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" | |
25 | #include "xfs_sb.h" | |
26 | #include "xfs_ag.h" | |
fe4fa4b8 DC |
27 | #include "xfs_mount.h" |
28 | #include "xfs_bmap_btree.h" | |
fe4fa4b8 DC |
29 | #include "xfs_inode.h" |
30 | #include "xfs_dinode.h" | |
31 | #include "xfs_error.h" | |
fe4fa4b8 DC |
32 | #include "xfs_filestream.h" |
33 | #include "xfs_vnodeops.h" | |
fe4fa4b8 | 34 | #include "xfs_inode_item.h" |
7d095257 | 35 | #include "xfs_quota.h" |
0b1b213f | 36 | #include "xfs_trace.h" |
1a387d3b | 37 | #include "xfs_fsops.h" |
fe4fa4b8 | 38 | |
a167b17e DC |
39 | #include <linux/kthread.h> |
40 | #include <linux/freezer.h> | |
41 | ||
c6d09b66 DC |
42 | struct workqueue_struct *xfs_syncd_wq; /* sync workqueue */ |
43 | ||
78ae5256 DC |
44 | /* |
45 | * The inode lookup is done in batches to keep the amount of lock traffic and | |
46 | * radix tree lookups to a minimum. The batch size is a trade off between | |
47 | * lookup reduction and stack usage. This is in the reclaim path, so we can't | |
48 | * be too greedy. | |
49 | */ | |
50 | #define XFS_LOOKUP_BATCH 32 | |
51 | ||
e13de955 DC |
52 | STATIC int |
53 | xfs_inode_ag_walk_grab( | |
54 | struct xfs_inode *ip) | |
55 | { | |
56 | struct inode *inode = VFS_I(ip); | |
57 | ||
1a3e8f3d DC |
58 | ASSERT(rcu_read_lock_held()); |
59 | ||
60 | /* | |
61 | * check for stale RCU freed inode | |
62 | * | |
63 | * If the inode has been reallocated, it doesn't matter if it's not in | |
64 | * the AG we are walking - we are walking for writeback, so if it | |
65 | * passes all the "valid inode" checks and is dirty, then we'll write | |
66 | * it back anyway. If it has been reallocated and still being | |
67 | * initialised, the XFS_INEW check below will catch it. | |
68 | */ | |
69 | spin_lock(&ip->i_flags_lock); | |
70 | if (!ip->i_ino) | |
71 | goto out_unlock_noent; | |
72 | ||
73 | /* avoid new or reclaimable inodes. Leave for reclaim code to flush */ | |
74 | if (__xfs_iflags_test(ip, XFS_INEW | XFS_IRECLAIMABLE | XFS_IRECLAIM)) | |
75 | goto out_unlock_noent; | |
76 | spin_unlock(&ip->i_flags_lock); | |
77 | ||
e13de955 DC |
78 | /* nothing to sync during shutdown */ |
79 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) | |
80 | return EFSCORRUPTED; | |
81 | ||
e13de955 DC |
82 | /* If we can't grab the inode, it must on it's way to reclaim. */ |
83 | if (!igrab(inode)) | |
84 | return ENOENT; | |
85 | ||
86 | if (is_bad_inode(inode)) { | |
87 | IRELE(ip); | |
88 | return ENOENT; | |
89 | } | |
90 | ||
91 | /* inode is valid */ | |
92 | return 0; | |
1a3e8f3d DC |
93 | |
94 | out_unlock_noent: | |
95 | spin_unlock(&ip->i_flags_lock); | |
96 | return ENOENT; | |
e13de955 DC |
97 | } |
98 | ||
75f3cb13 DC |
99 | STATIC int |
100 | xfs_inode_ag_walk( | |
101 | struct xfs_mount *mp, | |
5017e97d | 102 | struct xfs_perag *pag, |
75f3cb13 DC |
103 | int (*execute)(struct xfs_inode *ip, |
104 | struct xfs_perag *pag, int flags), | |
65d0f205 | 105 | int flags) |
75f3cb13 | 106 | { |
75f3cb13 DC |
107 | uint32_t first_index; |
108 | int last_error = 0; | |
109 | int skipped; | |
65d0f205 | 110 | int done; |
78ae5256 | 111 | int nr_found; |
75f3cb13 DC |
112 | |
113 | restart: | |
65d0f205 | 114 | done = 0; |
75f3cb13 DC |
115 | skipped = 0; |
116 | first_index = 0; | |
78ae5256 | 117 | nr_found = 0; |
75f3cb13 | 118 | do { |
78ae5256 | 119 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
75f3cb13 | 120 | int error = 0; |
78ae5256 | 121 | int i; |
75f3cb13 | 122 | |
1a3e8f3d | 123 | rcu_read_lock(); |
65d0f205 | 124 | nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, |
78ae5256 DC |
125 | (void **)batch, first_index, |
126 | XFS_LOOKUP_BATCH); | |
65d0f205 | 127 | if (!nr_found) { |
1a3e8f3d | 128 | rcu_read_unlock(); |
75f3cb13 | 129 | break; |
c8e20be0 | 130 | } |
75f3cb13 | 131 | |
65d0f205 | 132 | /* |
78ae5256 DC |
133 | * Grab the inodes before we drop the lock. if we found |
134 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 135 | */ |
78ae5256 DC |
136 | for (i = 0; i < nr_found; i++) { |
137 | struct xfs_inode *ip = batch[i]; | |
138 | ||
139 | if (done || xfs_inode_ag_walk_grab(ip)) | |
140 | batch[i] = NULL; | |
141 | ||
142 | /* | |
1a3e8f3d DC |
143 | * Update the index for the next lookup. Catch |
144 | * overflows into the next AG range which can occur if | |
145 | * we have inodes in the last block of the AG and we | |
146 | * are currently pointing to the last inode. | |
147 | * | |
148 | * Because we may see inodes that are from the wrong AG | |
149 | * due to RCU freeing and reallocation, only update the | |
150 | * index if it lies in this AG. It was a race that lead | |
151 | * us to see this inode, so another lookup from the | |
152 | * same index will not find it again. | |
78ae5256 | 153 | */ |
1a3e8f3d DC |
154 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno) |
155 | continue; | |
78ae5256 DC |
156 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
157 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
158 | done = 1; | |
e13de955 | 159 | } |
78ae5256 DC |
160 | |
161 | /* unlock now we've grabbed the inodes. */ | |
1a3e8f3d | 162 | rcu_read_unlock(); |
e13de955 | 163 | |
78ae5256 DC |
164 | for (i = 0; i < nr_found; i++) { |
165 | if (!batch[i]) | |
166 | continue; | |
167 | error = execute(batch[i], pag, flags); | |
168 | IRELE(batch[i]); | |
169 | if (error == EAGAIN) { | |
170 | skipped++; | |
171 | continue; | |
172 | } | |
173 | if (error && last_error != EFSCORRUPTED) | |
174 | last_error = error; | |
75f3cb13 | 175 | } |
c8e20be0 DC |
176 | |
177 | /* bail out if the filesystem is corrupted. */ | |
75f3cb13 DC |
178 | if (error == EFSCORRUPTED) |
179 | break; | |
180 | ||
78ae5256 | 181 | } while (nr_found && !done); |
75f3cb13 DC |
182 | |
183 | if (skipped) { | |
184 | delay(1); | |
185 | goto restart; | |
186 | } | |
75f3cb13 DC |
187 | return last_error; |
188 | } | |
189 | ||
fe588ed3 | 190 | int |
75f3cb13 DC |
191 | xfs_inode_ag_iterator( |
192 | struct xfs_mount *mp, | |
193 | int (*execute)(struct xfs_inode *ip, | |
194 | struct xfs_perag *pag, int flags), | |
65d0f205 | 195 | int flags) |
75f3cb13 | 196 | { |
16fd5367 | 197 | struct xfs_perag *pag; |
75f3cb13 DC |
198 | int error = 0; |
199 | int last_error = 0; | |
200 | xfs_agnumber_t ag; | |
201 | ||
16fd5367 | 202 | ag = 0; |
65d0f205 DC |
203 | while ((pag = xfs_perag_get(mp, ag))) { |
204 | ag = pag->pag_agno + 1; | |
205 | error = xfs_inode_ag_walk(mp, pag, execute, flags); | |
5017e97d | 206 | xfs_perag_put(pag); |
75f3cb13 DC |
207 | if (error) { |
208 | last_error = error; | |
209 | if (error == EFSCORRUPTED) | |
210 | break; | |
211 | } | |
212 | } | |
213 | return XFS_ERROR(last_error); | |
214 | } | |
215 | ||
5a34d5cd DC |
216 | STATIC int |
217 | xfs_sync_inode_data( | |
218 | struct xfs_inode *ip, | |
75f3cb13 | 219 | struct xfs_perag *pag, |
5a34d5cd DC |
220 | int flags) |
221 | { | |
222 | struct inode *inode = VFS_I(ip); | |
223 | struct address_space *mapping = inode->i_mapping; | |
224 | int error = 0; | |
225 | ||
226 | if (!mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) | |
227 | goto out_wait; | |
228 | ||
229 | if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED)) { | |
230 | if (flags & SYNC_TRYLOCK) | |
231 | goto out_wait; | |
232 | xfs_ilock(ip, XFS_IOLOCK_SHARED); | |
233 | } | |
234 | ||
235 | error = xfs_flush_pages(ip, 0, -1, (flags & SYNC_WAIT) ? | |
0cadda1c | 236 | 0 : XBF_ASYNC, FI_NONE); |
5a34d5cd DC |
237 | xfs_iunlock(ip, XFS_IOLOCK_SHARED); |
238 | ||
239 | out_wait: | |
b0710ccc | 240 | if (flags & SYNC_WAIT) |
5a34d5cd DC |
241 | xfs_ioend_wait(ip); |
242 | return error; | |
243 | } | |
244 | ||
845b6d0c CH |
245 | STATIC int |
246 | xfs_sync_inode_attr( | |
247 | struct xfs_inode *ip, | |
75f3cb13 | 248 | struct xfs_perag *pag, |
845b6d0c CH |
249 | int flags) |
250 | { | |
251 | int error = 0; | |
252 | ||
253 | xfs_ilock(ip, XFS_ILOCK_SHARED); | |
254 | if (xfs_inode_clean(ip)) | |
255 | goto out_unlock; | |
256 | if (!xfs_iflock_nowait(ip)) { | |
257 | if (!(flags & SYNC_WAIT)) | |
258 | goto out_unlock; | |
259 | xfs_iflock(ip); | |
260 | } | |
261 | ||
262 | if (xfs_inode_clean(ip)) { | |
263 | xfs_ifunlock(ip); | |
264 | goto out_unlock; | |
265 | } | |
266 | ||
c854363e | 267 | error = xfs_iflush(ip, flags); |
845b6d0c CH |
268 | |
269 | out_unlock: | |
270 | xfs_iunlock(ip, XFS_ILOCK_SHARED); | |
271 | return error; | |
272 | } | |
273 | ||
075fe102 CH |
274 | /* |
275 | * Write out pagecache data for the whole filesystem. | |
276 | */ | |
64c86149 | 277 | STATIC int |
075fe102 CH |
278 | xfs_sync_data( |
279 | struct xfs_mount *mp, | |
280 | int flags) | |
683a8970 | 281 | { |
075fe102 | 282 | int error; |
fe4fa4b8 | 283 | |
b0710ccc | 284 | ASSERT((flags & ~(SYNC_TRYLOCK|SYNC_WAIT)) == 0); |
fe4fa4b8 | 285 | |
65d0f205 | 286 | error = xfs_inode_ag_iterator(mp, xfs_sync_inode_data, flags); |
075fe102 CH |
287 | if (error) |
288 | return XFS_ERROR(error); | |
e9f1c6ee | 289 | |
a14a348b | 290 | xfs_log_force(mp, (flags & SYNC_WAIT) ? XFS_LOG_SYNC : 0); |
075fe102 CH |
291 | return 0; |
292 | } | |
e9f1c6ee | 293 | |
075fe102 CH |
294 | /* |
295 | * Write out inode metadata (attributes) for the whole filesystem. | |
296 | */ | |
64c86149 | 297 | STATIC int |
075fe102 CH |
298 | xfs_sync_attr( |
299 | struct xfs_mount *mp, | |
300 | int flags) | |
301 | { | |
302 | ASSERT((flags & ~SYNC_WAIT) == 0); | |
75f3cb13 | 303 | |
65d0f205 | 304 | return xfs_inode_ag_iterator(mp, xfs_sync_inode_attr, flags); |
fe4fa4b8 DC |
305 | } |
306 | ||
5d77c0dc | 307 | STATIC int |
2af75df7 | 308 | xfs_sync_fsdata( |
df308bcf | 309 | struct xfs_mount *mp) |
2af75df7 CH |
310 | { |
311 | struct xfs_buf *bp; | |
2af75df7 CH |
312 | |
313 | /* | |
df308bcf CH |
314 | * If the buffer is pinned then push on the log so we won't get stuck |
315 | * waiting in the write for someone, maybe ourselves, to flush the log. | |
316 | * | |
317 | * Even though we just pushed the log above, we did not have the | |
318 | * superblock buffer locked at that point so it can become pinned in | |
319 | * between there and here. | |
2af75df7 | 320 | */ |
df308bcf CH |
321 | bp = xfs_getsb(mp, 0); |
322 | if (XFS_BUF_ISPINNED(bp)) | |
323 | xfs_log_force(mp, 0); | |
2af75df7 | 324 | |
df308bcf | 325 | return xfs_bwrite(mp, bp); |
e9f1c6ee DC |
326 | } |
327 | ||
328 | /* | |
a4e4c4f4 DC |
329 | * When remounting a filesystem read-only or freezing the filesystem, we have |
330 | * two phases to execute. This first phase is syncing the data before we | |
331 | * quiesce the filesystem, and the second is flushing all the inodes out after | |
332 | * we've waited for all the transactions created by the first phase to | |
333 | * complete. The second phase ensures that the inodes are written to their | |
334 | * location on disk rather than just existing in transactions in the log. This | |
335 | * means after a quiesce there is no log replay required to write the inodes to | |
336 | * disk (this is the main difference between a sync and a quiesce). | |
337 | */ | |
338 | /* | |
339 | * First stage of freeze - no writers will make progress now we are here, | |
e9f1c6ee DC |
340 | * so we flush delwri and delalloc buffers here, then wait for all I/O to |
341 | * complete. Data is frozen at that point. Metadata is not frozen, | |
a4e4c4f4 DC |
342 | * transactions can still occur here so don't bother flushing the buftarg |
343 | * because it'll just get dirty again. | |
e9f1c6ee DC |
344 | */ |
345 | int | |
346 | xfs_quiesce_data( | |
347 | struct xfs_mount *mp) | |
348 | { | |
df308bcf | 349 | int error, error2 = 0; |
e9f1c6ee DC |
350 | |
351 | /* push non-blocking */ | |
075fe102 | 352 | xfs_sync_data(mp, 0); |
8b5403a6 | 353 | xfs_qm_sync(mp, SYNC_TRYLOCK); |
e9f1c6ee | 354 | |
c90b07e8 | 355 | /* push and block till complete */ |
b0710ccc | 356 | xfs_sync_data(mp, SYNC_WAIT); |
7d095257 | 357 | xfs_qm_sync(mp, SYNC_WAIT); |
e9f1c6ee | 358 | |
a4e4c4f4 | 359 | /* write superblock and hoover up shutdown errors */ |
df308bcf CH |
360 | error = xfs_sync_fsdata(mp); |
361 | ||
362 | /* make sure all delwri buffers are written out */ | |
363 | xfs_flush_buftarg(mp->m_ddev_targp, 1); | |
364 | ||
365 | /* mark the log as covered if needed */ | |
366 | if (xfs_log_need_covered(mp)) | |
c58efdb4 | 367 | error2 = xfs_fs_log_dummy(mp); |
e9f1c6ee | 368 | |
a4e4c4f4 | 369 | /* flush data-only devices */ |
e9f1c6ee DC |
370 | if (mp->m_rtdev_targp) |
371 | XFS_bflush(mp->m_rtdev_targp); | |
372 | ||
df308bcf | 373 | return error ? error : error2; |
2af75df7 CH |
374 | } |
375 | ||
76bf105c DC |
376 | STATIC void |
377 | xfs_quiesce_fs( | |
378 | struct xfs_mount *mp) | |
379 | { | |
380 | int count = 0, pincount; | |
381 | ||
c854363e | 382 | xfs_reclaim_inodes(mp, 0); |
76bf105c | 383 | xfs_flush_buftarg(mp->m_ddev_targp, 0); |
76bf105c DC |
384 | |
385 | /* | |
386 | * This loop must run at least twice. The first instance of the loop | |
387 | * will flush most meta data but that will generate more meta data | |
388 | * (typically directory updates). Which then must be flushed and | |
c854363e DC |
389 | * logged before we can write the unmount record. We also so sync |
390 | * reclaim of inodes to catch any that the above delwri flush skipped. | |
76bf105c DC |
391 | */ |
392 | do { | |
c854363e | 393 | xfs_reclaim_inodes(mp, SYNC_WAIT); |
075fe102 | 394 | xfs_sync_attr(mp, SYNC_WAIT); |
76bf105c DC |
395 | pincount = xfs_flush_buftarg(mp->m_ddev_targp, 1); |
396 | if (!pincount) { | |
397 | delay(50); | |
398 | count++; | |
399 | } | |
400 | } while (count < 2); | |
401 | } | |
402 | ||
403 | /* | |
404 | * Second stage of a quiesce. The data is already synced, now we have to take | |
405 | * care of the metadata. New transactions are already blocked, so we need to | |
406 | * wait for any remaining transactions to drain out before proceding. | |
407 | */ | |
408 | void | |
409 | xfs_quiesce_attr( | |
410 | struct xfs_mount *mp) | |
411 | { | |
412 | int error = 0; | |
413 | ||
414 | /* wait for all modifications to complete */ | |
415 | while (atomic_read(&mp->m_active_trans) > 0) | |
416 | delay(100); | |
417 | ||
418 | /* flush inodes and push all remaining buffers out to disk */ | |
419 | xfs_quiesce_fs(mp); | |
420 | ||
5e106572 FB |
421 | /* |
422 | * Just warn here till VFS can correctly support | |
423 | * read-only remount without racing. | |
424 | */ | |
425 | WARN_ON(atomic_read(&mp->m_active_trans) != 0); | |
76bf105c DC |
426 | |
427 | /* Push the superblock and write an unmount record */ | |
428 | error = xfs_log_sbcount(mp, 1); | |
429 | if (error) | |
4f10700a | 430 | xfs_warn(mp, "xfs_attr_quiesce: failed to log sb changes. " |
76bf105c DC |
431 | "Frozen image may not be consistent."); |
432 | xfs_log_unmount_write(mp); | |
433 | xfs_unmountfs_writesb(mp); | |
434 | } | |
435 | ||
c6d09b66 DC |
436 | static void |
437 | xfs_syncd_queue_sync( | |
438 | struct xfs_mount *mp) | |
439 | { | |
440 | queue_delayed_work(xfs_syncd_wq, &mp->m_sync_work, | |
441 | msecs_to_jiffies(xfs_syncd_centisecs * 10)); | |
442 | } | |
443 | ||
444 | /* | |
445 | * Every sync period we need to unpin all items, reclaim inodes and sync | |
446 | * disk quotas. We might need to cover the log to indicate that the | |
447 | * filesystem is idle and not frozen. | |
448 | */ | |
449 | STATIC void | |
450 | xfs_sync_worker( | |
451 | struct work_struct *work) | |
452 | { | |
453 | struct xfs_mount *mp = container_of(to_delayed_work(work), | |
454 | struct xfs_mount, m_sync_work); | |
455 | int error; | |
456 | ||
457 | if (!(mp->m_flags & XFS_MOUNT_RDONLY)) { | |
458 | /* dgc: errors ignored here */ | |
459 | if (mp->m_super->s_frozen == SB_UNFROZEN && | |
460 | xfs_log_need_covered(mp)) | |
461 | error = xfs_fs_log_dummy(mp); | |
462 | else | |
463 | xfs_log_force(mp, 0); | |
464 | xfs_reclaim_inodes(mp, 0); | |
465 | error = xfs_qm_sync(mp, SYNC_TRYLOCK); | |
466 | } | |
467 | ||
468 | /* queue us up again */ | |
469 | xfs_syncd_queue_sync(mp); | |
470 | } | |
471 | ||
89e4cb55 DC |
472 | /* |
473 | * Flush delayed allocate data, attempting to free up reserved space | |
474 | * from existing allocations. At this point a new allocation attempt | |
475 | * has failed with ENOSPC and we are in the process of scratching our | |
476 | * heads, looking about for more room. | |
477 | * | |
478 | * Queue a new data flush if there isn't one already in progress and | |
479 | * wait for completion of the flush. This means that we only ever have one | |
480 | * inode flush in progress no matter how many ENOSPC events are occurring and | |
481 | * so will prevent the system from bogging down due to every concurrent | |
482 | * ENOSPC event scanning all the active inodes in the system for writeback. | |
483 | */ | |
484 | void | |
485 | xfs_flush_inodes( | |
486 | struct xfs_inode *ip) | |
487 | { | |
488 | struct xfs_mount *mp = ip->i_mount; | |
489 | ||
490 | queue_work(xfs_syncd_wq, &mp->m_flush_work); | |
491 | flush_work_sync(&mp->m_flush_work); | |
492 | } | |
493 | ||
494 | STATIC void | |
495 | xfs_flush_worker( | |
496 | struct work_struct *work) | |
497 | { | |
498 | struct xfs_mount *mp = container_of(work, | |
499 | struct xfs_mount, m_flush_work); | |
500 | ||
501 | xfs_sync_data(mp, SYNC_TRYLOCK); | |
502 | xfs_sync_data(mp, SYNC_TRYLOCK | SYNC_WAIT); | |
503 | } | |
504 | ||
a167b17e DC |
505 | int |
506 | xfs_syncd_init( | |
507 | struct xfs_mount *mp) | |
508 | { | |
89e4cb55 | 509 | INIT_WORK(&mp->m_flush_work, xfs_flush_worker); |
c6d09b66 DC |
510 | INIT_DELAYED_WORK(&mp->m_sync_work, xfs_sync_worker); |
511 | xfs_syncd_queue_sync(mp); | |
512 | ||
a167b17e DC |
513 | return 0; |
514 | } | |
515 | ||
516 | void | |
517 | xfs_syncd_stop( | |
518 | struct xfs_mount *mp) | |
519 | { | |
c6d09b66 | 520 | cancel_delayed_work_sync(&mp->m_sync_work); |
89e4cb55 | 521 | cancel_work_sync(&mp->m_flush_work); |
a167b17e DC |
522 | } |
523 | ||
bc990f5c CH |
524 | void |
525 | __xfs_inode_set_reclaim_tag( | |
526 | struct xfs_perag *pag, | |
527 | struct xfs_inode *ip) | |
528 | { | |
529 | radix_tree_tag_set(&pag->pag_ici_root, | |
530 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino), | |
531 | XFS_ICI_RECLAIM_TAG); | |
16fd5367 DC |
532 | |
533 | if (!pag->pag_ici_reclaimable) { | |
534 | /* propagate the reclaim tag up into the perag radix tree */ | |
535 | spin_lock(&ip->i_mount->m_perag_lock); | |
536 | radix_tree_tag_set(&ip->i_mount->m_perag_tree, | |
537 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
538 | XFS_ICI_RECLAIM_TAG); | |
539 | spin_unlock(&ip->i_mount->m_perag_lock); | |
540 | trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno, | |
541 | -1, _RET_IP_); | |
542 | } | |
9bf729c0 | 543 | pag->pag_ici_reclaimable++; |
bc990f5c CH |
544 | } |
545 | ||
11654513 DC |
546 | /* |
547 | * We set the inode flag atomically with the radix tree tag. | |
548 | * Once we get tag lookups on the radix tree, this inode flag | |
549 | * can go away. | |
550 | */ | |
396beb85 DC |
551 | void |
552 | xfs_inode_set_reclaim_tag( | |
553 | xfs_inode_t *ip) | |
554 | { | |
5017e97d DC |
555 | struct xfs_mount *mp = ip->i_mount; |
556 | struct xfs_perag *pag; | |
396beb85 | 557 | |
5017e97d | 558 | pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino)); |
1a427ab0 | 559 | spin_lock(&pag->pag_ici_lock); |
396beb85 | 560 | spin_lock(&ip->i_flags_lock); |
bc990f5c | 561 | __xfs_inode_set_reclaim_tag(pag, ip); |
11654513 | 562 | __xfs_iflags_set(ip, XFS_IRECLAIMABLE); |
396beb85 | 563 | spin_unlock(&ip->i_flags_lock); |
1a427ab0 | 564 | spin_unlock(&pag->pag_ici_lock); |
5017e97d | 565 | xfs_perag_put(pag); |
396beb85 DC |
566 | } |
567 | ||
081003ff JW |
568 | STATIC void |
569 | __xfs_inode_clear_reclaim( | |
396beb85 DC |
570 | xfs_perag_t *pag, |
571 | xfs_inode_t *ip) | |
572 | { | |
9bf729c0 | 573 | pag->pag_ici_reclaimable--; |
16fd5367 DC |
574 | if (!pag->pag_ici_reclaimable) { |
575 | /* clear the reclaim tag from the perag radix tree */ | |
576 | spin_lock(&ip->i_mount->m_perag_lock); | |
577 | radix_tree_tag_clear(&ip->i_mount->m_perag_tree, | |
578 | XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino), | |
579 | XFS_ICI_RECLAIM_TAG); | |
580 | spin_unlock(&ip->i_mount->m_perag_lock); | |
581 | trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno, | |
582 | -1, _RET_IP_); | |
583 | } | |
396beb85 DC |
584 | } |
585 | ||
081003ff JW |
586 | void |
587 | __xfs_inode_clear_reclaim_tag( | |
588 | xfs_mount_t *mp, | |
589 | xfs_perag_t *pag, | |
590 | xfs_inode_t *ip) | |
591 | { | |
592 | radix_tree_tag_clear(&pag->pag_ici_root, | |
593 | XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG); | |
594 | __xfs_inode_clear_reclaim(pag, ip); | |
595 | } | |
596 | ||
e3a20c0b DC |
597 | /* |
598 | * Grab the inode for reclaim exclusively. | |
599 | * Return 0 if we grabbed it, non-zero otherwise. | |
600 | */ | |
601 | STATIC int | |
602 | xfs_reclaim_inode_grab( | |
603 | struct xfs_inode *ip, | |
604 | int flags) | |
605 | { | |
1a3e8f3d DC |
606 | ASSERT(rcu_read_lock_held()); |
607 | ||
608 | /* quick check for stale RCU freed inode */ | |
609 | if (!ip->i_ino) | |
610 | return 1; | |
e3a20c0b DC |
611 | |
612 | /* | |
1a3e8f3d | 613 | * do some unlocked checks first to avoid unnecessary lock traffic. |
e3a20c0b DC |
614 | * The first is a flush lock check, the second is a already in reclaim |
615 | * check. Only do these checks if we are not going to block on locks. | |
616 | */ | |
617 | if ((flags & SYNC_TRYLOCK) && | |
618 | (!ip->i_flush.done || __xfs_iflags_test(ip, XFS_IRECLAIM))) { | |
619 | return 1; | |
620 | } | |
621 | ||
622 | /* | |
623 | * The radix tree lock here protects a thread in xfs_iget from racing | |
624 | * with us starting reclaim on the inode. Once we have the | |
625 | * XFS_IRECLAIM flag set it will not touch us. | |
1a3e8f3d DC |
626 | * |
627 | * Due to RCU lookup, we may find inodes that have been freed and only | |
628 | * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that | |
629 | * aren't candidates for reclaim at all, so we must check the | |
630 | * XFS_IRECLAIMABLE is set first before proceeding to reclaim. | |
e3a20c0b DC |
631 | */ |
632 | spin_lock(&ip->i_flags_lock); | |
1a3e8f3d DC |
633 | if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) || |
634 | __xfs_iflags_test(ip, XFS_IRECLAIM)) { | |
635 | /* not a reclaim candidate. */ | |
e3a20c0b DC |
636 | spin_unlock(&ip->i_flags_lock); |
637 | return 1; | |
638 | } | |
639 | __xfs_iflags_set(ip, XFS_IRECLAIM); | |
640 | spin_unlock(&ip->i_flags_lock); | |
641 | return 0; | |
642 | } | |
643 | ||
777df5af DC |
644 | /* |
645 | * Inodes in different states need to be treated differently, and the return | |
646 | * value of xfs_iflush is not sufficient to get this right. The following table | |
647 | * lists the inode states and the reclaim actions necessary for non-blocking | |
648 | * reclaim: | |
649 | * | |
650 | * | |
651 | * inode state iflush ret required action | |
652 | * --------------- ---------- --------------- | |
653 | * bad - reclaim | |
654 | * shutdown EIO unpin and reclaim | |
655 | * clean, unpinned 0 reclaim | |
656 | * stale, unpinned 0 reclaim | |
c854363e DC |
657 | * clean, pinned(*) 0 requeue |
658 | * stale, pinned EAGAIN requeue | |
659 | * dirty, delwri ok 0 requeue | |
660 | * dirty, delwri blocked EAGAIN requeue | |
661 | * dirty, sync flush 0 reclaim | |
777df5af DC |
662 | * |
663 | * (*) dgc: I don't think the clean, pinned state is possible but it gets | |
664 | * handled anyway given the order of checks implemented. | |
665 | * | |
c854363e DC |
666 | * As can be seen from the table, the return value of xfs_iflush() is not |
667 | * sufficient to correctly decide the reclaim action here. The checks in | |
668 | * xfs_iflush() might look like duplicates, but they are not. | |
669 | * | |
670 | * Also, because we get the flush lock first, we know that any inode that has | |
671 | * been flushed delwri has had the flush completed by the time we check that | |
672 | * the inode is clean. The clean inode check needs to be done before flushing | |
673 | * the inode delwri otherwise we would loop forever requeuing clean inodes as | |
674 | * we cannot tell apart a successful delwri flush and a clean inode from the | |
675 | * return value of xfs_iflush(). | |
676 | * | |
677 | * Note that because the inode is flushed delayed write by background | |
678 | * writeback, the flush lock may already be held here and waiting on it can | |
679 | * result in very long latencies. Hence for sync reclaims, where we wait on the | |
680 | * flush lock, the caller should push out delayed write inodes first before | |
681 | * trying to reclaim them to minimise the amount of time spent waiting. For | |
682 | * background relaim, we just requeue the inode for the next pass. | |
683 | * | |
777df5af DC |
684 | * Hence the order of actions after gaining the locks should be: |
685 | * bad => reclaim | |
686 | * shutdown => unpin and reclaim | |
c854363e DC |
687 | * pinned, delwri => requeue |
688 | * pinned, sync => unpin | |
777df5af DC |
689 | * stale => reclaim |
690 | * clean => reclaim | |
c854363e DC |
691 | * dirty, delwri => flush and requeue |
692 | * dirty, sync => flush, wait and reclaim | |
777df5af | 693 | */ |
75f3cb13 | 694 | STATIC int |
c8e20be0 | 695 | xfs_reclaim_inode( |
75f3cb13 DC |
696 | struct xfs_inode *ip, |
697 | struct xfs_perag *pag, | |
c8e20be0 | 698 | int sync_mode) |
fce08f2f | 699 | { |
1bfd8d04 | 700 | int error; |
777df5af | 701 | |
1bfd8d04 DC |
702 | restart: |
703 | error = 0; | |
c8e20be0 | 704 | xfs_ilock(ip, XFS_ILOCK_EXCL); |
c854363e DC |
705 | if (!xfs_iflock_nowait(ip)) { |
706 | if (!(sync_mode & SYNC_WAIT)) | |
707 | goto out; | |
708 | xfs_iflock(ip); | |
709 | } | |
7a3be02b | 710 | |
777df5af DC |
711 | if (is_bad_inode(VFS_I(ip))) |
712 | goto reclaim; | |
713 | if (XFS_FORCED_SHUTDOWN(ip->i_mount)) { | |
714 | xfs_iunpin_wait(ip); | |
715 | goto reclaim; | |
716 | } | |
c854363e DC |
717 | if (xfs_ipincount(ip)) { |
718 | if (!(sync_mode & SYNC_WAIT)) { | |
719 | xfs_ifunlock(ip); | |
720 | goto out; | |
721 | } | |
777df5af | 722 | xfs_iunpin_wait(ip); |
c854363e | 723 | } |
777df5af DC |
724 | if (xfs_iflags_test(ip, XFS_ISTALE)) |
725 | goto reclaim; | |
726 | if (xfs_inode_clean(ip)) | |
727 | goto reclaim; | |
728 | ||
1bfd8d04 DC |
729 | /* |
730 | * Now we have an inode that needs flushing. | |
731 | * | |
732 | * We do a nonblocking flush here even if we are doing a SYNC_WAIT | |
733 | * reclaim as we can deadlock with inode cluster removal. | |
734 | * xfs_ifree_cluster() can lock the inode buffer before it locks the | |
735 | * ip->i_lock, and we are doing the exact opposite here. As a result, | |
736 | * doing a blocking xfs_itobp() to get the cluster buffer will result | |
737 | * in an ABBA deadlock with xfs_ifree_cluster(). | |
738 | * | |
739 | * As xfs_ifree_cluser() must gather all inodes that are active in the | |
740 | * cache to mark them stale, if we hit this case we don't actually want | |
741 | * to do IO here - we want the inode marked stale so we can simply | |
742 | * reclaim it. Hence if we get an EAGAIN error on a SYNC_WAIT flush, | |
743 | * just unlock the inode, back off and try again. Hopefully the next | |
744 | * pass through will see the stale flag set on the inode. | |
745 | */ | |
746 | error = xfs_iflush(ip, SYNC_TRYLOCK | sync_mode); | |
c854363e | 747 | if (sync_mode & SYNC_WAIT) { |
1bfd8d04 DC |
748 | if (error == EAGAIN) { |
749 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
750 | /* backoff longer than in xfs_ifree_cluster */ | |
751 | delay(2); | |
752 | goto restart; | |
753 | } | |
c854363e DC |
754 | xfs_iflock(ip); |
755 | goto reclaim; | |
c8e20be0 DC |
756 | } |
757 | ||
c854363e DC |
758 | /* |
759 | * When we have to flush an inode but don't have SYNC_WAIT set, we | |
760 | * flush the inode out using a delwri buffer and wait for the next | |
761 | * call into reclaim to find it in a clean state instead of waiting for | |
762 | * it now. We also don't return errors here - if the error is transient | |
763 | * then the next reclaim pass will flush the inode, and if the error | |
f1d486a3 | 764 | * is permanent then the next sync reclaim will reclaim the inode and |
c854363e DC |
765 | * pass on the error. |
766 | */ | |
f1d486a3 | 767 | if (error && error != EAGAIN && !XFS_FORCED_SHUTDOWN(ip->i_mount)) { |
4f10700a | 768 | xfs_warn(ip->i_mount, |
c854363e DC |
769 | "inode 0x%llx background reclaim flush failed with %d", |
770 | (long long)ip->i_ino, error); | |
771 | } | |
772 | out: | |
773 | xfs_iflags_clear(ip, XFS_IRECLAIM); | |
774 | xfs_iunlock(ip, XFS_ILOCK_EXCL); | |
775 | /* | |
776 | * We could return EAGAIN here to make reclaim rescan the inode tree in | |
777 | * a short while. However, this just burns CPU time scanning the tree | |
778 | * waiting for IO to complete and xfssyncd never goes back to the idle | |
779 | * state. Instead, return 0 to let the next scheduled background reclaim | |
780 | * attempt to reclaim the inode again. | |
781 | */ | |
782 | return 0; | |
783 | ||
777df5af DC |
784 | reclaim: |
785 | xfs_ifunlock(ip); | |
c8e20be0 | 786 | xfs_iunlock(ip, XFS_ILOCK_EXCL); |
2f11feab DC |
787 | |
788 | XFS_STATS_INC(xs_ig_reclaims); | |
789 | /* | |
790 | * Remove the inode from the per-AG radix tree. | |
791 | * | |
792 | * Because radix_tree_delete won't complain even if the item was never | |
793 | * added to the tree assert that it's been there before to catch | |
794 | * problems with the inode life time early on. | |
795 | */ | |
1a427ab0 | 796 | spin_lock(&pag->pag_ici_lock); |
2f11feab DC |
797 | if (!radix_tree_delete(&pag->pag_ici_root, |
798 | XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino))) | |
799 | ASSERT(0); | |
081003ff | 800 | __xfs_inode_clear_reclaim(pag, ip); |
1a427ab0 | 801 | spin_unlock(&pag->pag_ici_lock); |
2f11feab DC |
802 | |
803 | /* | |
804 | * Here we do an (almost) spurious inode lock in order to coordinate | |
805 | * with inode cache radix tree lookups. This is because the lookup | |
806 | * can reference the inodes in the cache without taking references. | |
807 | * | |
808 | * We make that OK here by ensuring that we wait until the inode is | |
809 | * unlocked after the lookup before we go ahead and free it. We get | |
810 | * both the ilock and the iolock because the code may need to drop the | |
811 | * ilock one but will still hold the iolock. | |
812 | */ | |
813 | xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL); | |
814 | xfs_qm_dqdetach(ip); | |
815 | xfs_iunlock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL); | |
816 | ||
817 | xfs_inode_free(ip); | |
c854363e DC |
818 | return error; |
819 | ||
7a3be02b DC |
820 | } |
821 | ||
65d0f205 DC |
822 | /* |
823 | * Walk the AGs and reclaim the inodes in them. Even if the filesystem is | |
824 | * corrupted, we still want to try to reclaim all the inodes. If we don't, | |
825 | * then a shut down during filesystem unmount reclaim walk leak all the | |
826 | * unreclaimed inodes. | |
827 | */ | |
828 | int | |
829 | xfs_reclaim_inodes_ag( | |
830 | struct xfs_mount *mp, | |
831 | int flags, | |
832 | int *nr_to_scan) | |
833 | { | |
834 | struct xfs_perag *pag; | |
835 | int error = 0; | |
836 | int last_error = 0; | |
837 | xfs_agnumber_t ag; | |
69b491c2 DC |
838 | int trylock = flags & SYNC_TRYLOCK; |
839 | int skipped; | |
65d0f205 | 840 | |
69b491c2 | 841 | restart: |
65d0f205 | 842 | ag = 0; |
69b491c2 | 843 | skipped = 0; |
65d0f205 DC |
844 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
845 | unsigned long first_index = 0; | |
846 | int done = 0; | |
e3a20c0b | 847 | int nr_found = 0; |
65d0f205 DC |
848 | |
849 | ag = pag->pag_agno + 1; | |
850 | ||
69b491c2 DC |
851 | if (trylock) { |
852 | if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) { | |
853 | skipped++; | |
f83282a8 | 854 | xfs_perag_put(pag); |
69b491c2 DC |
855 | continue; |
856 | } | |
857 | first_index = pag->pag_ici_reclaim_cursor; | |
858 | } else | |
859 | mutex_lock(&pag->pag_ici_reclaim_lock); | |
860 | ||
65d0f205 | 861 | do { |
e3a20c0b DC |
862 | struct xfs_inode *batch[XFS_LOOKUP_BATCH]; |
863 | int i; | |
65d0f205 | 864 | |
1a3e8f3d | 865 | rcu_read_lock(); |
e3a20c0b DC |
866 | nr_found = radix_tree_gang_lookup_tag( |
867 | &pag->pag_ici_root, | |
868 | (void **)batch, first_index, | |
869 | XFS_LOOKUP_BATCH, | |
65d0f205 DC |
870 | XFS_ICI_RECLAIM_TAG); |
871 | if (!nr_found) { | |
1a3e8f3d | 872 | rcu_read_unlock(); |
65d0f205 DC |
873 | break; |
874 | } | |
875 | ||
876 | /* | |
e3a20c0b DC |
877 | * Grab the inodes before we drop the lock. if we found |
878 | * nothing, nr == 0 and the loop will be skipped. | |
65d0f205 | 879 | */ |
e3a20c0b DC |
880 | for (i = 0; i < nr_found; i++) { |
881 | struct xfs_inode *ip = batch[i]; | |
882 | ||
883 | if (done || xfs_reclaim_inode_grab(ip, flags)) | |
884 | batch[i] = NULL; | |
885 | ||
886 | /* | |
887 | * Update the index for the next lookup. Catch | |
888 | * overflows into the next AG range which can | |
889 | * occur if we have inodes in the last block of | |
890 | * the AG and we are currently pointing to the | |
891 | * last inode. | |
1a3e8f3d DC |
892 | * |
893 | * Because we may see inodes that are from the | |
894 | * wrong AG due to RCU freeing and | |
895 | * reallocation, only update the index if it | |
896 | * lies in this AG. It was a race that lead us | |
897 | * to see this inode, so another lookup from | |
898 | * the same index will not find it again. | |
e3a20c0b | 899 | */ |
1a3e8f3d DC |
900 | if (XFS_INO_TO_AGNO(mp, ip->i_ino) != |
901 | pag->pag_agno) | |
902 | continue; | |
e3a20c0b DC |
903 | first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1); |
904 | if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino)) | |
905 | done = 1; | |
906 | } | |
65d0f205 | 907 | |
e3a20c0b | 908 | /* unlock now we've grabbed the inodes. */ |
1a3e8f3d | 909 | rcu_read_unlock(); |
e3a20c0b DC |
910 | |
911 | for (i = 0; i < nr_found; i++) { | |
912 | if (!batch[i]) | |
913 | continue; | |
914 | error = xfs_reclaim_inode(batch[i], pag, flags); | |
915 | if (error && last_error != EFSCORRUPTED) | |
916 | last_error = error; | |
917 | } | |
918 | ||
919 | *nr_to_scan -= XFS_LOOKUP_BATCH; | |
65d0f205 | 920 | |
e3a20c0b | 921 | } while (nr_found && !done && *nr_to_scan > 0); |
65d0f205 | 922 | |
69b491c2 DC |
923 | if (trylock && !done) |
924 | pag->pag_ici_reclaim_cursor = first_index; | |
925 | else | |
926 | pag->pag_ici_reclaim_cursor = 0; | |
927 | mutex_unlock(&pag->pag_ici_reclaim_lock); | |
65d0f205 DC |
928 | xfs_perag_put(pag); |
929 | } | |
69b491c2 DC |
930 | |
931 | /* | |
932 | * if we skipped any AG, and we still have scan count remaining, do | |
933 | * another pass this time using blocking reclaim semantics (i.e | |
934 | * waiting on the reclaim locks and ignoring the reclaim cursors). This | |
935 | * ensure that when we get more reclaimers than AGs we block rather | |
936 | * than spin trying to execute reclaim. | |
937 | */ | |
938 | if (trylock && skipped && *nr_to_scan > 0) { | |
939 | trylock = 0; | |
940 | goto restart; | |
941 | } | |
65d0f205 DC |
942 | return XFS_ERROR(last_error); |
943 | } | |
944 | ||
7a3be02b DC |
945 | int |
946 | xfs_reclaim_inodes( | |
947 | xfs_mount_t *mp, | |
7a3be02b DC |
948 | int mode) |
949 | { | |
65d0f205 DC |
950 | int nr_to_scan = INT_MAX; |
951 | ||
952 | return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan); | |
9bf729c0 DC |
953 | } |
954 | ||
955 | /* | |
956 | * Shrinker infrastructure. | |
9bf729c0 | 957 | */ |
9bf729c0 DC |
958 | static int |
959 | xfs_reclaim_inode_shrink( | |
7f8275d0 | 960 | struct shrinker *shrink, |
9bf729c0 DC |
961 | int nr_to_scan, |
962 | gfp_t gfp_mask) | |
963 | { | |
964 | struct xfs_mount *mp; | |
965 | struct xfs_perag *pag; | |
966 | xfs_agnumber_t ag; | |
16fd5367 | 967 | int reclaimable; |
9bf729c0 | 968 | |
70e60ce7 | 969 | mp = container_of(shrink, struct xfs_mount, m_inode_shrink); |
9bf729c0 DC |
970 | if (nr_to_scan) { |
971 | if (!(gfp_mask & __GFP_FS)) | |
972 | return -1; | |
973 | ||
e3a20c0b | 974 | xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK, &nr_to_scan); |
65d0f205 | 975 | /* terminate if we don't exhaust the scan */ |
70e60ce7 DC |
976 | if (nr_to_scan > 0) |
977 | return -1; | |
978 | } | |
9bf729c0 | 979 | |
16fd5367 DC |
980 | reclaimable = 0; |
981 | ag = 0; | |
65d0f205 DC |
982 | while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) { |
983 | ag = pag->pag_agno + 1; | |
70e60ce7 DC |
984 | reclaimable += pag->pag_ici_reclaimable; |
985 | xfs_perag_put(pag); | |
9bf729c0 | 986 | } |
9bf729c0 DC |
987 | return reclaimable; |
988 | } | |
989 | ||
9bf729c0 DC |
990 | void |
991 | xfs_inode_shrinker_register( | |
992 | struct xfs_mount *mp) | |
993 | { | |
70e60ce7 DC |
994 | mp->m_inode_shrink.shrink = xfs_reclaim_inode_shrink; |
995 | mp->m_inode_shrink.seeks = DEFAULT_SEEKS; | |
996 | register_shrinker(&mp->m_inode_shrink); | |
9bf729c0 DC |
997 | } |
998 | ||
999 | void | |
1000 | xfs_inode_shrinker_unregister( | |
1001 | struct xfs_mount *mp) | |
1002 | { | |
70e60ce7 | 1003 | unregister_shrinker(&mp->m_inode_shrink); |
fce08f2f | 1004 | } |