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" | |
21 | #include "xfs_bit.h" | |
22 | #include "xfs_log.h" | |
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 | ||
5a34d5cd DC |
219 | STATIC int |
220 | xfs_sync_inode_data( | |
221 | struct xfs_inode *ip, | |
75f3cb13 | 222 | struct xfs_perag *pag, |
5a34d5cd DC |
223 | int flags) |
224 | { | |
225 | struct inode *inode = VFS_I(ip); | |
226 | struct address_space *mapping = inode->i_mapping; | |
227 | int error = 0; | |
228 | ||
229 | if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) | |
4a06fd26 | 230 | return 0; |
5a34d5cd DC |
231 | |
232 | if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED)) { | |
233 | if (flags & SYNC_TRYLOCK) | |
4a06fd26 | 234 | return 0; |
5a34d5cd DC |
235 | xfs_ilock(ip, XFS_IOLOCK_SHARED); |
236 | } | |
237 | ||
238 | error = xfs_flush_pages(ip, 0, -1, (flags & SYNC_WAIT) ? | |
0cadda1c | 239 | 0 : XBF_ASYNC, FI_NONE); |
5a34d5cd | 240 | xfs_iunlock(ip, XFS_IOLOCK_SHARED); |
5a34d5cd DC |
241 | return error; |
242 | } | |
243 | ||
075fe102 CH |
244 | /* |
245 | * Write out pagecache data for the whole filesystem. | |
246 | */ | |
64c86149 | 247 | STATIC int |
075fe102 CH |
248 | xfs_sync_data( |
249 | struct xfs_mount *mp, | |
250 | int flags) | |
683a8970 | 251 | { |
075fe102 | 252 | int error; |
fe4fa4b8 | 253 | |
b0710ccc | 254 | ASSERT((flags & ~(SYNC_TRYLOCK|SYNC_WAIT)) == 0); |
fe4fa4b8 | 255 | |
65d0f205 | 256 | error = xfs_inode_ag_iterator(mp, xfs_sync_inode_data, flags); |
075fe102 CH |
257 | if (error) |
258 | return XFS_ERROR(error); | |
e9f1c6ee | 259 | |
a14a348b | 260 | xfs_log_force(mp, (flags & SYNC_WAIT) ? XFS_LOG_SYNC : 0); |
075fe102 CH |
261 | return 0; |
262 | } | |
e9f1c6ee | 263 | |
5d77c0dc | 264 | STATIC int |
2af75df7 | 265 | xfs_sync_fsdata( |
df308bcf | 266 | struct xfs_mount *mp) |
2af75df7 CH |
267 | { |
268 | struct xfs_buf *bp; | |
c2b006c1 | 269 | int error; |
2af75df7 CH |
270 | |
271 | /* | |
df308bcf CH |
272 | * If the buffer is pinned then push on the log so we won't get stuck |
273 | * waiting in the write for someone, maybe ourselves, to flush the log. | |
274 | * | |
275 | * Even though we just pushed the log above, we did not have the | |
276 | * superblock buffer locked at that point so it can become pinned in | |
277 | * between there and here. | |
2af75df7 | 278 | */ |
df308bcf | 279 | bp = xfs_getsb(mp, 0); |
811e64c7 | 280 | if (xfs_buf_ispinned(bp)) |
df308bcf | 281 | xfs_log_force(mp, 0); |
c2b006c1 CH |
282 | error = xfs_bwrite(bp); |
283 | xfs_buf_relse(bp); | |
284 | return error; | |
e9f1c6ee DC |
285 | } |
286 | ||
287 | /* | |
a4e4c4f4 DC |
288 | * When remounting a filesystem read-only or freezing the filesystem, we have |
289 | * two phases to execute. This first phase is syncing the data before we | |
290 | * quiesce the filesystem, and the second is flushing all the inodes out after | |
291 | * we've waited for all the transactions created by the first phase to | |
292 | * complete. The second phase ensures that the inodes are written to their | |
293 | * location on disk rather than just existing in transactions in the log. This | |
294 | * means after a quiesce there is no log replay required to write the inodes to | |
295 | * disk (this is the main difference between a sync and a quiesce). | |
296 | */ | |
297 | /* | |
298 | * First stage of freeze - no writers will make progress now we are here, | |
e9f1c6ee DC |
299 | * so we flush delwri and delalloc buffers here, then wait for all I/O to |
300 | * complete. Data is frozen at that point. Metadata is not frozen, | |
211e4d43 | 301 | * transactions can still occur here so don't bother emptying the AIL |
a4e4c4f4 | 302 | * because it'll just get dirty again. |
e9f1c6ee DC |
303 | */ |
304 | int | |
305 | xfs_quiesce_data( | |
306 | struct xfs_mount *mp) | |
307 | { | |
df308bcf | 308 | int error, error2 = 0; |
e9f1c6ee | 309 | |
34625c66 | 310 | /* force out the log */ |
33b8f7c2 CH |
311 | xfs_log_force(mp, XFS_LOG_SYNC); |
312 | ||
a4e4c4f4 | 313 | /* write superblock and hoover up shutdown errors */ |
df308bcf CH |
314 | error = xfs_sync_fsdata(mp); |
315 | ||
df308bcf CH |
316 | /* mark the log as covered if needed */ |
317 | if (xfs_log_need_covered(mp)) | |
c58efdb4 | 318 | error2 = xfs_fs_log_dummy(mp); |
e9f1c6ee | 319 | |
df308bcf | 320 | return error ? error : error2; |
2af75df7 CH |
321 | } |
322 | ||
76bf105c DC |
323 | /* |
324 | * Second stage of a quiesce. The data is already synced, now we have to take | |
325 | * care of the metadata. New transactions are already blocked, so we need to | |
25985edc | 326 | * wait for any remaining transactions to drain out before proceeding. |
76bf105c DC |
327 | */ |
328 | void | |
329 | xfs_quiesce_attr( | |
330 | struct xfs_mount *mp) | |
331 | { | |
332 | int error = 0; | |
333 | ||
334 | /* wait for all modifications to complete */ | |
335 | while (atomic_read(&mp->m_active_trans) > 0) | |
336 | delay(100); | |
337 | ||
211e4d43 CH |
338 | /* reclaim inodes to do any IO before the freeze completes */ |
339 | xfs_reclaim_inodes(mp, 0); | |
340 | xfs_reclaim_inodes(mp, SYNC_WAIT); | |
341 | ||
342 | /* flush all pending changes from the AIL */ | |
343 | xfs_ail_push_all_sync(mp->m_ail); | |
76bf105c | 344 | |
5e106572 FB |
345 | /* |
346 | * Just warn here till VFS can correctly support | |
347 | * read-only remount without racing. | |
348 | */ | |
349 | WARN_ON(atomic_read(&mp->m_active_trans) != 0); | |
76bf105c DC |
350 | |
351 | /* Push the superblock and write an unmount record */ | |
adab0f67 | 352 | error = xfs_log_sbcount(mp); |
76bf105c | 353 | if (error) |
4f10700a | 354 | xfs_warn(mp, "xfs_attr_quiesce: failed to log sb changes. " |
76bf105c DC |
355 | "Frozen image may not be consistent."); |
356 | xfs_log_unmount_write(mp); | |
211e4d43 CH |
357 | |
358 | /* | |
359 | * At this point we might have modified the superblock again and thus | |
360 | * added an item to the AIL, thus flush it again. | |
361 | */ | |
362 | xfs_ail_push_all_sync(mp->m_ail); | |
76bf105c DC |
363 | } |
364 | ||
c6d09b66 DC |
365 | static void |
366 | xfs_syncd_queue_sync( | |
367 | struct xfs_mount *mp) | |
a167b17e | 368 | { |
c6d09b66 DC |
369 | queue_delayed_work(xfs_syncd_wq, &mp->m_sync_work, |
370 | msecs_to_jiffies(xfs_syncd_centisecs * 10)); | |
a167b17e DC |
371 | } |
372 | ||
aacaa880 | 373 | /* |
df308bcf CH |
374 | * Every sync period we need to unpin all items, reclaim inodes and sync |
375 | * disk quotas. We might need to cover the log to indicate that the | |
1a387d3b | 376 | * filesystem is idle and not frozen. |
aacaa880 | 377 | */ |
a167b17e DC |
378 | STATIC void |
379 | xfs_sync_worker( | |
c6d09b66 | 380 | struct work_struct *work) |
a167b17e | 381 | { |
c6d09b66 DC |
382 | struct xfs_mount *mp = container_of(to_delayed_work(work), |
383 | struct xfs_mount, m_sync_work); | |
a167b17e DC |
384 | int error; |
385 | ||
8a00ebe4 DC |
386 | /* |
387 | * We shouldn't write/force the log if we are in the mount/unmount | |
388 | * process or on a read only filesystem. The workqueue still needs to be | |
389 | * active in both cases, however, because it is used for inode reclaim | |
390 | * during these times. hence use the MS_ACTIVE flag to avoid doing | |
391 | * anything in these periods. | |
392 | */ | |
393 | if (!(mp->m_super->s_flags & MS_ACTIVE) && | |
394 | !(mp->m_flags & XFS_MOUNT_RDONLY)) { | |
aacaa880 | 395 | /* dgc: errors ignored here */ |
1a387d3b DC |
396 | if (mp->m_super->s_frozen == SB_UNFROZEN && |
397 | xfs_log_need_covered(mp)) | |
c58efdb4 DC |
398 | error = xfs_fs_log_dummy(mp); |
399 | else | |
400 | xfs_log_force(mp, 0); | |
fd074841 DC |
401 | |
402 | /* start pushing all the metadata that is currently dirty */ | |
403 | xfs_ail_push_all(mp->m_ail); | |
aacaa880 | 404 | } |
c6d09b66 DC |
405 | |
406 | /* queue us up again */ | |
407 | xfs_syncd_queue_sync(mp); | |
a167b17e DC |
408 | } |
409 | ||
a7b339f1 DC |
410 | /* |
411 | * Queue a new inode reclaim pass if there are reclaimable inodes and there | |
412 | * isn't a reclaim pass already in progress. By default it runs every 5s based | |
413 | * on the xfs syncd work default of 30s. Perhaps this should have it's own | |
414 | * tunable, but that can be done if this method proves to be ineffective or too | |
415 | * aggressive. | |
416 | */ | |
417 | static void | |
418 | xfs_syncd_queue_reclaim( | |
419 | struct xfs_mount *mp) | |
a167b17e | 420 | { |
a167b17e | 421 | |
a7b339f1 DC |
422 | rcu_read_lock(); |
423 | if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) { | |
424 | queue_delayed_work(xfs_syncd_wq, &mp->m_reclaim_work, | |
425 | msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10)); | |
a167b17e | 426 | } |
a7b339f1 DC |
427 | rcu_read_unlock(); |
428 | } | |
a167b17e | 429 | |
a7b339f1 DC |
430 | /* |
431 | * This is a fast pass over the inode cache to try to get reclaim moving on as | |
432 | * many inodes as possible in a short period of time. It kicks itself every few | |
433 | * seconds, as well as being kicked by the inode cache shrinker when memory | |
434 | * goes low. It scans as quickly as possible avoiding locked inodes or those | |
435 | * already being flushed, and once done schedules a future pass. | |
436 | */ | |
437 | STATIC void | |
438 | xfs_reclaim_worker( | |
439 | struct work_struct *work) | |
440 | { | |
441 | struct xfs_mount *mp = container_of(to_delayed_work(work), | |
442 | struct xfs_mount, m_reclaim_work); | |
443 | ||
444 | xfs_reclaim_inodes(mp, SYNC_TRYLOCK); | |
445 | xfs_syncd_queue_reclaim(mp); | |
446 | } | |
447 | ||
89e4cb55 DC |
448 | /* |
449 | * Flush delayed allocate data, attempting to free up reserved space | |
450 | * from existing allocations. At this point a new allocation attempt | |
451 | * has failed with ENOSPC and we are in the process of scratching our | |
452 | * heads, looking about for more room. | |
453 | * | |
454 | * Queue a new data flush if there isn't one already in progress and | |
455 | * wait for completion of the flush. This means that we only ever have one | |
456 | * inode flush in progress no matter how many ENOSPC events are occurring and | |
457 | * so will prevent the system from bogging down due to every concurrent | |
458 | * ENOSPC event scanning all the active inodes in the system for writeback. | |
459 | */ | |
460 | void | |
461 | xfs_flush_inodes( | |
462 | struct xfs_inode *ip) | |
463 | { | |
464 | struct xfs_mount *mp = ip->i_mount; | |
465 | ||
466 | queue_work(xfs_syncd_wq, &mp->m_flush_work); | |
467 | flush_work_sync(&mp->m_flush_work); | |
468 | } | |
469 | ||
470 | STATIC void | |
471 | xfs_flush_worker( | |
472 | struct work_struct *work) | |
473 | { | |
474 | struct xfs_mount *mp = container_of(work, | |
475 | struct xfs_mount, m_flush_work); | |
476 | ||
477 | xfs_sync_data(mp, SYNC_TRYLOCK); | |
478 | xfs_sync_data(mp, SYNC_TRYLOCK | SYNC_WAIT); | |
a167b17e DC |
479 | } |
480 | ||
481 | int | |
482 | xfs_syncd_init( | |
483 | struct xfs_mount *mp) | |
484 | { | |
89e4cb55 | 485 | INIT_WORK(&mp->m_flush_work, xfs_flush_worker); |
c6d09b66 | 486 | INIT_DELAYED_WORK(&mp->m_sync_work, xfs_sync_worker); |
a7b339f1 DC |
487 | INIT_DELAYED_WORK(&mp->m_reclaim_work, xfs_reclaim_worker); |
488 | ||
c6d09b66 DC |
489 | xfs_syncd_queue_sync(mp); |
490 | ||
a167b17e DC |
491 | return 0; |
492 | } | |
493 | ||
494 | void | |
495 | xfs_syncd_stop( | |
496 | struct xfs_mount *mp) | |
497 | { | |
c6d09b66 | 498 | cancel_delayed_work_sync(&mp->m_sync_work); |
a7b339f1 | 499 | cancel_delayed_work_sync(&mp->m_reclaim_work); |
89e4cb55 | 500 | cancel_work_sync(&mp->m_flush_work); |
a167b17e DC |
501 | } |
502 | ||
bc990f5c CH |
503 | void |
504 | __xfs_inode_set_reclaim_tag( | |
505 | struct xfs_perag *pag, | |
506 | struct xfs_inode *ip) | |
507 | { | |
508 | radix_tree_tag_set(&pag->pag_ici_root, | |
509 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
510 | XFS_ICI_RECLAIM_TAG); | |
16fd5367 DC |
511 | |
512 | if (!pag->pag_ici_reclaimable) { | |
513 | /* propagate the reclaim tag up into the perag radix tree */ | |
514 | spin_lock(&ip->i_mount->m_perag_lock); | |
515 | radix_tree_tag_set(&ip->i_mount->m_perag_tree, | |
516 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
517 | XFS_ICI_RECLAIM_TAG); | |
518 | spin_unlock(&ip->i_mount->m_perag_lock); | |
a7b339f1 DC |
519 | |
520 | /* schedule periodic background inode reclaim */ | |
521 | xfs_syncd_queue_reclaim(ip->i_mount); | |
522 | ||
16fd5367 DC |
523 | trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno, |
524 | -1, _RET_IP_); | |
525 | } | |
9bf729c0 | 526 | pag->pag_ici_reclaimable++; |
bc990f5c CH |
527 | } |
528 | ||
11654513 DC |
529 | /* |
530 | * We set the inode flag atomically with the radix tree tag. | |
531 | * Once we get tag lookups on the radix tree, this inode flag | |
532 | * can go away. | |
533 | */ | |
396beb85 DC |
534 | void |
535 | xfs_inode_set_reclaim_tag( | |
536 | xfs_inode_t *ip) | |
537 | { | |
5017e97d DC |
538 | struct xfs_mount *mp = ip->i_mount; |
539 | struct xfs_perag *pag; | |
396beb85 | 540 | |
5017e97d | 541 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); |
1a427ab0 | 542 | spin_lock(&pag->pag_ici_lock); |
396beb85 | 543 | spin_lock(&ip->i_flags_lock); |
bc990f5c | 544 | __xfs_inode_set_reclaim_tag(pag, ip); |
11654513 | 545 | __xfs_iflags_set(ip, XFS_IRECLAIMABLE); |
396beb85 | 546 | spin_unlock(&ip->i_flags_lock); |
1a427ab0 | 547 | spin_unlock(&pag->pag_ici_lock); |
5017e97d | 548 | xfs_perag_put(pag); |
396beb85 DC |
549 | } |
550 | ||
081003ff JW |
551 | STATIC void |
552 | __xfs_inode_clear_reclaim( | |
396beb85 DC |
553 | xfs_perag_t *pag, |
554 | xfs_inode_t *ip) | |
555 | { | |
9bf729c0 | 556 | pag->pag_ici_reclaimable--; |
16fd5367 DC |
557 | if (!pag->pag_ici_reclaimable) { |
558 | /* clear the reclaim tag from the perag radix tree */ | |
559 | spin_lock(&ip->i_mount->m_perag_lock); | |
560 | radix_tree_tag_clear(&ip->i_mount->m_perag_tree, | |
561 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
562 | XFS_ICI_RECLAIM_TAG); | |
563 | spin_unlock(&ip->i_mount->m_perag_lock); | |
564 | trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno, | |
565 | -1, _RET_IP_); | |
566 | } | |
396beb85 DC |
567 | } |
568 | ||
081003ff JW |
569 | void |
570 | __xfs_inode_clear_reclaim_tag( | |
571 | xfs_mount_t *mp, | |
572 | xfs_perag_t *pag, | |
573 | xfs_inode_t *ip) | |
574 | { | |
575 | radix_tree_tag_clear(&pag->pag_ici_root, | |
576 | XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG); | |
577 | __xfs_inode_clear_reclaim(pag, ip); | |
578 | } | |
579 | ||
e3a20c0b DC |
580 | /* |
581 | * Grab the inode for reclaim exclusively. | |
582 | * Return 0 if we grabbed it, non-zero otherwise. | |
583 | */ | |
584 | STATIC int | |
585 | xfs_reclaim_inode_grab( | |
586 | struct xfs_inode *ip, | |
587 | int flags) | |
588 | { | |
1a3e8f3d DC |
589 | ASSERT(rcu_read_lock_held()); |
590 | ||
591 | /* quick check for stale RCU freed inode */ | |
592 | if (!ip->i_ino) | |
593 | return 1; | |
e3a20c0b DC |
594 | |
595 | /* | |
474fce06 CH |
596 | * If we are asked for non-blocking operation, do unlocked checks to |
597 | * see if the inode already is being flushed or in reclaim to avoid | |
598 | * lock traffic. | |
e3a20c0b DC |
599 | */ |
600 | if ((flags & SYNC_TRYLOCK) && | |
474fce06 | 601 | __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM)) |
e3a20c0b | 602 | return 1; |
e3a20c0b DC |
603 | |
604 | /* | |
605 | * The radix tree lock here protects a thread in xfs_iget from racing | |
606 | * with us starting reclaim on the inode. Once we have the | |
607 | * XFS_IRECLAIM flag set it will not touch us. | |
1a3e8f3d DC |
608 | * |
609 | * Due to RCU lookup, we may find inodes that have been freed and only | |
610 | * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that | |
611 | * aren't candidates for reclaim at all, so we must check the | |
612 | * XFS_IRECLAIMABLE is set first before proceeding to reclaim. | |
e3a20c0b DC |
613 | */ |
614 | spin_lock(&ip->i_flags_lock); | |
1a3e8f3d DC |
615 | if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) || |
616 | __xfs_iflags_test(ip, XFS_IRECLAIM)) { | |
617 | /* not a reclaim candidate. */ | |
e3a20c0b DC |
618 | spin_unlock(&ip->i_flags_lock); |
619 | return 1; | |
620 | } | |
621 | __xfs_iflags_set(ip, XFS_IRECLAIM); | |
622 | spin_unlock(&ip->i_flags_lock); | |
623 | return 0; | |
624 | } | |
625 | ||
777df5af | 626 | /* |
8a48088f CH |
627 | * Inodes in different states need to be treated differently. The following |
628 | * table lists the inode states and the reclaim actions necessary: | |
777df5af DC |
629 | * |
630 | * inode state iflush ret required action | |
631 | * --------------- ---------- --------------- | |
632 | * bad - reclaim | |
633 | * shutdown EIO unpin and reclaim | |
634 | * clean, unpinned 0 reclaim | |
635 | * stale, unpinned 0 reclaim | |
c854363e DC |
636 | * clean, pinned(*) 0 requeue |
637 | * stale, pinned EAGAIN requeue | |
8a48088f CH |
638 | * dirty, async - requeue |
639 | * dirty, sync 0 reclaim | |
777df5af DC |
640 | * |
641 | * (*) dgc: I don't think the clean, pinned state is possible but it gets | |
642 | * handled anyway given the order of checks implemented. | |
643 | * | |
c854363e DC |
644 | * Also, because we get the flush lock first, we know that any inode that has |
645 | * been flushed delwri has had the flush completed by the time we check that | |
8a48088f | 646 | * the inode is clean. |
c854363e | 647 | * |
8a48088f CH |
648 | * Note that because the inode is flushed delayed write by AIL pushing, the |
649 | * flush lock may already be held here and waiting on it can result in very | |
650 | * long latencies. Hence for sync reclaims, where we wait on the flush lock, | |
651 | * the caller should push the AIL first before trying to reclaim inodes to | |
652 | * minimise the amount of time spent waiting. For background relaim, we only | |
653 | * bother to reclaim clean inodes anyway. | |
c854363e | 654 | * |
777df5af DC |
655 | * Hence the order of actions after gaining the locks should be: |
656 | * bad => reclaim | |
657 | * shutdown => unpin and reclaim | |
8a48088f | 658 | * pinned, async => requeue |
c854363e | 659 | * pinned, sync => unpin |
777df5af DC |
660 | * stale => reclaim |
661 | * clean => reclaim | |
8a48088f | 662 | * dirty, async => requeue |
c854363e | 663 | * dirty, sync => flush, wait and reclaim |
777df5af | 664 | */ |
75f3cb13 | 665 | STATIC int |
c8e20be0 | 666 | xfs_reclaim_inode( |
75f3cb13 DC |
667 | struct xfs_inode *ip, |
668 | struct xfs_perag *pag, | |
c8e20be0 | 669 | int sync_mode) |
fce08f2f | 670 | { |
4c46819a CH |
671 | struct xfs_buf *bp = NULL; |
672 | int error; | |
777df5af | 673 | |
1bfd8d04 DC |
674 | restart: |
675 | error = 0; | |
c8e20be0 | 676 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
c854363e DC |
677 | if (!xfs_iflock_nowait(ip)) { |
678 | if (!(sync_mode & SYNC_WAIT)) | |
679 | goto out; | |
680 | xfs_iflock(ip); | |
681 | } | |
7a3be02b | 682 | |
777df5af DC |
683 | if (is_bad_inode(VFS_I(ip))) |
684 | goto reclaim; | |
685 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) { | |
686 | xfs_iunpin_wait(ip); | |
32ce90a4 | 687 | xfs_iflush_abort(ip); |
777df5af DC |
688 | goto reclaim; |
689 | } | |
c854363e | 690 | if (xfs_ipincount(ip)) { |
8a48088f CH |
691 | if (!(sync_mode & SYNC_WAIT)) |
692 | goto out_ifunlock; | |
777df5af | 693 | xfs_iunpin_wait(ip); |
c854363e | 694 | } |
777df5af DC |
695 | if (xfs_iflags_test(ip, XFS_ISTALE)) |
696 | goto reclaim; | |
697 | if (xfs_inode_clean(ip)) | |
698 | goto reclaim; | |
699 | ||
8a48088f CH |
700 | /* |
701 | * Never flush out dirty data during non-blocking reclaim, as it would | |
702 | * just contend with AIL pushing trying to do the same job. | |
703 | */ | |
704 | if (!(sync_mode & SYNC_WAIT)) | |
705 | goto out_ifunlock; | |
706 | ||
1bfd8d04 DC |
707 | /* |
708 | * Now we have an inode that needs flushing. | |
709 | * | |
4c46819a | 710 | * Note that xfs_iflush will never block on the inode buffer lock, as |
1bfd8d04 | 711 | * xfs_ifree_cluster() can lock the inode buffer before it locks the |
4c46819a CH |
712 | * ip->i_lock, and we are doing the exact opposite here. As a result, |
713 | * doing a blocking xfs_itobp() to get the cluster buffer would result | |
1bfd8d04 DC |
714 | * in an ABBA deadlock with xfs_ifree_cluster(). |
715 | * | |
716 | * As xfs_ifree_cluser() must gather all inodes that are active in the | |
717 | * cache to mark them stale, if we hit this case we don't actually want | |
718 | * to do IO here - we want the inode marked stale so we can simply | |
4c46819a CH |
719 | * reclaim it. Hence if we get an EAGAIN error here, just unlock the |
720 | * inode, back off and try again. Hopefully the next pass through will | |
721 | * see the stale flag set on the inode. | |
1bfd8d04 | 722 | */ |
4c46819a | 723 | error = xfs_iflush(ip, &bp); |
8a48088f CH |
724 | if (error == EAGAIN) { |
725 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
726 | /* backoff longer than in xfs_ifree_cluster */ | |
727 | delay(2); | |
728 | goto restart; | |
c854363e | 729 | } |
c854363e | 730 | |
4c46819a CH |
731 | if (!error) { |
732 | error = xfs_bwrite(bp); | |
733 | xfs_buf_relse(bp); | |
734 | } | |
735 | ||
736 | xfs_iflock(ip); | |
777df5af DC |
737 | reclaim: |
738 | xfs_ifunlock(ip); | |
c8e20be0 | 739 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
740 | |
741 | XFS_STATS_INC(xs_ig_reclaims); | |
742 | /* | |
743 | * Remove the inode from the per-AG radix tree. | |
744 | * | |
745 | * Because radix_tree_delete won't complain even if the item was never | |
746 | * added to the tree assert that it's been there before to catch | |
747 | * problems with the inode life time early on. | |
748 | */ | |
1a427ab0 | 749 | spin_lock(&pag->pag_ici_lock); |
2f11feab DC |
750 | if (!radix_tree_delete(&pag->pag_ici_root, |
751 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino))) | |
752 | ASSERT(0); | |
081003ff | 753 | __xfs_inode_clear_reclaim(pag, ip); |
1a427ab0 | 754 | spin_unlock(&pag->pag_ici_lock); |
2f11feab DC |
755 | |
756 | /* | |
757 | * Here we do an (almost) spurious inode lock in order to coordinate | |
758 | * with inode cache radix tree lookups. This is because the lookup | |
759 | * can reference the inodes in the cache without taking references. | |
760 | * | |
761 | * We make that OK here by ensuring that we wait until the inode is | |
ad637a10 | 762 | * unlocked after the lookup before we go ahead and free it. |
2f11feab | 763 | */ |
ad637a10 | 764 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
2f11feab | 765 | xfs_qm_dqdetach(ip); |
ad637a10 | 766 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
767 | |
768 | xfs_inode_free(ip); | |
ad637a10 | 769 | return error; |
8a48088f CH |
770 | |
771 | out_ifunlock: | |
772 | xfs_ifunlock(ip); | |
773 | out: | |
774 | xfs_iflags_clear(ip, XFS_IRECLAIM); | |
775 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
776 | /* | |
777 | * We could return EAGAIN here to make reclaim rescan the inode tree in | |
778 | * a short while. However, this just burns CPU time scanning the tree | |
779 | * waiting for IO to complete and xfssyncd never goes back to the idle | |
780 | * state. Instead, return 0 to let the next scheduled background reclaim | |
781 | * attempt to reclaim the inode again. | |
782 | */ | |
783 | return 0; | |
7a3be02b DC |
784 | } |
785 | ||
65d0f205 DC |
786 | /* |
787 | * Walk the AGs and reclaim the inodes in them. Even if the filesystem is | |
788 | * corrupted, we still want to try to reclaim all the inodes. If we don't, | |
789 | * then a shut down during filesystem unmount reclaim walk leak all the | |
790 | * unreclaimed inodes. | |
791 | */ | |
792 | int | |
793 | xfs_reclaim_inodes_ag( | |
794 | struct xfs_mount *mp, | |
795 | int flags, | |
796 | int *nr_to_scan) | |
797 | { | |
798 | struct xfs_perag *pag; | |
799 | int error = 0; | |
800 | int last_error = 0; | |
801 | xfs_agnumber_t ag; | |
69b491c2 DC |
802 | int trylock = flags & SYNC_TRYLOCK; |
803 | int skipped; | |
65d0f205 | 804 | |
69b491c2 | 805 | restart: |
65d0f205 | 806 | ag = 0; |
69b491c2 | 807 | skipped = 0; |
65d0f205 DC |
808 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
809 | unsigned long first_index = 0; | |
810 | int done = 0; | |
e3a20c0b | 811 | int nr_found = 0; |
65d0f205 DC |
812 | |
813 | ag = pag->pag_agno + 1; | |
814 | ||
69b491c2 DC |
815 | if (trylock) { |
816 | if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) { | |
817 | skipped++; | |
f83282a8 | 818 | xfs_perag_put(pag); |
69b491c2 DC |
819 | continue; |
820 | } | |
821 | first_index = pag->pag_ici_reclaim_cursor; | |
822 | } else | |
823 | mutex_lock(&pag->pag_ici_reclaim_lock); | |
824 | ||
65d0f205 | 825 | do { |
e3a20c0b DC |
826 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
827 | int i; | |
65d0f205 | 828 | |
1a3e8f3d | 829 | rcu_read_lock(); |
e3a20c0b DC |
830 | nr_found = radix_tree_gang_lookup_tag( |
831 | &pag->pag_ici_root, | |
832 | (void **)batch, first_index, | |
833 | XFS_LOOKUP_BATCH, | |
65d0f205 DC |
834 | XFS_ICI_RECLAIM_TAG); |
835 | if (!nr_found) { | |
b2232219 | 836 | done = 1; |
1a3e8f3d | 837 | rcu_read_unlock(); |
65d0f205 DC |
838 | break; |
839 | } | |
840 | ||
841 | /* | |
e3a20c0b DC |
842 | * Grab the inodes before we drop the lock. if we found |
843 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 844 | */ |
e3a20c0b DC |
845 | for (i = 0; i < nr_found; i++) { |
846 | struct xfs_inode *ip = batch[i]; | |
847 | ||
848 | if (done || xfs_reclaim_inode_grab(ip, flags)) | |
849 | batch[i] = NULL; | |
850 | ||
851 | /* | |
852 | * Update the index for the next lookup. Catch | |
853 | * overflows into the next AG range which can | |
854 | * occur if we have inodes in the last block of | |
855 | * the AG and we are currently pointing to the | |
856 | * last inode. | |
1a3e8f3d DC |
857 | * |
858 | * Because we may see inodes that are from the | |
859 | * wrong AG due to RCU freeing and | |
860 | * reallocation, only update the index if it | |
861 | * lies in this AG. It was a race that lead us | |
862 | * to see this inode, so another lookup from | |
863 | * the same index will not find it again. | |
e3a20c0b | 864 | */ |
1a3e8f3d DC |
865 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != |
866 | pag->pag_agno) | |
867 | continue; | |
e3a20c0b DC |
868 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
869 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
870 | done = 1; | |
871 | } | |
65d0f205 | 872 | |
e3a20c0b | 873 | /* unlock now we've grabbed the inodes. */ |
1a3e8f3d | 874 | rcu_read_unlock(); |
e3a20c0b DC |
875 | |
876 | for (i = 0; i < nr_found; i++) { | |
877 | if (!batch[i]) | |
878 | continue; | |
879 | error = xfs_reclaim_inode(batch[i], pag, flags); | |
880 | if (error && last_error != EFSCORRUPTED) | |
881 | last_error = error; | |
882 | } | |
883 | ||
884 | *nr_to_scan -= XFS_LOOKUP_BATCH; | |
65d0f205 | 885 | |
8daaa831 DC |
886 | cond_resched(); |
887 | ||
e3a20c0b | 888 | } while (nr_found && !done && *nr_to_scan > 0); |
65d0f205 | 889 | |
69b491c2 DC |
890 | if (trylock && !done) |
891 | pag->pag_ici_reclaim_cursor = first_index; | |
892 | else | |
893 | pag->pag_ici_reclaim_cursor = 0; | |
894 | mutex_unlock(&pag->pag_ici_reclaim_lock); | |
65d0f205 DC |
895 | xfs_perag_put(pag); |
896 | } | |
69b491c2 DC |
897 | |
898 | /* | |
899 | * if we skipped any AG, and we still have scan count remaining, do | |
900 | * another pass this time using blocking reclaim semantics (i.e | |
901 | * waiting on the reclaim locks and ignoring the reclaim cursors). This | |
902 | * ensure that when we get more reclaimers than AGs we block rather | |
903 | * than spin trying to execute reclaim. | |
904 | */ | |
8daaa831 | 905 | if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) { |
69b491c2 DC |
906 | trylock = 0; |
907 | goto restart; | |
908 | } | |
65d0f205 DC |
909 | return XFS_ERROR(last_error); |
910 | } | |
911 | ||
7a3be02b DC |
912 | int |
913 | xfs_reclaim_inodes( | |
914 | xfs_mount_t *mp, | |
7a3be02b DC |
915 | int mode) |
916 | { | |
65d0f205 DC |
917 | int nr_to_scan = INT_MAX; |
918 | ||
919 | return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan); | |
9bf729c0 DC |
920 | } |
921 | ||
922 | /* | |
8daaa831 | 923 | * Scan a certain number of inodes for reclaim. |
a7b339f1 DC |
924 | * |
925 | * When called we make sure that there is a background (fast) inode reclaim in | |
8daaa831 | 926 | * progress, while we will throttle the speed of reclaim via doing synchronous |
a7b339f1 DC |
927 | * reclaim of inodes. That means if we come across dirty inodes, we wait for |
928 | * them to be cleaned, which we hope will not be very long due to the | |
929 | * background walker having already kicked the IO off on those dirty inodes. | |
9bf729c0 | 930 | */ |
8daaa831 DC |
931 | void |
932 | xfs_reclaim_inodes_nr( | |
933 | struct xfs_mount *mp, | |
934 | int nr_to_scan) | |
9bf729c0 | 935 | { |
8daaa831 DC |
936 | /* kick background reclaimer and push the AIL */ |
937 | xfs_syncd_queue_reclaim(mp); | |
938 | xfs_ail_push_all(mp->m_ail); | |
a7b339f1 | 939 | |
8daaa831 DC |
940 | xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan); |
941 | } | |
9bf729c0 | 942 | |
8daaa831 DC |
943 | /* |
944 | * Return the number of reclaimable inodes in the filesystem for | |
945 | * the shrinker to determine how much to reclaim. | |
946 | */ | |
947 | int | |
948 | xfs_reclaim_inodes_count( | |
949 | struct xfs_mount *mp) | |
950 | { | |
951 | struct xfs_perag *pag; | |
952 | xfs_agnumber_t ag = 0; | |
953 | int reclaimable = 0; | |
9bf729c0 | 954 | |
65d0f205 DC |
955 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
956 | ag = pag->pag_agno + 1; | |
70e60ce7 DC |
957 | reclaimable += pag->pag_ici_reclaimable; |
958 | xfs_perag_put(pag); | |
9bf729c0 | 959 | } |
9bf729c0 DC |
960 | return reclaimable; |
961 | } | |
962 |