110 | int skipped; | |
65d0f205 | 111 | int done; |
78ae5256 | 112 | int nr_found; |
75f3cb13 DC |
113 | |
114 | restart: | |
65d0f205 | 115 | done = 0; |
75f3cb13 DC |
116 | skipped = 0; |
117 | first_index = 0; | |
78ae5256 | 118 | nr_found = 0; |
75f3cb13 | 119 | do { |
78ae5256 | 120 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
75f3cb13 | 121 | int error = 0; |
78ae5256 | 122 | int i; |
75f3cb13 | 123 | |
1a3e8f3d | 124 | rcu_read_lock(); |
65d0f205 | 125 | nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, |
78ae5256 DC |
126 | (void **)batch, first_index, |
127 | XFS_LOOKUP_BATCH); | |
65d0f205 | 128 | if (!nr_found) { |
1a3e8f3d | 129 | rcu_read_unlock(); |
75f3cb13 | 130 | break; |
c8e20be0 | 131 | } |
75f3cb13 | 132 | |
65d0f205 | 133 | /* |
78ae5256 DC |
134 | * Grab the inodes before we drop the lock. if we found |
135 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 136 | */ |
78ae5256 DC |
137 | for (i = 0; i < nr_found; i++) { |
138 | struct xfs_inode *ip = batch[i]; | |
139 | ||
140 | if (done || xfs_inode_ag_walk_grab(ip)) | |
141 | batch[i] = NULL; | |
142 | ||
143 | /* | |
1a3e8f3d DC |
144 | * Update the index for the next lookup. Catch |
145 | * overflows into the next AG range which can occur if | |
146 | * we have inodes in the last block of the AG and we | |
147 | * are currently pointing to the last inode. | |
148 | * | |
149 | * Because we may see inodes that are from the wrong AG | |
150 | * due to RCU freeing and reallocation, only update the | |
151 | * index if it lies in this AG. It was a race that lead | |
152 | * us to see this inode, so another lookup from the | |
153 | * same index will not find it again. | |
78ae5256 | 154 | */ |
1a3e8f3d DC |
155 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno) |
156 | continue; | |
78ae5256 DC |
157 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
158 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
159 | done = 1; | |
e13de955 | 160 | } |
78ae5256 DC |
161 | |
162 | /* unlock now we've grabbed the inodes. */ | |
1a3e8f3d | 163 | rcu_read_unlock(); |
e13de955 | 164 | |
78ae5256 DC |
165 | for (i = 0; i < nr_found; i++) { |
166 | if (!batch[i]) | |
167 | continue; | |
168 | error = execute(batch[i], pag, flags); | |
169 | IRELE(batch[i]); | |
170 | if (error == EAGAIN) { | |
171 | skipped++; | |
172 | continue; | |
173 | } | |
174 | if (error && last_error != EFSCORRUPTED) | |
175 | last_error = error; | |
75f3cb13 | 176 | } |
c8e20be0 DC |
177 | |
178 | /* bail out if the filesystem is corrupted. */ | |
75f3cb13 DC |
179 | if (error == EFSCORRUPTED) |
180 | break; | |
181 | ||
8daaa831 DC |
182 | cond_resched(); |
183 | ||
78ae5256 | 184 | } while (nr_found && !done); |
75f3cb13 DC |
185 | |
186 | if (skipped) { | |
187 | delay(1); | |
188 | goto restart; | |
189 | } | |
75f3cb13 DC |
190 | return last_error; |
191 | } | |
192 | ||
fe588ed3 | 193 | int |
75f3cb13 DC |
194 | xfs_inode_ag_iterator( |
195 | struct xfs_mount *mp, | |
196 | int (*execute)(struct xfs_inode *ip, | |
197 | struct xfs_perag *pag, int flags), | |
65d0f205 | 198 | int flags) |
75f3cb13 | 199 | { |
16fd5367 | 200 | struct xfs_perag *pag; |
75f3cb13 DC |
201 | int error = 0; |
202 | int last_error = 0; | |
203 | xfs_agnumber_t ag; | |
204 | ||
16fd5367 | 205 | ag = 0; |
65d0f205 DC |
206 | while ((pag = xfs_perag_get(mp, ag))) { |
207 | ag = pag->pag_agno + 1; | |
208 | error = xfs_inode_ag_walk(mp, pag, execute, flags); | |
5017e97d | 209 | xfs_perag_put(pag); |
75f3cb13 DC |
210 | if (error) { |
211 | last_error = error; | |
212 | if (error == EFSCORRUPTED) | |
213 | break; | |
214 | } | |
215 | } | |
216 | return XFS_ERROR(last_error); | |
217 | } | |
218 | ||
5d77c0dc | 219 | STATIC int |
2af75df7 | 220 | xfs_sync_fsdata( |
df308bcf | 221 | struct xfs_mount *mp) |
2af75df7 CH |
222 | { |
223 | struct xfs_buf *bp; | |
c2b006c1 | 224 | int error; |
2af75df7 CH |
225 | |
226 | /* | |
df308bcf CH |
227 | * If the buffer is pinned then push on the log so we won't get stuck |
228 | * waiting in the write for someone, maybe ourselves, to flush the log. | |
229 | * | |
230 | * Even though we just pushed the log above, we did not have the | |
231 | * superblock buffer locked at that point so it can become pinned in | |
232 | * between there and here. | |
2af75df7 | 233 | */ |
df308bcf | 234 | bp = xfs_getsb(mp, 0); |
811e64c7 | 235 | if (xfs_buf_ispinned(bp)) |
df308bcf | 236 | xfs_log_force(mp, 0); |
c2b006c1 CH |
237 | error = xfs_bwrite(bp); |
238 | xfs_buf_relse(bp); | |
239 | return error; | |
e9f1c6ee DC |
240 | } |
241 | ||
242 | /* | |
a4e4c4f4 DC |
243 | * When remounting a filesystem read-only or freezing the filesystem, we have |
244 | * two phases to execute. This first phase is syncing the data before we | |
245 | * quiesce the filesystem, and the second is flushing all the inodes out after | |
246 | * we've waited for all the transactions created by the first phase to | |
247 | * complete. The second phase ensures that the inodes are written to their | |
248 | * location on disk rather than just existing in transactions in the log. This | |
249 | * means after a quiesce there is no log replay required to write the inodes to | |
250 | * disk (this is the main difference between a sync and a quiesce). | |
251 | */ | |
252 | /* | |
253 | * First stage of freeze - no writers will make progress now we are here, | |
e9f1c6ee DC |
254 | * so we flush delwri and delalloc buffers here, then wait for all I/O to |
255 | * complete. Data is frozen at that point. Metadata is not frozen, | |
211e4d43 | 256 | * transactions can still occur here so don't bother emptying the AIL |
a4e4c4f4 | 257 | * because it'll just get dirty again. |
e9f1c6ee DC |
258 | */ |
259 | int | |
260 | xfs_quiesce_data( | |
261 | struct xfs_mount *mp) | |
262 | { | |
df308bcf | 263 | int error, error2 = 0; |
e9f1c6ee | 264 | |
34625c66 | 265 | /* force out the log */ |
33b8f7c2 CH |
266 | xfs_log_force(mp, XFS_LOG_SYNC); |
267 | ||
a4e4c4f4 | 268 | /* write superblock and hoover up shutdown errors */ |
df308bcf CH |
269 | error = xfs_sync_fsdata(mp); |
270 | ||
df308bcf CH |
271 | /* mark the log as covered if needed */ |
272 | if (xfs_log_need_covered(mp)) | |
c58efdb4 | 273 | error2 = xfs_fs_log_dummy(mp); |
e9f1c6ee | 274 | |
df308bcf | 275 | return error ? error : error2; |
2af75df7 CH |
276 | } |
277 | ||
76bf105c DC |
278 | /* |
279 | * Second stage of a quiesce. The data is already synced, now we have to take | |
280 | * care of the metadata. New transactions are already blocked, so we need to | |
25985edc | 281 | * wait for any remaining transactions to drain out before proceeding. |
7f7bebef DC |
282 | * |
283 | * Note: this stops background sync work - the callers must ensure it is started | |
284 | * again when appropriate. | |
76bf105c DC |
285 | */ |
286 | void | |
287 | xfs_quiesce_attr( | |
288 | struct xfs_mount *mp) | |
289 | { | |
290 | int error = 0; | |
291 | ||
292 | /* wait for all modifications to complete */ | |
293 | while (atomic_read(&mp->m_active_trans) > 0) | |
294 | delay(100); | |
295 | ||
211e4d43 CH |
296 | /* reclaim inodes to do any IO before the freeze completes */ |
297 | xfs_reclaim_inodes(mp, 0); | |
298 | xfs_reclaim_inodes(mp, SYNC_WAIT); | |
299 | ||
300 | /* flush all pending changes from the AIL */ | |
301 | xfs_ail_push_all_sync(mp->m_ail); | |
76bf105c | 302 | |
f661f1e0 DC |
303 | /* stop background log work */ |
304 | cancel_delayed_work_sync(&mp->m_log->l_work); | |
7f7bebef | 305 | |
5e106572 FB |
306 | /* |
307 | * Just warn here till VFS can correctly support | |
308 | * read-only remount without racing. | |
309 | */ | |
310 | WARN_ON(atomic_read(&mp->m_active_trans) != 0); | |
76bf105c DC |
311 | |
312 | /* Push the superblock and write an unmount record */ | |
adab0f67 | 313 | error = xfs_log_sbcount(mp); |
76bf105c | 314 | if (error) |
4f10700a | 315 | xfs_warn(mp, "xfs_attr_quiesce: failed to log sb changes. " |
76bf105c DC |
316 | "Frozen image may not be consistent."); |
317 | xfs_log_unmount_write(mp); | |
211e4d43 CH |
318 | |
319 | /* | |
320 | * At this point we might have modified the superblock again and thus | |
321 | * added an item to the AIL, thus flush it again. | |
322 | */ | |
323 | xfs_ail_push_all_sync(mp->m_ail); | |
9a57fa8e MT |
324 | |
325 | /* | |
326 | * The superblock buffer is uncached and xfsaild_push() will lock and | |
327 | * set the XBF_ASYNC flag on the buffer. We cannot do xfs_buf_iowait() | |
328 | * here but a lock on the superblock buffer will block until iodone() | |
329 | * has completed. | |
330 | */ | |
331 | xfs_buf_lock(mp->m_sb_bp); | |
332 | xfs_buf_unlock(mp->m_sb_bp); | |
76bf105c DC |
333 | } |
334 | ||
a7b339f1 DC |
335 | /* |
336 | * Queue a new inode reclaim pass if there are reclaimable inodes and there | |
337 | * isn't a reclaim pass already in progress. By default it runs every 5s based | |
338 | * on the xfs syncd work default of 30s. Perhaps this should have it's own | |
339 | * tunable, but that can be done if this method proves to be ineffective or too | |
340 | * aggressive. | |
341 | */ | |
342 | static void | |
343 | xfs_syncd_queue_reclaim( | |
344 | struct xfs_mount *mp) | |
a167b17e | 345 | { |
a167b17e | 346 | |
a7b339f1 DC |
347 | rcu_read_lock(); |
348 | if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) { | |
349 | queue_delayed_work(xfs_syncd_wq, &mp->m_reclaim_work, | |
350 | msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10)); | |
a167b17e | 351 | } |
a7b339f1 DC |
352 | rcu_read_unlock(); |
353 | } | |
a167b17e | 354 | |
a7b339f1 DC |
355 | /* |
356 | * This is a fast pass over the inode cache to try to get reclaim moving on as | |
357 | * many inodes as possible in a short period of time. It kicks itself every few | |
358 | * seconds, as well as being kicked by the inode cache shrinker when memory | |
359 | * goes low. It scans as quickly as possible avoiding locked inodes or those | |
360 | * already being flushed, and once done schedules a future pass. | |
361 | */ | |
33c7a2bc | 362 | void |
a7b339f1 DC |
363 | xfs_reclaim_worker( |
364 | struct work_struct *work) | |
365 | { | |
366 | struct xfs_mount *mp = container_of(to_delayed_work(work), | |
367 | struct xfs_mount, m_reclaim_work); | |
368 | ||
369 | xfs_reclaim_inodes(mp, SYNC_TRYLOCK); | |
370 | xfs_syncd_queue_reclaim(mp); | |
371 | } | |
372 | ||
bc990f5c CH |
373 | void |
374 | __xfs_inode_set_reclaim_tag( | |
375 | struct xfs_perag *pag, | |
376 | struct xfs_inode *ip) | |
377 | { | |
378 | radix_tree_tag_set(&pag->pag_ici_root, | |
379 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
380 | XFS_ICI_RECLAIM_TAG); | |
16fd5367 DC |
381 | |
382 | if (!pag->pag_ici_reclaimable) { | |
383 | /* propagate the reclaim tag up into the perag radix tree */ | |
384 | spin_lock(&ip->i_mount->m_perag_lock); | |
385 | radix_tree_tag_set(&ip->i_mount->m_perag_tree, | |
386 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
387 | XFS_ICI_RECLAIM_TAG); | |
388 | spin_unlock(&ip->i_mount->m_perag_lock); | |
a7b339f1 DC |
389 | |
390 | /* schedule periodic background inode reclaim */ | |
391 | xfs_syncd_queue_reclaim(ip->i_mount); | |
392 | ||
16fd5367 DC |
393 | trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno, |
394 | -1, _RET_IP_); | |
395 | } | |
9bf729c0 | 396 | pag->pag_ici_reclaimable++; |
bc990f5c CH |
397 | } |
398 | ||
11654513 DC |
399 | /* |
400 | * We set the inode flag atomically with the radix tree tag. | |
401 | * Once we get tag lookups on the radix tree, this inode flag | |
402 | * can go away. | |
403 | */ | |
396beb85 DC |
404 | void |
405 | xfs_inode_set_reclaim_tag( | |
406 | xfs_inode_t *ip) | |
407 | { | |
5017e97d DC |
408 | struct xfs_mount *mp = ip->i_mount; |
409 | struct xfs_perag *pag; | |
396beb85 | 410 | |
5017e97d | 411 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); |
1a427ab0 | 412 | spin_lock(&pag->pag_ici_lock); |
396beb85 | 413 | spin_lock(&ip->i_flags_lock); |
bc990f5c | 414 | __xfs_inode_set_reclaim_tag(pag, ip); |
11654513 | 415 | __xfs_iflags_set(ip, XFS_IRECLAIMABLE); |
396beb85 | 416 | spin_unlock(&ip->i_flags_lock); |
1a427ab0 | 417 | spin_unlock(&pag->pag_ici_lock); |
5017e97d | 418 | xfs_perag_put(pag); |
396beb85 DC |
419 | } |
420 | ||
081003ff JW |
421 | STATIC void |
422 | __xfs_inode_clear_reclaim( | |
396beb85 DC |
423 | xfs_perag_t *pag, |
424 | xfs_inode_t *ip) | |
425 | { | |
9bf729c0 | 426 | pag->pag_ici_reclaimable--; |
16fd5367 DC |
427 | if (!pag->pag_ici_reclaimable) { |
428 | /* clear the reclaim tag from the perag radix tree */ | |
429 | spin_lock(&ip->i_mount->m_perag_lock); | |
430 | radix_tree_tag_clear(&ip->i_mount->m_perag_tree, | |
431 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
432 | XFS_ICI_RECLAIM_TAG); | |
433 | spin_unlock(&ip->i_mount->m_perag_lock); | |
434 | trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno, | |
435 | -1, _RET_IP_); | |
436 | } | |
396beb85 DC |
437 | } |
438 | ||
081003ff JW |
439 | void |
440 | __xfs_inode_clear_reclaim_tag( | |
441 | xfs_mount_t *mp, | |
442 | xfs_perag_t *pag, | |
443 | xfs_inode_t *ip) | |
444 | { | |
445 | radix_tree_tag_clear(&pag->pag_ici_root, | |
446 | XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG); | |
447 | __xfs_inode_clear_reclaim(pag, ip); | |
448 | } | |
449 | ||
e3a20c0b DC |
450 | /* |
451 | * Grab the inode for reclaim exclusively. | |
452 | * Return 0 if we grabbed it, non-zero otherwise. | |
453 | */ | |
454 | STATIC int | |
455 | xfs_reclaim_inode_grab( | |
456 | struct xfs_inode *ip, | |
457 | int flags) | |
458 | { | |
1a3e8f3d DC |
459 | ASSERT(rcu_read_lock_held()); |
460 | ||
461 | /* quick check for stale RCU freed inode */ | |
462 | if (!ip->i_ino) | |
463 | return 1; | |
e3a20c0b DC |
464 | |
465 | /* | |
474fce06 CH |
466 | * If we are asked for non-blocking operation, do unlocked checks to |
467 | * see if the inode already is being flushed or in reclaim to avoid | |
468 | * lock traffic. | |
e3a20c0b DC |
469 | */ |
470 | if ((flags & SYNC_TRYLOCK) && | |
474fce06 | 471 | __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM)) |
e3a20c0b | 472 | return 1; |
e3a20c0b DC |
473 | |
474 | /* | |
475 | * The radix tree lock here protects a thread in xfs_iget from racing | |
476 | * with us starting reclaim on the inode. Once we have the | |
477 | * XFS_IRECLAIM flag set it will not touch us. | |
1a3e8f3d DC |
478 | * |
479 | * Due to RCU lookup, we may find inodes that have been freed and only | |
480 | * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that | |
481 | * aren't candidates for reclaim at all, so we must check the | |
482 | * XFS_IRECLAIMABLE is set first before proceeding to reclaim. | |
e3a20c0b DC |
483 | */ |
484 | spin_lock(&ip->i_flags_lock); | |
1a3e8f3d DC |
485 | if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) || |
486 | __xfs_iflags_test(ip, XFS_IRECLAIM)) { | |
487 | /* not a reclaim candidate. */ | |
e3a20c0b DC |
488 | spin_unlock(&ip->i_flags_lock); |
489 | return 1; | |
490 | } | |
491 | __xfs_iflags_set(ip, XFS_IRECLAIM); | |
492 | spin_unlock(&ip->i_flags_lock); | |
493 | return 0; | |
494 | } | |
495 | ||
777df5af | 496 | /* |
8a48088f CH |
497 | * Inodes in different states need to be treated differently. The following |
498 | * table lists the inode states and the reclaim actions necessary: | |
777df5af DC |
499 | * |
500 | * inode state iflush ret required action | |
501 | * --------------- ---------- --------------- | |
502 | * bad - reclaim | |
503 | * shutdown EIO unpin and reclaim | |
504 | * clean, unpinned 0 reclaim | |
505 | * stale, unpinned 0 reclaim | |
c854363e DC |
506 | * clean, pinned(*) 0 requeue |
507 | * stale, pinned EAGAIN requeue | |
8a48088f CH |
508 | * dirty, async - requeue |
509 | * dirty, sync 0 reclaim | |
777df5af DC |
510 | * |
511 | * (*) dgc: I don't think the clean, pinned state is possible but it gets | |
512 | * handled anyway given the order of checks implemented. | |
513 | * | |
c854363e DC |
514 | * Also, because we get the flush lock first, we know that any inode that has |
515 | * been flushed delwri has had the flush completed by the time we check that | |
8a48088f | 516 | * the inode is clean. |
c854363e | 517 | * |
8a48088f CH |
518 | * Note that because the inode is flushed delayed write by AIL pushing, the |
519 | * flush lock may already be held here and waiting on it can result in very | |
520 | * long latencies. Hence for sync reclaims, where we wait on the flush lock, | |
521 | * the caller should push the AIL first before trying to reclaim inodes to | |
522 | * minimise the amount of time spent waiting. For background relaim, we only | |
523 | * bother to reclaim clean inodes anyway. | |
c854363e | 524 | * |
777df5af DC |
525 | * Hence the order of actions after gaining the locks should be: |
526 | * bad => reclaim | |
527 | * shutdown => unpin and reclaim | |
8a48088f | 528 | * pinned, async => requeue |
c854363e | 529 | * pinned, sync => unpin |
777df5af DC |
530 | * stale => reclaim |
531 | * clean => reclaim | |
8a48088f | 532 | * dirty, async => requeue |
c854363e | 533 | * dirty, sync => flush, wait and reclaim |
777df5af | 534 | */ |
75f3cb13 | 535 | STATIC int |
c8e20be0 | 536 | xfs_reclaim_inode( |
75f3cb13 DC |
537 | struct xfs_inode *ip, |
538 | struct xfs_perag *pag, | |
c8e20be0 | 539 | int sync_mode) |
fce08f2f | 540 | { |
4c46819a CH |
541 | struct xfs_buf *bp = NULL; |
542 | int error; | |
777df5af | 543 | |
1bfd8d04 DC |
544 | restart: |
545 | error = 0; | |
c8e20be0 | 546 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
c854363e DC |
547 | if (!xfs_iflock_nowait(ip)) { |
548 | if (!(sync_mode & SYNC_WAIT)) | |
549 | goto out; | |
550 | xfs_iflock(ip); | |
551 | } | |
7a3be02b | 552 | |
777df5af DC |
553 | if (is_bad_inode(VFS_I(ip))) |
554 | goto reclaim; | |
555 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) { | |
556 | xfs_iunpin_wait(ip); | |
04913fdd | 557 | xfs_iflush_abort(ip, false); |
777df5af DC |
558 | goto reclaim; |
559 | } | |
c854363e | 560 | if (xfs_ipincount(ip)) { |
8a48088f CH |
561 | if (!(sync_mode & SYNC_WAIT)) |
562 | goto out_ifunlock; | |
777df5af | 563 | xfs_iunpin_wait(ip); |
c854363e | 564 | } |
777df5af DC |
565 | if (xfs_iflags_test(ip, XFS_ISTALE)) |
566 | goto reclaim; | |
567 | if (xfs_inode_clean(ip)) | |
568 | goto reclaim; | |
569 | ||
8a48088f CH |
570 | /* |
571 | * Never flush out dirty data during non-blocking reclaim, as it would | |
572 | * just contend with AIL pushing trying to do the same job. | |
573 | */ | |
574 | if (!(sync_mode & SYNC_WAIT)) | |
575 | goto out_ifunlock; | |
576 | ||
1bfd8d04 DC |
577 | /* |
578 | * Now we have an inode that needs flushing. | |
579 | * | |
4c46819a | 580 | * Note that xfs_iflush will never block on the inode buffer lock, as |
1bfd8d04 | 581 | * xfs_ifree_cluster() can lock the inode buffer before it locks the |
4c46819a | 582 | * ip->i_lock, and we are doing the exact opposite here. As a result, |
475ee413 CH |
583 | * doing a blocking xfs_imap_to_bp() to get the cluster buffer would |
584 | * result in an ABBA deadlock with xfs_ifree_cluster(). | |
1bfd8d04 DC |
585 | * |
586 | * As xfs_ifree_cluser() must gather all inodes that are active in the | |
587 | * cache to mark them stale, if we hit this case we don't actually want | |
588 | * to do IO here - we want the inode marked stale so we can simply | |
4c46819a CH |
589 | * reclaim it. Hence if we get an EAGAIN error here, just unlock the |
590 | * inode, back off and try again. Hopefully the next pass through will | |
591 | * see the stale flag set on the inode. | |
1bfd8d04 | 592 | */ |
4c46819a | 593 | error = xfs_iflush(ip, &bp); |
8a48088f CH |
594 | if (error == EAGAIN) { |
595 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
596 | /* backoff longer than in xfs_ifree_cluster */ | |
597 | delay(2); | |
598 | goto restart; | |
c854363e | 599 | } |
c854363e | 600 | |
4c46819a CH |
601 | if (!error) { |
602 | error = xfs_bwrite(bp); | |
603 | xfs_buf_relse(bp); | |
604 | } | |
605 | ||
606 | xfs_iflock(ip); | |
777df5af DC |
607 | reclaim: |
608 | xfs_ifunlock(ip); | |
c8e20be0 | 609 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
610 | |
611 | XFS_STATS_INC(xs_ig_reclaims); | |
612 | /* | |
613 | * Remove the inode from the per-AG radix tree. | |
614 | * | |
615 | * Because radix_tree_delete won't complain even if the item was never | |
616 | * added to the tree assert that it's been there before to catch | |
617 | * problems with the inode life time early on. | |
618 | */ | |
1a427ab0 | 619 | spin_lock(&pag->pag_ici_lock); |
2f11feab DC |
620 | if (!radix_tree_delete(&pag->pag_ici_root, |
621 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino))) | |
622 | ASSERT(0); | |
081003ff | 623 | __xfs_inode_clear_reclaim(pag, ip); |
1a427ab0 | 624 | spin_unlock(&pag->pag_ici_lock); |
2f11feab DC |
625 | |
626 | /* | |
627 | * Here we do an (almost) spurious inode lock in order to coordinate | |
628 | * with inode cache radix tree lookups. This is because the lookup | |
629 | * can reference the inodes in the cache without taking references. | |
630 | * | |
631 | * We make that OK here by ensuring that we wait until the inode is | |
ad637a10 | 632 | * unlocked after the lookup before we go ahead and free it. |
2f11feab | 633 | */ |
ad637a10 | 634 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
2f11feab | 635 | xfs_qm_dqdetach(ip); |
ad637a10 | 636 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
637 | |
638 | xfs_inode_free(ip); | |
ad637a10 | 639 | return error; |
8a48088f CH |
640 | |
641 | out_ifunlock: | |
642 | xfs_ifunlock(ip); | |
643 | out: | |
644 | xfs_iflags_clear(ip, XFS_IRECLAIM); | |
645 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
646 | /* | |
647 | * We could return EAGAIN here to make reclaim rescan the inode tree in | |
648 | * a short while. However, this just burns CPU time scanning the tree | |
649 | * waiting for IO to complete and xfssyncd never goes back to the idle | |
650 | * state. Instead, return 0 to let the next scheduled background reclaim | |
651 | * attempt to reclaim the inode again. | |
652 | */ | |
653 | return 0; | |
7a3be02b DC |
654 | } |
655 | ||
65d0f205 DC |
656 | /* |
657 | * Walk the AGs and reclaim the inodes in them. Even if the filesystem is | |
658 | * corrupted, we still want to try to reclaim all the inodes. If we don't, | |
659 | * then a shut down during filesystem unmount reclaim walk leak all the | |
660 | * unreclaimed inodes. | |
661 | */ | |
662 | int | |
663 | xfs_reclaim_inodes_ag( | |
664 | struct xfs_mount *mp, | |
665 | int flags, | |
666 | int *nr_to_scan) | |
667 | { | |
668 | struct xfs_perag *pag; | |
669 | int error = 0; | |
670 | int last_error = 0; | |
671 | xfs_agnumber_t ag; | |
69b491c2 DC |
672 | int trylock = flags & SYNC_TRYLOCK; |
673 | int skipped; | |
65d0f205 | 674 | |
69b491c2 | 675 | restart: |
65d0f205 | 676 | ag = 0; |
69b491c2 | 677 | skipped = 0; |
65d0f205 DC |
678 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
679 | unsigned long first_index = 0; | |
680 | int done = 0; | |
e3a20c0b | 681 | int nr_found = 0; |
65d0f205 DC |
682 | |
683 | ag = pag->pag_agno + 1; | |
684 | ||
69b491c2 DC |
685 | if (trylock) { |
686 | if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) { | |
687 | skipped++; | |
f83282a8 | 688 | xfs_perag_put(pag); |
69b491c2 DC |
689 | continue; |
690 | } | |
691 | first_index = pag->pag_ici_reclaim_cursor; | |
692 | } else | |
693 | mutex_lock(&pag->pag_ici_reclaim_lock); | |
694 | ||
65d0f205 | 695 | do { |
e3a20c0b DC |
696 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
697 | int i; | |
65d0f205 | 698 | |
1a3e8f3d | 699 | rcu_read_lock(); |
e3a20c0b DC |
700 | nr_found = radix_tree_gang_lookup_tag( |
701 | &pag->pag_ici_root, | |
702 | (void **)batch, first_index, | |
703 | XFS_LOOKUP_BATCH, | |
65d0f205 DC |
704 | XFS_ICI_RECLAIM_TAG); |
705 | if (!nr_found) { | |
b2232219 | 706 | done = 1; |
1a3e8f3d | 707 | rcu_read_unlock(); |
65d0f205 DC |
708 | break; |
709 | } | |
710 | ||
711 | /* | |
e3a20c0b DC |
712 | * Grab the inodes before we drop the lock. if we found |
713 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 714 | */ |
e3a20c0b DC |
715 | for (i = 0; i < nr_found; i++) { |
716 | struct xfs_inode *ip = batch[i]; | |
717 | ||
718 | if (done || xfs_reclaim_inode_grab(ip, flags)) | |
719 | batch[i] = NULL; | |
720 | ||
721 | /* | |
722 | * Update the index for the next lookup. Catch | |
723 | * overflows into the next AG range which can | |
724 | * occur if we have inodes in the last block of | |
725 | * the AG and we are currently pointing to the | |
726 | * last inode. | |
1a3e8f3d DC |
727 | * |
728 | * Because we may see inodes that are from the | |
729 | * wrong AG due to RCU freeing and | |
730 | * reallocation, only update the index if it | |
731 | * lies in this AG. It was a race that lead us | |
732 | * to see this inode, so another lookup from | |
733 | * the same index will not find it again. | |
e3a20c0b | 734 | */ |
1a3e8f3d DC |
735 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != |
736 | pag->pag_agno) | |
737 | continue; | |
e3a20c0b DC |
738 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
739 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
740 | done = 1; | |
741 | } | |
65d0f205 | 742 | |
e3a20c0b | 743 | /* unlock now we've grabbed the inodes. */ |
1a3e8f3d | 744 | rcu_read_unlock(); |
e3a20c0b DC |
745 | |
746 | for (i = 0; i < nr_found; i++) { | |
747 | if (!batch[i]) | |
748 | continue; | |
749 | error = xfs_reclaim_inode(batch[i], pag, flags); | |
750 | if (error && last_error != EFSCORRUPTED) | |
751 | last_error = error; | |
752 | } | |
753 | ||
754 | *nr_to_scan -= XFS_LOOKUP_BATCH; | |
65d0f205 | 755 | |
8daaa831 DC |
756 | cond_resched(); |
757 | ||
e3a20c0b | 758 | } while (nr_found && !done && *nr_to_scan > 0); |
65d0f205 | 759 | |
69b491c2 DC |
760 | if (trylock && !done) |
761 | pag->pag_ici_reclaim_cursor = first_index; | |
762 | else | |
763 | pag->pag_ici_reclaim_cursor = 0; | |
764 | mutex_unlock(&pag->pag_ici_reclaim_lock); | |
65d0f205 DC |
765 | xfs_perag_put(pag); |
766 | } | |
69b491c2 DC |
767 | |
768 | /* | |
769 | * if we skipped any AG, and we still have scan count remaining, do | |
770 | * another pass this time using blocking reclaim semantics (i.e | |
771 | * waiting on the reclaim locks and ignoring the reclaim cursors). This | |
772 | * ensure that when we get more reclaimers than AGs we block rather | |
773 | * than spin trying to execute reclaim. | |
774 | */ | |
8daaa831 | 775 | if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) { |
69b491c2 DC |
776 | trylock = 0; |
777 | goto restart; | |
778 | } | |
65d0f205 DC |
779 | return XFS_ERROR(last_error); |
780 | } | |
781 | ||
7a3be02b DC |
782 | int |
783 | xfs_reclaim_inodes( | |
784 | xfs_mount_t *mp, | |
7a3be02b DC |
785 | int mode) |
786 | { | |
65d0f205 DC |
787 | int nr_to_scan = INT_MAX; |
788 | ||
789 | return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan); | |
9bf729c0 DC |
790 | } |
791 | ||
792 | /* | |
8daaa831 | 793 | * Scan a certain number of inodes for reclaim. |
a7b339f1 DC |
794 | * |
795 | * When called we make sure that there is a background (fast) inode reclaim in | |
8daaa831 | 796 | * progress, while we will throttle the speed of reclaim via doing synchronous |
a7b339f1 DC |
797 | * reclaim of inodes. That means if we come across dirty inodes, we wait for |
798 | * them to be cleaned, which we hope will not be very long due to the | |
799 | * background walker having already kicked the IO off on those dirty inodes. | |
9bf729c0 | 800 | */ |
8daaa831 DC |
801 | void |
802 | xfs_reclaim_inodes_nr( | |
803 | struct xfs_mount *mp, | |
804 | int nr_to_scan) | |
9bf729c0 | 805 | { |
8daaa831 DC |
806 | /* kick background reclaimer and push the AIL */ |
807 | xfs_syncd_queue_reclaim(mp); | |
808 | xfs_ail_push_all(mp->m_ail); | |
a7b339f1 | 809 | |
8daaa831 DC |
810 | xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan); |
811 | } | |
9bf729c0 | 812 | |
8daaa831 DC |
813 | /* |
814 | * Return the number of reclaimable inodes in the filesystem for | |
815 | * the shrinker to determine how much to reclaim. | |
816 | */ | |
817 | int | |
818 | xfs_reclaim_inodes_count( | |
819 | struct xfs_mount *mp) | |
820 | { | |
821 | struct xfs_perag *pag; | |
822 | xfs_agnumber_t ag = 0; | |
823 | int reclaimable = 0; | |
9bf729c0 | 824 | |
65d0f205 DC |
825 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
826 | ag = pag->pag_agno + 1; | |
70e60ce7 DC |
827 | reclaimable += pag->pag_ici_reclaimable; |
828 | xfs_perag_put(pag); | |
9bf729c0 | 829 | } |
9bf729c0 DC |
830 | return reclaimable; |
831 | } | |
832 |