[XFS] Use the inode tree for finding dirty inodes
[linux-2.6.git] / fs / xfs / linux-2.6 / xfs_sync.c
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_bmap_btree.h"
31 #include "xfs_alloc_btree.h"
32 #include "xfs_ialloc_btree.h"
33 #include "xfs_btree.h"
34 #include "xfs_dir2_sf.h"
35 #include "xfs_attr_sf.h"
36 #include "xfs_inode.h"
37 #include "xfs_dinode.h"
38 #include "xfs_error.h"
39 #include "xfs_mru_cache.h"
40 #include "xfs_filestream.h"
41 #include "xfs_vnodeops.h"
42 #include "xfs_utils.h"
43 #include "xfs_buf_item.h"
44 #include "xfs_inode_item.h"
45 #include "xfs_rw.h"
46
47 #include <linux/kthread.h>
48 #include <linux/freezer.h>
49
50 /*
51  * xfs_sync flushes any pending I/O to file system vfsp.
52  *
53  * This routine is called by vfs_sync() to make sure that things make it
54  * out to disk eventually, on sync() system calls to flush out everything,
55  * and when the file system is unmounted.  For the vfs_sync() case, all
56  * we really need to do is sync out the log to make all of our meta-data
57  * updates permanent (except for timestamps).  For calls from pflushd(),
58  * dirty pages are kept moving by calling pdflush() on the inodes
59  * containing them.  We also flush the inodes that we can lock without
60  * sleeping and the superblock if we can lock it without sleeping from
61  * vfs_sync() so that items at the tail of the log are always moving out.
62  *
63  * Flags:
64  *      SYNC_BDFLUSH - We're being called from vfs_sync() so we don't want
65  *                     to sleep if we can help it.  All we really need
66  *                     to do is ensure that the log is synced at least
67  *                     periodically.  We also push the inodes and
68  *                     superblock if we can lock them without sleeping
69  *                      and they are not pinned.
70  *      SYNC_ATTR    - We need to flush the inodes.  If SYNC_BDFLUSH is not
71  *                     set, then we really want to lock each inode and flush
72  *                     it.
73  *      SYNC_WAIT    - All the flushes that take place in this call should
74  *                     be synchronous.
75  *      SYNC_DELWRI  - This tells us to push dirty pages associated with
76  *                     inodes.  SYNC_WAIT and SYNC_BDFLUSH are used to
77  *                     determine if they should be flushed sync, async, or
78  *                     delwri.
79  *      SYNC_CLOSE   - This flag is passed when the system is being
80  *                     unmounted.  We should sync and invalidate everything.
81  *      SYNC_FSDATA  - This indicates that the caller would like to make
82  *                     sure the superblock is safe on disk.  We can ensure
83  *                     this by simply making sure the log gets flushed
84  *                     if SYNC_BDFLUSH is set, and by actually writing it
85  *                     out otherwise.
86  *      SYNC_IOWAIT  - The caller wants us to wait for all data I/O to complete
87  *                     before we return (including direct I/O). Forms the drain
88  *                     side of the write barrier needed to safely quiesce the
89  *                     filesystem.
90  *
91  */
92 int
93 xfs_sync(
94         xfs_mount_t     *mp,
95         int             flags)
96 {
97         int             error;
98
99         /*
100          * Get the Quota Manager to flush the dquots.
101          *
102          * If XFS quota support is not enabled or this filesystem
103          * instance does not use quotas XFS_QM_DQSYNC will always
104          * return zero.
105          */
106         error = XFS_QM_DQSYNC(mp, flags);
107         if (error) {
108                 /*
109                  * If we got an IO error, we will be shutting down.
110                  * So, there's nothing more for us to do here.
111                  */
112                 ASSERT(error != EIO || XFS_FORCED_SHUTDOWN(mp));
113                 if (XFS_FORCED_SHUTDOWN(mp))
114                         return XFS_ERROR(error);
115         }
116
117         if (flags & SYNC_IOWAIT)
118                 xfs_filestream_flush(mp);
119
120         return xfs_syncsub(mp, flags, NULL);
121 }
122
123 /*
124  * Sync all the inodes in the given AG according to the
125  * direction given by the flags.
126  */
127 STATIC int
128 xfs_sync_inodes_ag(
129         xfs_mount_t     *mp,
130         int             ag,
131         int             flags,
132         int             *bypassed)
133 {
134         xfs_inode_t     *ip = NULL;
135         struct inode    *vp = NULL;
136         xfs_perag_t     *pag = &mp->m_perag[ag];
137         boolean_t       vnode_refed = B_FALSE;
138         int             nr_found;
139         int             first_index = 0;
140         int             error = 0;
141         int             last_error = 0;
142         int             fflag = XFS_B_ASYNC;
143         int             lock_flags = XFS_ILOCK_SHARED;
144
145         if (flags & SYNC_DELWRI)
146                 fflag = XFS_B_DELWRI;
147         if (flags & SYNC_WAIT)
148                 fflag = 0;              /* synchronous overrides all */
149
150         if (flags & (SYNC_DELWRI | SYNC_CLOSE)) {
151                 /*
152                  * We need the I/O lock if we're going to call any of
153                  * the flush/inval routines.
154                  */
155                 lock_flags |= XFS_IOLOCK_SHARED;
156         }
157
158         do {
159                 /*
160                  * use a gang lookup to find the next inode in the tree
161                  * as the tree is sparse and a gang lookup walks to find
162                  * the number of objects requested.
163                  */
164                 read_lock(&pag->pag_ici_lock);
165                 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root,
166                                 (void**)&ip, first_index, 1);
167
168                 if (!nr_found) {
169                         read_unlock(&pag->pag_ici_lock);
170                         break;
171                 }
172
173                 /* update the index for the next lookup */
174                 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino + 1);
175
176                 /*
177                  * skip inodes in reclaim. Let xfs_syncsub do that for
178                  * us so we don't need to worry.
179                  */
180                 vp = VFS_I(ip);
181                 if (!vp) {
182                         read_unlock(&pag->pag_ici_lock);
183                         continue;
184                 }
185
186                 /* bad inodes are dealt with elsewhere */
187                 if (VN_BAD(vp)) {
188                         read_unlock(&pag->pag_ici_lock);
189                         continue;
190                 }
191
192                 /* nothing to sync during shutdown */
193                 if (XFS_FORCED_SHUTDOWN(mp) && !(flags & SYNC_CLOSE)) {
194                         read_unlock(&pag->pag_ici_lock);
195                         return 0;
196                 }
197
198                 /*
199                  * The inode lock here actually coordinates with the almost
200                  * spurious inode lock in xfs_ireclaim() to prevent the vnode
201                  * we handle here without a reference from being freed while we
202                  * reference it.  If we lock the inode while it's on the mount
203                  * list here, then the spurious inode lock in xfs_ireclaim()
204                  * after the inode is pulled from the mount list will sleep
205                  * until we release it here.  This keeps the vnode from being
206                  * freed while we reference it.
207                  */
208                 if (xfs_ilock_nowait(ip, lock_flags) == 0) {
209                         vp = vn_grab(vp);
210                         read_unlock(&pag->pag_ici_lock);
211                         if (!vp)
212                                 continue;
213                         xfs_ilock(ip, lock_flags);
214
215                         ASSERT(vp == VFS_I(ip));
216                         ASSERT(ip->i_mount == mp);
217
218                         vnode_refed = B_TRUE;
219                 } else {
220                         /* safe to unlock here as we have a reference */
221                         read_unlock(&pag->pag_ici_lock);
222                 }
223                 /*
224                  * If we have to flush data or wait for I/O completion
225                  * we need to drop the ilock that we currently hold.
226                  * If we need to drop the lock, insert a marker if we
227                  * have not already done so.
228                  */
229                 if (flags & SYNC_CLOSE) {
230                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
231                         if (XFS_FORCED_SHUTDOWN(mp))
232                                 xfs_tosspages(ip, 0, -1, FI_REMAPF);
233                         else
234                                 error = xfs_flushinval_pages(ip, 0, -1,
235                                                         FI_REMAPF);
236                         /* wait for I/O on freeze */
237                         if (flags & SYNC_IOWAIT)
238                                 vn_iowait(ip);
239
240                         xfs_ilock(ip, XFS_ILOCK_SHARED);
241                 }
242
243                 if ((flags & SYNC_DELWRI) && VN_DIRTY(vp)) {
244                         xfs_iunlock(ip, XFS_ILOCK_SHARED);
245                         error = xfs_flush_pages(ip, 0, -1, fflag, FI_NONE);
246                         if (flags & SYNC_IOWAIT)
247                                 vn_iowait(ip);
248                         xfs_ilock(ip, XFS_ILOCK_SHARED);
249                 }
250
251                 if ((flags & SYNC_ATTR) && !xfs_inode_clean(ip)) {
252                         if (flags & SYNC_WAIT) {
253                                 xfs_iflock(ip);
254                                 if (!xfs_inode_clean(ip))
255                                         error = xfs_iflush(ip, XFS_IFLUSH_SYNC);
256                                 else
257                                         xfs_ifunlock(ip);
258                         } else if (xfs_iflock_nowait(ip)) {
259                                 if (!xfs_inode_clean(ip))
260                                         error = xfs_iflush(ip, XFS_IFLUSH_DELWRI);
261                                 else
262                                         xfs_ifunlock(ip);
263                         } else if (bypassed) {
264                                 (*bypassed)++;
265                         }
266                 }
267
268                 if (lock_flags)
269                         xfs_iunlock(ip, lock_flags);
270
271                 if (vnode_refed) {
272                         IRELE(ip);
273                         vnode_refed = B_FALSE;
274                 }
275
276                 if (error)
277                         last_error = error;
278                 /*
279                  * bail out if the filesystem is corrupted.
280                  */
281                 if (error == EFSCORRUPTED)
282                         return XFS_ERROR(error);
283
284         } while (nr_found);
285
286         return last_error;
287 }
288
289 int
290 xfs_sync_inodes(
291         xfs_mount_t     *mp,
292         int             flags,
293         int             *bypassed)
294 {
295         int             error;
296         int             last_error;
297         int             i;
298
299         if (bypassed)
300                 *bypassed = 0;
301         if (mp->m_flags & XFS_MOUNT_RDONLY)
302                 return 0;
303         error = 0;
304         last_error = 0;
305
306         for (i = 0; i < mp->m_sb.sb_agcount; i++) {
307                 if (!mp->m_perag[i].pag_ici_init)
308                         continue;
309                 error = xfs_sync_inodes_ag(mp, i, flags, bypassed);
310                 if (error)
311                         last_error = error;
312                 if (error == EFSCORRUPTED)
313                         break;
314         }
315         return XFS_ERROR(last_error);
316 }
317
318 /*
319  * xfs sync routine for internal use
320  *
321  * This routine supports all of the flags defined for the generic vfs_sync
322  * interface as explained above under xfs_sync.
323  *
324  */
325 int
326 xfs_syncsub(
327         xfs_mount_t     *mp,
328         int             flags,
329         int             *bypassed)
330 {
331         int             error = 0;
332         int             last_error = 0;
333         uint            log_flags = XFS_LOG_FORCE;
334         xfs_buf_t       *bp;
335         xfs_buf_log_item_t      *bip;
336
337         /*
338          * Sync out the log.  This ensures that the log is periodically
339          * flushed even if there is not enough activity to fill it up.
340          */
341         if (flags & SYNC_WAIT)
342                 log_flags |= XFS_LOG_SYNC;
343
344         xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
345
346         if (flags & (SYNC_ATTR|SYNC_DELWRI)) {
347                 if (flags & SYNC_BDFLUSH)
348                         xfs_finish_reclaim_all(mp, 1, XFS_IFLUSH_DELWRI_ELSE_ASYNC);
349                 else
350                         error = xfs_sync_inodes(mp, flags, bypassed);
351         }
352
353         /*
354          * Flushing out dirty data above probably generated more
355          * log activity, so if this isn't vfs_sync() then flush
356          * the log again.
357          */
358         if (flags & SYNC_DELWRI) {
359                 xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
360         }
361
362         if (flags & SYNC_FSDATA) {
363                 /*
364                  * If this is vfs_sync() then only sync the superblock
365                  * if we can lock it without sleeping and it is not pinned.
366                  */
367                 if (flags & SYNC_BDFLUSH) {
368                         bp = xfs_getsb(mp, XFS_BUF_TRYLOCK);
369                         if (bp != NULL) {
370                                 bip = XFS_BUF_FSPRIVATE(bp,xfs_buf_log_item_t*);
371                                 if ((bip != NULL) &&
372                                     xfs_buf_item_dirty(bip)) {
373                                         if (!(XFS_BUF_ISPINNED(bp))) {
374                                                 XFS_BUF_ASYNC(bp);
375                                                 error = xfs_bwrite(mp, bp);
376                                         } else {
377                                                 xfs_buf_relse(bp);
378                                         }
379                                 } else {
380                                         xfs_buf_relse(bp);
381                                 }
382                         }
383                 } else {
384                         bp = xfs_getsb(mp, 0);
385                         /*
386                          * If the buffer is pinned then push on the log so
387                          * we won't get stuck waiting in the write for
388                          * someone, maybe ourselves, to flush the log.
389                          * Even though we just pushed the log above, we
390                          * did not have the superblock buffer locked at
391                          * that point so it can become pinned in between
392                          * there and here.
393                          */
394                         if (XFS_BUF_ISPINNED(bp))
395                                 xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
396                         if (flags & SYNC_WAIT)
397                                 XFS_BUF_UNASYNC(bp);
398                         else
399                                 XFS_BUF_ASYNC(bp);
400                         error = xfs_bwrite(mp, bp);
401                 }
402                 if (error) {
403                         last_error = error;
404                 }
405         }
406
407         /*
408          * Now check to see if the log needs a "dummy" transaction.
409          */
410         if (!(flags & SYNC_REMOUNT) && xfs_log_need_covered(mp)) {
411                 xfs_trans_t *tp;
412                 xfs_inode_t *ip;
413
414                 /*
415                  * Put a dummy transaction in the log to tell
416                  * recovery that all others are OK.
417                  */
418                 tp = xfs_trans_alloc(mp, XFS_TRANS_DUMMY1);
419                 if ((error = xfs_trans_reserve(tp, 0,
420                                 XFS_ICHANGE_LOG_RES(mp),
421                                 0, 0, 0)))  {
422                         xfs_trans_cancel(tp, 0);
423                         return error;
424                 }
425
426                 ip = mp->m_rootip;
427                 xfs_ilock(ip, XFS_ILOCK_EXCL);
428
429                 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
430                 xfs_trans_ihold(tp, ip);
431                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
432                 error = xfs_trans_commit(tp, 0);
433                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
434                 xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
435         }
436
437         /*
438          * When shutting down, we need to insure that the AIL is pushed
439          * to disk or the filesystem can appear corrupt from the PROM.
440          */
441         if ((flags & (SYNC_CLOSE|SYNC_WAIT)) == (SYNC_CLOSE|SYNC_WAIT)) {
442                 XFS_bflush(mp->m_ddev_targp);
443                 if (mp->m_rtdev_targp) {
444                         XFS_bflush(mp->m_rtdev_targp);
445                 }
446         }
447
448         return XFS_ERROR(last_error);
449 }
450
451 /*
452  * Enqueue a work item to be picked up by the vfs xfssyncd thread.
453  * Doing this has two advantages:
454  * - It saves on stack space, which is tight in certain situations
455  * - It can be used (with care) as a mechanism to avoid deadlocks.
456  * Flushing while allocating in a full filesystem requires both.
457  */
458 STATIC void
459 xfs_syncd_queue_work(
460         struct xfs_mount *mp,
461         void            *data,
462         void            (*syncer)(struct xfs_mount *, void *))
463 {
464         struct bhv_vfs_sync_work *work;
465
466         work = kmem_alloc(sizeof(struct bhv_vfs_sync_work), KM_SLEEP);
467         INIT_LIST_HEAD(&work->w_list);
468         work->w_syncer = syncer;
469         work->w_data = data;
470         work->w_mount = mp;
471         spin_lock(&mp->m_sync_lock);
472         list_add_tail(&work->w_list, &mp->m_sync_list);
473         spin_unlock(&mp->m_sync_lock);
474         wake_up_process(mp->m_sync_task);
475 }
476
477 /*
478  * Flush delayed allocate data, attempting to free up reserved space
479  * from existing allocations.  At this point a new allocation attempt
480  * has failed with ENOSPC and we are in the process of scratching our
481  * heads, looking about for more room...
482  */
483 STATIC void
484 xfs_flush_inode_work(
485         struct xfs_mount *mp,
486         void            *arg)
487 {
488         struct inode    *inode = arg;
489         filemap_flush(inode->i_mapping);
490         iput(inode);
491 }
492
493 void
494 xfs_flush_inode(
495         xfs_inode_t     *ip)
496 {
497         struct inode    *inode = VFS_I(ip);
498
499         igrab(inode);
500         xfs_syncd_queue_work(ip->i_mount, inode, xfs_flush_inode_work);
501         delay(msecs_to_jiffies(500));
502 }
503
504 /*
505  * This is the "bigger hammer" version of xfs_flush_inode_work...
506  * (IOW, "If at first you don't succeed, use a Bigger Hammer").
507  */
508 STATIC void
509 xfs_flush_device_work(
510         struct xfs_mount *mp,
511         void            *arg)
512 {
513         struct inode    *inode = arg;
514         sync_blockdev(mp->m_super->s_bdev);
515         iput(inode);
516 }
517
518 void
519 xfs_flush_device(
520         xfs_inode_t     *ip)
521 {
522         struct inode    *inode = VFS_I(ip);
523
524         igrab(inode);
525         xfs_syncd_queue_work(ip->i_mount, inode, xfs_flush_device_work);
526         delay(msecs_to_jiffies(500));
527         xfs_log_force(ip->i_mount, (xfs_lsn_t)0, XFS_LOG_FORCE|XFS_LOG_SYNC);
528 }
529
530 STATIC void
531 xfs_sync_worker(
532         struct xfs_mount *mp,
533         void            *unused)
534 {
535         int             error;
536
537         if (!(mp->m_flags & XFS_MOUNT_RDONLY))
538                 error = xfs_sync(mp, SYNC_FSDATA | SYNC_BDFLUSH | SYNC_ATTR);
539         mp->m_sync_seq++;
540         wake_up(&mp->m_wait_single_sync_task);
541 }
542
543 STATIC int
544 xfssyncd(
545         void                    *arg)
546 {
547         struct xfs_mount        *mp = arg;
548         long                    timeleft;
549         bhv_vfs_sync_work_t     *work, *n;
550         LIST_HEAD               (tmp);
551
552         set_freezable();
553         timeleft = xfs_syncd_centisecs * msecs_to_jiffies(10);
554         for (;;) {
555                 timeleft = schedule_timeout_interruptible(timeleft);
556                 /* swsusp */
557                 try_to_freeze();
558                 if (kthread_should_stop() && list_empty(&mp->m_sync_list))
559                         break;
560
561                 spin_lock(&mp->m_sync_lock);
562                 /*
563                  * We can get woken by laptop mode, to do a sync -
564                  * that's the (only!) case where the list would be
565                  * empty with time remaining.
566                  */
567                 if (!timeleft || list_empty(&mp->m_sync_list)) {
568                         if (!timeleft)
569                                 timeleft = xfs_syncd_centisecs *
570                                                         msecs_to_jiffies(10);
571                         INIT_LIST_HEAD(&mp->m_sync_work.w_list);
572                         list_add_tail(&mp->m_sync_work.w_list,
573                                         &mp->m_sync_list);
574                 }
575                 list_for_each_entry_safe(work, n, &mp->m_sync_list, w_list)
576                         list_move(&work->w_list, &tmp);
577                 spin_unlock(&mp->m_sync_lock);
578
579                 list_for_each_entry_safe(work, n, &tmp, w_list) {
580                         (*work->w_syncer)(mp, work->w_data);
581                         list_del(&work->w_list);
582                         if (work == &mp->m_sync_work)
583                                 continue;
584                         kmem_free(work);
585                 }
586         }
587
588         return 0;
589 }
590
591 int
592 xfs_syncd_init(
593         struct xfs_mount        *mp)
594 {
595         mp->m_sync_work.w_syncer = xfs_sync_worker;
596         mp->m_sync_work.w_mount = mp;
597         mp->m_sync_task = kthread_run(xfssyncd, mp, "xfssyncd");
598         if (IS_ERR(mp->m_sync_task))
599                 return -PTR_ERR(mp->m_sync_task);
600         return 0;
601 }
602
603 void
604 xfs_syncd_stop(
605         struct xfs_mount        *mp)
606 {
607         kthread_stop(mp->m_sync_task);
608 }
609