blob: 9bd7e895a4e28433750979b4f705c795e4b4878f [file] [log] [blame]
David Chinnerfe4fa4b2008-10-30 17:06:08 +11001/*
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"
Dave Chinnerfd074842011-04-08 12:45:07 +100025#include "xfs_trans_priv.h"
David Chinnerfe4fa4b2008-10-30 17:06:08 +110026#include "xfs_sb.h"
27#include "xfs_ag.h"
David Chinnerfe4fa4b2008-10-30 17:06:08 +110028#include "xfs_mount.h"
29#include "xfs_bmap_btree.h"
David Chinnerfe4fa4b2008-10-30 17:06:08 +110030#include "xfs_inode.h"
31#include "xfs_dinode.h"
32#include "xfs_error.h"
David Chinnerfe4fa4b2008-10-30 17:06:08 +110033#include "xfs_filestream.h"
34#include "xfs_vnodeops.h"
David Chinnerfe4fa4b2008-10-30 17:06:08 +110035#include "xfs_inode_item.h"
Christoph Hellwig7d095252009-06-08 15:33:32 +020036#include "xfs_quota.h"
Christoph Hellwig0b1b2132009-12-14 23:14:59 +000037#include "xfs_trace.h"
Dave Chinner1a387d32010-08-24 11:46:31 +100038#include "xfs_fsops.h"
David Chinnerfe4fa4b2008-10-30 17:06:08 +110039
David Chinnera167b172008-10-30 17:06:18 +110040#include <linux/kthread.h>
41#include <linux/freezer.h>
42
Dave Chinnerc6d09b62011-04-08 12:45:07 +100043struct workqueue_struct *xfs_syncd_wq; /* sync workqueue */
44
Dave Chinner78ae5252010-09-28 12:28:19 +100045/*
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
Dave Chinnere13de952010-09-28 12:28:06 +100053STATIC int
54xfs_inode_ag_walk_grab(
55 struct xfs_inode *ip)
56{
57 struct inode *inode = VFS_I(ip);
58
Dave Chinner1a3e8f32010-12-17 17:29:43 +110059 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
Dave Chinnere13de952010-09-28 12:28:06 +100079 /* nothing to sync during shutdown */
80 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
81 return EFSCORRUPTED;
82
Dave Chinnere13de952010-09-28 12:28:06 +100083 /* 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;
Dave Chinner1a3e8f32010-12-17 17:29:43 +110094
95out_unlock_noent:
96 spin_unlock(&ip->i_flags_lock);
97 return ENOENT;
Dave Chinnere13de952010-09-28 12:28:06 +100098}
99
Dave Chinner75f3cb12009-06-08 15:35:14 +0200100STATIC int
101xfs_inode_ag_walk(
102 struct xfs_mount *mp,
Dave Chinner5017e972010-01-11 11:47:40 +0000103 struct xfs_perag *pag,
Dave Chinner75f3cb12009-06-08 15:35:14 +0200104 int (*execute)(struct xfs_inode *ip,
105 struct xfs_perag *pag, int flags),
Dave Chinner65d0f202010-09-24 18:40:15 +1000106 int flags)
Dave Chinner75f3cb12009-06-08 15:35:14 +0200107{
Dave Chinner75f3cb12009-06-08 15:35:14 +0200108 uint32_t first_index;
109 int last_error = 0;
110 int skipped;
Dave Chinner65d0f202010-09-24 18:40:15 +1000111 int done;
Dave Chinner78ae5252010-09-28 12:28:19 +1000112 int nr_found;
Dave Chinner75f3cb12009-06-08 15:35:14 +0200113
114restart:
Dave Chinner65d0f202010-09-24 18:40:15 +1000115 done = 0;
Dave Chinner75f3cb12009-06-08 15:35:14 +0200116 skipped = 0;
117 first_index = 0;
Dave Chinner78ae5252010-09-28 12:28:19 +1000118 nr_found = 0;
Dave Chinner75f3cb12009-06-08 15:35:14 +0200119 do {
Dave Chinner78ae5252010-09-28 12:28:19 +1000120 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
Dave Chinner75f3cb12009-06-08 15:35:14 +0200121 int error = 0;
Dave Chinner78ae5252010-09-28 12:28:19 +1000122 int i;
Dave Chinner75f3cb12009-06-08 15:35:14 +0200123
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100124 rcu_read_lock();
Dave Chinner65d0f202010-09-24 18:40:15 +1000125 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
Dave Chinner78ae5252010-09-28 12:28:19 +1000126 (void **)batch, first_index,
127 XFS_LOOKUP_BATCH);
Dave Chinner65d0f202010-09-24 18:40:15 +1000128 if (!nr_found) {
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100129 rcu_read_unlock();
Dave Chinner75f3cb12009-06-08 15:35:14 +0200130 break;
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000131 }
Dave Chinner75f3cb12009-06-08 15:35:14 +0200132
Dave Chinner65d0f202010-09-24 18:40:15 +1000133 /*
Dave Chinner78ae5252010-09-28 12:28:19 +1000134 * Grab the inodes before we drop the lock. if we found
135 * nothing, nr == 0 and the loop will be skipped.
Dave Chinner65d0f202010-09-24 18:40:15 +1000136 */
Dave Chinner78ae5252010-09-28 12:28:19 +1000137 for (i = 0; i < nr_found; i++) {
138 struct xfs_inode *ip = batch[i];
Dave Chinner65d0f202010-09-24 18:40:15 +1000139
Dave Chinner78ae5252010-09-28 12:28:19 +1000140 if (done || xfs_inode_ag_walk_grab(ip))
141 batch[i] = NULL;
142
143 /*
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100144 * 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.
Dave Chinner78ae5252010-09-28 12:28:19 +1000154 */
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100155 if (XFS_INO_TO_AGNO(mp, ip->i_ino) != pag->pag_agno)
156 continue;
Dave Chinner78ae5252010-09-28 12:28:19 +1000157 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;
Dave Chinnere13de952010-09-28 12:28:06 +1000160 }
Dave Chinner78ae5252010-09-28 12:28:19 +1000161
162 /* unlock now we've grabbed the inodes. */
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100163 rcu_read_unlock();
Dave Chinnere13de952010-09-28 12:28:06 +1000164
Dave Chinner78ae5252010-09-28 12:28:19 +1000165 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;
Dave Chinner75f3cb12009-06-08 15:35:14 +0200176 }
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000177
178 /* bail out if the filesystem is corrupted. */
Dave Chinner75f3cb12009-06-08 15:35:14 +0200179 if (error == EFSCORRUPTED)
180 break;
181
Dave Chinner8daaa832011-07-08 14:14:46 +1000182 cond_resched();
183
Dave Chinner78ae5252010-09-28 12:28:19 +1000184 } while (nr_found && !done);
Dave Chinner75f3cb12009-06-08 15:35:14 +0200185
186 if (skipped) {
187 delay(1);
188 goto restart;
189 }
Dave Chinner75f3cb12009-06-08 15:35:14 +0200190 return last_error;
191}
192
Christoph Hellwigfe588ed2009-06-08 15:35:27 +0200193int
Dave Chinner75f3cb12009-06-08 15:35:14 +0200194xfs_inode_ag_iterator(
195 struct xfs_mount *mp,
196 int (*execute)(struct xfs_inode *ip,
197 struct xfs_perag *pag, int flags),
Dave Chinner65d0f202010-09-24 18:40:15 +1000198 int flags)
Dave Chinner75f3cb12009-06-08 15:35:14 +0200199{
Dave Chinner16fd5362010-07-20 09:43:39 +1000200 struct xfs_perag *pag;
Dave Chinner75f3cb12009-06-08 15:35:14 +0200201 int error = 0;
202 int last_error = 0;
203 xfs_agnumber_t ag;
204
Dave Chinner16fd5362010-07-20 09:43:39 +1000205 ag = 0;
Dave Chinner65d0f202010-09-24 18:40:15 +1000206 while ((pag = xfs_perag_get(mp, ag))) {
207 ag = pag->pag_agno + 1;
208 error = xfs_inode_ag_walk(mp, pag, execute, flags);
Dave Chinner5017e972010-01-11 11:47:40 +0000209 xfs_perag_put(pag);
Dave Chinner75f3cb12009-06-08 15:35:14 +0200210 if (error) {
211 last_error = error;
212 if (error == EFSCORRUPTED)
213 break;
214 }
215 }
216 return XFS_ERROR(last_error);
217}
218
Dave Chinner5a34d5c2009-06-08 15:35:03 +0200219STATIC int
220xfs_sync_inode_data(
221 struct xfs_inode *ip,
Dave Chinner75f3cb12009-06-08 15:35:14 +0200222 struct xfs_perag *pag,
Dave Chinner5a34d5c2009-06-08 15:35:03 +0200223 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))
230 goto out_wait;
231
232 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED)) {
233 if (flags & SYNC_TRYLOCK)
234 goto out_wait;
235 xfs_ilock(ip, XFS_IOLOCK_SHARED);
236 }
237
238 error = xfs_flush_pages(ip, 0, -1, (flags & SYNC_WAIT) ?
Christoph Hellwig0cadda12010-01-19 09:56:44 +0000239 0 : XBF_ASYNC, FI_NONE);
Dave Chinner5a34d5c2009-06-08 15:35:03 +0200240 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
241
242 out_wait:
Christoph Hellwigb0710cc2009-06-08 15:37:11 +0200243 if (flags & SYNC_WAIT)
Dave Chinner5a34d5c2009-06-08 15:35:03 +0200244 xfs_ioend_wait(ip);
245 return error;
246}
247
Christoph Hellwig845b6d02009-06-08 15:35:05 +0200248STATIC int
249xfs_sync_inode_attr(
250 struct xfs_inode *ip,
Dave Chinner75f3cb12009-06-08 15:35:14 +0200251 struct xfs_perag *pag,
Christoph Hellwig845b6d02009-06-08 15:35:05 +0200252 int flags)
253{
254 int error = 0;
255
256 xfs_ilock(ip, XFS_ILOCK_SHARED);
257 if (xfs_inode_clean(ip))
258 goto out_unlock;
259 if (!xfs_iflock_nowait(ip)) {
260 if (!(flags & SYNC_WAIT))
261 goto out_unlock;
262 xfs_iflock(ip);
263 }
264
265 if (xfs_inode_clean(ip)) {
266 xfs_ifunlock(ip);
267 goto out_unlock;
268 }
269
Dave Chinnerc8543632010-02-06 12:39:36 +1100270 error = xfs_iflush(ip, flags);
Christoph Hellwig845b6d02009-06-08 15:35:05 +0200271
Dave Chinneree58abd2011-04-21 09:34:26 +0000272 /*
273 * We don't want to try again on non-blocking flushes that can't run
274 * again immediately. If an inode really must be written, then that's
275 * what the SYNC_WAIT flag is for.
276 */
277 if (error == EAGAIN) {
278 ASSERT(!(flags & SYNC_WAIT));
279 error = 0;
280 }
281
Christoph Hellwig845b6d02009-06-08 15:35:05 +0200282 out_unlock:
283 xfs_iunlock(ip, XFS_ILOCK_SHARED);
284 return error;
285}
286
Christoph Hellwig075fe102009-06-08 15:35:48 +0200287/*
288 * Write out pagecache data for the whole filesystem.
289 */
Christoph Hellwig64c86142010-06-24 11:45:34 +1000290STATIC int
Christoph Hellwig075fe102009-06-08 15:35:48 +0200291xfs_sync_data(
292 struct xfs_mount *mp,
293 int flags)
David Chinnerfe4fa4b2008-10-30 17:06:08 +1100294{
Christoph Hellwig075fe102009-06-08 15:35:48 +0200295 int error;
David Chinnerfe4fa4b2008-10-30 17:06:08 +1100296
Christoph Hellwigb0710cc2009-06-08 15:37:11 +0200297 ASSERT((flags & ~(SYNC_TRYLOCK|SYNC_WAIT)) == 0);
David Chinnerfe4fa4b2008-10-30 17:06:08 +1100298
Dave Chinner65d0f202010-09-24 18:40:15 +1000299 error = xfs_inode_ag_iterator(mp, xfs_sync_inode_data, flags);
Christoph Hellwig075fe102009-06-08 15:35:48 +0200300 if (error)
301 return XFS_ERROR(error);
David Chinnere9f1c6e2008-10-30 17:15:50 +1100302
Christoph Hellwiga14a3482010-01-19 09:56:46 +0000303 xfs_log_force(mp, (flags & SYNC_WAIT) ? XFS_LOG_SYNC : 0);
Christoph Hellwig075fe102009-06-08 15:35:48 +0200304 return 0;
305}
David Chinnere9f1c6e2008-10-30 17:15:50 +1100306
Christoph Hellwig075fe102009-06-08 15:35:48 +0200307/*
308 * Write out inode metadata (attributes) for the whole filesystem.
309 */
Christoph Hellwig64c86142010-06-24 11:45:34 +1000310STATIC int
Christoph Hellwig075fe102009-06-08 15:35:48 +0200311xfs_sync_attr(
312 struct xfs_mount *mp,
313 int flags)
314{
315 ASSERT((flags & ~SYNC_WAIT) == 0);
Dave Chinner75f3cb12009-06-08 15:35:14 +0200316
Dave Chinner65d0f202010-09-24 18:40:15 +1000317 return xfs_inode_ag_iterator(mp, xfs_sync_inode_attr, flags);
David Chinnerfe4fa4b2008-10-30 17:06:08 +1100318}
319
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100320STATIC int
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100321xfs_sync_fsdata(
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000322 struct xfs_mount *mp)
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100323{
324 struct xfs_buf *bp;
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100325
326 /*
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000327 * If the buffer is pinned then push on the log so we won't get stuck
328 * waiting in the write for someone, maybe ourselves, to flush the log.
329 *
330 * Even though we just pushed the log above, we did not have the
331 * superblock buffer locked at that point so it can become pinned in
332 * between there and here.
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100333 */
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000334 bp = xfs_getsb(mp, 0);
335 if (XFS_BUF_ISPINNED(bp))
336 xfs_log_force(mp, 0);
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100337
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000338 return xfs_bwrite(mp, bp);
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100339}
340
David Chinnerfe4fa4b2008-10-30 17:06:08 +1100341/*
David Chinnera4e4c4f2008-10-30 17:16:11 +1100342 * When remounting a filesystem read-only or freezing the filesystem, we have
343 * two phases to execute. This first phase is syncing the data before we
344 * quiesce the filesystem, and the second is flushing all the inodes out after
345 * we've waited for all the transactions created by the first phase to
346 * complete. The second phase ensures that the inodes are written to their
347 * location on disk rather than just existing in transactions in the log. This
348 * means after a quiesce there is no log replay required to write the inodes to
349 * disk (this is the main difference between a sync and a quiesce).
350 */
351/*
352 * First stage of freeze - no writers will make progress now we are here,
David Chinnere9f1c6e2008-10-30 17:15:50 +1100353 * so we flush delwri and delalloc buffers here, then wait for all I/O to
354 * complete. Data is frozen at that point. Metadata is not frozen,
David Chinnera4e4c4f2008-10-30 17:16:11 +1100355 * transactions can still occur here so don't bother flushing the buftarg
356 * because it'll just get dirty again.
David Chinnere9f1c6e2008-10-30 17:15:50 +1100357 */
358int
359xfs_quiesce_data(
360 struct xfs_mount *mp)
361{
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000362 int error, error2 = 0;
David Chinnere9f1c6e2008-10-30 17:15:50 +1100363
364 /* push non-blocking */
Christoph Hellwig075fe102009-06-08 15:35:48 +0200365 xfs_sync_data(mp, 0);
Christoph Hellwig8b5403a2009-06-08 15:37:16 +0200366 xfs_qm_sync(mp, SYNC_TRYLOCK);
David Chinnere9f1c6e2008-10-30 17:15:50 +1100367
Dave Chinnerc90b07e2009-10-06 20:29:27 +0000368 /* push and block till complete */
Christoph Hellwigb0710cc2009-06-08 15:37:11 +0200369 xfs_sync_data(mp, SYNC_WAIT);
Christoph Hellwig7d095252009-06-08 15:33:32 +0200370 xfs_qm_sync(mp, SYNC_WAIT);
David Chinnere9f1c6e2008-10-30 17:15:50 +1100371
David Chinnera4e4c4f2008-10-30 17:16:11 +1100372 /* write superblock and hoover up shutdown errors */
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000373 error = xfs_sync_fsdata(mp);
374
375 /* make sure all delwri buffers are written out */
376 xfs_flush_buftarg(mp->m_ddev_targp, 1);
377
378 /* mark the log as covered if needed */
379 if (xfs_log_need_covered(mp))
Dave Chinnerc58efdb2011-01-04 04:49:29 +0000380 error2 = xfs_fs_log_dummy(mp);
David Chinnere9f1c6e2008-10-30 17:15:50 +1100381
David Chinnera4e4c4f2008-10-30 17:16:11 +1100382 /* flush data-only devices */
David Chinnere9f1c6e2008-10-30 17:15:50 +1100383 if (mp->m_rtdev_targp)
384 XFS_bflush(mp->m_rtdev_targp);
385
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000386 return error ? error : error2;
David Chinnere9f1c6e2008-10-30 17:15:50 +1100387}
388
David Chinner76bf1052008-10-30 17:16:21 +1100389STATIC void
390xfs_quiesce_fs(
391 struct xfs_mount *mp)
392{
393 int count = 0, pincount;
394
Dave Chinnerc8543632010-02-06 12:39:36 +1100395 xfs_reclaim_inodes(mp, 0);
David Chinner76bf1052008-10-30 17:16:21 +1100396 xfs_flush_buftarg(mp->m_ddev_targp, 0);
David Chinner76bf1052008-10-30 17:16:21 +1100397
398 /*
399 * This loop must run at least twice. The first instance of the loop
400 * will flush most meta data but that will generate more meta data
401 * (typically directory updates). Which then must be flushed and
Dave Chinnerc8543632010-02-06 12:39:36 +1100402 * logged before we can write the unmount record. We also so sync
403 * reclaim of inodes to catch any that the above delwri flush skipped.
David Chinner76bf1052008-10-30 17:16:21 +1100404 */
405 do {
Dave Chinnerc8543632010-02-06 12:39:36 +1100406 xfs_reclaim_inodes(mp, SYNC_WAIT);
Christoph Hellwig075fe102009-06-08 15:35:48 +0200407 xfs_sync_attr(mp, SYNC_WAIT);
David Chinner76bf1052008-10-30 17:16:21 +1100408 pincount = xfs_flush_buftarg(mp->m_ddev_targp, 1);
409 if (!pincount) {
410 delay(50);
411 count++;
412 }
413 } while (count < 2);
414}
415
416/*
417 * Second stage of a quiesce. The data is already synced, now we have to take
418 * care of the metadata. New transactions are already blocked, so we need to
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300419 * wait for any remaining transactions to drain out before proceeding.
David Chinner76bf1052008-10-30 17:16:21 +1100420 */
421void
422xfs_quiesce_attr(
423 struct xfs_mount *mp)
424{
425 int error = 0;
426
427 /* wait for all modifications to complete */
428 while (atomic_read(&mp->m_active_trans) > 0)
429 delay(100);
430
431 /* flush inodes and push all remaining buffers out to disk */
432 xfs_quiesce_fs(mp);
433
Felix Blyakher5e106572009-01-22 21:34:05 -0600434 /*
435 * Just warn here till VFS can correctly support
436 * read-only remount without racing.
437 */
438 WARN_ON(atomic_read(&mp->m_active_trans) != 0);
David Chinner76bf1052008-10-30 17:16:21 +1100439
440 /* Push the superblock and write an unmount record */
441 error = xfs_log_sbcount(mp, 1);
442 if (error)
Dave Chinner4f107002011-03-07 10:00:35 +1100443 xfs_warn(mp, "xfs_attr_quiesce: failed to log sb changes. "
David Chinner76bf1052008-10-30 17:16:21 +1100444 "Frozen image may not be consistent.");
445 xfs_log_unmount_write(mp);
446 xfs_unmountfs_writesb(mp);
447}
448
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000449static void
450xfs_syncd_queue_sync(
451 struct xfs_mount *mp)
David Chinnera167b172008-10-30 17:06:18 +1100452{
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000453 queue_delayed_work(xfs_syncd_wq, &mp->m_sync_work,
454 msecs_to_jiffies(xfs_syncd_centisecs * 10));
David Chinnera167b172008-10-30 17:06:18 +1100455}
456
David Chinneraacaa882008-10-30 17:15:29 +1100457/*
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000458 * Every sync period we need to unpin all items, reclaim inodes and sync
459 * disk quotas. We might need to cover the log to indicate that the
Dave Chinner1a387d32010-08-24 11:46:31 +1000460 * filesystem is idle and not frozen.
David Chinneraacaa882008-10-30 17:15:29 +1100461 */
David Chinnera167b172008-10-30 17:06:18 +1100462STATIC void
463xfs_sync_worker(
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000464 struct work_struct *work)
David Chinnera167b172008-10-30 17:06:18 +1100465{
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000466 struct xfs_mount *mp = container_of(to_delayed_work(work),
467 struct xfs_mount, m_sync_work);
David Chinnera167b172008-10-30 17:06:18 +1100468 int error;
469
David Chinneraacaa882008-10-30 17:15:29 +1100470 if (!(mp->m_flags & XFS_MOUNT_RDONLY)) {
David Chinneraacaa882008-10-30 17:15:29 +1100471 /* dgc: errors ignored here */
Dave Chinner1a387d32010-08-24 11:46:31 +1000472 if (mp->m_super->s_frozen == SB_UNFROZEN &&
473 xfs_log_need_covered(mp))
Dave Chinnerc58efdb2011-01-04 04:49:29 +0000474 error = xfs_fs_log_dummy(mp);
475 else
476 xfs_log_force(mp, 0);
Dave Chinnerc58efdb2011-01-04 04:49:29 +0000477 error = xfs_qm_sync(mp, SYNC_TRYLOCK);
Dave Chinnerfd074842011-04-08 12:45:07 +1000478
479 /* start pushing all the metadata that is currently dirty */
480 xfs_ail_push_all(mp->m_ail);
David Chinneraacaa882008-10-30 17:15:29 +1100481 }
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000482
483 /* queue us up again */
484 xfs_syncd_queue_sync(mp);
David Chinnera167b172008-10-30 17:06:18 +1100485}
486
Dave Chinner89e4cb52011-04-08 12:45:07 +1000487/*
Dave Chinnera7b339f2011-04-08 12:45:07 +1000488 * Queue a new inode reclaim pass if there are reclaimable inodes and there
489 * isn't a reclaim pass already in progress. By default it runs every 5s based
490 * on the xfs syncd work default of 30s. Perhaps this should have it's own
491 * tunable, but that can be done if this method proves to be ineffective or too
492 * aggressive.
493 */
494static void
495xfs_syncd_queue_reclaim(
496 struct xfs_mount *mp)
David Chinnera167b172008-10-30 17:06:18 +1100497{
David Chinnera167b172008-10-30 17:06:18 +1100498
Dave Chinnera7b339f2011-04-08 12:45:07 +1000499 /*
500 * We can have inodes enter reclaim after we've shut down the syncd
501 * workqueue during unmount, so don't allow reclaim work to be queued
502 * during unmount.
503 */
504 if (!(mp->m_super->s_flags & MS_ACTIVE))
505 return;
David Chinnera167b172008-10-30 17:06:18 +1100506
Dave Chinnera7b339f2011-04-08 12:45:07 +1000507 rcu_read_lock();
508 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
509 queue_delayed_work(xfs_syncd_wq, &mp->m_reclaim_work,
510 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
David Chinnera167b172008-10-30 17:06:18 +1100511 }
Dave Chinnera7b339f2011-04-08 12:45:07 +1000512 rcu_read_unlock();
513}
David Chinnera167b172008-10-30 17:06:18 +1100514
Dave Chinnera7b339f2011-04-08 12:45:07 +1000515/*
516 * This is a fast pass over the inode cache to try to get reclaim moving on as
517 * many inodes as possible in a short period of time. It kicks itself every few
518 * seconds, as well as being kicked by the inode cache shrinker when memory
519 * goes low. It scans as quickly as possible avoiding locked inodes or those
520 * already being flushed, and once done schedules a future pass.
521 */
522STATIC void
523xfs_reclaim_worker(
524 struct work_struct *work)
525{
526 struct xfs_mount *mp = container_of(to_delayed_work(work),
527 struct xfs_mount, m_reclaim_work);
528
529 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
530 xfs_syncd_queue_reclaim(mp);
531}
532
533/*
Dave Chinner89e4cb52011-04-08 12:45:07 +1000534 * Flush delayed allocate data, attempting to free up reserved space
535 * from existing allocations. At this point a new allocation attempt
536 * has failed with ENOSPC and we are in the process of scratching our
537 * heads, looking about for more room.
538 *
539 * Queue a new data flush if there isn't one already in progress and
540 * wait for completion of the flush. This means that we only ever have one
541 * inode flush in progress no matter how many ENOSPC events are occurring and
542 * so will prevent the system from bogging down due to every concurrent
543 * ENOSPC event scanning all the active inodes in the system for writeback.
544 */
545void
546xfs_flush_inodes(
547 struct xfs_inode *ip)
548{
549 struct xfs_mount *mp = ip->i_mount;
550
551 queue_work(xfs_syncd_wq, &mp->m_flush_work);
552 flush_work_sync(&mp->m_flush_work);
553}
554
555STATIC void
556xfs_flush_worker(
557 struct work_struct *work)
558{
559 struct xfs_mount *mp = container_of(work,
560 struct xfs_mount, m_flush_work);
561
562 xfs_sync_data(mp, SYNC_TRYLOCK);
563 xfs_sync_data(mp, SYNC_TRYLOCK | SYNC_WAIT);
David Chinnera167b172008-10-30 17:06:18 +1100564}
565
566int
567xfs_syncd_init(
568 struct xfs_mount *mp)
569{
Dave Chinner89e4cb52011-04-08 12:45:07 +1000570 INIT_WORK(&mp->m_flush_work, xfs_flush_worker);
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000571 INIT_DELAYED_WORK(&mp->m_sync_work, xfs_sync_worker);
Dave Chinnera7b339f2011-04-08 12:45:07 +1000572 INIT_DELAYED_WORK(&mp->m_reclaim_work, xfs_reclaim_worker);
573
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000574 xfs_syncd_queue_sync(mp);
Dave Chinnera7b339f2011-04-08 12:45:07 +1000575 xfs_syncd_queue_reclaim(mp);
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000576
David Chinnera167b172008-10-30 17:06:18 +1100577 return 0;
578}
579
580void
581xfs_syncd_stop(
582 struct xfs_mount *mp)
583{
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000584 cancel_delayed_work_sync(&mp->m_sync_work);
Dave Chinnera7b339f2011-04-08 12:45:07 +1000585 cancel_delayed_work_sync(&mp->m_reclaim_work);
Dave Chinner89e4cb52011-04-08 12:45:07 +1000586 cancel_work_sync(&mp->m_flush_work);
David Chinnera167b172008-10-30 17:06:18 +1100587}
588
Christoph Hellwigbc990f52009-08-16 20:36:34 -0400589void
590__xfs_inode_set_reclaim_tag(
591 struct xfs_perag *pag,
592 struct xfs_inode *ip)
593{
594 radix_tree_tag_set(&pag->pag_ici_root,
595 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
596 XFS_ICI_RECLAIM_TAG);
Dave Chinner16fd5362010-07-20 09:43:39 +1000597
598 if (!pag->pag_ici_reclaimable) {
599 /* propagate the reclaim tag up into the perag radix tree */
600 spin_lock(&ip->i_mount->m_perag_lock);
601 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
602 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
603 XFS_ICI_RECLAIM_TAG);
604 spin_unlock(&ip->i_mount->m_perag_lock);
Dave Chinnera7b339f2011-04-08 12:45:07 +1000605
606 /* schedule periodic background inode reclaim */
607 xfs_syncd_queue_reclaim(ip->i_mount);
608
Dave Chinner16fd5362010-07-20 09:43:39 +1000609 trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno,
610 -1, _RET_IP_);
611 }
Dave Chinner9bf729c2010-04-29 09:55:50 +1000612 pag->pag_ici_reclaimable++;
Christoph Hellwigbc990f52009-08-16 20:36:34 -0400613}
614
David Chinner11654512008-10-30 17:37:49 +1100615/*
616 * We set the inode flag atomically with the radix tree tag.
617 * Once we get tag lookups on the radix tree, this inode flag
618 * can go away.
619 */
David Chinner396beb82008-10-30 17:37:26 +1100620void
621xfs_inode_set_reclaim_tag(
622 xfs_inode_t *ip)
623{
Dave Chinner5017e972010-01-11 11:47:40 +0000624 struct xfs_mount *mp = ip->i_mount;
625 struct xfs_perag *pag;
David Chinner396beb82008-10-30 17:37:26 +1100626
Dave Chinner5017e972010-01-11 11:47:40 +0000627 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
Dave Chinner1a427ab2010-12-16 17:08:41 +1100628 spin_lock(&pag->pag_ici_lock);
David Chinner396beb82008-10-30 17:37:26 +1100629 spin_lock(&ip->i_flags_lock);
Christoph Hellwigbc990f52009-08-16 20:36:34 -0400630 __xfs_inode_set_reclaim_tag(pag, ip);
David Chinner11654512008-10-30 17:37:49 +1100631 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
David Chinner396beb82008-10-30 17:37:26 +1100632 spin_unlock(&ip->i_flags_lock);
Dave Chinner1a427ab2010-12-16 17:08:41 +1100633 spin_unlock(&pag->pag_ici_lock);
Dave Chinner5017e972010-01-11 11:47:40 +0000634 xfs_perag_put(pag);
David Chinner396beb82008-10-30 17:37:26 +1100635}
636
Johannes Weiner081003f2010-10-01 07:43:54 +0000637STATIC void
638__xfs_inode_clear_reclaim(
David Chinner396beb82008-10-30 17:37:26 +1100639 xfs_perag_t *pag,
640 xfs_inode_t *ip)
641{
Dave Chinner9bf729c2010-04-29 09:55:50 +1000642 pag->pag_ici_reclaimable--;
Dave Chinner16fd5362010-07-20 09:43:39 +1000643 if (!pag->pag_ici_reclaimable) {
644 /* clear the reclaim tag from the perag radix tree */
645 spin_lock(&ip->i_mount->m_perag_lock);
646 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
647 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
648 XFS_ICI_RECLAIM_TAG);
649 spin_unlock(&ip->i_mount->m_perag_lock);
650 trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno,
651 -1, _RET_IP_);
652 }
David Chinner396beb82008-10-30 17:37:26 +1100653}
654
Johannes Weiner081003f2010-10-01 07:43:54 +0000655void
656__xfs_inode_clear_reclaim_tag(
657 xfs_mount_t *mp,
658 xfs_perag_t *pag,
659 xfs_inode_t *ip)
660{
661 radix_tree_tag_clear(&pag->pag_ici_root,
662 XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
663 __xfs_inode_clear_reclaim(pag, ip);
664}
665
Dave Chinner777df5a2010-02-06 12:37:26 +1100666/*
Dave Chinnere3a20c02010-09-24 19:51:50 +1000667 * Grab the inode for reclaim exclusively.
668 * Return 0 if we grabbed it, non-zero otherwise.
669 */
670STATIC int
671xfs_reclaim_inode_grab(
672 struct xfs_inode *ip,
673 int flags)
674{
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100675 ASSERT(rcu_read_lock_held());
676
677 /* quick check for stale RCU freed inode */
678 if (!ip->i_ino)
679 return 1;
Dave Chinnere3a20c02010-09-24 19:51:50 +1000680
681 /*
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100682 * do some unlocked checks first to avoid unnecessary lock traffic.
Dave Chinnere3a20c02010-09-24 19:51:50 +1000683 * The first is a flush lock check, the second is a already in reclaim
684 * check. Only do these checks if we are not going to block on locks.
685 */
686 if ((flags & SYNC_TRYLOCK) &&
687 (!ip->i_flush.done || __xfs_iflags_test(ip, XFS_IRECLAIM))) {
688 return 1;
689 }
690
691 /*
692 * The radix tree lock here protects a thread in xfs_iget from racing
693 * with us starting reclaim on the inode. Once we have the
694 * XFS_IRECLAIM flag set it will not touch us.
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100695 *
696 * Due to RCU lookup, we may find inodes that have been freed and only
697 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
698 * aren't candidates for reclaim at all, so we must check the
699 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
Dave Chinnere3a20c02010-09-24 19:51:50 +1000700 */
701 spin_lock(&ip->i_flags_lock);
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100702 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
703 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
704 /* not a reclaim candidate. */
Dave Chinnere3a20c02010-09-24 19:51:50 +1000705 spin_unlock(&ip->i_flags_lock);
706 return 1;
707 }
708 __xfs_iflags_set(ip, XFS_IRECLAIM);
709 spin_unlock(&ip->i_flags_lock);
710 return 0;
711}
712
713/*
Dave Chinner777df5a2010-02-06 12:37:26 +1100714 * Inodes in different states need to be treated differently, and the return
715 * value of xfs_iflush is not sufficient to get this right. The following table
716 * lists the inode states and the reclaim actions necessary for non-blocking
717 * reclaim:
718 *
719 *
720 * inode state iflush ret required action
721 * --------------- ---------- ---------------
722 * bad - reclaim
723 * shutdown EIO unpin and reclaim
724 * clean, unpinned 0 reclaim
725 * stale, unpinned 0 reclaim
Dave Chinnerc8543632010-02-06 12:39:36 +1100726 * clean, pinned(*) 0 requeue
727 * stale, pinned EAGAIN requeue
728 * dirty, delwri ok 0 requeue
729 * dirty, delwri blocked EAGAIN requeue
730 * dirty, sync flush 0 reclaim
Dave Chinner777df5a2010-02-06 12:37:26 +1100731 *
732 * (*) dgc: I don't think the clean, pinned state is possible but it gets
733 * handled anyway given the order of checks implemented.
734 *
Dave Chinnerc8543632010-02-06 12:39:36 +1100735 * As can be seen from the table, the return value of xfs_iflush() is not
736 * sufficient to correctly decide the reclaim action here. The checks in
737 * xfs_iflush() might look like duplicates, but they are not.
738 *
739 * Also, because we get the flush lock first, we know that any inode that has
740 * been flushed delwri has had the flush completed by the time we check that
741 * the inode is clean. The clean inode check needs to be done before flushing
742 * the inode delwri otherwise we would loop forever requeuing clean inodes as
743 * we cannot tell apart a successful delwri flush and a clean inode from the
744 * return value of xfs_iflush().
745 *
746 * Note that because the inode is flushed delayed write by background
747 * writeback, the flush lock may already be held here and waiting on it can
748 * result in very long latencies. Hence for sync reclaims, where we wait on the
749 * flush lock, the caller should push out delayed write inodes first before
750 * trying to reclaim them to minimise the amount of time spent waiting. For
751 * background relaim, we just requeue the inode for the next pass.
752 *
Dave Chinner777df5a2010-02-06 12:37:26 +1100753 * Hence the order of actions after gaining the locks should be:
754 * bad => reclaim
755 * shutdown => unpin and reclaim
Dave Chinnerc8543632010-02-06 12:39:36 +1100756 * pinned, delwri => requeue
757 * pinned, sync => unpin
Dave Chinner777df5a2010-02-06 12:37:26 +1100758 * stale => reclaim
759 * clean => reclaim
Dave Chinnerc8543632010-02-06 12:39:36 +1100760 * dirty, delwri => flush and requeue
761 * dirty, sync => flush, wait and reclaim
Dave Chinner777df5a2010-02-06 12:37:26 +1100762 */
Dave Chinner75f3cb12009-06-08 15:35:14 +0200763STATIC int
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000764xfs_reclaim_inode(
Dave Chinner75f3cb12009-06-08 15:35:14 +0200765 struct xfs_inode *ip,
766 struct xfs_perag *pag,
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000767 int sync_mode)
David Chinner7a3be022008-10-30 17:37:37 +1100768{
Dave Chinner1bfd8d02011-03-26 09:13:55 +1100769 int error;
Dave Chinner777df5a2010-02-06 12:37:26 +1100770
Dave Chinner1bfd8d02011-03-26 09:13:55 +1100771restart:
772 error = 0;
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000773 xfs_ilock(ip, XFS_ILOCK_EXCL);
Dave Chinnerc8543632010-02-06 12:39:36 +1100774 if (!xfs_iflock_nowait(ip)) {
775 if (!(sync_mode & SYNC_WAIT))
776 goto out;
777 xfs_iflock(ip);
778 }
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000779
Dave Chinner777df5a2010-02-06 12:37:26 +1100780 if (is_bad_inode(VFS_I(ip)))
781 goto reclaim;
782 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
783 xfs_iunpin_wait(ip);
784 goto reclaim;
785 }
Dave Chinnerc8543632010-02-06 12:39:36 +1100786 if (xfs_ipincount(ip)) {
787 if (!(sync_mode & SYNC_WAIT)) {
788 xfs_ifunlock(ip);
789 goto out;
790 }
Dave Chinner777df5a2010-02-06 12:37:26 +1100791 xfs_iunpin_wait(ip);
Dave Chinnerc8543632010-02-06 12:39:36 +1100792 }
Dave Chinner777df5a2010-02-06 12:37:26 +1100793 if (xfs_iflags_test(ip, XFS_ISTALE))
794 goto reclaim;
795 if (xfs_inode_clean(ip))
796 goto reclaim;
797
Dave Chinner1bfd8d02011-03-26 09:13:55 +1100798 /*
799 * Now we have an inode that needs flushing.
800 *
801 * We do a nonblocking flush here even if we are doing a SYNC_WAIT
802 * reclaim as we can deadlock with inode cluster removal.
803 * xfs_ifree_cluster() can lock the inode buffer before it locks the
804 * ip->i_lock, and we are doing the exact opposite here. As a result,
805 * doing a blocking xfs_itobp() to get the cluster buffer will result
806 * in an ABBA deadlock with xfs_ifree_cluster().
807 *
808 * As xfs_ifree_cluser() must gather all inodes that are active in the
809 * cache to mark them stale, if we hit this case we don't actually want
810 * to do IO here - we want the inode marked stale so we can simply
811 * reclaim it. Hence if we get an EAGAIN error on a SYNC_WAIT flush,
812 * just unlock the inode, back off and try again. Hopefully the next
813 * pass through will see the stale flag set on the inode.
814 */
815 error = xfs_iflush(ip, SYNC_TRYLOCK | sync_mode);
Dave Chinnerc8543632010-02-06 12:39:36 +1100816 if (sync_mode & SYNC_WAIT) {
Dave Chinner1bfd8d02011-03-26 09:13:55 +1100817 if (error == EAGAIN) {
818 xfs_iunlock(ip, XFS_ILOCK_EXCL);
819 /* backoff longer than in xfs_ifree_cluster */
820 delay(2);
821 goto restart;
822 }
Dave Chinnerc8543632010-02-06 12:39:36 +1100823 xfs_iflock(ip);
824 goto reclaim;
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000825 }
826
Dave Chinnerc8543632010-02-06 12:39:36 +1100827 /*
828 * When we have to flush an inode but don't have SYNC_WAIT set, we
829 * flush the inode out using a delwri buffer and wait for the next
830 * call into reclaim to find it in a clean state instead of waiting for
831 * it now. We also don't return errors here - if the error is transient
832 * then the next reclaim pass will flush the inode, and if the error
Dave Chinnerf1d486a2010-04-13 15:06:45 +1000833 * is permanent then the next sync reclaim will reclaim the inode and
Dave Chinnerc8543632010-02-06 12:39:36 +1100834 * pass on the error.
835 */
Dave Chinnerf1d486a2010-04-13 15:06:45 +1000836 if (error && error != EAGAIN && !XFS_FORCED_SHUTDOWN(ip->i_mount)) {
Dave Chinner4f107002011-03-07 10:00:35 +1100837 xfs_warn(ip->i_mount,
Dave Chinnerc8543632010-02-06 12:39:36 +1100838 "inode 0x%llx background reclaim flush failed with %d",
839 (long long)ip->i_ino, error);
840 }
841out:
842 xfs_iflags_clear(ip, XFS_IRECLAIM);
843 xfs_iunlock(ip, XFS_ILOCK_EXCL);
844 /*
845 * We could return EAGAIN here to make reclaim rescan the inode tree in
846 * a short while. However, this just burns CPU time scanning the tree
847 * waiting for IO to complete and xfssyncd never goes back to the idle
848 * state. Instead, return 0 to let the next scheduled background reclaim
849 * attempt to reclaim the inode again.
850 */
851 return 0;
852
Dave Chinner777df5a2010-02-06 12:37:26 +1100853reclaim:
854 xfs_ifunlock(ip);
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000855 xfs_iunlock(ip, XFS_ILOCK_EXCL);
Dave Chinner2f11fea2010-07-20 17:53:25 +1000856
857 XFS_STATS_INC(xs_ig_reclaims);
858 /*
859 * Remove the inode from the per-AG radix tree.
860 *
861 * Because radix_tree_delete won't complain even if the item was never
862 * added to the tree assert that it's been there before to catch
863 * problems with the inode life time early on.
864 */
Dave Chinner1a427ab2010-12-16 17:08:41 +1100865 spin_lock(&pag->pag_ici_lock);
Dave Chinner2f11fea2010-07-20 17:53:25 +1000866 if (!radix_tree_delete(&pag->pag_ici_root,
867 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino)))
868 ASSERT(0);
Johannes Weiner081003f2010-10-01 07:43:54 +0000869 __xfs_inode_clear_reclaim(pag, ip);
Dave Chinner1a427ab2010-12-16 17:08:41 +1100870 spin_unlock(&pag->pag_ici_lock);
Dave Chinner2f11fea2010-07-20 17:53:25 +1000871
872 /*
873 * Here we do an (almost) spurious inode lock in order to coordinate
874 * with inode cache radix tree lookups. This is because the lookup
875 * can reference the inodes in the cache without taking references.
876 *
877 * We make that OK here by ensuring that we wait until the inode is
878 * unlocked after the lookup before we go ahead and free it. We get
879 * both the ilock and the iolock because the code may need to drop the
880 * ilock one but will still hold the iolock.
881 */
882 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
883 xfs_qm_dqdetach(ip);
884 xfs_iunlock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
885
886 xfs_inode_free(ip);
Dave Chinnerc8543632010-02-06 12:39:36 +1100887 return error;
888
David Chinner7a3be022008-10-30 17:37:37 +1100889}
890
Dave Chinner65d0f202010-09-24 18:40:15 +1000891/*
892 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
893 * corrupted, we still want to try to reclaim all the inodes. If we don't,
894 * then a shut down during filesystem unmount reclaim walk leak all the
895 * unreclaimed inodes.
896 */
897int
898xfs_reclaim_inodes_ag(
899 struct xfs_mount *mp,
900 int flags,
901 int *nr_to_scan)
902{
903 struct xfs_perag *pag;
904 int error = 0;
905 int last_error = 0;
906 xfs_agnumber_t ag;
Dave Chinner69b491c2010-09-27 11:09:51 +1000907 int trylock = flags & SYNC_TRYLOCK;
908 int skipped;
Dave Chinner65d0f202010-09-24 18:40:15 +1000909
Dave Chinner69b491c2010-09-27 11:09:51 +1000910restart:
Dave Chinner65d0f202010-09-24 18:40:15 +1000911 ag = 0;
Dave Chinner69b491c2010-09-27 11:09:51 +1000912 skipped = 0;
Dave Chinner65d0f202010-09-24 18:40:15 +1000913 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
914 unsigned long first_index = 0;
915 int done = 0;
Dave Chinnere3a20c02010-09-24 19:51:50 +1000916 int nr_found = 0;
Dave Chinner65d0f202010-09-24 18:40:15 +1000917
918 ag = pag->pag_agno + 1;
919
Dave Chinner69b491c2010-09-27 11:09:51 +1000920 if (trylock) {
921 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
922 skipped++;
Dave Chinnerf83282a2010-11-08 08:55:04 +0000923 xfs_perag_put(pag);
Dave Chinner69b491c2010-09-27 11:09:51 +1000924 continue;
925 }
926 first_index = pag->pag_ici_reclaim_cursor;
927 } else
928 mutex_lock(&pag->pag_ici_reclaim_lock);
929
Dave Chinner65d0f202010-09-24 18:40:15 +1000930 do {
Dave Chinnere3a20c02010-09-24 19:51:50 +1000931 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
932 int i;
Dave Chinner65d0f202010-09-24 18:40:15 +1000933
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100934 rcu_read_lock();
Dave Chinnere3a20c02010-09-24 19:51:50 +1000935 nr_found = radix_tree_gang_lookup_tag(
936 &pag->pag_ici_root,
937 (void **)batch, first_index,
938 XFS_LOOKUP_BATCH,
Dave Chinner65d0f202010-09-24 18:40:15 +1000939 XFS_ICI_RECLAIM_TAG);
940 if (!nr_found) {
Dave Chinnerb2232212011-05-06 02:54:04 +0000941 done = 1;
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100942 rcu_read_unlock();
Dave Chinner65d0f202010-09-24 18:40:15 +1000943 break;
944 }
945
946 /*
Dave Chinnere3a20c02010-09-24 19:51:50 +1000947 * Grab the inodes before we drop the lock. if we found
948 * nothing, nr == 0 and the loop will be skipped.
Dave Chinner65d0f202010-09-24 18:40:15 +1000949 */
Dave Chinnere3a20c02010-09-24 19:51:50 +1000950 for (i = 0; i < nr_found; i++) {
951 struct xfs_inode *ip = batch[i];
Dave Chinner65d0f202010-09-24 18:40:15 +1000952
Dave Chinnere3a20c02010-09-24 19:51:50 +1000953 if (done || xfs_reclaim_inode_grab(ip, flags))
954 batch[i] = NULL;
Dave Chinner65d0f202010-09-24 18:40:15 +1000955
Dave Chinnere3a20c02010-09-24 19:51:50 +1000956 /*
957 * Update the index for the next lookup. Catch
958 * overflows into the next AG range which can
959 * occur if we have inodes in the last block of
960 * the AG and we are currently pointing to the
961 * last inode.
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100962 *
963 * Because we may see inodes that are from the
964 * wrong AG due to RCU freeing and
965 * reallocation, only update the index if it
966 * lies in this AG. It was a race that lead us
967 * to see this inode, so another lookup from
968 * the same index will not find it again.
Dave Chinnere3a20c02010-09-24 19:51:50 +1000969 */
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100970 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
971 pag->pag_agno)
972 continue;
Dave Chinnere3a20c02010-09-24 19:51:50 +1000973 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
974 if (first_index < XFS_INO_TO_AGINO(mp, ip->i_ino))
975 done = 1;
976 }
977
978 /* unlock now we've grabbed the inodes. */
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100979 rcu_read_unlock();
Dave Chinnere3a20c02010-09-24 19:51:50 +1000980
981 for (i = 0; i < nr_found; i++) {
982 if (!batch[i])
983 continue;
984 error = xfs_reclaim_inode(batch[i], pag, flags);
985 if (error && last_error != EFSCORRUPTED)
986 last_error = error;
987 }
988
989 *nr_to_scan -= XFS_LOOKUP_BATCH;
990
Dave Chinner8daaa832011-07-08 14:14:46 +1000991 cond_resched();
992
Dave Chinnere3a20c02010-09-24 19:51:50 +1000993 } while (nr_found && !done && *nr_to_scan > 0);
Dave Chinner65d0f202010-09-24 18:40:15 +1000994
Dave Chinner69b491c2010-09-27 11:09:51 +1000995 if (trylock && !done)
996 pag->pag_ici_reclaim_cursor = first_index;
997 else
998 pag->pag_ici_reclaim_cursor = 0;
999 mutex_unlock(&pag->pag_ici_reclaim_lock);
Dave Chinner65d0f202010-09-24 18:40:15 +10001000 xfs_perag_put(pag);
1001 }
Dave Chinner69b491c2010-09-27 11:09:51 +10001002
1003 /*
1004 * if we skipped any AG, and we still have scan count remaining, do
1005 * another pass this time using blocking reclaim semantics (i.e
1006 * waiting on the reclaim locks and ignoring the reclaim cursors). This
1007 * ensure that when we get more reclaimers than AGs we block rather
1008 * than spin trying to execute reclaim.
1009 */
Dave Chinner8daaa832011-07-08 14:14:46 +10001010 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
Dave Chinner69b491c2010-09-27 11:09:51 +10001011 trylock = 0;
1012 goto restart;
1013 }
Dave Chinner65d0f202010-09-24 18:40:15 +10001014 return XFS_ERROR(last_error);
1015}
1016
David Chinnerfce08f22008-10-30 17:37:03 +11001017int
David Chinner1dc33182008-10-30 17:37:15 +11001018xfs_reclaim_inodes(
David Chinnerfce08f22008-10-30 17:37:03 +11001019 xfs_mount_t *mp,
David Chinnerfce08f22008-10-30 17:37:03 +11001020 int mode)
1021{
Dave Chinner65d0f202010-09-24 18:40:15 +10001022 int nr_to_scan = INT_MAX;
1023
1024 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
Dave Chinner9bf729c2010-04-29 09:55:50 +10001025}
1026
1027/*
Dave Chinner8daaa832011-07-08 14:14:46 +10001028 * Scan a certain number of inodes for reclaim.
Dave Chinnera7b339f2011-04-08 12:45:07 +10001029 *
1030 * When called we make sure that there is a background (fast) inode reclaim in
Dave Chinner8daaa832011-07-08 14:14:46 +10001031 * progress, while we will throttle the speed of reclaim via doing synchronous
Dave Chinnera7b339f2011-04-08 12:45:07 +10001032 * reclaim of inodes. That means if we come across dirty inodes, we wait for
1033 * them to be cleaned, which we hope will not be very long due to the
1034 * background walker having already kicked the IO off on those dirty inodes.
Dave Chinner9bf729c2010-04-29 09:55:50 +10001035 */
Dave Chinner8daaa832011-07-08 14:14:46 +10001036void
1037xfs_reclaim_inodes_nr(
1038 struct xfs_mount *mp,
1039 int nr_to_scan)
Dave Chinner9bf729c2010-04-29 09:55:50 +10001040{
Dave Chinner8daaa832011-07-08 14:14:46 +10001041 /* kick background reclaimer and push the AIL */
1042 xfs_syncd_queue_reclaim(mp);
1043 xfs_ail_push_all(mp->m_ail);
Dave Chinner9bf729c2010-04-29 09:55:50 +10001044
Dave Chinner8daaa832011-07-08 14:14:46 +10001045 xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
1046}
Dave Chinnera7b339f2011-04-08 12:45:07 +10001047
Dave Chinner8daaa832011-07-08 14:14:46 +10001048/*
1049 * Return the number of reclaimable inodes in the filesystem for
1050 * the shrinker to determine how much to reclaim.
1051 */
1052int
1053xfs_reclaim_inodes_count(
1054 struct xfs_mount *mp)
1055{
1056 struct xfs_perag *pag;
1057 xfs_agnumber_t ag = 0;
1058 int reclaimable = 0;
Dave Chinner9bf729c2010-04-29 09:55:50 +10001059
Dave Chinner65d0f202010-09-24 18:40:15 +10001060 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
1061 ag = pag->pag_agno + 1;
Dave Chinner70e60ce2010-07-20 08:07:02 +10001062 reclaimable += pag->pag_ici_reclaimable;
1063 xfs_perag_put(pag);
Dave Chinner9bf729c2010-04-29 09:55:50 +10001064 }
Dave Chinner9bf729c2010-04-29 09:55:50 +10001065 return reclaimable;
1066}
1067