blob: 85d03e6a2677cbbe31450eaa909d4738686b3dad [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))
Christoph Hellwig4a06fd22011-08-23 08:28:13 +0000230 return 0;
Dave Chinner5a34d5c2009-06-08 15:35:03 +0200231
232 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED)) {
233 if (flags & SYNC_TRYLOCK)
Christoph Hellwig4a06fd22011-08-23 08:28:13 +0000234 return 0;
Dave Chinner5a34d5c2009-06-08 15:35:03 +0200235 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);
Dave Chinner5a34d5c2009-06-08 15:35:03 +0200241 return error;
242}
243
Christoph Hellwig075fe102009-06-08 15:35:48 +0200244/*
245 * Write out pagecache data for the whole filesystem.
246 */
Christoph Hellwig64c86142010-06-24 11:45:34 +1000247STATIC int
Christoph Hellwig075fe102009-06-08 15:35:48 +0200248xfs_sync_data(
249 struct xfs_mount *mp,
250 int flags)
David Chinnerfe4fa4b2008-10-30 17:06:08 +1100251{
Christoph Hellwig075fe102009-06-08 15:35:48 +0200252 int error;
David Chinnerfe4fa4b2008-10-30 17:06:08 +1100253
Christoph Hellwigb0710cc2009-06-08 15:37:11 +0200254 ASSERT((flags & ~(SYNC_TRYLOCK|SYNC_WAIT)) == 0);
David Chinnerfe4fa4b2008-10-30 17:06:08 +1100255
Dave Chinner65d0f202010-09-24 18:40:15 +1000256 error = xfs_inode_ag_iterator(mp, xfs_sync_inode_data, flags);
Christoph Hellwig075fe102009-06-08 15:35:48 +0200257 if (error)
258 return XFS_ERROR(error);
David Chinnere9f1c6e2008-10-30 17:15:50 +1100259
Christoph Hellwiga14a3482010-01-19 09:56:46 +0000260 xfs_log_force(mp, (flags & SYNC_WAIT) ? XFS_LOG_SYNC : 0);
Christoph Hellwig075fe102009-06-08 15:35:48 +0200261 return 0;
262}
David Chinnere9f1c6e2008-10-30 17:15:50 +1100263
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100264STATIC int
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100265xfs_sync_fsdata(
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000266 struct xfs_mount *mp)
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100267{
268 struct xfs_buf *bp;
Christoph Hellwigc2b006c2011-08-23 08:28:07 +0000269 int error;
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100270
271 /*
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000272 * 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.
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100278 */
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000279 bp = xfs_getsb(mp, 0);
Chandra Seetharaman811e64c2011-07-22 23:40:27 +0000280 if (xfs_buf_ispinned(bp))
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000281 xfs_log_force(mp, 0);
Christoph Hellwigc2b006c2011-08-23 08:28:07 +0000282 error = xfs_bwrite(bp);
283 xfs_buf_relse(bp);
284 return error;
Christoph Hellwig2af75df2008-10-30 17:14:53 +1100285}
286
David Chinnerfe4fa4b2008-10-30 17:06:08 +1100287/*
David Chinnera4e4c4f2008-10-30 17:16:11 +1100288 * 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,
David Chinnere9f1c6e2008-10-30 17:15:50 +1100299 * 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,
Christoph Hellwig211e4d42012-04-23 15:58:34 +1000301 * transactions can still occur here so don't bother emptying the AIL
David Chinnera4e4c4f2008-10-30 17:16:11 +1100302 * because it'll just get dirty again.
David Chinnere9f1c6e2008-10-30 17:15:50 +1100303 */
304int
305xfs_quiesce_data(
306 struct xfs_mount *mp)
307{
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000308 int error, error2 = 0;
David Chinnere9f1c6e2008-10-30 17:15:50 +1100309
Christoph Hellwig34625c62011-12-06 21:58:12 +0000310 /* force out the log */
Christoph Hellwig33b8f7c2011-07-08 14:34:39 +0200311 xfs_log_force(mp, XFS_LOG_SYNC);
312
David Chinnera4e4c4f2008-10-30 17:16:11 +1100313 /* write superblock and hoover up shutdown errors */
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000314 error = xfs_sync_fsdata(mp);
315
316 /* make sure all delwri buffers are written out */
317 xfs_flush_buftarg(mp->m_ddev_targp, 1);
318
319 /* mark the log as covered if needed */
320 if (xfs_log_need_covered(mp))
Dave Chinnerc58efdb2011-01-04 04:49:29 +0000321 error2 = xfs_fs_log_dummy(mp);
David Chinnere9f1c6e2008-10-30 17:15:50 +1100322
David Chinnera4e4c4f2008-10-30 17:16:11 +1100323 /* flush data-only devices */
David Chinnere9f1c6e2008-10-30 17:15:50 +1100324 if (mp->m_rtdev_targp)
Christoph Hellwiga9add832011-10-10 16:52:52 +0000325 xfs_flush_buftarg(mp->m_rtdev_targp, 1);
David Chinnere9f1c6e2008-10-30 17:15:50 +1100326
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000327 return error ? error : error2;
David Chinnere9f1c6e2008-10-30 17:15:50 +1100328}
329
David Chinner76bf1052008-10-30 17:16:21 +1100330/*
331 * Second stage of a quiesce. The data is already synced, now we have to take
332 * care of the metadata. New transactions are already blocked, so we need to
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300333 * wait for any remaining transactions to drain out before proceeding.
David Chinner76bf1052008-10-30 17:16:21 +1100334 */
335void
336xfs_quiesce_attr(
337 struct xfs_mount *mp)
338{
339 int error = 0;
340
341 /* wait for all modifications to complete */
342 while (atomic_read(&mp->m_active_trans) > 0)
343 delay(100);
344
Christoph Hellwig211e4d42012-04-23 15:58:34 +1000345 /* reclaim inodes to do any IO before the freeze completes */
346 xfs_reclaim_inodes(mp, 0);
347 xfs_reclaim_inodes(mp, SYNC_WAIT);
348
349 /* flush all pending changes from the AIL */
350 xfs_ail_push_all_sync(mp->m_ail);
David Chinner76bf1052008-10-30 17:16:21 +1100351
Felix Blyakher5e106572009-01-22 21:34:05 -0600352 /*
353 * Just warn here till VFS can correctly support
354 * read-only remount without racing.
355 */
356 WARN_ON(atomic_read(&mp->m_active_trans) != 0);
David Chinner76bf1052008-10-30 17:16:21 +1100357
358 /* Push the superblock and write an unmount record */
Chandra Seetharamanadab0f62011-06-29 22:10:14 +0000359 error = xfs_log_sbcount(mp);
David Chinner76bf1052008-10-30 17:16:21 +1100360 if (error)
Dave Chinner4f107002011-03-07 10:00:35 +1100361 xfs_warn(mp, "xfs_attr_quiesce: failed to log sb changes. "
David Chinner76bf1052008-10-30 17:16:21 +1100362 "Frozen image may not be consistent.");
363 xfs_log_unmount_write(mp);
Christoph Hellwig211e4d42012-04-23 15:58:34 +1000364
365 /*
366 * At this point we might have modified the superblock again and thus
367 * added an item to the AIL, thus flush it again.
368 */
369 xfs_ail_push_all_sync(mp->m_ail);
David Chinner76bf1052008-10-30 17:16:21 +1100370}
371
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000372static void
373xfs_syncd_queue_sync(
374 struct xfs_mount *mp)
David Chinnera167b172008-10-30 17:06:18 +1100375{
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000376 queue_delayed_work(xfs_syncd_wq, &mp->m_sync_work,
377 msecs_to_jiffies(xfs_syncd_centisecs * 10));
David Chinnera167b172008-10-30 17:06:18 +1100378}
379
David Chinneraacaa882008-10-30 17:15:29 +1100380/*
Christoph Hellwigdf308bc2010-03-12 10:59:16 +0000381 * Every sync period we need to unpin all items, reclaim inodes and sync
382 * disk quotas. We might need to cover the log to indicate that the
Dave Chinner1a387d32010-08-24 11:46:31 +1000383 * filesystem is idle and not frozen.
David Chinneraacaa882008-10-30 17:15:29 +1100384 */
David Chinnera167b172008-10-30 17:06:18 +1100385STATIC void
386xfs_sync_worker(
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000387 struct work_struct *work)
David Chinnera167b172008-10-30 17:06:18 +1100388{
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000389 struct xfs_mount *mp = container_of(to_delayed_work(work),
390 struct xfs_mount, m_sync_work);
David Chinnera167b172008-10-30 17:06:18 +1100391 int error;
392
Dave Chinner8a00ebe2012-04-13 12:10:44 +0000393 /*
394 * We shouldn't write/force the log if we are in the mount/unmount
395 * process or on a read only filesystem. The workqueue still needs to be
396 * active in both cases, however, because it is used for inode reclaim
397 * during these times. hence use the MS_ACTIVE flag to avoid doing
398 * anything in these periods.
399 */
400 if (!(mp->m_super->s_flags & MS_ACTIVE) &&
401 !(mp->m_flags & XFS_MOUNT_RDONLY)) {
David Chinneraacaa882008-10-30 17:15:29 +1100402 /* dgc: errors ignored here */
Dave Chinner1a387d32010-08-24 11:46:31 +1000403 if (mp->m_super->s_frozen == SB_UNFROZEN &&
404 xfs_log_need_covered(mp))
Dave Chinnerc58efdb2011-01-04 04:49:29 +0000405 error = xfs_fs_log_dummy(mp);
406 else
407 xfs_log_force(mp, 0);
Dave Chinnerfd074842011-04-08 12:45:07 +1000408
409 /* start pushing all the metadata that is currently dirty */
410 xfs_ail_push_all(mp->m_ail);
David Chinneraacaa882008-10-30 17:15:29 +1100411 }
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000412
413 /* queue us up again */
414 xfs_syncd_queue_sync(mp);
David Chinnera167b172008-10-30 17:06:18 +1100415}
416
Dave Chinner89e4cb52011-04-08 12:45:07 +1000417/*
Dave Chinnera7b339f2011-04-08 12:45:07 +1000418 * Queue a new inode reclaim pass if there are reclaimable inodes and there
419 * isn't a reclaim pass already in progress. By default it runs every 5s based
420 * on the xfs syncd work default of 30s. Perhaps this should have it's own
421 * tunable, but that can be done if this method proves to be ineffective or too
422 * aggressive.
423 */
424static void
425xfs_syncd_queue_reclaim(
426 struct xfs_mount *mp)
David Chinnera167b172008-10-30 17:06:18 +1100427{
David Chinnera167b172008-10-30 17:06:18 +1100428
Dave Chinnera7b339f2011-04-08 12:45:07 +1000429 rcu_read_lock();
430 if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG)) {
431 queue_delayed_work(xfs_syncd_wq, &mp->m_reclaim_work,
432 msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
David Chinnera167b172008-10-30 17:06:18 +1100433 }
Dave Chinnera7b339f2011-04-08 12:45:07 +1000434 rcu_read_unlock();
435}
David Chinnera167b172008-10-30 17:06:18 +1100436
Dave Chinnera7b339f2011-04-08 12:45:07 +1000437/*
438 * This is a fast pass over the inode cache to try to get reclaim moving on as
439 * many inodes as possible in a short period of time. It kicks itself every few
440 * seconds, as well as being kicked by the inode cache shrinker when memory
441 * goes low. It scans as quickly as possible avoiding locked inodes or those
442 * already being flushed, and once done schedules a future pass.
443 */
444STATIC void
445xfs_reclaim_worker(
446 struct work_struct *work)
447{
448 struct xfs_mount *mp = container_of(to_delayed_work(work),
449 struct xfs_mount, m_reclaim_work);
450
451 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
452 xfs_syncd_queue_reclaim(mp);
453}
454
455/*
Dave Chinner89e4cb52011-04-08 12:45:07 +1000456 * Flush delayed allocate data, attempting to free up reserved space
457 * from existing allocations. At this point a new allocation attempt
458 * has failed with ENOSPC and we are in the process of scratching our
459 * heads, looking about for more room.
460 *
461 * Queue a new data flush if there isn't one already in progress and
462 * wait for completion of the flush. This means that we only ever have one
463 * inode flush in progress no matter how many ENOSPC events are occurring and
464 * so will prevent the system from bogging down due to every concurrent
465 * ENOSPC event scanning all the active inodes in the system for writeback.
466 */
467void
468xfs_flush_inodes(
469 struct xfs_inode *ip)
470{
471 struct xfs_mount *mp = ip->i_mount;
472
473 queue_work(xfs_syncd_wq, &mp->m_flush_work);
474 flush_work_sync(&mp->m_flush_work);
475}
476
477STATIC void
478xfs_flush_worker(
479 struct work_struct *work)
480{
481 struct xfs_mount *mp = container_of(work,
482 struct xfs_mount, m_flush_work);
483
484 xfs_sync_data(mp, SYNC_TRYLOCK);
485 xfs_sync_data(mp, SYNC_TRYLOCK | SYNC_WAIT);
David Chinnera167b172008-10-30 17:06:18 +1100486}
487
488int
489xfs_syncd_init(
490 struct xfs_mount *mp)
491{
Dave Chinner89e4cb52011-04-08 12:45:07 +1000492 INIT_WORK(&mp->m_flush_work, xfs_flush_worker);
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000493 INIT_DELAYED_WORK(&mp->m_sync_work, xfs_sync_worker);
Dave Chinnera7b339f2011-04-08 12:45:07 +1000494 INIT_DELAYED_WORK(&mp->m_reclaim_work, xfs_reclaim_worker);
495
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000496 xfs_syncd_queue_sync(mp);
497
David Chinnera167b172008-10-30 17:06:18 +1100498 return 0;
499}
500
501void
502xfs_syncd_stop(
503 struct xfs_mount *mp)
504{
Dave Chinnerc6d09b62011-04-08 12:45:07 +1000505 cancel_delayed_work_sync(&mp->m_sync_work);
Dave Chinnera7b339f2011-04-08 12:45:07 +1000506 cancel_delayed_work_sync(&mp->m_reclaim_work);
Dave Chinner89e4cb52011-04-08 12:45:07 +1000507 cancel_work_sync(&mp->m_flush_work);
David Chinnera167b172008-10-30 17:06:18 +1100508}
509
Christoph Hellwigbc990f52009-08-16 20:36:34 -0400510void
511__xfs_inode_set_reclaim_tag(
512 struct xfs_perag *pag,
513 struct xfs_inode *ip)
514{
515 radix_tree_tag_set(&pag->pag_ici_root,
516 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
517 XFS_ICI_RECLAIM_TAG);
Dave Chinner16fd5362010-07-20 09:43:39 +1000518
519 if (!pag->pag_ici_reclaimable) {
520 /* propagate the reclaim tag up into the perag radix tree */
521 spin_lock(&ip->i_mount->m_perag_lock);
522 radix_tree_tag_set(&ip->i_mount->m_perag_tree,
523 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
524 XFS_ICI_RECLAIM_TAG);
525 spin_unlock(&ip->i_mount->m_perag_lock);
Dave Chinnera7b339f2011-04-08 12:45:07 +1000526
527 /* schedule periodic background inode reclaim */
528 xfs_syncd_queue_reclaim(ip->i_mount);
529
Dave Chinner16fd5362010-07-20 09:43:39 +1000530 trace_xfs_perag_set_reclaim(ip->i_mount, pag->pag_agno,
531 -1, _RET_IP_);
532 }
Dave Chinner9bf729c2010-04-29 09:55:50 +1000533 pag->pag_ici_reclaimable++;
Christoph Hellwigbc990f52009-08-16 20:36:34 -0400534}
535
David Chinner11654512008-10-30 17:37:49 +1100536/*
537 * We set the inode flag atomically with the radix tree tag.
538 * Once we get tag lookups on the radix tree, this inode flag
539 * can go away.
540 */
David Chinner396beb82008-10-30 17:37:26 +1100541void
542xfs_inode_set_reclaim_tag(
543 xfs_inode_t *ip)
544{
Dave Chinner5017e972010-01-11 11:47:40 +0000545 struct xfs_mount *mp = ip->i_mount;
546 struct xfs_perag *pag;
David Chinner396beb82008-10-30 17:37:26 +1100547
Dave Chinner5017e972010-01-11 11:47:40 +0000548 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
Dave Chinner1a427ab2010-12-16 17:08:41 +1100549 spin_lock(&pag->pag_ici_lock);
David Chinner396beb82008-10-30 17:37:26 +1100550 spin_lock(&ip->i_flags_lock);
Christoph Hellwigbc990f52009-08-16 20:36:34 -0400551 __xfs_inode_set_reclaim_tag(pag, ip);
David Chinner11654512008-10-30 17:37:49 +1100552 __xfs_iflags_set(ip, XFS_IRECLAIMABLE);
David Chinner396beb82008-10-30 17:37:26 +1100553 spin_unlock(&ip->i_flags_lock);
Dave Chinner1a427ab2010-12-16 17:08:41 +1100554 spin_unlock(&pag->pag_ici_lock);
Dave Chinner5017e972010-01-11 11:47:40 +0000555 xfs_perag_put(pag);
David Chinner396beb82008-10-30 17:37:26 +1100556}
557
Johannes Weiner081003f2010-10-01 07:43:54 +0000558STATIC void
559__xfs_inode_clear_reclaim(
David Chinner396beb82008-10-30 17:37:26 +1100560 xfs_perag_t *pag,
561 xfs_inode_t *ip)
562{
Dave Chinner9bf729c2010-04-29 09:55:50 +1000563 pag->pag_ici_reclaimable--;
Dave Chinner16fd5362010-07-20 09:43:39 +1000564 if (!pag->pag_ici_reclaimable) {
565 /* clear the reclaim tag from the perag radix tree */
566 spin_lock(&ip->i_mount->m_perag_lock);
567 radix_tree_tag_clear(&ip->i_mount->m_perag_tree,
568 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
569 XFS_ICI_RECLAIM_TAG);
570 spin_unlock(&ip->i_mount->m_perag_lock);
571 trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno,
572 -1, _RET_IP_);
573 }
David Chinner396beb82008-10-30 17:37:26 +1100574}
575
Johannes Weiner081003f2010-10-01 07:43:54 +0000576void
577__xfs_inode_clear_reclaim_tag(
578 xfs_mount_t *mp,
579 xfs_perag_t *pag,
580 xfs_inode_t *ip)
581{
582 radix_tree_tag_clear(&pag->pag_ici_root,
583 XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
584 __xfs_inode_clear_reclaim(pag, ip);
585}
586
Dave Chinner777df5a2010-02-06 12:37:26 +1100587/*
Dave Chinnere3a20c02010-09-24 19:51:50 +1000588 * Grab the inode for reclaim exclusively.
589 * Return 0 if we grabbed it, non-zero otherwise.
590 */
591STATIC int
592xfs_reclaim_inode_grab(
593 struct xfs_inode *ip,
594 int flags)
595{
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100596 ASSERT(rcu_read_lock_held());
597
598 /* quick check for stale RCU freed inode */
599 if (!ip->i_ino)
600 return 1;
Dave Chinnere3a20c02010-09-24 19:51:50 +1000601
602 /*
Christoph Hellwig474fce02011-12-18 20:00:09 +0000603 * If we are asked for non-blocking operation, do unlocked checks to
604 * see if the inode already is being flushed or in reclaim to avoid
605 * lock traffic.
Dave Chinnere3a20c02010-09-24 19:51:50 +1000606 */
607 if ((flags & SYNC_TRYLOCK) &&
Christoph Hellwig474fce02011-12-18 20:00:09 +0000608 __xfs_iflags_test(ip, XFS_IFLOCK | XFS_IRECLAIM))
Dave Chinnere3a20c02010-09-24 19:51:50 +1000609 return 1;
Dave Chinnere3a20c02010-09-24 19:51:50 +1000610
611 /*
612 * The radix tree lock here protects a thread in xfs_iget from racing
613 * with us starting reclaim on the inode. Once we have the
614 * XFS_IRECLAIM flag set it will not touch us.
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100615 *
616 * Due to RCU lookup, we may find inodes that have been freed and only
617 * have XFS_IRECLAIM set. Indeed, we may see reallocated inodes that
618 * aren't candidates for reclaim at all, so we must check the
619 * XFS_IRECLAIMABLE is set first before proceeding to reclaim.
Dave Chinnere3a20c02010-09-24 19:51:50 +1000620 */
621 spin_lock(&ip->i_flags_lock);
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100622 if (!__xfs_iflags_test(ip, XFS_IRECLAIMABLE) ||
623 __xfs_iflags_test(ip, XFS_IRECLAIM)) {
624 /* not a reclaim candidate. */
Dave Chinnere3a20c02010-09-24 19:51:50 +1000625 spin_unlock(&ip->i_flags_lock);
626 return 1;
627 }
628 __xfs_iflags_set(ip, XFS_IRECLAIM);
629 spin_unlock(&ip->i_flags_lock);
630 return 0;
631}
632
633/*
Dave Chinner777df5a2010-02-06 12:37:26 +1100634 * Inodes in different states need to be treated differently, and the return
635 * value of xfs_iflush is not sufficient to get this right. The following table
636 * lists the inode states and the reclaim actions necessary for non-blocking
637 * reclaim:
638 *
639 *
640 * inode state iflush ret required action
641 * --------------- ---------- ---------------
642 * bad - reclaim
643 * shutdown EIO unpin and reclaim
644 * clean, unpinned 0 reclaim
645 * stale, unpinned 0 reclaim
Dave Chinnerc8543632010-02-06 12:39:36 +1100646 * clean, pinned(*) 0 requeue
647 * stale, pinned EAGAIN requeue
648 * dirty, delwri ok 0 requeue
649 * dirty, delwri blocked EAGAIN requeue
650 * dirty, sync flush 0 reclaim
Dave Chinner777df5a2010-02-06 12:37:26 +1100651 *
652 * (*) dgc: I don't think the clean, pinned state is possible but it gets
653 * handled anyway given the order of checks implemented.
654 *
Dave Chinnerc8543632010-02-06 12:39:36 +1100655 * As can be seen from the table, the return value of xfs_iflush() is not
656 * sufficient to correctly decide the reclaim action here. The checks in
657 * xfs_iflush() might look like duplicates, but they are not.
658 *
659 * Also, because we get the flush lock first, we know that any inode that has
660 * been flushed delwri has had the flush completed by the time we check that
661 * the inode is clean. The clean inode check needs to be done before flushing
662 * the inode delwri otherwise we would loop forever requeuing clean inodes as
663 * we cannot tell apart a successful delwri flush and a clean inode from the
664 * return value of xfs_iflush().
665 *
666 * Note that because the inode is flushed delayed write by background
667 * writeback, the flush lock may already be held here and waiting on it can
668 * result in very long latencies. Hence for sync reclaims, where we wait on the
669 * flush lock, the caller should push out delayed write inodes first before
670 * trying to reclaim them to minimise the amount of time spent waiting. For
671 * background relaim, we just requeue the inode for the next pass.
672 *
Dave Chinner777df5a2010-02-06 12:37:26 +1100673 * Hence the order of actions after gaining the locks should be:
674 * bad => reclaim
675 * shutdown => unpin and reclaim
Dave Chinnerc8543632010-02-06 12:39:36 +1100676 * pinned, delwri => requeue
677 * pinned, sync => unpin
Dave Chinner777df5a2010-02-06 12:37:26 +1100678 * stale => reclaim
679 * clean => reclaim
Dave Chinnerc8543632010-02-06 12:39:36 +1100680 * dirty, delwri => flush and requeue
681 * dirty, sync => flush, wait and reclaim
Dave Chinner777df5a2010-02-06 12:37:26 +1100682 */
Dave Chinner75f3cb12009-06-08 15:35:14 +0200683STATIC int
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000684xfs_reclaim_inode(
Dave Chinner75f3cb12009-06-08 15:35:14 +0200685 struct xfs_inode *ip,
686 struct xfs_perag *pag,
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000687 int sync_mode)
David Chinner7a3be022008-10-30 17:37:37 +1100688{
Dave Chinner1bfd8d02011-03-26 09:13:55 +1100689 int error;
Dave Chinner777df5a2010-02-06 12:37:26 +1100690
Dave Chinner1bfd8d02011-03-26 09:13:55 +1100691restart:
692 error = 0;
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000693 xfs_ilock(ip, XFS_ILOCK_EXCL);
Dave Chinnerc8543632010-02-06 12:39:36 +1100694 if (!xfs_iflock_nowait(ip)) {
695 if (!(sync_mode & SYNC_WAIT))
696 goto out;
Christoph Hellwig4dd2cb42011-11-29 12:06:14 -0600697
698 /*
699 * If we only have a single dirty inode in a cluster there is
700 * a fair chance that the AIL push may have pushed it into
701 * the buffer, but xfsbufd won't touch it until 30 seconds
702 * from now, and thus we will lock up here.
703 *
704 * Promote the inode buffer to the front of the delwri list
705 * and wake up xfsbufd now.
706 */
707 xfs_promote_inode(ip);
Dave Chinnerc8543632010-02-06 12:39:36 +1100708 xfs_iflock(ip);
709 }
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000710
Dave Chinner777df5a2010-02-06 12:37:26 +1100711 if (is_bad_inode(VFS_I(ip)))
712 goto reclaim;
713 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
714 xfs_iunpin_wait(ip);
Christoph Hellwig32ce90a2012-04-23 15:58:32 +1000715 xfs_iflush_abort(ip);
Dave Chinner777df5a2010-02-06 12:37:26 +1100716 goto reclaim;
717 }
Dave Chinnerc8543632010-02-06 12:39:36 +1100718 if (xfs_ipincount(ip)) {
719 if (!(sync_mode & SYNC_WAIT)) {
720 xfs_ifunlock(ip);
721 goto out;
722 }
Dave Chinner777df5a2010-02-06 12:37:26 +1100723 xfs_iunpin_wait(ip);
Dave Chinnerc8543632010-02-06 12:39:36 +1100724 }
Dave Chinner777df5a2010-02-06 12:37:26 +1100725 if (xfs_iflags_test(ip, XFS_ISTALE))
726 goto reclaim;
727 if (xfs_inode_clean(ip))
728 goto reclaim;
729
Dave Chinner1bfd8d02011-03-26 09:13:55 +1100730 /*
731 * Now we have an inode that needs flushing.
732 *
733 * We do a nonblocking flush here even if we are doing a SYNC_WAIT
734 * reclaim as we can deadlock with inode cluster removal.
735 * xfs_ifree_cluster() can lock the inode buffer before it locks the
736 * ip->i_lock, and we are doing the exact opposite here. As a result,
737 * doing a blocking xfs_itobp() to get the cluster buffer will result
738 * in an ABBA deadlock with xfs_ifree_cluster().
739 *
740 * As xfs_ifree_cluser() must gather all inodes that are active in the
741 * cache to mark them stale, if we hit this case we don't actually want
742 * to do IO here - we want the inode marked stale so we can simply
743 * reclaim it. Hence if we get an EAGAIN error on a SYNC_WAIT flush,
744 * just unlock the inode, back off and try again. Hopefully the next
745 * pass through will see the stale flag set on the inode.
746 */
747 error = xfs_iflush(ip, SYNC_TRYLOCK | sync_mode);
Dave Chinnerc8543632010-02-06 12:39:36 +1100748 if (sync_mode & SYNC_WAIT) {
Dave Chinner1bfd8d02011-03-26 09:13:55 +1100749 if (error == EAGAIN) {
750 xfs_iunlock(ip, XFS_ILOCK_EXCL);
751 /* backoff longer than in xfs_ifree_cluster */
752 delay(2);
753 goto restart;
754 }
Dave Chinnerc8543632010-02-06 12:39:36 +1100755 xfs_iflock(ip);
756 goto reclaim;
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000757 }
758
Dave Chinnerc8543632010-02-06 12:39:36 +1100759 /*
760 * When we have to flush an inode but don't have SYNC_WAIT set, we
761 * flush the inode out using a delwri buffer and wait for the next
762 * call into reclaim to find it in a clean state instead of waiting for
763 * it now. We also don't return errors here - if the error is transient
764 * then the next reclaim pass will flush the inode, and if the error
Dave Chinnerf1d486a2010-04-13 15:06:45 +1000765 * is permanent then the next sync reclaim will reclaim the inode and
Dave Chinnerc8543632010-02-06 12:39:36 +1100766 * pass on the error.
767 */
Dave Chinnerf1d486a2010-04-13 15:06:45 +1000768 if (error && error != EAGAIN && !XFS_FORCED_SHUTDOWN(ip->i_mount)) {
Dave Chinner4f107002011-03-07 10:00:35 +1100769 xfs_warn(ip->i_mount,
Dave Chinnerc8543632010-02-06 12:39:36 +1100770 "inode 0x%llx background reclaim flush failed with %d",
771 (long long)ip->i_ino, error);
772 }
773out:
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;
784
Dave Chinner777df5a2010-02-06 12:37:26 +1100785reclaim:
786 xfs_ifunlock(ip);
Dave Chinnerc8e20be2010-01-10 23:51:45 +0000787 xfs_iunlock(ip, XFS_ILOCK_EXCL);
Dave Chinner2f11fea2010-07-20 17:53:25 +1000788
789 XFS_STATS_INC(xs_ig_reclaims);
790 /*
791 * Remove the inode from the per-AG radix tree.
792 *
793 * Because radix_tree_delete won't complain even if the item was never
794 * added to the tree assert that it's been there before to catch
795 * problems with the inode life time early on.
796 */
Dave Chinner1a427ab2010-12-16 17:08:41 +1100797 spin_lock(&pag->pag_ici_lock);
Dave Chinner2f11fea2010-07-20 17:53:25 +1000798 if (!radix_tree_delete(&pag->pag_ici_root,
799 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino)))
800 ASSERT(0);
Johannes Weiner081003f2010-10-01 07:43:54 +0000801 __xfs_inode_clear_reclaim(pag, ip);
Dave Chinner1a427ab2010-12-16 17:08:41 +1100802 spin_unlock(&pag->pag_ici_lock);
Dave Chinner2f11fea2010-07-20 17:53:25 +1000803
804 /*
805 * Here we do an (almost) spurious inode lock in order to coordinate
806 * with inode cache radix tree lookups. This is because the lookup
807 * can reference the inodes in the cache without taking references.
808 *
809 * We make that OK here by ensuring that we wait until the inode is
Alex Elderad637a12012-02-16 22:01:00 +0000810 * unlocked after the lookup before we go ahead and free it.
Dave Chinner2f11fea2010-07-20 17:53:25 +1000811 */
Alex Elderad637a12012-02-16 22:01:00 +0000812 xfs_ilock(ip, XFS_ILOCK_EXCL);
Dave Chinner2f11fea2010-07-20 17:53:25 +1000813 xfs_qm_dqdetach(ip);
Alex Elderad637a12012-02-16 22:01:00 +0000814 xfs_iunlock(ip, XFS_ILOCK_EXCL);
Dave Chinner2f11fea2010-07-20 17:53:25 +1000815
816 xfs_inode_free(ip);
Dave Chinnerc8543632010-02-06 12:39:36 +1100817
Alex Elderad637a12012-02-16 22:01:00 +0000818 return error;
David Chinner7a3be022008-10-30 17:37:37 +1100819}
820
Dave Chinner65d0f202010-09-24 18:40:15 +1000821/*
822 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
823 * corrupted, we still want to try to reclaim all the inodes. If we don't,
824 * then a shut down during filesystem unmount reclaim walk leak all the
825 * unreclaimed inodes.
826 */
827int
828xfs_reclaim_inodes_ag(
829 struct xfs_mount *mp,
830 int flags,
831 int *nr_to_scan)
832{
833 struct xfs_perag *pag;
834 int error = 0;
835 int last_error = 0;
836 xfs_agnumber_t ag;
Dave Chinner69b491c2010-09-27 11:09:51 +1000837 int trylock = flags & SYNC_TRYLOCK;
838 int skipped;
Dave Chinner65d0f202010-09-24 18:40:15 +1000839
Dave Chinner69b491c2010-09-27 11:09:51 +1000840restart:
Dave Chinner65d0f202010-09-24 18:40:15 +1000841 ag = 0;
Dave Chinner69b491c2010-09-27 11:09:51 +1000842 skipped = 0;
Dave Chinner65d0f202010-09-24 18:40:15 +1000843 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
844 unsigned long first_index = 0;
845 int done = 0;
Dave Chinnere3a20c02010-09-24 19:51:50 +1000846 int nr_found = 0;
Dave Chinner65d0f202010-09-24 18:40:15 +1000847
848 ag = pag->pag_agno + 1;
849
Dave Chinner69b491c2010-09-27 11:09:51 +1000850 if (trylock) {
851 if (!mutex_trylock(&pag->pag_ici_reclaim_lock)) {
852 skipped++;
Dave Chinnerf83282a2010-11-08 08:55:04 +0000853 xfs_perag_put(pag);
Dave Chinner69b491c2010-09-27 11:09:51 +1000854 continue;
855 }
856 first_index = pag->pag_ici_reclaim_cursor;
857 } else
858 mutex_lock(&pag->pag_ici_reclaim_lock);
859
Dave Chinner65d0f202010-09-24 18:40:15 +1000860 do {
Dave Chinnere3a20c02010-09-24 19:51:50 +1000861 struct xfs_inode *batch[XFS_LOOKUP_BATCH];
862 int i;
Dave Chinner65d0f202010-09-24 18:40:15 +1000863
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100864 rcu_read_lock();
Dave Chinnere3a20c02010-09-24 19:51:50 +1000865 nr_found = radix_tree_gang_lookup_tag(
866 &pag->pag_ici_root,
867 (void **)batch, first_index,
868 XFS_LOOKUP_BATCH,
Dave Chinner65d0f202010-09-24 18:40:15 +1000869 XFS_ICI_RECLAIM_TAG);
870 if (!nr_found) {
Dave Chinnerb2232212011-05-06 02:54:04 +0000871 done = 1;
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100872 rcu_read_unlock();
Dave Chinner65d0f202010-09-24 18:40:15 +1000873 break;
874 }
875
876 /*
Dave Chinnere3a20c02010-09-24 19:51:50 +1000877 * Grab the inodes before we drop the lock. if we found
878 * nothing, nr == 0 and the loop will be skipped.
Dave Chinner65d0f202010-09-24 18:40:15 +1000879 */
Dave Chinnere3a20c02010-09-24 19:51:50 +1000880 for (i = 0; i < nr_found; i++) {
881 struct xfs_inode *ip = batch[i];
Dave Chinner65d0f202010-09-24 18:40:15 +1000882
Dave Chinnere3a20c02010-09-24 19:51:50 +1000883 if (done || xfs_reclaim_inode_grab(ip, flags))
884 batch[i] = NULL;
Dave Chinner65d0f202010-09-24 18:40:15 +1000885
Dave Chinnere3a20c02010-09-24 19:51:50 +1000886 /*
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.
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100892 *
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.
Dave Chinnere3a20c02010-09-24 19:51:50 +1000899 */
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100900 if (XFS_INO_TO_AGNO(mp, ip->i_ino) !=
901 pag->pag_agno)
902 continue;
Dave Chinnere3a20c02010-09-24 19:51:50 +1000903 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 }
907
908 /* unlock now we've grabbed the inodes. */
Dave Chinner1a3e8f32010-12-17 17:29:43 +1100909 rcu_read_unlock();
Dave Chinnere3a20c02010-09-24 19:51:50 +1000910
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;
920
Dave Chinner8daaa832011-07-08 14:14:46 +1000921 cond_resched();
922
Dave Chinnere3a20c02010-09-24 19:51:50 +1000923 } while (nr_found && !done && *nr_to_scan > 0);
Dave Chinner65d0f202010-09-24 18:40:15 +1000924
Dave Chinner69b491c2010-09-27 11:09:51 +1000925 if (trylock && !done)
926 pag->pag_ici_reclaim_cursor = first_index;
927 else
928 pag->pag_ici_reclaim_cursor = 0;
929 mutex_unlock(&pag->pag_ici_reclaim_lock);
Dave Chinner65d0f202010-09-24 18:40:15 +1000930 xfs_perag_put(pag);
931 }
Dave Chinner69b491c2010-09-27 11:09:51 +1000932
933 /*
934 * if we skipped any AG, and we still have scan count remaining, do
935 * another pass this time using blocking reclaim semantics (i.e
936 * waiting on the reclaim locks and ignoring the reclaim cursors). This
937 * ensure that when we get more reclaimers than AGs we block rather
938 * than spin trying to execute reclaim.
939 */
Dave Chinner8daaa832011-07-08 14:14:46 +1000940 if (skipped && (flags & SYNC_WAIT) && *nr_to_scan > 0) {
Dave Chinner69b491c2010-09-27 11:09:51 +1000941 trylock = 0;
942 goto restart;
943 }
Dave Chinner65d0f202010-09-24 18:40:15 +1000944 return XFS_ERROR(last_error);
945}
946
David Chinnerfce08f22008-10-30 17:37:03 +1100947int
David Chinner1dc33182008-10-30 17:37:15 +1100948xfs_reclaim_inodes(
David Chinnerfce08f22008-10-30 17:37:03 +1100949 xfs_mount_t *mp,
David Chinnerfce08f22008-10-30 17:37:03 +1100950 int mode)
951{
Dave Chinner65d0f202010-09-24 18:40:15 +1000952 int nr_to_scan = INT_MAX;
953
954 return xfs_reclaim_inodes_ag(mp, mode, &nr_to_scan);
Dave Chinner9bf729c2010-04-29 09:55:50 +1000955}
956
957/*
Dave Chinner8daaa832011-07-08 14:14:46 +1000958 * Scan a certain number of inodes for reclaim.
Dave Chinnera7b339f2011-04-08 12:45:07 +1000959 *
960 * When called we make sure that there is a background (fast) inode reclaim in
Dave Chinner8daaa832011-07-08 14:14:46 +1000961 * progress, while we will throttle the speed of reclaim via doing synchronous
Dave Chinnera7b339f2011-04-08 12:45:07 +1000962 * reclaim of inodes. That means if we come across dirty inodes, we wait for
963 * them to be cleaned, which we hope will not be very long due to the
964 * background walker having already kicked the IO off on those dirty inodes.
Dave Chinner9bf729c2010-04-29 09:55:50 +1000965 */
Dave Chinner8daaa832011-07-08 14:14:46 +1000966void
967xfs_reclaim_inodes_nr(
968 struct xfs_mount *mp,
969 int nr_to_scan)
Dave Chinner9bf729c2010-04-29 09:55:50 +1000970{
Dave Chinner8daaa832011-07-08 14:14:46 +1000971 /* kick background reclaimer and push the AIL */
972 xfs_syncd_queue_reclaim(mp);
973 xfs_ail_push_all(mp->m_ail);
Dave Chinner9bf729c2010-04-29 09:55:50 +1000974
Dave Chinner8daaa832011-07-08 14:14:46 +1000975 xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
976}
Dave Chinnera7b339f2011-04-08 12:45:07 +1000977
Dave Chinner8daaa832011-07-08 14:14:46 +1000978/*
979 * Return the number of reclaimable inodes in the filesystem for
980 * the shrinker to determine how much to reclaim.
981 */
982int
983xfs_reclaim_inodes_count(
984 struct xfs_mount *mp)
985{
986 struct xfs_perag *pag;
987 xfs_agnumber_t ag = 0;
988 int reclaimable = 0;
Dave Chinner9bf729c2010-04-29 09:55:50 +1000989
Dave Chinner65d0f202010-09-24 18:40:15 +1000990 while ((pag = xfs_perag_get_tag(mp, ag, XFS_ICI_RECLAIM_TAG))) {
991 ag = pag->pag_agno + 1;
Dave Chinner70e60ce2010-07-20 08:07:02 +1000992 reclaimable += pag->pag_ici_reclaimable;
993 xfs_perag_put(pag);
Dave Chinner9bf729c2010-04-29 09:55:50 +1000994 }
Dave Chinner9bf729c2010-04-29 09:55:50 +1000995 return reclaimable;
996}
997