Linux 2.6.28-rc5
[cris-mirror.git] / fs / xfs / xfs_vfsops.c
blob439dd3939dda043d172c5a9b0285085afc042534
1 /*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
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.
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.
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
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_da_btree.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_ialloc_btree.h"
33 #include "xfs_alloc_btree.h"
34 #include "xfs_dir2_sf.h"
35 #include "xfs_attr_sf.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
38 #include "xfs_inode_item.h"
39 #include "xfs_btree.h"
40 #include "xfs_alloc.h"
41 #include "xfs_ialloc.h"
42 #include "xfs_quota.h"
43 #include "xfs_error.h"
44 #include "xfs_bmap.h"
45 #include "xfs_rw.h"
46 #include "xfs_buf_item.h"
47 #include "xfs_log_priv.h"
48 #include "xfs_dir2_trace.h"
49 #include "xfs_extfree_item.h"
50 #include "xfs_acl.h"
51 #include "xfs_attr.h"
52 #include "xfs_clnt.h"
53 #include "xfs_mru_cache.h"
54 #include "xfs_filestream.h"
55 #include "xfs_fsops.h"
56 #include "xfs_vnodeops.h"
57 #include "xfs_vfsops.h"
58 #include "xfs_utils.h"
61 STATIC void
62 xfs_quiesce_fs(
63 xfs_mount_t *mp)
65 int count = 0, pincount;
67 xfs_flush_buftarg(mp->m_ddev_targp, 0);
68 xfs_finish_reclaim_all(mp, 0);
70 /* This loop must run at least twice.
71 * The first instance of the loop will flush
72 * most meta data but that will generate more
73 * meta data (typically directory updates).
74 * Which then must be flushed and logged before
75 * we can write the unmount record.
77 do {
78 xfs_syncsub(mp, SYNC_INODE_QUIESCE, NULL);
79 pincount = xfs_flush_buftarg(mp->m_ddev_targp, 1);
80 if (!pincount) {
81 delay(50);
82 count++;
84 } while (count < 2);
88 * Second stage of a quiesce. The data is already synced, now we have to take
89 * care of the metadata. New transactions are already blocked, so we need to
90 * wait for any remaining transactions to drain out before proceding.
92 void
93 xfs_attr_quiesce(
94 xfs_mount_t *mp)
96 int error = 0;
98 /* wait for all modifications to complete */
99 while (atomic_read(&mp->m_active_trans) > 0)
100 delay(100);
102 /* flush inodes and push all remaining buffers out to disk */
103 xfs_quiesce_fs(mp);
105 ASSERT_ALWAYS(atomic_read(&mp->m_active_trans) == 0);
107 /* Push the superblock and write an unmount record */
108 error = xfs_log_sbcount(mp, 1);
109 if (error)
110 xfs_fs_cmn_err(CE_WARN, mp,
111 "xfs_attr_quiesce: failed to log sb changes. "
112 "Frozen image may not be consistent.");
113 xfs_log_unmount_write(mp);
114 xfs_unmountfs_writesb(mp);
118 * xfs_unmount_flush implements a set of flush operation on special
119 * inodes, which are needed as a separate set of operations so that
120 * they can be called as part of relocation process.
123 xfs_unmount_flush(
124 xfs_mount_t *mp, /* Mount structure we are getting
125 rid of. */
126 int relocation) /* Called from vfs relocation. */
128 xfs_inode_t *rip = mp->m_rootip;
129 xfs_inode_t *rbmip;
130 xfs_inode_t *rsumip = NULL;
131 int error;
133 xfs_ilock(rip, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
134 xfs_iflock(rip);
137 * Flush out the real time inodes.
139 if ((rbmip = mp->m_rbmip) != NULL) {
140 xfs_ilock(rbmip, XFS_ILOCK_EXCL);
141 xfs_iflock(rbmip);
142 error = xfs_iflush(rbmip, XFS_IFLUSH_SYNC);
143 xfs_iunlock(rbmip, XFS_ILOCK_EXCL);
145 if (error == EFSCORRUPTED)
146 goto fscorrupt_out;
148 ASSERT(vn_count(VFS_I(rbmip)) == 1);
150 rsumip = mp->m_rsumip;
151 xfs_ilock(rsumip, XFS_ILOCK_EXCL);
152 xfs_iflock(rsumip);
153 error = xfs_iflush(rsumip, XFS_IFLUSH_SYNC);
154 xfs_iunlock(rsumip, XFS_ILOCK_EXCL);
156 if (error == EFSCORRUPTED)
157 goto fscorrupt_out;
159 ASSERT(vn_count(VFS_I(rsumip)) == 1);
163 * Synchronously flush root inode to disk
165 error = xfs_iflush(rip, XFS_IFLUSH_SYNC);
166 if (error == EFSCORRUPTED)
167 goto fscorrupt_out2;
169 if (vn_count(VFS_I(rip)) != 1 && !relocation) {
170 xfs_iunlock(rip, XFS_ILOCK_EXCL);
171 return XFS_ERROR(EBUSY);
175 * Release dquot that rootinode, rbmino and rsumino might be holding,
176 * flush and purge the quota inodes.
178 error = XFS_QM_UNMOUNT(mp);
179 if (error == EFSCORRUPTED)
180 goto fscorrupt_out2;
182 if (rbmip) {
183 IRELE(rbmip);
184 IRELE(rsumip);
187 xfs_iunlock(rip, XFS_ILOCK_EXCL);
188 return 0;
190 fscorrupt_out:
191 xfs_ifunlock(rip);
193 fscorrupt_out2:
194 xfs_iunlock(rip, XFS_ILOCK_EXCL);
196 return XFS_ERROR(EFSCORRUPTED);
200 * xfs_sync flushes any pending I/O to file system vfsp.
202 * This routine is called by vfs_sync() to make sure that things make it
203 * out to disk eventually, on sync() system calls to flush out everything,
204 * and when the file system is unmounted. For the vfs_sync() case, all
205 * we really need to do is sync out the log to make all of our meta-data
206 * updates permanent (except for timestamps). For calls from pflushd(),
207 * dirty pages are kept moving by calling pdflush() on the inodes
208 * containing them. We also flush the inodes that we can lock without
209 * sleeping and the superblock if we can lock it without sleeping from
210 * vfs_sync() so that items at the tail of the log are always moving out.
212 * Flags:
213 * SYNC_BDFLUSH - We're being called from vfs_sync() so we don't want
214 * to sleep if we can help it. All we really need
215 * to do is ensure that the log is synced at least
216 * periodically. We also push the inodes and
217 * superblock if we can lock them without sleeping
218 * and they are not pinned.
219 * SYNC_ATTR - We need to flush the inodes. If SYNC_BDFLUSH is not
220 * set, then we really want to lock each inode and flush
221 * it.
222 * SYNC_WAIT - All the flushes that take place in this call should
223 * be synchronous.
224 * SYNC_DELWRI - This tells us to push dirty pages associated with
225 * inodes. SYNC_WAIT and SYNC_BDFLUSH are used to
226 * determine if they should be flushed sync, async, or
227 * delwri.
228 * SYNC_CLOSE - This flag is passed when the system is being
229 * unmounted. We should sync and invalidate everything.
230 * SYNC_FSDATA - This indicates that the caller would like to make
231 * sure the superblock is safe on disk. We can ensure
232 * this by simply making sure the log gets flushed
233 * if SYNC_BDFLUSH is set, and by actually writing it
234 * out otherwise.
235 * SYNC_IOWAIT - The caller wants us to wait for all data I/O to complete
236 * before we return (including direct I/O). Forms the drain
237 * side of the write barrier needed to safely quiesce the
238 * filesystem.
242 xfs_sync(
243 xfs_mount_t *mp,
244 int flags)
246 int error;
249 * Get the Quota Manager to flush the dquots.
251 * If XFS quota support is not enabled or this filesystem
252 * instance does not use quotas XFS_QM_DQSYNC will always
253 * return zero.
255 error = XFS_QM_DQSYNC(mp, flags);
256 if (error) {
258 * If we got an IO error, we will be shutting down.
259 * So, there's nothing more for us to do here.
261 ASSERT(error != EIO || XFS_FORCED_SHUTDOWN(mp));
262 if (XFS_FORCED_SHUTDOWN(mp))
263 return XFS_ERROR(error);
266 if (flags & SYNC_IOWAIT)
267 xfs_filestream_flush(mp);
269 return xfs_syncsub(mp, flags, NULL);
273 * xfs sync routine for internal use
275 * This routine supports all of the flags defined for the generic vfs_sync
276 * interface as explained above under xfs_sync.
280 xfs_sync_inodes(
281 xfs_mount_t *mp,
282 int flags,
283 int *bypassed)
285 xfs_inode_t *ip = NULL;
286 struct inode *vp = NULL;
287 int error;
288 int last_error;
289 uint64_t fflag;
290 uint lock_flags;
291 uint base_lock_flags;
292 boolean_t mount_locked;
293 boolean_t vnode_refed;
294 int preempt;
295 xfs_iptr_t *ipointer;
296 #ifdef DEBUG
297 boolean_t ipointer_in = B_FALSE;
299 #define IPOINTER_SET ipointer_in = B_TRUE
300 #define IPOINTER_CLR ipointer_in = B_FALSE
301 #else
302 #define IPOINTER_SET
303 #define IPOINTER_CLR
304 #endif
307 /* Insert a marker record into the inode list after inode ip. The list
308 * must be locked when this is called. After the call the list will no
309 * longer be locked.
311 #define IPOINTER_INSERT(ip, mp) { \
312 ASSERT(ipointer_in == B_FALSE); \
313 ipointer->ip_mnext = ip->i_mnext; \
314 ipointer->ip_mprev = ip; \
315 ip->i_mnext = (xfs_inode_t *)ipointer; \
316 ipointer->ip_mnext->i_mprev = (xfs_inode_t *)ipointer; \
317 preempt = 0; \
318 XFS_MOUNT_IUNLOCK(mp); \
319 mount_locked = B_FALSE; \
320 IPOINTER_SET; \
323 /* Remove the marker from the inode list. If the marker was the only item
324 * in the list then there are no remaining inodes and we should zero out
325 * the whole list. If we are the current head of the list then move the head
326 * past us.
328 #define IPOINTER_REMOVE(ip, mp) { \
329 ASSERT(ipointer_in == B_TRUE); \
330 if (ipointer->ip_mnext != (xfs_inode_t *)ipointer) { \
331 ip = ipointer->ip_mnext; \
332 ip->i_mprev = ipointer->ip_mprev; \
333 ipointer->ip_mprev->i_mnext = ip; \
334 if (mp->m_inodes == (xfs_inode_t *)ipointer) { \
335 mp->m_inodes = ip; \
337 } else { \
338 ASSERT(mp->m_inodes == (xfs_inode_t *)ipointer); \
339 mp->m_inodes = NULL; \
340 ip = NULL; \
342 IPOINTER_CLR; \
345 #define XFS_PREEMPT_MASK 0x7f
347 ASSERT(!(flags & SYNC_BDFLUSH));
349 if (bypassed)
350 *bypassed = 0;
351 if (mp->m_flags & XFS_MOUNT_RDONLY)
352 return 0;
353 error = 0;
354 last_error = 0;
355 preempt = 0;
357 /* Allocate a reference marker */
358 ipointer = (xfs_iptr_t *)kmem_zalloc(sizeof(xfs_iptr_t), KM_SLEEP);
360 fflag = XFS_B_ASYNC; /* default is don't wait */
361 if (flags & SYNC_DELWRI)
362 fflag = XFS_B_DELWRI;
363 if (flags & SYNC_WAIT)
364 fflag = 0; /* synchronous overrides all */
366 base_lock_flags = XFS_ILOCK_SHARED;
367 if (flags & (SYNC_DELWRI | SYNC_CLOSE)) {
369 * We need the I/O lock if we're going to call any of
370 * the flush/inval routines.
372 base_lock_flags |= XFS_IOLOCK_SHARED;
375 XFS_MOUNT_ILOCK(mp);
377 ip = mp->m_inodes;
379 mount_locked = B_TRUE;
380 vnode_refed = B_FALSE;
382 IPOINTER_CLR;
384 do {
385 ASSERT(ipointer_in == B_FALSE);
386 ASSERT(vnode_refed == B_FALSE);
388 lock_flags = base_lock_flags;
391 * There were no inodes in the list, just break out
392 * of the loop.
394 if (ip == NULL) {
395 break;
399 * We found another sync thread marker - skip it
401 if (ip->i_mount == NULL) {
402 ip = ip->i_mnext;
403 continue;
406 vp = VFS_I(ip);
409 * If the vnode is gone then this is being torn down,
410 * call reclaim if it is flushed, else let regular flush
411 * code deal with it later in the loop.
414 if (vp == NULL) {
415 /* Skip ones already in reclaim */
416 if (ip->i_flags & XFS_IRECLAIM) {
417 ip = ip->i_mnext;
418 continue;
420 if (xfs_ilock_nowait(ip, XFS_ILOCK_EXCL) == 0) {
421 ip = ip->i_mnext;
422 } else if ((xfs_ipincount(ip) == 0) &&
423 xfs_iflock_nowait(ip)) {
424 IPOINTER_INSERT(ip, mp);
426 xfs_finish_reclaim(ip, 1,
427 XFS_IFLUSH_DELWRI_ELSE_ASYNC);
429 XFS_MOUNT_ILOCK(mp);
430 mount_locked = B_TRUE;
431 IPOINTER_REMOVE(ip, mp);
432 } else {
433 xfs_iunlock(ip, XFS_ILOCK_EXCL);
434 ip = ip->i_mnext;
436 continue;
439 if (VN_BAD(vp)) {
440 ip = ip->i_mnext;
441 continue;
444 if (XFS_FORCED_SHUTDOWN(mp) && !(flags & SYNC_CLOSE)) {
445 XFS_MOUNT_IUNLOCK(mp);
446 kmem_free(ipointer);
447 return 0;
451 * Try to lock without sleeping. We're out of order with
452 * the inode list lock here, so if we fail we need to drop
453 * the mount lock and try again. If we're called from
454 * bdflush() here, then don't bother.
456 * The inode lock here actually coordinates with the
457 * almost spurious inode lock in xfs_ireclaim() to prevent
458 * the vnode we handle here without a reference from
459 * being freed while we reference it. If we lock the inode
460 * while it's on the mount list here, then the spurious inode
461 * lock in xfs_ireclaim() after the inode is pulled from
462 * the mount list will sleep until we release it here.
463 * This keeps the vnode from being freed while we reference
464 * it.
466 if (xfs_ilock_nowait(ip, lock_flags) == 0) {
467 if (vp == NULL) {
468 ip = ip->i_mnext;
469 continue;
472 vp = vn_grab(vp);
473 if (vp == NULL) {
474 ip = ip->i_mnext;
475 continue;
478 IPOINTER_INSERT(ip, mp);
479 xfs_ilock(ip, lock_flags);
481 ASSERT(vp == VFS_I(ip));
482 ASSERT(ip->i_mount == mp);
484 vnode_refed = B_TRUE;
487 /* From here on in the loop we may have a marker record
488 * in the inode list.
492 * If we have to flush data or wait for I/O completion
493 * we need to drop the ilock that we currently hold.
494 * If we need to drop the lock, insert a marker if we
495 * have not already done so.
497 if ((flags & (SYNC_CLOSE|SYNC_IOWAIT)) ||
498 ((flags & SYNC_DELWRI) && VN_DIRTY(vp))) {
499 if (mount_locked) {
500 IPOINTER_INSERT(ip, mp);
502 xfs_iunlock(ip, XFS_ILOCK_SHARED);
504 if (flags & SYNC_CLOSE) {
505 /* Shutdown case. Flush and invalidate. */
506 if (XFS_FORCED_SHUTDOWN(mp))
507 xfs_tosspages(ip, 0, -1,
508 FI_REMAPF);
509 else
510 error = xfs_flushinval_pages(ip,
511 0, -1, FI_REMAPF);
512 } else if ((flags & SYNC_DELWRI) && VN_DIRTY(vp)) {
513 error = xfs_flush_pages(ip, 0,
514 -1, fflag, FI_NONE);
518 * When freezing, we need to wait ensure all I/O (including direct
519 * I/O) is complete to ensure no further data modification can take
520 * place after this point
522 if (flags & SYNC_IOWAIT)
523 vn_iowait(ip);
525 xfs_ilock(ip, XFS_ILOCK_SHARED);
528 if ((flags & SYNC_ATTR) &&
529 (ip->i_update_core ||
530 (ip->i_itemp && ip->i_itemp->ili_format.ilf_fields))) {
531 if (mount_locked)
532 IPOINTER_INSERT(ip, mp);
534 if (flags & SYNC_WAIT) {
535 xfs_iflock(ip);
536 error = xfs_iflush(ip, XFS_IFLUSH_SYNC);
539 * If we can't acquire the flush lock, then the inode
540 * is already being flushed so don't bother waiting.
542 * If we can lock it then do a delwri flush so we can
543 * combine multiple inode flushes in each disk write.
545 } else if (xfs_iflock_nowait(ip)) {
546 error = xfs_iflush(ip, XFS_IFLUSH_DELWRI);
547 } else if (bypassed) {
548 (*bypassed)++;
552 if (lock_flags != 0) {
553 xfs_iunlock(ip, lock_flags);
556 if (vnode_refed) {
558 * If we had to take a reference on the vnode
559 * above, then wait until after we've unlocked
560 * the inode to release the reference. This is
561 * because we can be already holding the inode
562 * lock when IRELE() calls xfs_inactive().
564 * Make sure to drop the mount lock before calling
565 * IRELE() so that we don't trip over ourselves if
566 * we have to go for the mount lock again in the
567 * inactive code.
569 if (mount_locked) {
570 IPOINTER_INSERT(ip, mp);
573 IRELE(ip);
575 vnode_refed = B_FALSE;
578 if (error) {
579 last_error = error;
583 * bail out if the filesystem is corrupted.
585 if (error == EFSCORRUPTED) {
586 if (!mount_locked) {
587 XFS_MOUNT_ILOCK(mp);
588 IPOINTER_REMOVE(ip, mp);
590 XFS_MOUNT_IUNLOCK(mp);
591 ASSERT(ipointer_in == B_FALSE);
592 kmem_free(ipointer);
593 return XFS_ERROR(error);
596 /* Let other threads have a chance at the mount lock
597 * if we have looped many times without dropping the
598 * lock.
600 if ((++preempt & XFS_PREEMPT_MASK) == 0) {
601 if (mount_locked) {
602 IPOINTER_INSERT(ip, mp);
606 if (mount_locked == B_FALSE) {
607 XFS_MOUNT_ILOCK(mp);
608 mount_locked = B_TRUE;
609 IPOINTER_REMOVE(ip, mp);
610 continue;
613 ASSERT(ipointer_in == B_FALSE);
614 ip = ip->i_mnext;
616 } while (ip != mp->m_inodes);
618 XFS_MOUNT_IUNLOCK(mp);
620 ASSERT(ipointer_in == B_FALSE);
622 kmem_free(ipointer);
623 return XFS_ERROR(last_error);
627 * xfs sync routine for internal use
629 * This routine supports all of the flags defined for the generic vfs_sync
630 * interface as explained above under xfs_sync.
634 xfs_syncsub(
635 xfs_mount_t *mp,
636 int flags,
637 int *bypassed)
639 int error = 0;
640 int last_error = 0;
641 uint log_flags = XFS_LOG_FORCE;
642 xfs_buf_t *bp;
643 xfs_buf_log_item_t *bip;
646 * Sync out the log. This ensures that the log is periodically
647 * flushed even if there is not enough activity to fill it up.
649 if (flags & SYNC_WAIT)
650 log_flags |= XFS_LOG_SYNC;
652 xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
654 if (flags & (SYNC_ATTR|SYNC_DELWRI)) {
655 if (flags & SYNC_BDFLUSH)
656 xfs_finish_reclaim_all(mp, 1);
657 else
658 error = xfs_sync_inodes(mp, flags, bypassed);
662 * Flushing out dirty data above probably generated more
663 * log activity, so if this isn't vfs_sync() then flush
664 * the log again.
666 if (flags & SYNC_DELWRI) {
667 xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
670 if (flags & SYNC_FSDATA) {
672 * If this is vfs_sync() then only sync the superblock
673 * if we can lock it without sleeping and it is not pinned.
675 if (flags & SYNC_BDFLUSH) {
676 bp = xfs_getsb(mp, XFS_BUF_TRYLOCK);
677 if (bp != NULL) {
678 bip = XFS_BUF_FSPRIVATE(bp,xfs_buf_log_item_t*);
679 if ((bip != NULL) &&
680 xfs_buf_item_dirty(bip)) {
681 if (!(XFS_BUF_ISPINNED(bp))) {
682 XFS_BUF_ASYNC(bp);
683 error = xfs_bwrite(mp, bp);
684 } else {
685 xfs_buf_relse(bp);
687 } else {
688 xfs_buf_relse(bp);
691 } else {
692 bp = xfs_getsb(mp, 0);
694 * If the buffer is pinned then push on the log so
695 * we won't get stuck waiting in the write for
696 * someone, maybe ourselves, to flush the log.
697 * Even though we just pushed the log above, we
698 * did not have the superblock buffer locked at
699 * that point so it can become pinned in between
700 * there and here.
702 if (XFS_BUF_ISPINNED(bp))
703 xfs_log_force(mp, (xfs_lsn_t)0, XFS_LOG_FORCE);
704 if (flags & SYNC_WAIT)
705 XFS_BUF_UNASYNC(bp);
706 else
707 XFS_BUF_ASYNC(bp);
708 error = xfs_bwrite(mp, bp);
710 if (error) {
711 last_error = error;
716 * Now check to see if the log needs a "dummy" transaction.
718 if (!(flags & SYNC_REMOUNT) && xfs_log_need_covered(mp)) {
719 xfs_trans_t *tp;
720 xfs_inode_t *ip;
723 * Put a dummy transaction in the log to tell
724 * recovery that all others are OK.
726 tp = xfs_trans_alloc(mp, XFS_TRANS_DUMMY1);
727 if ((error = xfs_trans_reserve(tp, 0,
728 XFS_ICHANGE_LOG_RES(mp),
729 0, 0, 0))) {
730 xfs_trans_cancel(tp, 0);
731 return error;
734 ip = mp->m_rootip;
735 xfs_ilock(ip, XFS_ILOCK_EXCL);
737 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
738 xfs_trans_ihold(tp, ip);
739 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
740 error = xfs_trans_commit(tp, 0);
741 xfs_iunlock(ip, XFS_ILOCK_EXCL);
742 xfs_log_force(mp, (xfs_lsn_t)0, log_flags);
746 * When shutting down, we need to insure that the AIL is pushed
747 * to disk or the filesystem can appear corrupt from the PROM.
749 if ((flags & (SYNC_CLOSE|SYNC_WAIT)) == (SYNC_CLOSE|SYNC_WAIT)) {
750 XFS_bflush(mp->m_ddev_targp);
751 if (mp->m_rtdev_targp) {
752 XFS_bflush(mp->m_rtdev_targp);
756 return XFS_ERROR(last_error);