2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
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
20 #include "xfs_shared.h"
21 #include "xfs_format.h"
22 #include "xfs_log_format.h"
23 #include "xfs_trans_resv.h"
24 #include "xfs_mount.h"
25 #include "xfs_error.h"
26 #include "xfs_trans.h"
27 #include "xfs_trans_priv.h"
29 #include "xfs_log_priv.h"
30 #include "xfs_log_recover.h"
31 #include "xfs_inode.h"
32 #include "xfs_trace.h"
33 #include "xfs_fsops.h"
34 #include "xfs_cksum.h"
35 #include "xfs_sysfs.h"
38 kmem_zone_t
*xfs_log_ticket_zone
;
40 /* Local miscellaneous function prototypes */
44 struct xlog_ticket
*ticket
,
45 struct xlog_in_core
**iclog
,
46 xfs_lsn_t
*commitlsnp
);
51 struct xfs_buftarg
*log_target
,
52 xfs_daddr_t blk_offset
,
61 struct xlog_in_core
*iclog
);
66 /* local state machine functions */
67 STATIC
void xlog_state_done_syncing(xlog_in_core_t
*iclog
, int);
69 xlog_state_do_callback(
72 struct xlog_in_core
*iclog
);
74 xlog_state_get_iclog_space(
77 struct xlog_in_core
**iclog
,
78 struct xlog_ticket
*ticket
,
82 xlog_state_release_iclog(
84 struct xlog_in_core
*iclog
);
86 xlog_state_switch_iclogs(
88 struct xlog_in_core
*iclog
,
93 struct xlog_in_core
*iclog
);
100 xlog_regrant_reserve_log_space(
102 struct xlog_ticket
*ticket
);
104 xlog_ungrant_log_space(
106 struct xlog_ticket
*ticket
);
110 xlog_verify_dest_ptr(
114 xlog_verify_grant_tail(
119 struct xlog_in_core
*iclog
,
123 xlog_verify_tail_lsn(
125 struct xlog_in_core
*iclog
,
128 #define xlog_verify_dest_ptr(a,b)
129 #define xlog_verify_grant_tail(a)
130 #define xlog_verify_iclog(a,b,c,d)
131 #define xlog_verify_tail_lsn(a,b,c)
139 xlog_grant_sub_space(
144 int64_t head_val
= atomic64_read(head
);
150 xlog_crack_grant_head_val(head_val
, &cycle
, &space
);
154 space
+= log
->l_logsize
;
159 new = xlog_assign_grant_head_val(cycle
, space
);
160 head_val
= atomic64_cmpxchg(head
, old
, new);
161 } while (head_val
!= old
);
165 xlog_grant_add_space(
170 int64_t head_val
= atomic64_read(head
);
177 xlog_crack_grant_head_val(head_val
, &cycle
, &space
);
179 tmp
= log
->l_logsize
- space
;
188 new = xlog_assign_grant_head_val(cycle
, space
);
189 head_val
= atomic64_cmpxchg(head
, old
, new);
190 } while (head_val
!= old
);
194 xlog_grant_head_init(
195 struct xlog_grant_head
*head
)
197 xlog_assign_grant_head(&head
->grant
, 1, 0);
198 INIT_LIST_HEAD(&head
->waiters
);
199 spin_lock_init(&head
->lock
);
203 xlog_grant_head_wake_all(
204 struct xlog_grant_head
*head
)
206 struct xlog_ticket
*tic
;
208 spin_lock(&head
->lock
);
209 list_for_each_entry(tic
, &head
->waiters
, t_queue
)
210 wake_up_process(tic
->t_task
);
211 spin_unlock(&head
->lock
);
215 xlog_ticket_reservation(
217 struct xlog_grant_head
*head
,
218 struct xlog_ticket
*tic
)
220 if (head
== &log
->l_write_head
) {
221 ASSERT(tic
->t_flags
& XLOG_TIC_PERM_RESERV
);
222 return tic
->t_unit_res
;
224 if (tic
->t_flags
& XLOG_TIC_PERM_RESERV
)
225 return tic
->t_unit_res
* tic
->t_cnt
;
227 return tic
->t_unit_res
;
232 xlog_grant_head_wake(
234 struct xlog_grant_head
*head
,
237 struct xlog_ticket
*tic
;
240 list_for_each_entry(tic
, &head
->waiters
, t_queue
) {
241 need_bytes
= xlog_ticket_reservation(log
, head
, tic
);
242 if (*free_bytes
< need_bytes
)
245 *free_bytes
-= need_bytes
;
246 trace_xfs_log_grant_wake_up(log
, tic
);
247 wake_up_process(tic
->t_task
);
254 xlog_grant_head_wait(
256 struct xlog_grant_head
*head
,
257 struct xlog_ticket
*tic
,
258 int need_bytes
) __releases(&head
->lock
)
259 __acquires(&head
->lock
)
261 list_add_tail(&tic
->t_queue
, &head
->waiters
);
264 if (XLOG_FORCED_SHUTDOWN(log
))
266 xlog_grant_push_ail(log
, need_bytes
);
268 __set_current_state(TASK_UNINTERRUPTIBLE
);
269 spin_unlock(&head
->lock
);
271 XFS_STATS_INC(xs_sleep_logspace
);
273 trace_xfs_log_grant_sleep(log
, tic
);
275 trace_xfs_log_grant_wake(log
, tic
);
277 spin_lock(&head
->lock
);
278 if (XLOG_FORCED_SHUTDOWN(log
))
280 } while (xlog_space_left(log
, &head
->grant
) < need_bytes
);
282 list_del_init(&tic
->t_queue
);
285 list_del_init(&tic
->t_queue
);
290 * Atomically get the log space required for a log ticket.
292 * Once a ticket gets put onto head->waiters, it will only return after the
293 * needed reservation is satisfied.
295 * This function is structured so that it has a lock free fast path. This is
296 * necessary because every new transaction reservation will come through this
297 * path. Hence any lock will be globally hot if we take it unconditionally on
300 * As tickets are only ever moved on and off head->waiters under head->lock, we
301 * only need to take that lock if we are going to add the ticket to the queue
302 * and sleep. We can avoid taking the lock if the ticket was never added to
303 * head->waiters because the t_queue list head will be empty and we hold the
304 * only reference to it so it can safely be checked unlocked.
307 xlog_grant_head_check(
309 struct xlog_grant_head
*head
,
310 struct xlog_ticket
*tic
,
316 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
319 * If there are other waiters on the queue then give them a chance at
320 * logspace before us. Wake up the first waiters, if we do not wake
321 * up all the waiters then go to sleep waiting for more free space,
322 * otherwise try to get some space for this transaction.
324 *need_bytes
= xlog_ticket_reservation(log
, head
, tic
);
325 free_bytes
= xlog_space_left(log
, &head
->grant
);
326 if (!list_empty_careful(&head
->waiters
)) {
327 spin_lock(&head
->lock
);
328 if (!xlog_grant_head_wake(log
, head
, &free_bytes
) ||
329 free_bytes
< *need_bytes
) {
330 error
= xlog_grant_head_wait(log
, head
, tic
,
333 spin_unlock(&head
->lock
);
334 } else if (free_bytes
< *need_bytes
) {
335 spin_lock(&head
->lock
);
336 error
= xlog_grant_head_wait(log
, head
, tic
, *need_bytes
);
337 spin_unlock(&head
->lock
);
344 xlog_tic_reset_res(xlog_ticket_t
*tic
)
347 tic
->t_res_arr_sum
= 0;
348 tic
->t_res_num_ophdrs
= 0;
352 xlog_tic_add_region(xlog_ticket_t
*tic
, uint len
, uint type
)
354 if (tic
->t_res_num
== XLOG_TIC_LEN_MAX
) {
355 /* add to overflow and start again */
356 tic
->t_res_o_flow
+= tic
->t_res_arr_sum
;
358 tic
->t_res_arr_sum
= 0;
361 tic
->t_res_arr
[tic
->t_res_num
].r_len
= len
;
362 tic
->t_res_arr
[tic
->t_res_num
].r_type
= type
;
363 tic
->t_res_arr_sum
+= len
;
368 * Replenish the byte reservation required by moving the grant write head.
372 struct xfs_mount
*mp
,
373 struct xlog_ticket
*tic
)
375 struct xlog
*log
= mp
->m_log
;
379 if (XLOG_FORCED_SHUTDOWN(log
))
382 XFS_STATS_INC(xs_try_logspace
);
385 * This is a new transaction on the ticket, so we need to change the
386 * transaction ID so that the next transaction has a different TID in
387 * the log. Just add one to the existing tid so that we can see chains
388 * of rolling transactions in the log easily.
392 xlog_grant_push_ail(log
, tic
->t_unit_res
);
394 tic
->t_curr_res
= tic
->t_unit_res
;
395 xlog_tic_reset_res(tic
);
400 trace_xfs_log_regrant(log
, tic
);
402 error
= xlog_grant_head_check(log
, &log
->l_write_head
, tic
,
407 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, need_bytes
);
408 trace_xfs_log_regrant_exit(log
, tic
);
409 xlog_verify_grant_tail(log
);
414 * If we are failing, make sure the ticket doesn't have any current
415 * reservations. We don't want to add this back when the ticket/
416 * transaction gets cancelled.
419 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
424 * Reserve log space and return a ticket corresponding the reservation.
426 * Each reservation is going to reserve extra space for a log record header.
427 * When writes happen to the on-disk log, we don't subtract the length of the
428 * log record header from any reservation. By wasting space in each
429 * reservation, we prevent over allocation problems.
433 struct xfs_mount
*mp
,
436 struct xlog_ticket
**ticp
,
441 struct xlog
*log
= mp
->m_log
;
442 struct xlog_ticket
*tic
;
446 ASSERT(client
== XFS_TRANSACTION
|| client
== XFS_LOG
);
448 if (XLOG_FORCED_SHUTDOWN(log
))
451 XFS_STATS_INC(xs_try_logspace
);
453 ASSERT(*ticp
== NULL
);
454 tic
= xlog_ticket_alloc(log
, unit_bytes
, cnt
, client
, permanent
,
455 KM_SLEEP
| KM_MAYFAIL
);
459 tic
->t_trans_type
= t_type
;
462 xlog_grant_push_ail(log
, tic
->t_cnt
? tic
->t_unit_res
* tic
->t_cnt
465 trace_xfs_log_reserve(log
, tic
);
467 error
= xlog_grant_head_check(log
, &log
->l_reserve_head
, tic
,
472 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
, need_bytes
);
473 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, need_bytes
);
474 trace_xfs_log_reserve_exit(log
, tic
);
475 xlog_verify_grant_tail(log
);
480 * If we are failing, make sure the ticket doesn't have any current
481 * reservations. We don't want to add this back when the ticket/
482 * transaction gets cancelled.
485 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
493 * 1. currblock field gets updated at startup and after in-core logs
494 * marked as with WANT_SYNC.
498 * This routine is called when a user of a log manager ticket is done with
499 * the reservation. If the ticket was ever used, then a commit record for
500 * the associated transaction is written out as a log operation header with
501 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
502 * a given ticket. If the ticket was one with a permanent reservation, then
503 * a few operations are done differently. Permanent reservation tickets by
504 * default don't release the reservation. They just commit the current
505 * transaction with the belief that the reservation is still needed. A flag
506 * must be passed in before permanent reservations are actually released.
507 * When these type of tickets are not released, they need to be set into
508 * the inited state again. By doing this, a start record will be written
509 * out when the next write occurs.
513 struct xfs_mount
*mp
,
514 struct xlog_ticket
*ticket
,
515 struct xlog_in_core
**iclog
,
518 struct xlog
*log
= mp
->m_log
;
521 if (XLOG_FORCED_SHUTDOWN(log
) ||
523 * If nothing was ever written, don't write out commit record.
524 * If we get an error, just continue and give back the log ticket.
526 (((ticket
->t_flags
& XLOG_TIC_INITED
) == 0) &&
527 (xlog_commit_record(log
, ticket
, iclog
, &lsn
)))) {
528 lsn
= (xfs_lsn_t
) -1;
534 trace_xfs_log_done_nonperm(log
, ticket
);
537 * Release ticket if not permanent reservation or a specific
538 * request has been made to release a permanent reservation.
540 xlog_ungrant_log_space(log
, ticket
);
542 trace_xfs_log_done_perm(log
, ticket
);
544 xlog_regrant_reserve_log_space(log
, ticket
);
545 /* If this ticket was a permanent reservation and we aren't
546 * trying to release it, reset the inited flags; so next time
547 * we write, a start record will be written out.
549 ticket
->t_flags
|= XLOG_TIC_INITED
;
552 xfs_log_ticket_put(ticket
);
557 * Attaches a new iclog I/O completion callback routine during
558 * transaction commit. If the log is in error state, a non-zero
559 * return code is handed back and the caller is responsible for
560 * executing the callback at an appropriate time.
564 struct xfs_mount
*mp
,
565 struct xlog_in_core
*iclog
,
566 xfs_log_callback_t
*cb
)
570 spin_lock(&iclog
->ic_callback_lock
);
571 abortflg
= (iclog
->ic_state
& XLOG_STATE_IOERROR
);
573 ASSERT_ALWAYS((iclog
->ic_state
== XLOG_STATE_ACTIVE
) ||
574 (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
));
576 *(iclog
->ic_callback_tail
) = cb
;
577 iclog
->ic_callback_tail
= &(cb
->cb_next
);
579 spin_unlock(&iclog
->ic_callback_lock
);
584 xfs_log_release_iclog(
585 struct xfs_mount
*mp
,
586 struct xlog_in_core
*iclog
)
588 if (xlog_state_release_iclog(mp
->m_log
, iclog
)) {
589 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
597 * Mount a log filesystem
599 * mp - ubiquitous xfs mount point structure
600 * log_target - buftarg of on-disk log device
601 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
602 * num_bblocks - Number of BBSIZE blocks in on-disk log
604 * Return error or zero.
609 xfs_buftarg_t
*log_target
,
610 xfs_daddr_t blk_offset
,
616 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
)) {
617 xfs_notice(mp
, "Mounting V%d Filesystem",
618 XFS_SB_VERSION_NUM(&mp
->m_sb
));
621 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
622 XFS_SB_VERSION_NUM(&mp
->m_sb
));
623 ASSERT(mp
->m_flags
& XFS_MOUNT_RDONLY
);
626 mp
->m_log
= xlog_alloc_log(mp
, log_target
, blk_offset
, num_bblks
);
627 if (IS_ERR(mp
->m_log
)) {
628 error
= PTR_ERR(mp
->m_log
);
633 * Validate the given log space and drop a critical message via syslog
634 * if the log size is too small that would lead to some unexpected
635 * situations in transaction log space reservation stage.
637 * Note: we can't just reject the mount if the validation fails. This
638 * would mean that people would have to downgrade their kernel just to
639 * remedy the situation as there is no way to grow the log (short of
640 * black magic surgery with xfs_db).
642 * We can, however, reject mounts for CRC format filesystems, as the
643 * mkfs binary being used to make the filesystem should never create a
644 * filesystem with a log that is too small.
646 min_logfsbs
= xfs_log_calc_minimum_size(mp
);
648 if (mp
->m_sb
.sb_logblocks
< min_logfsbs
) {
650 "Log size %d blocks too small, minimum size is %d blocks",
651 mp
->m_sb
.sb_logblocks
, min_logfsbs
);
653 } else if (mp
->m_sb
.sb_logblocks
> XFS_MAX_LOG_BLOCKS
) {
655 "Log size %d blocks too large, maximum size is %lld blocks",
656 mp
->m_sb
.sb_logblocks
, XFS_MAX_LOG_BLOCKS
);
658 } else if (XFS_FSB_TO_B(mp
, mp
->m_sb
.sb_logblocks
) > XFS_MAX_LOG_BYTES
) {
660 "log size %lld bytes too large, maximum size is %lld bytes",
661 XFS_FSB_TO_B(mp
, mp
->m_sb
.sb_logblocks
),
666 if (xfs_sb_version_hascrc(&mp
->m_sb
)) {
667 xfs_crit(mp
, "AAIEEE! Log failed size checks. Abort!");
671 xfs_crit(mp
, "Log size out of supported range.");
673 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
677 * Initialize the AIL now we have a log.
679 error
= xfs_trans_ail_init(mp
);
681 xfs_warn(mp
, "AIL initialisation failed: error %d", error
);
684 mp
->m_log
->l_ailp
= mp
->m_ail
;
687 * skip log recovery on a norecovery mount. pretend it all
690 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
)) {
691 int readonly
= (mp
->m_flags
& XFS_MOUNT_RDONLY
);
694 mp
->m_flags
&= ~XFS_MOUNT_RDONLY
;
696 error
= xlog_recover(mp
->m_log
);
699 mp
->m_flags
|= XFS_MOUNT_RDONLY
;
701 xfs_warn(mp
, "log mount/recovery failed: error %d",
703 xlog_recover_cancel(mp
->m_log
);
704 goto out_destroy_ail
;
708 error
= xfs_sysfs_init(&mp
->m_log
->l_kobj
, &xfs_log_ktype
, &mp
->m_kobj
,
711 goto out_destroy_ail
;
713 /* Normal transactions can now occur */
714 mp
->m_log
->l_flags
&= ~XLOG_ACTIVE_RECOVERY
;
717 * Now the log has been fully initialised and we know were our
718 * space grant counters are, we can initialise the permanent ticket
719 * needed for delayed logging to work.
721 xlog_cil_init_post_recovery(mp
->m_log
);
726 xfs_trans_ail_destroy(mp
);
728 xlog_dealloc_log(mp
->m_log
);
734 * Finish the recovery of the file system. This is separate from the
735 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
736 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
739 * If we finish recovery successfully, start the background log work. If we are
740 * not doing recovery, then we have a RO filesystem and we don't need to start
744 xfs_log_mount_finish(
745 struct xfs_mount
*mp
)
749 if (mp
->m_flags
& XFS_MOUNT_NORECOVERY
) {
750 ASSERT(mp
->m_flags
& XFS_MOUNT_RDONLY
);
754 error
= xlog_recover_finish(mp
->m_log
);
756 xfs_log_work_queue(mp
);
762 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
766 xfs_log_mount_cancel(
767 struct xfs_mount
*mp
)
771 error
= xlog_recover_cancel(mp
->m_log
);
778 * Final log writes as part of unmount.
780 * Mark the filesystem clean as unmount happens. Note that during relocation
781 * this routine needs to be executed as part of source-bag while the
782 * deallocation must not be done until source-end.
786 * Unmount record used to have a string "Unmount filesystem--" in the
787 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
788 * We just write the magic number now since that particular field isn't
789 * currently architecture converted and "Unmount" is a bit foo.
790 * As far as I know, there weren't any dependencies on the old behaviour.
794 xfs_log_unmount_write(xfs_mount_t
*mp
)
796 struct xlog
*log
= mp
->m_log
;
797 xlog_in_core_t
*iclog
;
799 xlog_in_core_t
*first_iclog
;
801 xlog_ticket_t
*tic
= NULL
;
806 * Don't write out unmount record on read-only mounts.
807 * Or, if we are doing a forced umount (typically because of IO errors).
809 if (mp
->m_flags
& XFS_MOUNT_RDONLY
)
812 error
= _xfs_log_force(mp
, XFS_LOG_SYNC
, NULL
);
813 ASSERT(error
|| !(XLOG_FORCED_SHUTDOWN(log
)));
816 first_iclog
= iclog
= log
->l_iclog
;
818 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
819 ASSERT(iclog
->ic_state
& XLOG_STATE_ACTIVE
);
820 ASSERT(iclog
->ic_offset
== 0);
822 iclog
= iclog
->ic_next
;
823 } while (iclog
!= first_iclog
);
825 if (! (XLOG_FORCED_SHUTDOWN(log
))) {
826 error
= xfs_log_reserve(mp
, 600, 1, &tic
,
827 XFS_LOG
, 0, XLOG_UNMOUNT_REC_TYPE
);
829 /* the data section must be 32 bit size aligned */
833 __uint32_t pad2
; /* may as well make it 64 bits */
835 .magic
= XLOG_UNMOUNT_TYPE
,
837 struct xfs_log_iovec reg
= {
839 .i_len
= sizeof(magic
),
840 .i_type
= XLOG_REG_TYPE_UNMOUNT
,
842 struct xfs_log_vec vec
= {
847 /* remove inited flag, and account for space used */
849 tic
->t_curr_res
-= sizeof(magic
);
850 error
= xlog_write(log
, &vec
, tic
, &lsn
,
851 NULL
, XLOG_UNMOUNT_TRANS
);
853 * At this point, we're umounting anyway,
854 * so there's no point in transitioning log state
855 * to IOERROR. Just continue...
860 xfs_alert(mp
, "%s: unmount record failed", __func__
);
863 spin_lock(&log
->l_icloglock
);
864 iclog
= log
->l_iclog
;
865 atomic_inc(&iclog
->ic_refcnt
);
866 xlog_state_want_sync(log
, iclog
);
867 spin_unlock(&log
->l_icloglock
);
868 error
= xlog_state_release_iclog(log
, iclog
);
870 spin_lock(&log
->l_icloglock
);
871 if (!(iclog
->ic_state
== XLOG_STATE_ACTIVE
||
872 iclog
->ic_state
== XLOG_STATE_DIRTY
)) {
873 if (!XLOG_FORCED_SHUTDOWN(log
)) {
874 xlog_wait(&iclog
->ic_force_wait
,
877 spin_unlock(&log
->l_icloglock
);
880 spin_unlock(&log
->l_icloglock
);
883 trace_xfs_log_umount_write(log
, tic
);
884 xlog_ungrant_log_space(log
, tic
);
885 xfs_log_ticket_put(tic
);
889 * We're already in forced_shutdown mode, couldn't
890 * even attempt to write out the unmount transaction.
892 * Go through the motions of sync'ing and releasing
893 * the iclog, even though no I/O will actually happen,
894 * we need to wait for other log I/Os that may already
895 * be in progress. Do this as a separate section of
896 * code so we'll know if we ever get stuck here that
897 * we're in this odd situation of trying to unmount
898 * a file system that went into forced_shutdown as
899 * the result of an unmount..
901 spin_lock(&log
->l_icloglock
);
902 iclog
= log
->l_iclog
;
903 atomic_inc(&iclog
->ic_refcnt
);
905 xlog_state_want_sync(log
, iclog
);
906 spin_unlock(&log
->l_icloglock
);
907 error
= xlog_state_release_iclog(log
, iclog
);
909 spin_lock(&log
->l_icloglock
);
911 if ( ! ( iclog
->ic_state
== XLOG_STATE_ACTIVE
912 || iclog
->ic_state
== XLOG_STATE_DIRTY
913 || iclog
->ic_state
== XLOG_STATE_IOERROR
) ) {
915 xlog_wait(&iclog
->ic_force_wait
,
918 spin_unlock(&log
->l_icloglock
);
923 } /* xfs_log_unmount_write */
926 * Empty the log for unmount/freeze.
928 * To do this, we first need to shut down the background log work so it is not
929 * trying to cover the log as we clean up. We then need to unpin all objects in
930 * the log so we can then flush them out. Once they have completed their IO and
931 * run the callbacks removing themselves from the AIL, we can write the unmount
936 struct xfs_mount
*mp
)
938 cancel_delayed_work_sync(&mp
->m_log
->l_work
);
939 xfs_log_force(mp
, XFS_LOG_SYNC
);
942 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
943 * will push it, xfs_wait_buftarg() will not wait for it. Further,
944 * xfs_buf_iowait() cannot be used because it was pushed with the
945 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
946 * the IO to complete.
948 xfs_ail_push_all_sync(mp
->m_ail
);
949 xfs_wait_buftarg(mp
->m_ddev_targp
);
950 xfs_buf_lock(mp
->m_sb_bp
);
951 xfs_buf_unlock(mp
->m_sb_bp
);
953 xfs_log_unmount_write(mp
);
957 * Shut down and release the AIL and Log.
959 * During unmount, we need to ensure we flush all the dirty metadata objects
960 * from the AIL so that the log is empty before we write the unmount record to
961 * the log. Once this is done, we can tear down the AIL and the log.
965 struct xfs_mount
*mp
)
969 xfs_trans_ail_destroy(mp
);
971 xfs_sysfs_del(&mp
->m_log
->l_kobj
);
973 xlog_dealloc_log(mp
->m_log
);
978 struct xfs_mount
*mp
,
979 struct xfs_log_item
*item
,
981 const struct xfs_item_ops
*ops
)
983 item
->li_mountp
= mp
;
984 item
->li_ailp
= mp
->m_ail
;
985 item
->li_type
= type
;
989 INIT_LIST_HEAD(&item
->li_ail
);
990 INIT_LIST_HEAD(&item
->li_cil
);
994 * Wake up processes waiting for log space after we have moved the log tail.
998 struct xfs_mount
*mp
)
1000 struct xlog
*log
= mp
->m_log
;
1003 if (XLOG_FORCED_SHUTDOWN(log
))
1006 if (!list_empty_careful(&log
->l_write_head
.waiters
)) {
1007 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
1009 spin_lock(&log
->l_write_head
.lock
);
1010 free_bytes
= xlog_space_left(log
, &log
->l_write_head
.grant
);
1011 xlog_grant_head_wake(log
, &log
->l_write_head
, &free_bytes
);
1012 spin_unlock(&log
->l_write_head
.lock
);
1015 if (!list_empty_careful(&log
->l_reserve_head
.waiters
)) {
1016 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
1018 spin_lock(&log
->l_reserve_head
.lock
);
1019 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
1020 xlog_grant_head_wake(log
, &log
->l_reserve_head
, &free_bytes
);
1021 spin_unlock(&log
->l_reserve_head
.lock
);
1026 * Determine if we have a transaction that has gone to disk that needs to be
1027 * covered. To begin the transition to the idle state firstly the log needs to
1028 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1029 * we start attempting to cover the log.
1031 * Only if we are then in a state where covering is needed, the caller is
1032 * informed that dummy transactions are required to move the log into the idle
1035 * If there are any items in the AIl or CIL, then we do not want to attempt to
1036 * cover the log as we may be in a situation where there isn't log space
1037 * available to run a dummy transaction and this can lead to deadlocks when the
1038 * tail of the log is pinned by an item that is modified in the CIL. Hence
1039 * there's no point in running a dummy transaction at this point because we
1040 * can't start trying to idle the log until both the CIL and AIL are empty.
1043 xfs_log_need_covered(xfs_mount_t
*mp
)
1045 struct xlog
*log
= mp
->m_log
;
1048 if (!xfs_fs_writable(mp
, SB_FREEZE_WRITE
))
1051 if (!xlog_cil_empty(log
))
1054 spin_lock(&log
->l_icloglock
);
1055 switch (log
->l_covered_state
) {
1056 case XLOG_STATE_COVER_DONE
:
1057 case XLOG_STATE_COVER_DONE2
:
1058 case XLOG_STATE_COVER_IDLE
:
1060 case XLOG_STATE_COVER_NEED
:
1061 case XLOG_STATE_COVER_NEED2
:
1062 if (xfs_ail_min_lsn(log
->l_ailp
))
1064 if (!xlog_iclogs_empty(log
))
1068 if (log
->l_covered_state
== XLOG_STATE_COVER_NEED
)
1069 log
->l_covered_state
= XLOG_STATE_COVER_DONE
;
1071 log
->l_covered_state
= XLOG_STATE_COVER_DONE2
;
1077 spin_unlock(&log
->l_icloglock
);
1082 * We may be holding the log iclog lock upon entering this routine.
1085 xlog_assign_tail_lsn_locked(
1086 struct xfs_mount
*mp
)
1088 struct xlog
*log
= mp
->m_log
;
1089 struct xfs_log_item
*lip
;
1092 assert_spin_locked(&mp
->m_ail
->xa_lock
);
1095 * To make sure we always have a valid LSN for the log tail we keep
1096 * track of the last LSN which was committed in log->l_last_sync_lsn,
1097 * and use that when the AIL was empty.
1099 lip
= xfs_ail_min(mp
->m_ail
);
1101 tail_lsn
= lip
->li_lsn
;
1103 tail_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
1104 trace_xfs_log_assign_tail_lsn(log
, tail_lsn
);
1105 atomic64_set(&log
->l_tail_lsn
, tail_lsn
);
1110 xlog_assign_tail_lsn(
1111 struct xfs_mount
*mp
)
1115 spin_lock(&mp
->m_ail
->xa_lock
);
1116 tail_lsn
= xlog_assign_tail_lsn_locked(mp
);
1117 spin_unlock(&mp
->m_ail
->xa_lock
);
1123 * Return the space in the log between the tail and the head. The head
1124 * is passed in the cycle/bytes formal parms. In the special case where
1125 * the reserve head has wrapped passed the tail, this calculation is no
1126 * longer valid. In this case, just return 0 which means there is no space
1127 * in the log. This works for all places where this function is called
1128 * with the reserve head. Of course, if the write head were to ever
1129 * wrap the tail, we should blow up. Rather than catch this case here,
1130 * we depend on other ASSERTions in other parts of the code. XXXmiken
1132 * This code also handles the case where the reservation head is behind
1133 * the tail. The details of this case are described below, but the end
1134 * result is that we return the size of the log as the amount of space left.
1147 xlog_crack_grant_head(head
, &head_cycle
, &head_bytes
);
1148 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_bytes
);
1149 tail_bytes
= BBTOB(tail_bytes
);
1150 if (tail_cycle
== head_cycle
&& head_bytes
>= tail_bytes
)
1151 free_bytes
= log
->l_logsize
- (head_bytes
- tail_bytes
);
1152 else if (tail_cycle
+ 1 < head_cycle
)
1154 else if (tail_cycle
< head_cycle
) {
1155 ASSERT(tail_cycle
== (head_cycle
- 1));
1156 free_bytes
= tail_bytes
- head_bytes
;
1159 * The reservation head is behind the tail.
1160 * In this case we just want to return the size of the
1161 * log as the amount of space left.
1163 xfs_alert(log
->l_mp
, "xlog_space_left: head behind tail");
1164 xfs_alert(log
->l_mp
,
1165 " tail_cycle = %d, tail_bytes = %d",
1166 tail_cycle
, tail_bytes
);
1167 xfs_alert(log
->l_mp
,
1168 " GH cycle = %d, GH bytes = %d",
1169 head_cycle
, head_bytes
);
1171 free_bytes
= log
->l_logsize
;
1178 * Log function which is called when an io completes.
1180 * The log manager needs its own routine, in order to control what
1181 * happens with the buffer after the write completes.
1184 xlog_iodone(xfs_buf_t
*bp
)
1186 struct xlog_in_core
*iclog
= bp
->b_fspriv
;
1187 struct xlog
*l
= iclog
->ic_log
;
1191 * Race to shutdown the filesystem if we see an error.
1193 if (XFS_TEST_ERROR(bp
->b_error
, l
->l_mp
,
1194 XFS_ERRTAG_IODONE_IOERR
, XFS_RANDOM_IODONE_IOERR
)) {
1195 xfs_buf_ioerror_alert(bp
, __func__
);
1197 xfs_force_shutdown(l
->l_mp
, SHUTDOWN_LOG_IO_ERROR
);
1199 * This flag will be propagated to the trans-committed
1200 * callback routines to let them know that the log-commit
1203 aborted
= XFS_LI_ABORTED
;
1204 } else if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
1205 aborted
= XFS_LI_ABORTED
;
1208 /* log I/O is always issued ASYNC */
1209 ASSERT(XFS_BUF_ISASYNC(bp
));
1210 xlog_state_done_syncing(iclog
, aborted
);
1213 * drop the buffer lock now that we are done. Nothing references
1214 * the buffer after this, so an unmount waiting on this lock can now
1215 * tear it down safely. As such, it is unsafe to reference the buffer
1216 * (bp) after the unlock as we could race with it being freed.
1222 * Return size of each in-core log record buffer.
1224 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1226 * If the filesystem blocksize is too large, we may need to choose a
1227 * larger size since the directory code currently logs entire blocks.
1231 xlog_get_iclog_buffer_size(
1232 struct xfs_mount
*mp
,
1238 if (mp
->m_logbufs
<= 0)
1239 log
->l_iclog_bufs
= XLOG_MAX_ICLOGS
;
1241 log
->l_iclog_bufs
= mp
->m_logbufs
;
1244 * Buffer size passed in from mount system call.
1246 if (mp
->m_logbsize
> 0) {
1247 size
= log
->l_iclog_size
= mp
->m_logbsize
;
1248 log
->l_iclog_size_log
= 0;
1250 log
->l_iclog_size_log
++;
1254 if (xfs_sb_version_haslogv2(&mp
->m_sb
)) {
1255 /* # headers = size / 32k
1256 * one header holds cycles from 32k of data
1259 xhdrs
= mp
->m_logbsize
/ XLOG_HEADER_CYCLE_SIZE
;
1260 if (mp
->m_logbsize
% XLOG_HEADER_CYCLE_SIZE
)
1262 log
->l_iclog_hsize
= xhdrs
<< BBSHIFT
;
1263 log
->l_iclog_heads
= xhdrs
;
1265 ASSERT(mp
->m_logbsize
<= XLOG_BIG_RECORD_BSIZE
);
1266 log
->l_iclog_hsize
= BBSIZE
;
1267 log
->l_iclog_heads
= 1;
1272 /* All machines use 32kB buffers by default. */
1273 log
->l_iclog_size
= XLOG_BIG_RECORD_BSIZE
;
1274 log
->l_iclog_size_log
= XLOG_BIG_RECORD_BSHIFT
;
1276 /* the default log size is 16k or 32k which is one header sector */
1277 log
->l_iclog_hsize
= BBSIZE
;
1278 log
->l_iclog_heads
= 1;
1281 /* are we being asked to make the sizes selected above visible? */
1282 if (mp
->m_logbufs
== 0)
1283 mp
->m_logbufs
= log
->l_iclog_bufs
;
1284 if (mp
->m_logbsize
== 0)
1285 mp
->m_logbsize
= log
->l_iclog_size
;
1286 } /* xlog_get_iclog_buffer_size */
1291 struct xfs_mount
*mp
)
1293 queue_delayed_work(mp
->m_log_workqueue
, &mp
->m_log
->l_work
,
1294 msecs_to_jiffies(xfs_syncd_centisecs
* 10));
1298 * Every sync period we need to unpin all items in the AIL and push them to
1299 * disk. If there is nothing dirty, then we might need to cover the log to
1300 * indicate that the filesystem is idle.
1304 struct work_struct
*work
)
1306 struct xlog
*log
= container_of(to_delayed_work(work
),
1307 struct xlog
, l_work
);
1308 struct xfs_mount
*mp
= log
->l_mp
;
1310 /* dgc: errors ignored - not fatal and nowhere to report them */
1311 if (xfs_log_need_covered(mp
)) {
1313 * Dump a transaction into the log that contains no real change.
1314 * This is needed to stamp the current tail LSN into the log
1315 * during the covering operation.
1317 * We cannot use an inode here for this - that will push dirty
1318 * state back up into the VFS and then periodic inode flushing
1319 * will prevent log covering from making progress. Hence we
1320 * synchronously log the superblock instead to ensure the
1321 * superblock is immediately unpinned and can be written back.
1323 xfs_sync_sb(mp
, true);
1325 xfs_log_force(mp
, 0);
1327 /* start pushing all the metadata that is currently dirty */
1328 xfs_ail_push_all(mp
->m_ail
);
1330 /* queue us up again */
1331 xfs_log_work_queue(mp
);
1335 * This routine initializes some of the log structure for a given mount point.
1336 * Its primary purpose is to fill in enough, so recovery can occur. However,
1337 * some other stuff may be filled in too.
1339 STATIC
struct xlog
*
1341 struct xfs_mount
*mp
,
1342 struct xfs_buftarg
*log_target
,
1343 xfs_daddr_t blk_offset
,
1347 xlog_rec_header_t
*head
;
1348 xlog_in_core_t
**iclogp
;
1349 xlog_in_core_t
*iclog
, *prev_iclog
=NULL
;
1352 int error
= -ENOMEM
;
1355 log
= kmem_zalloc(sizeof(struct xlog
), KM_MAYFAIL
);
1357 xfs_warn(mp
, "Log allocation failed: No memory!");
1362 log
->l_targ
= log_target
;
1363 log
->l_logsize
= BBTOB(num_bblks
);
1364 log
->l_logBBstart
= blk_offset
;
1365 log
->l_logBBsize
= num_bblks
;
1366 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
1367 log
->l_flags
|= XLOG_ACTIVE_RECOVERY
;
1368 INIT_DELAYED_WORK(&log
->l_work
, xfs_log_worker
);
1370 log
->l_prev_block
= -1;
1371 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1372 xlog_assign_atomic_lsn(&log
->l_tail_lsn
, 1, 0);
1373 xlog_assign_atomic_lsn(&log
->l_last_sync_lsn
, 1, 0);
1374 log
->l_curr_cycle
= 1; /* 0 is bad since this is initial value */
1376 xlog_grant_head_init(&log
->l_reserve_head
);
1377 xlog_grant_head_init(&log
->l_write_head
);
1379 error
= -EFSCORRUPTED
;
1380 if (xfs_sb_version_hassector(&mp
->m_sb
)) {
1381 log2_size
= mp
->m_sb
.sb_logsectlog
;
1382 if (log2_size
< BBSHIFT
) {
1383 xfs_warn(mp
, "Log sector size too small (0x%x < 0x%x)",
1384 log2_size
, BBSHIFT
);
1388 log2_size
-= BBSHIFT
;
1389 if (log2_size
> mp
->m_sectbb_log
) {
1390 xfs_warn(mp
, "Log sector size too large (0x%x > 0x%x)",
1391 log2_size
, mp
->m_sectbb_log
);
1395 /* for larger sector sizes, must have v2 or external log */
1396 if (log2_size
&& log
->l_logBBstart
> 0 &&
1397 !xfs_sb_version_haslogv2(&mp
->m_sb
)) {
1399 "log sector size (0x%x) invalid for configuration.",
1404 log
->l_sectBBsize
= 1 << log2_size
;
1406 xlog_get_iclog_buffer_size(mp
, log
);
1409 * Use a NULL block for the extra log buffer used during splits so that
1410 * it will trigger errors if we ever try to do IO on it without first
1411 * having set it up properly.
1414 bp
= xfs_buf_alloc(mp
->m_logdev_targp
, XFS_BUF_DADDR_NULL
,
1415 BTOBB(log
->l_iclog_size
), 0);
1420 * The iclogbuf buffer locks are held over IO but we are not going to do
1421 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1422 * when appropriately.
1424 ASSERT(xfs_buf_islocked(bp
));
1427 /* use high priority wq for log I/O completion */
1428 bp
->b_ioend_wq
= mp
->m_log_workqueue
;
1429 bp
->b_iodone
= xlog_iodone
;
1432 spin_lock_init(&log
->l_icloglock
);
1433 init_waitqueue_head(&log
->l_flush_wait
);
1435 iclogp
= &log
->l_iclog
;
1437 * The amount of memory to allocate for the iclog structure is
1438 * rather funky due to the way the structure is defined. It is
1439 * done this way so that we can use different sizes for machines
1440 * with different amounts of memory. See the definition of
1441 * xlog_in_core_t in xfs_log_priv.h for details.
1443 ASSERT(log
->l_iclog_size
>= 4096);
1444 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
1445 *iclogp
= kmem_zalloc(sizeof(xlog_in_core_t
), KM_MAYFAIL
);
1447 goto out_free_iclog
;
1450 iclog
->ic_prev
= prev_iclog
;
1453 bp
= xfs_buf_get_uncached(mp
->m_logdev_targp
,
1454 BTOBB(log
->l_iclog_size
), 0);
1456 goto out_free_iclog
;
1458 ASSERT(xfs_buf_islocked(bp
));
1461 /* use high priority wq for log I/O completion */
1462 bp
->b_ioend_wq
= mp
->m_log_workqueue
;
1463 bp
->b_iodone
= xlog_iodone
;
1465 iclog
->ic_data
= bp
->b_addr
;
1467 log
->l_iclog_bak
[i
] = &iclog
->ic_header
;
1469 head
= &iclog
->ic_header
;
1470 memset(head
, 0, sizeof(xlog_rec_header_t
));
1471 head
->h_magicno
= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
);
1472 head
->h_version
= cpu_to_be32(
1473 xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) ? 2 : 1);
1474 head
->h_size
= cpu_to_be32(log
->l_iclog_size
);
1476 head
->h_fmt
= cpu_to_be32(XLOG_FMT
);
1477 memcpy(&head
->h_fs_uuid
, &mp
->m_sb
.sb_uuid
, sizeof(uuid_t
));
1479 iclog
->ic_size
= BBTOB(bp
->b_length
) - log
->l_iclog_hsize
;
1480 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
1481 iclog
->ic_log
= log
;
1482 atomic_set(&iclog
->ic_refcnt
, 0);
1483 spin_lock_init(&iclog
->ic_callback_lock
);
1484 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
1485 iclog
->ic_datap
= (char *)iclog
->ic_data
+ log
->l_iclog_hsize
;
1487 init_waitqueue_head(&iclog
->ic_force_wait
);
1488 init_waitqueue_head(&iclog
->ic_write_wait
);
1490 iclogp
= &iclog
->ic_next
;
1492 *iclogp
= log
->l_iclog
; /* complete ring */
1493 log
->l_iclog
->ic_prev
= prev_iclog
; /* re-write 1st prev ptr */
1495 error
= xlog_cil_init(log
);
1497 goto out_free_iclog
;
1501 for (iclog
= log
->l_iclog
; iclog
; iclog
= prev_iclog
) {
1502 prev_iclog
= iclog
->ic_next
;
1504 xfs_buf_free(iclog
->ic_bp
);
1507 spinlock_destroy(&log
->l_icloglock
);
1508 xfs_buf_free(log
->l_xbuf
);
1512 return ERR_PTR(error
);
1513 } /* xlog_alloc_log */
1517 * Write out the commit record of a transaction associated with the given
1518 * ticket. Return the lsn of the commit record.
1523 struct xlog_ticket
*ticket
,
1524 struct xlog_in_core
**iclog
,
1525 xfs_lsn_t
*commitlsnp
)
1527 struct xfs_mount
*mp
= log
->l_mp
;
1529 struct xfs_log_iovec reg
= {
1532 .i_type
= XLOG_REG_TYPE_COMMIT
,
1534 struct xfs_log_vec vec
= {
1539 ASSERT_ALWAYS(iclog
);
1540 error
= xlog_write(log
, &vec
, ticket
, commitlsnp
, iclog
,
1543 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
1548 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1549 * log space. This code pushes on the lsn which would supposedly free up
1550 * the 25% which we want to leave free. We may need to adopt a policy which
1551 * pushes on an lsn which is further along in the log once we reach the high
1552 * water mark. In this manner, we would be creating a low water mark.
1555 xlog_grant_push_ail(
1559 xfs_lsn_t threshold_lsn
= 0;
1560 xfs_lsn_t last_sync_lsn
;
1563 int threshold_block
;
1564 int threshold_cycle
;
1567 ASSERT(BTOBB(need_bytes
) < log
->l_logBBsize
);
1569 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
1570 free_blocks
= BTOBBT(free_bytes
);
1573 * Set the threshold for the minimum number of free blocks in the
1574 * log to the maximum of what the caller needs, one quarter of the
1575 * log, and 256 blocks.
1577 free_threshold
= BTOBB(need_bytes
);
1578 free_threshold
= MAX(free_threshold
, (log
->l_logBBsize
>> 2));
1579 free_threshold
= MAX(free_threshold
, 256);
1580 if (free_blocks
>= free_threshold
)
1583 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &threshold_cycle
,
1585 threshold_block
+= free_threshold
;
1586 if (threshold_block
>= log
->l_logBBsize
) {
1587 threshold_block
-= log
->l_logBBsize
;
1588 threshold_cycle
+= 1;
1590 threshold_lsn
= xlog_assign_lsn(threshold_cycle
,
1593 * Don't pass in an lsn greater than the lsn of the last
1594 * log record known to be on disk. Use a snapshot of the last sync lsn
1595 * so that it doesn't change between the compare and the set.
1597 last_sync_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
1598 if (XFS_LSN_CMP(threshold_lsn
, last_sync_lsn
) > 0)
1599 threshold_lsn
= last_sync_lsn
;
1602 * Get the transaction layer to kick the dirty buffers out to
1603 * disk asynchronously. No point in trying to do this if
1604 * the filesystem is shutting down.
1606 if (!XLOG_FORCED_SHUTDOWN(log
))
1607 xfs_ail_push(log
->l_ailp
, threshold_lsn
);
1611 * Stamp cycle number in every block
1616 struct xlog_in_core
*iclog
,
1620 int size
= iclog
->ic_offset
+ roundoff
;
1624 cycle_lsn
= CYCLE_LSN_DISK(iclog
->ic_header
.h_lsn
);
1626 dp
= iclog
->ic_datap
;
1627 for (i
= 0; i
< BTOBB(size
); i
++) {
1628 if (i
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
))
1630 iclog
->ic_header
.h_cycle_data
[i
] = *(__be32
*)dp
;
1631 *(__be32
*)dp
= cycle_lsn
;
1635 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
1636 xlog_in_core_2_t
*xhdr
= iclog
->ic_data
;
1638 for ( ; i
< BTOBB(size
); i
++) {
1639 j
= i
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
1640 k
= i
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
1641 xhdr
[j
].hic_xheader
.xh_cycle_data
[k
] = *(__be32
*)dp
;
1642 *(__be32
*)dp
= cycle_lsn
;
1646 for (i
= 1; i
< log
->l_iclog_heads
; i
++)
1647 xhdr
[i
].hic_xheader
.xh_cycle
= cycle_lsn
;
1652 * Calculate the checksum for a log buffer.
1654 * This is a little more complicated than it should be because the various
1655 * headers and the actual data are non-contiguous.
1660 struct xlog_rec_header
*rhead
,
1666 /* first generate the crc for the record header ... */
1667 crc
= xfs_start_cksum((char *)rhead
,
1668 sizeof(struct xlog_rec_header
),
1669 offsetof(struct xlog_rec_header
, h_crc
));
1671 /* ... then for additional cycle data for v2 logs ... */
1672 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
)) {
1673 union xlog_in_core2
*xhdr
= (union xlog_in_core2
*)rhead
;
1677 xheads
= size
/ XLOG_HEADER_CYCLE_SIZE
;
1678 if (size
% XLOG_HEADER_CYCLE_SIZE
)
1681 for (i
= 1; i
< xheads
; i
++) {
1682 crc
= crc32c(crc
, &xhdr
[i
].hic_xheader
,
1683 sizeof(struct xlog_rec_ext_header
));
1687 /* ... and finally for the payload */
1688 crc
= crc32c(crc
, dp
, size
);
1690 return xfs_end_cksum(crc
);
1694 * The bdstrat callback function for log bufs. This gives us a central
1695 * place to trap bufs in case we get hit by a log I/O error and need to
1696 * shutdown. Actually, in practice, even when we didn't get a log error,
1697 * we transition the iclogs to IOERROR state *after* flushing all existing
1698 * iclogs to disk. This is because we don't want anymore new transactions to be
1699 * started or completed afterwards.
1701 * We lock the iclogbufs here so that we can serialise against IO completion
1702 * during unmount. We might be processing a shutdown triggered during unmount,
1703 * and that can occur asynchronously to the unmount thread, and hence we need to
1704 * ensure that completes before tearing down the iclogbufs. Hence we need to
1705 * hold the buffer lock across the log IO to acheive that.
1711 struct xlog_in_core
*iclog
= bp
->b_fspriv
;
1714 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
1715 xfs_buf_ioerror(bp
, -EIO
);
1719 * It would seem logical to return EIO here, but we rely on
1720 * the log state machine to propagate I/O errors instead of
1721 * doing it here. Similarly, IO completion will unlock the
1722 * buffer, so we don't do it here.
1732 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1733 * fashion. Previously, we should have moved the current iclog
1734 * ptr in the log to point to the next available iclog. This allows further
1735 * write to continue while this code syncs out an iclog ready to go.
1736 * Before an in-core log can be written out, the data section must be scanned
1737 * to save away the 1st word of each BBSIZE block into the header. We replace
1738 * it with the current cycle count. Each BBSIZE block is tagged with the
1739 * cycle count because there in an implicit assumption that drives will
1740 * guarantee that entire 512 byte blocks get written at once. In other words,
1741 * we can't have part of a 512 byte block written and part not written. By
1742 * tagging each block, we will know which blocks are valid when recovering
1743 * after an unclean shutdown.
1745 * This routine is single threaded on the iclog. No other thread can be in
1746 * this routine with the same iclog. Changing contents of iclog can there-
1747 * fore be done without grabbing the state machine lock. Updating the global
1748 * log will require grabbing the lock though.
1750 * The entire log manager uses a logical block numbering scheme. Only
1751 * log_sync (and then only bwrite()) know about the fact that the log may
1752 * not start with block zero on a given device. The log block start offset
1753 * is added immediately before calling bwrite().
1759 struct xlog_in_core
*iclog
)
1763 uint count
; /* byte count of bwrite */
1764 uint count_init
; /* initial count before roundup */
1765 int roundoff
; /* roundoff to BB or stripe */
1766 int split
= 0; /* split write into two regions */
1768 int v2
= xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
);
1771 XFS_STATS_INC(xs_log_writes
);
1772 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
1774 /* Add for LR header */
1775 count_init
= log
->l_iclog_hsize
+ iclog
->ic_offset
;
1777 /* Round out the log write size */
1778 if (v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1) {
1779 /* we have a v2 stripe unit to use */
1780 count
= XLOG_LSUNITTOB(log
, XLOG_BTOLSUNIT(log
, count_init
));
1782 count
= BBTOB(BTOBB(count_init
));
1784 roundoff
= count
- count_init
;
1785 ASSERT(roundoff
>= 0);
1786 ASSERT((v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1 &&
1787 roundoff
< log
->l_mp
->m_sb
.sb_logsunit
)
1789 (log
->l_mp
->m_sb
.sb_logsunit
<= 1 &&
1790 roundoff
< BBTOB(1)));
1792 /* move grant heads by roundoff in sync */
1793 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
, roundoff
);
1794 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, roundoff
);
1796 /* put cycle number in every block */
1797 xlog_pack_data(log
, iclog
, roundoff
);
1799 /* real byte length */
1800 size
= iclog
->ic_offset
;
1803 iclog
->ic_header
.h_len
= cpu_to_be32(size
);
1806 XFS_BUF_SET_ADDR(bp
, BLOCK_LSN(be64_to_cpu(iclog
->ic_header
.h_lsn
)));
1808 XFS_STATS_ADD(xs_log_blocks
, BTOBB(count
));
1810 /* Do we need to split this write into 2 parts? */
1811 if (XFS_BUF_ADDR(bp
) + BTOBB(count
) > log
->l_logBBsize
) {
1814 split
= count
- (BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
)));
1815 count
= BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
));
1816 iclog
->ic_bwritecnt
= 2;
1819 * Bump the cycle numbers at the start of each block in the
1820 * part of the iclog that ends up in the buffer that gets
1821 * written to the start of the log.
1823 * Watch out for the header magic number case, though.
1825 dptr
= (char *)&iclog
->ic_header
+ count
;
1826 for (i
= 0; i
< split
; i
+= BBSIZE
) {
1827 __uint32_t cycle
= be32_to_cpu(*(__be32
*)dptr
);
1828 if (++cycle
== XLOG_HEADER_MAGIC_NUM
)
1830 *(__be32
*)dptr
= cpu_to_be32(cycle
);
1835 iclog
->ic_bwritecnt
= 1;
1838 /* calculcate the checksum */
1839 iclog
->ic_header
.h_crc
= xlog_cksum(log
, &iclog
->ic_header
,
1840 iclog
->ic_datap
, size
);
1842 bp
->b_io_length
= BTOBB(count
);
1843 bp
->b_fspriv
= iclog
;
1844 XFS_BUF_ZEROFLAGS(bp
);
1846 bp
->b_flags
|= XBF_SYNCIO
;
1848 if (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
) {
1849 bp
->b_flags
|= XBF_FUA
;
1852 * Flush the data device before flushing the log to make
1853 * sure all meta data written back from the AIL actually made
1854 * it to disk before stamping the new log tail LSN into the
1855 * log buffer. For an external log we need to issue the
1856 * flush explicitly, and unfortunately synchronously here;
1857 * for an internal log we can simply use the block layer
1858 * state machine for preflushes.
1860 if (log
->l_mp
->m_logdev_targp
!= log
->l_mp
->m_ddev_targp
)
1861 xfs_blkdev_issue_flush(log
->l_mp
->m_ddev_targp
);
1863 bp
->b_flags
|= XBF_FLUSH
;
1866 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1867 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1869 xlog_verify_iclog(log
, iclog
, count
, true);
1871 /* account for log which doesn't start at block #0 */
1872 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1874 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1879 error
= xlog_bdstrat(bp
);
1881 xfs_buf_ioerror_alert(bp
, "xlog_sync");
1885 bp
= iclog
->ic_log
->l_xbuf
;
1886 XFS_BUF_SET_ADDR(bp
, 0); /* logical 0 */
1887 xfs_buf_associate_memory(bp
,
1888 (char *)&iclog
->ic_header
+ count
, split
);
1889 bp
->b_fspriv
= iclog
;
1890 XFS_BUF_ZEROFLAGS(bp
);
1892 bp
->b_flags
|= XBF_SYNCIO
;
1893 if (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
)
1894 bp
->b_flags
|= XBF_FUA
;
1896 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1897 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1899 /* account for internal log which doesn't start at block #0 */
1900 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1902 error
= xlog_bdstrat(bp
);
1904 xfs_buf_ioerror_alert(bp
, "xlog_sync (split)");
1912 * Deallocate a log structure
1918 xlog_in_core_t
*iclog
, *next_iclog
;
1921 xlog_cil_destroy(log
);
1924 * Cycle all the iclogbuf locks to make sure all log IO completion
1925 * is done before we tear down these buffers.
1927 iclog
= log
->l_iclog
;
1928 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
1929 xfs_buf_lock(iclog
->ic_bp
);
1930 xfs_buf_unlock(iclog
->ic_bp
);
1931 iclog
= iclog
->ic_next
;
1935 * Always need to ensure that the extra buffer does not point to memory
1936 * owned by another log buffer before we free it. Also, cycle the lock
1937 * first to ensure we've completed IO on it.
1939 xfs_buf_lock(log
->l_xbuf
);
1940 xfs_buf_unlock(log
->l_xbuf
);
1941 xfs_buf_set_empty(log
->l_xbuf
, BTOBB(log
->l_iclog_size
));
1942 xfs_buf_free(log
->l_xbuf
);
1944 iclog
= log
->l_iclog
;
1945 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
1946 xfs_buf_free(iclog
->ic_bp
);
1947 next_iclog
= iclog
->ic_next
;
1951 spinlock_destroy(&log
->l_icloglock
);
1953 log
->l_mp
->m_log
= NULL
;
1955 } /* xlog_dealloc_log */
1958 * Update counters atomically now that memcpy is done.
1962 xlog_state_finish_copy(
1964 struct xlog_in_core
*iclog
,
1968 spin_lock(&log
->l_icloglock
);
1970 be32_add_cpu(&iclog
->ic_header
.h_num_logops
, record_cnt
);
1971 iclog
->ic_offset
+= copy_bytes
;
1973 spin_unlock(&log
->l_icloglock
);
1974 } /* xlog_state_finish_copy */
1980 * print out info relating to regions written which consume
1985 struct xfs_mount
*mp
,
1986 struct xlog_ticket
*ticket
)
1989 uint ophdr_spc
= ticket
->t_res_num_ophdrs
* (uint
)sizeof(xlog_op_header_t
);
1991 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1992 static char *res_type_str
[XLOG_REG_TYPE_MAX
] = {
2013 static char *trans_type_str
[XFS_TRANS_TYPE_MAX
] = {
2056 xfs_warn(mp
, "xlog_write: reservation summary:");
2057 xfs_warn(mp
, " trans type = %s (%u)",
2058 ((ticket
->t_trans_type
<= 0 ||
2059 ticket
->t_trans_type
> XFS_TRANS_TYPE_MAX
) ?
2060 "bad-trans-type" : trans_type_str
[ticket
->t_trans_type
-1]),
2061 ticket
->t_trans_type
);
2062 xfs_warn(mp
, " unit res = %d bytes",
2063 ticket
->t_unit_res
);
2064 xfs_warn(mp
, " current res = %d bytes",
2065 ticket
->t_curr_res
);
2066 xfs_warn(mp
, " total reg = %u bytes (o/flow = %u bytes)",
2067 ticket
->t_res_arr_sum
, ticket
->t_res_o_flow
);
2068 xfs_warn(mp
, " ophdrs = %u (ophdr space = %u bytes)",
2069 ticket
->t_res_num_ophdrs
, ophdr_spc
);
2070 xfs_warn(mp
, " ophdr + reg = %u bytes",
2071 ticket
->t_res_arr_sum
+ ticket
->t_res_o_flow
+ ophdr_spc
);
2072 xfs_warn(mp
, " num regions = %u",
2075 for (i
= 0; i
< ticket
->t_res_num
; i
++) {
2076 uint r_type
= ticket
->t_res_arr
[i
].r_type
;
2077 xfs_warn(mp
, "region[%u]: %s - %u bytes", i
,
2078 ((r_type
<= 0 || r_type
> XLOG_REG_TYPE_MAX
) ?
2079 "bad-rtype" : res_type_str
[r_type
-1]),
2080 ticket
->t_res_arr
[i
].r_len
);
2083 xfs_alert_tag(mp
, XFS_PTAG_LOGRES
,
2084 "xlog_write: reservation ran out. Need to up reservation");
2085 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
2089 * Calculate the potential space needed by the log vector. Each region gets
2090 * its own xlog_op_header_t and may need to be double word aligned.
2093 xlog_write_calc_vec_length(
2094 struct xlog_ticket
*ticket
,
2095 struct xfs_log_vec
*log_vector
)
2097 struct xfs_log_vec
*lv
;
2102 /* acct for start rec of xact */
2103 if (ticket
->t_flags
& XLOG_TIC_INITED
)
2106 for (lv
= log_vector
; lv
; lv
= lv
->lv_next
) {
2107 /* we don't write ordered log vectors */
2108 if (lv
->lv_buf_len
== XFS_LOG_VEC_ORDERED
)
2111 headers
+= lv
->lv_niovecs
;
2113 for (i
= 0; i
< lv
->lv_niovecs
; i
++) {
2114 struct xfs_log_iovec
*vecp
= &lv
->lv_iovecp
[i
];
2117 xlog_tic_add_region(ticket
, vecp
->i_len
, vecp
->i_type
);
2121 ticket
->t_res_num_ophdrs
+= headers
;
2122 len
+= headers
* sizeof(struct xlog_op_header
);
2128 * If first write for transaction, insert start record We can't be trying to
2129 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2132 xlog_write_start_rec(
2133 struct xlog_op_header
*ophdr
,
2134 struct xlog_ticket
*ticket
)
2136 if (!(ticket
->t_flags
& XLOG_TIC_INITED
))
2139 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
2140 ophdr
->oh_clientid
= ticket
->t_clientid
;
2142 ophdr
->oh_flags
= XLOG_START_TRANS
;
2145 ticket
->t_flags
&= ~XLOG_TIC_INITED
;
2147 return sizeof(struct xlog_op_header
);
2150 static xlog_op_header_t
*
2151 xlog_write_setup_ophdr(
2153 struct xlog_op_header
*ophdr
,
2154 struct xlog_ticket
*ticket
,
2157 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
2158 ophdr
->oh_clientid
= ticket
->t_clientid
;
2161 /* are we copying a commit or unmount record? */
2162 ophdr
->oh_flags
= flags
;
2165 * We've seen logs corrupted with bad transaction client ids. This
2166 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2167 * and shut down the filesystem.
2169 switch (ophdr
->oh_clientid
) {
2170 case XFS_TRANSACTION
:
2176 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2177 ophdr
->oh_clientid
, ticket
);
2185 * Set up the parameters of the region copy into the log. This has
2186 * to handle region write split across multiple log buffers - this
2187 * state is kept external to this function so that this code can
2188 * be written in an obvious, self documenting manner.
2191 xlog_write_setup_copy(
2192 struct xlog_ticket
*ticket
,
2193 struct xlog_op_header
*ophdr
,
2194 int space_available
,
2198 int *last_was_partial_copy
,
2199 int *bytes_consumed
)
2203 still_to_copy
= space_required
- *bytes_consumed
;
2204 *copy_off
= *bytes_consumed
;
2206 if (still_to_copy
<= space_available
) {
2207 /* write of region completes here */
2208 *copy_len
= still_to_copy
;
2209 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
2210 if (*last_was_partial_copy
)
2211 ophdr
->oh_flags
|= (XLOG_END_TRANS
|XLOG_WAS_CONT_TRANS
);
2212 *last_was_partial_copy
= 0;
2213 *bytes_consumed
= 0;
2217 /* partial write of region, needs extra log op header reservation */
2218 *copy_len
= space_available
;
2219 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
2220 ophdr
->oh_flags
|= XLOG_CONTINUE_TRANS
;
2221 if (*last_was_partial_copy
)
2222 ophdr
->oh_flags
|= XLOG_WAS_CONT_TRANS
;
2223 *bytes_consumed
+= *copy_len
;
2224 (*last_was_partial_copy
)++;
2226 /* account for new log op header */
2227 ticket
->t_curr_res
-= sizeof(struct xlog_op_header
);
2228 ticket
->t_res_num_ophdrs
++;
2230 return sizeof(struct xlog_op_header
);
2234 xlog_write_copy_finish(
2236 struct xlog_in_core
*iclog
,
2241 int *partial_copy_len
,
2243 struct xlog_in_core
**commit_iclog
)
2245 if (*partial_copy
) {
2247 * This iclog has already been marked WANT_SYNC by
2248 * xlog_state_get_iclog_space.
2250 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
2253 return xlog_state_release_iclog(log
, iclog
);
2257 *partial_copy_len
= 0;
2259 if (iclog
->ic_size
- log_offset
<= sizeof(xlog_op_header_t
)) {
2260 /* no more space in this iclog - push it. */
2261 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
2265 spin_lock(&log
->l_icloglock
);
2266 xlog_state_want_sync(log
, iclog
);
2267 spin_unlock(&log
->l_icloglock
);
2270 return xlog_state_release_iclog(log
, iclog
);
2271 ASSERT(flags
& XLOG_COMMIT_TRANS
);
2272 *commit_iclog
= iclog
;
2279 * Write some region out to in-core log
2281 * This will be called when writing externally provided regions or when
2282 * writing out a commit record for a given transaction.
2284 * General algorithm:
2285 * 1. Find total length of this write. This may include adding to the
2286 * lengths passed in.
2287 * 2. Check whether we violate the tickets reservation.
2288 * 3. While writing to this iclog
2289 * A. Reserve as much space in this iclog as can get
2290 * B. If this is first write, save away start lsn
2291 * C. While writing this region:
2292 * 1. If first write of transaction, write start record
2293 * 2. Write log operation header (header per region)
2294 * 3. Find out if we can fit entire region into this iclog
2295 * 4. Potentially, verify destination memcpy ptr
2296 * 5. Memcpy (partial) region
2297 * 6. If partial copy, release iclog; otherwise, continue
2298 * copying more regions into current iclog
2299 * 4. Mark want sync bit (in simulation mode)
2300 * 5. Release iclog for potential flush to on-disk log.
2303 * 1. Panic if reservation is overrun. This should never happen since
2304 * reservation amounts are generated internal to the filesystem.
2306 * 1. Tickets are single threaded data structures.
2307 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2308 * syncing routine. When a single log_write region needs to span
2309 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2310 * on all log operation writes which don't contain the end of the
2311 * region. The XLOG_END_TRANS bit is used for the in-core log
2312 * operation which contains the end of the continued log_write region.
2313 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2314 * we don't really know exactly how much space will be used. As a result,
2315 * we don't update ic_offset until the end when we know exactly how many
2316 * bytes have been written out.
2321 struct xfs_log_vec
*log_vector
,
2322 struct xlog_ticket
*ticket
,
2323 xfs_lsn_t
*start_lsn
,
2324 struct xlog_in_core
**commit_iclog
,
2327 struct xlog_in_core
*iclog
= NULL
;
2328 struct xfs_log_iovec
*vecp
;
2329 struct xfs_log_vec
*lv
;
2332 int partial_copy
= 0;
2333 int partial_copy_len
= 0;
2341 len
= xlog_write_calc_vec_length(ticket
, log_vector
);
2344 * Region headers and bytes are already accounted for.
2345 * We only need to take into account start records and
2346 * split regions in this function.
2348 if (ticket
->t_flags
& XLOG_TIC_INITED
)
2349 ticket
->t_curr_res
-= sizeof(xlog_op_header_t
);
2352 * Commit record headers need to be accounted for. These
2353 * come in as separate writes so are easy to detect.
2355 if (flags
& (XLOG_COMMIT_TRANS
| XLOG_UNMOUNT_TRANS
))
2356 ticket
->t_curr_res
-= sizeof(xlog_op_header_t
);
2358 if (ticket
->t_curr_res
< 0)
2359 xlog_print_tic_res(log
->l_mp
, ticket
);
2363 vecp
= lv
->lv_iovecp
;
2364 while (lv
&& (!lv
->lv_niovecs
|| index
< lv
->lv_niovecs
)) {
2368 error
= xlog_state_get_iclog_space(log
, len
, &iclog
, ticket
,
2369 &contwr
, &log_offset
);
2373 ASSERT(log_offset
<= iclog
->ic_size
- 1);
2374 ptr
= iclog
->ic_datap
+ log_offset
;
2376 /* start_lsn is the first lsn written to. That's all we need. */
2378 *start_lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
2381 * This loop writes out as many regions as can fit in the amount
2382 * of space which was allocated by xlog_state_get_iclog_space().
2384 while (lv
&& (!lv
->lv_niovecs
|| index
< lv
->lv_niovecs
)) {
2385 struct xfs_log_iovec
*reg
;
2386 struct xlog_op_header
*ophdr
;
2390 bool ordered
= false;
2392 /* ordered log vectors have no regions to write */
2393 if (lv
->lv_buf_len
== XFS_LOG_VEC_ORDERED
) {
2394 ASSERT(lv
->lv_niovecs
== 0);
2400 ASSERT(reg
->i_len
% sizeof(__int32_t
) == 0);
2401 ASSERT((unsigned long)ptr
% sizeof(__int32_t
) == 0);
2403 start_rec_copy
= xlog_write_start_rec(ptr
, ticket
);
2404 if (start_rec_copy
) {
2406 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2410 ophdr
= xlog_write_setup_ophdr(log
, ptr
, ticket
, flags
);
2414 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2415 sizeof(struct xlog_op_header
));
2417 len
+= xlog_write_setup_copy(ticket
, ophdr
,
2418 iclog
->ic_size
-log_offset
,
2420 ©_off
, ©_len
,
2423 xlog_verify_dest_ptr(log
, ptr
);
2426 ASSERT(copy_len
>= 0);
2427 memcpy(ptr
, reg
->i_addr
+ copy_off
, copy_len
);
2428 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
, copy_len
);
2430 copy_len
+= start_rec_copy
+ sizeof(xlog_op_header_t
);
2432 data_cnt
+= contwr
? copy_len
: 0;
2434 error
= xlog_write_copy_finish(log
, iclog
, flags
,
2435 &record_cnt
, &data_cnt
,
2444 * if we had a partial copy, we need to get more iclog
2445 * space but we don't want to increment the region
2446 * index because there is still more is this region to
2449 * If we completed writing this region, and we flushed
2450 * the iclog (indicated by resetting of the record
2451 * count), then we also need to get more log space. If
2452 * this was the last record, though, we are done and
2458 if (++index
== lv
->lv_niovecs
) {
2463 vecp
= lv
->lv_iovecp
;
2465 if (record_cnt
== 0 && ordered
== false) {
2475 xlog_state_finish_copy(log
, iclog
, record_cnt
, data_cnt
);
2477 return xlog_state_release_iclog(log
, iclog
);
2479 ASSERT(flags
& XLOG_COMMIT_TRANS
);
2480 *commit_iclog
= iclog
;
2485 /*****************************************************************************
2487 * State Machine functions
2489 *****************************************************************************
2492 /* Clean iclogs starting from the head. This ordering must be
2493 * maintained, so an iclog doesn't become ACTIVE beyond one that
2494 * is SYNCING. This is also required to maintain the notion that we use
2495 * a ordered wait queue to hold off would be writers to the log when every
2496 * iclog is trying to sync to disk.
2498 * State Change: DIRTY -> ACTIVE
2501 xlog_state_clean_log(
2504 xlog_in_core_t
*iclog
;
2507 iclog
= log
->l_iclog
;
2509 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
2510 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
2511 iclog
->ic_offset
= 0;
2512 ASSERT(iclog
->ic_callback
== NULL
);
2514 * If the number of ops in this iclog indicate it just
2515 * contains the dummy transaction, we can
2516 * change state into IDLE (the second time around).
2517 * Otherwise we should change the state into
2519 * We don't need to cover the dummy.
2522 (be32_to_cpu(iclog
->ic_header
.h_num_logops
) ==
2527 * We have two dirty iclogs so start over
2528 * This could also be num of ops indicates
2529 * this is not the dummy going out.
2533 iclog
->ic_header
.h_num_logops
= 0;
2534 memset(iclog
->ic_header
.h_cycle_data
, 0,
2535 sizeof(iclog
->ic_header
.h_cycle_data
));
2536 iclog
->ic_header
.h_lsn
= 0;
2537 } else if (iclog
->ic_state
== XLOG_STATE_ACTIVE
)
2540 break; /* stop cleaning */
2541 iclog
= iclog
->ic_next
;
2542 } while (iclog
!= log
->l_iclog
);
2544 /* log is locked when we are called */
2546 * Change state for the dummy log recording.
2547 * We usually go to NEED. But we go to NEED2 if the changed indicates
2548 * we are done writing the dummy record.
2549 * If we are done with the second dummy recored (DONE2), then
2553 switch (log
->l_covered_state
) {
2554 case XLOG_STATE_COVER_IDLE
:
2555 case XLOG_STATE_COVER_NEED
:
2556 case XLOG_STATE_COVER_NEED2
:
2557 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2560 case XLOG_STATE_COVER_DONE
:
2562 log
->l_covered_state
= XLOG_STATE_COVER_NEED2
;
2564 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2567 case XLOG_STATE_COVER_DONE2
:
2569 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
2571 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2578 } /* xlog_state_clean_log */
2581 xlog_get_lowest_lsn(
2584 xlog_in_core_t
*lsn_log
;
2585 xfs_lsn_t lowest_lsn
, lsn
;
2587 lsn_log
= log
->l_iclog
;
2590 if (!(lsn_log
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
))) {
2591 lsn
= be64_to_cpu(lsn_log
->ic_header
.h_lsn
);
2592 if ((lsn
&& !lowest_lsn
) ||
2593 (XFS_LSN_CMP(lsn
, lowest_lsn
) < 0)) {
2597 lsn_log
= lsn_log
->ic_next
;
2598 } while (lsn_log
!= log
->l_iclog
);
2604 xlog_state_do_callback(
2607 struct xlog_in_core
*ciclog
)
2609 xlog_in_core_t
*iclog
;
2610 xlog_in_core_t
*first_iclog
; /* used to know when we've
2611 * processed all iclogs once */
2612 xfs_log_callback_t
*cb
, *cb_next
;
2614 xfs_lsn_t lowest_lsn
;
2615 int ioerrors
; /* counter: iclogs with errors */
2616 int loopdidcallbacks
; /* flag: inner loop did callbacks*/
2617 int funcdidcallbacks
; /* flag: function did callbacks */
2618 int repeats
; /* for issuing console warnings if
2619 * looping too many times */
2622 spin_lock(&log
->l_icloglock
);
2623 first_iclog
= iclog
= log
->l_iclog
;
2625 funcdidcallbacks
= 0;
2630 * Scan all iclogs starting with the one pointed to by the
2631 * log. Reset this starting point each time the log is
2632 * unlocked (during callbacks).
2634 * Keep looping through iclogs until one full pass is made
2635 * without running any callbacks.
2637 first_iclog
= log
->l_iclog
;
2638 iclog
= log
->l_iclog
;
2639 loopdidcallbacks
= 0;
2644 /* skip all iclogs in the ACTIVE & DIRTY states */
2645 if (iclog
->ic_state
&
2646 (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
)) {
2647 iclog
= iclog
->ic_next
;
2652 * Between marking a filesystem SHUTDOWN and stopping
2653 * the log, we do flush all iclogs to disk (if there
2654 * wasn't a log I/O error). So, we do want things to
2655 * go smoothly in case of just a SHUTDOWN w/o a
2658 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
2660 * Can only perform callbacks in order. Since
2661 * this iclog is not in the DONE_SYNC/
2662 * DO_CALLBACK state, we skip the rest and
2663 * just try to clean up. If we set our iclog
2664 * to DO_CALLBACK, we will not process it when
2665 * we retry since a previous iclog is in the
2666 * CALLBACK and the state cannot change since
2667 * we are holding the l_icloglock.
2669 if (!(iclog
->ic_state
&
2670 (XLOG_STATE_DONE_SYNC
|
2671 XLOG_STATE_DO_CALLBACK
))) {
2672 if (ciclog
&& (ciclog
->ic_state
==
2673 XLOG_STATE_DONE_SYNC
)) {
2674 ciclog
->ic_state
= XLOG_STATE_DO_CALLBACK
;
2679 * We now have an iclog that is in either the
2680 * DO_CALLBACK or DONE_SYNC states. The other
2681 * states (WANT_SYNC, SYNCING, or CALLBACK were
2682 * caught by the above if and are going to
2683 * clean (i.e. we aren't doing their callbacks)
2688 * We will do one more check here to see if we
2689 * have chased our tail around.
2692 lowest_lsn
= xlog_get_lowest_lsn(log
);
2694 XFS_LSN_CMP(lowest_lsn
,
2695 be64_to_cpu(iclog
->ic_header
.h_lsn
)) < 0) {
2696 iclog
= iclog
->ic_next
;
2697 continue; /* Leave this iclog for
2701 iclog
->ic_state
= XLOG_STATE_CALLBACK
;
2705 * Completion of a iclog IO does not imply that
2706 * a transaction has completed, as transactions
2707 * can be large enough to span many iclogs. We
2708 * cannot change the tail of the log half way
2709 * through a transaction as this may be the only
2710 * transaction in the log and moving th etail to
2711 * point to the middle of it will prevent
2712 * recovery from finding the start of the
2713 * transaction. Hence we should only update the
2714 * last_sync_lsn if this iclog contains
2715 * transaction completion callbacks on it.
2717 * We have to do this before we drop the
2718 * icloglock to ensure we are the only one that
2721 ASSERT(XFS_LSN_CMP(atomic64_read(&log
->l_last_sync_lsn
),
2722 be64_to_cpu(iclog
->ic_header
.h_lsn
)) <= 0);
2723 if (iclog
->ic_callback
)
2724 atomic64_set(&log
->l_last_sync_lsn
,
2725 be64_to_cpu(iclog
->ic_header
.h_lsn
));
2730 spin_unlock(&log
->l_icloglock
);
2733 * Keep processing entries in the callback list until
2734 * we come around and it is empty. We need to
2735 * atomically see that the list is empty and change the
2736 * state to DIRTY so that we don't miss any more
2737 * callbacks being added.
2739 spin_lock(&iclog
->ic_callback_lock
);
2740 cb
= iclog
->ic_callback
;
2742 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
2743 iclog
->ic_callback
= NULL
;
2744 spin_unlock(&iclog
->ic_callback_lock
);
2746 /* perform callbacks in the order given */
2747 for (; cb
; cb
= cb_next
) {
2748 cb_next
= cb
->cb_next
;
2749 cb
->cb_func(cb
->cb_arg
, aborted
);
2751 spin_lock(&iclog
->ic_callback_lock
);
2752 cb
= iclog
->ic_callback
;
2758 spin_lock(&log
->l_icloglock
);
2759 ASSERT(iclog
->ic_callback
== NULL
);
2760 spin_unlock(&iclog
->ic_callback_lock
);
2761 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
))
2762 iclog
->ic_state
= XLOG_STATE_DIRTY
;
2765 * Transition from DIRTY to ACTIVE if applicable.
2766 * NOP if STATE_IOERROR.
2768 xlog_state_clean_log(log
);
2770 /* wake up threads waiting in xfs_log_force() */
2771 wake_up_all(&iclog
->ic_force_wait
);
2773 iclog
= iclog
->ic_next
;
2774 } while (first_iclog
!= iclog
);
2776 if (repeats
> 5000) {
2777 flushcnt
+= repeats
;
2780 "%s: possible infinite loop (%d iterations)",
2781 __func__
, flushcnt
);
2783 } while (!ioerrors
&& loopdidcallbacks
);
2786 * make one last gasp attempt to see if iclogs are being left in
2790 if (funcdidcallbacks
) {
2791 first_iclog
= iclog
= log
->l_iclog
;
2793 ASSERT(iclog
->ic_state
!= XLOG_STATE_DO_CALLBACK
);
2795 * Terminate the loop if iclogs are found in states
2796 * which will cause other threads to clean up iclogs.
2798 * SYNCING - i/o completion will go through logs
2799 * DONE_SYNC - interrupt thread should be waiting for
2801 * IOERROR - give up hope all ye who enter here
2803 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
||
2804 iclog
->ic_state
== XLOG_STATE_SYNCING
||
2805 iclog
->ic_state
== XLOG_STATE_DONE_SYNC
||
2806 iclog
->ic_state
== XLOG_STATE_IOERROR
)
2808 iclog
= iclog
->ic_next
;
2809 } while (first_iclog
!= iclog
);
2813 if (log
->l_iclog
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_IOERROR
))
2815 spin_unlock(&log
->l_icloglock
);
2818 wake_up_all(&log
->l_flush_wait
);
2823 * Finish transitioning this iclog to the dirty state.
2825 * Make sure that we completely execute this routine only when this is
2826 * the last call to the iclog. There is a good chance that iclog flushes,
2827 * when we reach the end of the physical log, get turned into 2 separate
2828 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2829 * routine. By using the reference count bwritecnt, we guarantee that only
2830 * the second completion goes through.
2832 * Callbacks could take time, so they are done outside the scope of the
2833 * global state machine log lock.
2836 xlog_state_done_syncing(
2837 xlog_in_core_t
*iclog
,
2840 struct xlog
*log
= iclog
->ic_log
;
2842 spin_lock(&log
->l_icloglock
);
2844 ASSERT(iclog
->ic_state
== XLOG_STATE_SYNCING
||
2845 iclog
->ic_state
== XLOG_STATE_IOERROR
);
2846 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
2847 ASSERT(iclog
->ic_bwritecnt
== 1 || iclog
->ic_bwritecnt
== 2);
2851 * If we got an error, either on the first buffer, or in the case of
2852 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2853 * and none should ever be attempted to be written to disk
2856 if (iclog
->ic_state
!= XLOG_STATE_IOERROR
) {
2857 if (--iclog
->ic_bwritecnt
== 1) {
2858 spin_unlock(&log
->l_icloglock
);
2861 iclog
->ic_state
= XLOG_STATE_DONE_SYNC
;
2865 * Someone could be sleeping prior to writing out the next
2866 * iclog buffer, we wake them all, one will get to do the
2867 * I/O, the others get to wait for the result.
2869 wake_up_all(&iclog
->ic_write_wait
);
2870 spin_unlock(&log
->l_icloglock
);
2871 xlog_state_do_callback(log
, aborted
, iclog
); /* also cleans log */
2872 } /* xlog_state_done_syncing */
2876 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2877 * sleep. We wait on the flush queue on the head iclog as that should be
2878 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2879 * we will wait here and all new writes will sleep until a sync completes.
2881 * The in-core logs are used in a circular fashion. They are not used
2882 * out-of-order even when an iclog past the head is free.
2885 * * log_offset where xlog_write() can start writing into the in-core
2887 * * in-core log pointer to which xlog_write() should write.
2888 * * boolean indicating this is a continued write to an in-core log.
2889 * If this is the last write, then the in-core log's offset field
2890 * needs to be incremented, depending on the amount of data which
2894 xlog_state_get_iclog_space(
2897 struct xlog_in_core
**iclogp
,
2898 struct xlog_ticket
*ticket
,
2899 int *continued_write
,
2903 xlog_rec_header_t
*head
;
2904 xlog_in_core_t
*iclog
;
2908 spin_lock(&log
->l_icloglock
);
2909 if (XLOG_FORCED_SHUTDOWN(log
)) {
2910 spin_unlock(&log
->l_icloglock
);
2914 iclog
= log
->l_iclog
;
2915 if (iclog
->ic_state
!= XLOG_STATE_ACTIVE
) {
2916 XFS_STATS_INC(xs_log_noiclogs
);
2918 /* Wait for log writes to have flushed */
2919 xlog_wait(&log
->l_flush_wait
, &log
->l_icloglock
);
2923 head
= &iclog
->ic_header
;
2925 atomic_inc(&iclog
->ic_refcnt
); /* prevents sync */
2926 log_offset
= iclog
->ic_offset
;
2928 /* On the 1st write to an iclog, figure out lsn. This works
2929 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2930 * committing to. If the offset is set, that's how many blocks
2933 if (log_offset
== 0) {
2934 ticket
->t_curr_res
-= log
->l_iclog_hsize
;
2935 xlog_tic_add_region(ticket
,
2937 XLOG_REG_TYPE_LRHEADER
);
2938 head
->h_cycle
= cpu_to_be32(log
->l_curr_cycle
);
2939 head
->h_lsn
= cpu_to_be64(
2940 xlog_assign_lsn(log
->l_curr_cycle
, log
->l_curr_block
));
2941 ASSERT(log
->l_curr_block
>= 0);
2944 /* If there is enough room to write everything, then do it. Otherwise,
2945 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2946 * bit is on, so this will get flushed out. Don't update ic_offset
2947 * until you know exactly how many bytes get copied. Therefore, wait
2948 * until later to update ic_offset.
2950 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2951 * can fit into remaining data section.
2953 if (iclog
->ic_size
- iclog
->ic_offset
< 2*sizeof(xlog_op_header_t
)) {
2954 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2957 * If I'm the only one writing to this iclog, sync it to disk.
2958 * We need to do an atomic compare and decrement here to avoid
2959 * racing with concurrent atomic_dec_and_lock() calls in
2960 * xlog_state_release_iclog() when there is more than one
2961 * reference to the iclog.
2963 if (!atomic_add_unless(&iclog
->ic_refcnt
, -1, 1)) {
2964 /* we are the only one */
2965 spin_unlock(&log
->l_icloglock
);
2966 error
= xlog_state_release_iclog(log
, iclog
);
2970 spin_unlock(&log
->l_icloglock
);
2975 /* Do we have enough room to write the full amount in the remainder
2976 * of this iclog? Or must we continue a write on the next iclog and
2977 * mark this iclog as completely taken? In the case where we switch
2978 * iclogs (to mark it taken), this particular iclog will release/sync
2979 * to disk in xlog_write().
2981 if (len
<= iclog
->ic_size
- iclog
->ic_offset
) {
2982 *continued_write
= 0;
2983 iclog
->ic_offset
+= len
;
2985 *continued_write
= 1;
2986 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2990 ASSERT(iclog
->ic_offset
<= iclog
->ic_size
);
2991 spin_unlock(&log
->l_icloglock
);
2993 *logoffsetp
= log_offset
;
2995 } /* xlog_state_get_iclog_space */
2997 /* The first cnt-1 times through here we don't need to
2998 * move the grant write head because the permanent
2999 * reservation has reserved cnt times the unit amount.
3000 * Release part of current permanent unit reservation and
3001 * reset current reservation to be one units worth. Also
3002 * move grant reservation head forward.
3005 xlog_regrant_reserve_log_space(
3007 struct xlog_ticket
*ticket
)
3009 trace_xfs_log_regrant_reserve_enter(log
, ticket
);
3011 if (ticket
->t_cnt
> 0)
3014 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
,
3015 ticket
->t_curr_res
);
3016 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
,
3017 ticket
->t_curr_res
);
3018 ticket
->t_curr_res
= ticket
->t_unit_res
;
3019 xlog_tic_reset_res(ticket
);
3021 trace_xfs_log_regrant_reserve_sub(log
, ticket
);
3023 /* just return if we still have some of the pre-reserved space */
3024 if (ticket
->t_cnt
> 0)
3027 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
,
3028 ticket
->t_unit_res
);
3030 trace_xfs_log_regrant_reserve_exit(log
, ticket
);
3032 ticket
->t_curr_res
= ticket
->t_unit_res
;
3033 xlog_tic_reset_res(ticket
);
3034 } /* xlog_regrant_reserve_log_space */
3038 * Give back the space left from a reservation.
3040 * All the information we need to make a correct determination of space left
3041 * is present. For non-permanent reservations, things are quite easy. The
3042 * count should have been decremented to zero. We only need to deal with the
3043 * space remaining in the current reservation part of the ticket. If the
3044 * ticket contains a permanent reservation, there may be left over space which
3045 * needs to be released. A count of N means that N-1 refills of the current
3046 * reservation can be done before we need to ask for more space. The first
3047 * one goes to fill up the first current reservation. Once we run out of
3048 * space, the count will stay at zero and the only space remaining will be
3049 * in the current reservation field.
3052 xlog_ungrant_log_space(
3054 struct xlog_ticket
*ticket
)
3058 if (ticket
->t_cnt
> 0)
3061 trace_xfs_log_ungrant_enter(log
, ticket
);
3062 trace_xfs_log_ungrant_sub(log
, ticket
);
3065 * If this is a permanent reservation ticket, we may be able to free
3066 * up more space based on the remaining count.
3068 bytes
= ticket
->t_curr_res
;
3069 if (ticket
->t_cnt
> 0) {
3070 ASSERT(ticket
->t_flags
& XLOG_TIC_PERM_RESERV
);
3071 bytes
+= ticket
->t_unit_res
*ticket
->t_cnt
;
3074 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
, bytes
);
3075 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
, bytes
);
3077 trace_xfs_log_ungrant_exit(log
, ticket
);
3079 xfs_log_space_wake(log
->l_mp
);
3083 * Flush iclog to disk if this is the last reference to the given iclog and
3084 * the WANT_SYNC bit is set.
3086 * When this function is entered, the iclog is not necessarily in the
3087 * WANT_SYNC state. It may be sitting around waiting to get filled.
3092 xlog_state_release_iclog(
3094 struct xlog_in_core
*iclog
)
3096 int sync
= 0; /* do we sync? */
3098 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3101 ASSERT(atomic_read(&iclog
->ic_refcnt
) > 0);
3102 if (!atomic_dec_and_lock(&iclog
->ic_refcnt
, &log
->l_icloglock
))
3105 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3106 spin_unlock(&log
->l_icloglock
);
3109 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
||
3110 iclog
->ic_state
== XLOG_STATE_WANT_SYNC
);
3112 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
) {
3113 /* update tail before writing to iclog */
3114 xfs_lsn_t tail_lsn
= xlog_assign_tail_lsn(log
->l_mp
);
3116 iclog
->ic_state
= XLOG_STATE_SYNCING
;
3117 iclog
->ic_header
.h_tail_lsn
= cpu_to_be64(tail_lsn
);
3118 xlog_verify_tail_lsn(log
, iclog
, tail_lsn
);
3119 /* cycle incremented when incrementing curr_block */
3121 spin_unlock(&log
->l_icloglock
);
3124 * We let the log lock go, so it's possible that we hit a log I/O
3125 * error or some other SHUTDOWN condition that marks the iclog
3126 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3127 * this iclog has consistent data, so we ignore IOERROR
3128 * flags after this point.
3131 return xlog_sync(log
, iclog
);
3133 } /* xlog_state_release_iclog */
3137 * This routine will mark the current iclog in the ring as WANT_SYNC
3138 * and move the current iclog pointer to the next iclog in the ring.
3139 * When this routine is called from xlog_state_get_iclog_space(), the
3140 * exact size of the iclog has not yet been determined. All we know is
3141 * that every data block. We have run out of space in this log record.
3144 xlog_state_switch_iclogs(
3146 struct xlog_in_core
*iclog
,
3149 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
);
3151 eventual_size
= iclog
->ic_offset
;
3152 iclog
->ic_state
= XLOG_STATE_WANT_SYNC
;
3153 iclog
->ic_header
.h_prev_block
= cpu_to_be32(log
->l_prev_block
);
3154 log
->l_prev_block
= log
->l_curr_block
;
3155 log
->l_prev_cycle
= log
->l_curr_cycle
;
3157 /* roll log?: ic_offset changed later */
3158 log
->l_curr_block
+= BTOBB(eventual_size
)+BTOBB(log
->l_iclog_hsize
);
3160 /* Round up to next log-sunit */
3161 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) &&
3162 log
->l_mp
->m_sb
.sb_logsunit
> 1) {
3163 __uint32_t sunit_bb
= BTOBB(log
->l_mp
->m_sb
.sb_logsunit
);
3164 log
->l_curr_block
= roundup(log
->l_curr_block
, sunit_bb
);
3167 if (log
->l_curr_block
>= log
->l_logBBsize
) {
3168 log
->l_curr_cycle
++;
3169 if (log
->l_curr_cycle
== XLOG_HEADER_MAGIC_NUM
)
3170 log
->l_curr_cycle
++;
3171 log
->l_curr_block
-= log
->l_logBBsize
;
3172 ASSERT(log
->l_curr_block
>= 0);
3174 ASSERT(iclog
== log
->l_iclog
);
3175 log
->l_iclog
= iclog
->ic_next
;
3176 } /* xlog_state_switch_iclogs */
3179 * Write out all data in the in-core log as of this exact moment in time.
3181 * Data may be written to the in-core log during this call. However,
3182 * we don't guarantee this data will be written out. A change from past
3183 * implementation means this routine will *not* write out zero length LRs.
3185 * Basically, we try and perform an intelligent scan of the in-core logs.
3186 * If we determine there is no flushable data, we just return. There is no
3187 * flushable data if:
3189 * 1. the current iclog is active and has no data; the previous iclog
3190 * is in the active or dirty state.
3191 * 2. the current iclog is drity, and the previous iclog is in the
3192 * active or dirty state.
3196 * 1. the current iclog is not in the active nor dirty state.
3197 * 2. the current iclog dirty, and the previous iclog is not in the
3198 * active nor dirty state.
3199 * 3. the current iclog is active, and there is another thread writing
3200 * to this particular iclog.
3201 * 4. a) the current iclog is active and has no other writers
3202 * b) when we return from flushing out this iclog, it is still
3203 * not in the active nor dirty state.
3207 struct xfs_mount
*mp
,
3211 struct xlog
*log
= mp
->m_log
;
3212 struct xlog_in_core
*iclog
;
3215 XFS_STATS_INC(xs_log_force
);
3217 xlog_cil_force(log
);
3219 spin_lock(&log
->l_icloglock
);
3221 iclog
= log
->l_iclog
;
3222 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3223 spin_unlock(&log
->l_icloglock
);
3227 /* If the head iclog is not active nor dirty, we just attach
3228 * ourselves to the head and go to sleep.
3230 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
3231 iclog
->ic_state
== XLOG_STATE_DIRTY
) {
3233 * If the head is dirty or (active and empty), then
3234 * we need to look at the previous iclog. If the previous
3235 * iclog is active or dirty we are done. There is nothing
3236 * to sync out. Otherwise, we attach ourselves to the
3237 * previous iclog and go to sleep.
3239 if (iclog
->ic_state
== XLOG_STATE_DIRTY
||
3240 (atomic_read(&iclog
->ic_refcnt
) == 0
3241 && iclog
->ic_offset
== 0)) {
3242 iclog
= iclog
->ic_prev
;
3243 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
3244 iclog
->ic_state
== XLOG_STATE_DIRTY
)
3249 if (atomic_read(&iclog
->ic_refcnt
) == 0) {
3250 /* We are the only one with access to this
3251 * iclog. Flush it out now. There should
3252 * be a roundoff of zero to show that someone
3253 * has already taken care of the roundoff from
3254 * the previous sync.
3256 atomic_inc(&iclog
->ic_refcnt
);
3257 lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
3258 xlog_state_switch_iclogs(log
, iclog
, 0);
3259 spin_unlock(&log
->l_icloglock
);
3261 if (xlog_state_release_iclog(log
, iclog
))
3266 spin_lock(&log
->l_icloglock
);
3267 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) == lsn
&&
3268 iclog
->ic_state
!= XLOG_STATE_DIRTY
)
3273 /* Someone else is writing to this iclog.
3274 * Use its call to flush out the data. However,
3275 * the other thread may not force out this LR,
3276 * so we mark it WANT_SYNC.
3278 xlog_state_switch_iclogs(log
, iclog
, 0);
3284 /* By the time we come around again, the iclog could've been filled
3285 * which would give it another lsn. If we have a new lsn, just
3286 * return because the relevant data has been flushed.
3289 if (flags
& XFS_LOG_SYNC
) {
3291 * We must check if we're shutting down here, before
3292 * we wait, while we're holding the l_icloglock.
3293 * Then we check again after waking up, in case our
3294 * sleep was disturbed by a bad news.
3296 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3297 spin_unlock(&log
->l_icloglock
);
3300 XFS_STATS_INC(xs_log_force_sleep
);
3301 xlog_wait(&iclog
->ic_force_wait
, &log
->l_icloglock
);
3303 * No need to grab the log lock here since we're
3304 * only deciding whether or not to return EIO
3305 * and the memory read should be atomic.
3307 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3314 spin_unlock(&log
->l_icloglock
);
3320 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3321 * about errors or whether the log was flushed or not. This is the normal
3322 * interface to use when trying to unpin items or move the log forward.
3331 trace_xfs_log_force(mp
, 0);
3332 error
= _xfs_log_force(mp
, flags
, NULL
);
3334 xfs_warn(mp
, "%s: error %d returned.", __func__
, error
);
3338 * Force the in-core log to disk for a specific LSN.
3340 * Find in-core log with lsn.
3341 * If it is in the DIRTY state, just return.
3342 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3343 * state and go to sleep or return.
3344 * If it is in any other state, go to sleep or return.
3346 * Synchronous forces are implemented with a signal variable. All callers
3347 * to force a given lsn to disk will wait on a the sv attached to the
3348 * specific in-core log. When given in-core log finally completes its
3349 * write to disk, that thread will wake up all threads waiting on the
3354 struct xfs_mount
*mp
,
3359 struct xlog
*log
= mp
->m_log
;
3360 struct xlog_in_core
*iclog
;
3361 int already_slept
= 0;
3365 XFS_STATS_INC(xs_log_force
);
3367 lsn
= xlog_cil_force_lsn(log
, lsn
);
3368 if (lsn
== NULLCOMMITLSN
)
3372 spin_lock(&log
->l_icloglock
);
3373 iclog
= log
->l_iclog
;
3374 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3375 spin_unlock(&log
->l_icloglock
);
3380 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) != lsn
) {
3381 iclog
= iclog
->ic_next
;
3385 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
3386 spin_unlock(&log
->l_icloglock
);
3390 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3392 * We sleep here if we haven't already slept (e.g.
3393 * this is the first time we've looked at the correct
3394 * iclog buf) and the buffer before us is going to
3395 * be sync'ed. The reason for this is that if we
3396 * are doing sync transactions here, by waiting for
3397 * the previous I/O to complete, we can allow a few
3398 * more transactions into this iclog before we close
3401 * Otherwise, we mark the buffer WANT_SYNC, and bump
3402 * up the refcnt so we can release the log (which
3403 * drops the ref count). The state switch keeps new
3404 * transaction commits from using this buffer. When
3405 * the current commits finish writing into the buffer,
3406 * the refcount will drop to zero and the buffer will
3409 if (!already_slept
&&
3410 (iclog
->ic_prev
->ic_state
&
3411 (XLOG_STATE_WANT_SYNC
| XLOG_STATE_SYNCING
))) {
3412 ASSERT(!(iclog
->ic_state
& XLOG_STATE_IOERROR
));
3414 XFS_STATS_INC(xs_log_force_sleep
);
3416 xlog_wait(&iclog
->ic_prev
->ic_write_wait
,
3423 atomic_inc(&iclog
->ic_refcnt
);
3424 xlog_state_switch_iclogs(log
, iclog
, 0);
3425 spin_unlock(&log
->l_icloglock
);
3426 if (xlog_state_release_iclog(log
, iclog
))
3430 spin_lock(&log
->l_icloglock
);
3433 if ((flags
& XFS_LOG_SYNC
) && /* sleep */
3435 (XLOG_STATE_ACTIVE
| XLOG_STATE_DIRTY
))) {
3437 * Don't wait on completion if we know that we've
3438 * gotten a log write error.
3440 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3441 spin_unlock(&log
->l_icloglock
);
3444 XFS_STATS_INC(xs_log_force_sleep
);
3445 xlog_wait(&iclog
->ic_force_wait
, &log
->l_icloglock
);
3447 * No need to grab the log lock here since we're
3448 * only deciding whether or not to return EIO
3449 * and the memory read should be atomic.
3451 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3456 } else { /* just return */
3457 spin_unlock(&log
->l_icloglock
);
3461 } while (iclog
!= log
->l_iclog
);
3463 spin_unlock(&log
->l_icloglock
);
3468 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3469 * about errors or whether the log was flushed or not. This is the normal
3470 * interface to use when trying to unpin items or move the log forward.
3480 trace_xfs_log_force(mp
, lsn
);
3481 error
= _xfs_log_force_lsn(mp
, lsn
, flags
, NULL
);
3483 xfs_warn(mp
, "%s: error %d returned.", __func__
, error
);
3487 * Called when we want to mark the current iclog as being ready to sync to
3491 xlog_state_want_sync(
3493 struct xlog_in_core
*iclog
)
3495 assert_spin_locked(&log
->l_icloglock
);
3497 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3498 xlog_state_switch_iclogs(log
, iclog
, 0);
3500 ASSERT(iclog
->ic_state
&
3501 (XLOG_STATE_WANT_SYNC
|XLOG_STATE_IOERROR
));
3506 /*****************************************************************************
3510 *****************************************************************************
3514 * Free a used ticket when its refcount falls to zero.
3518 xlog_ticket_t
*ticket
)
3520 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3521 if (atomic_dec_and_test(&ticket
->t_ref
))
3522 kmem_zone_free(xfs_log_ticket_zone
, ticket
);
3527 xlog_ticket_t
*ticket
)
3529 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3530 atomic_inc(&ticket
->t_ref
);
3535 * Figure out the total log space unit (in bytes) that would be
3536 * required for a log ticket.
3539 xfs_log_calc_unit_res(
3540 struct xfs_mount
*mp
,
3543 struct xlog
*log
= mp
->m_log
;
3548 * Permanent reservations have up to 'cnt'-1 active log operations
3549 * in the log. A unit in this case is the amount of space for one
3550 * of these log operations. Normal reservations have a cnt of 1
3551 * and their unit amount is the total amount of space required.
3553 * The following lines of code account for non-transaction data
3554 * which occupy space in the on-disk log.
3556 * Normal form of a transaction is:
3557 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3558 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3560 * We need to account for all the leadup data and trailer data
3561 * around the transaction data.
3562 * And then we need to account for the worst case in terms of using
3564 * The worst case will happen if:
3565 * - the placement of the transaction happens to be such that the
3566 * roundoff is at its maximum
3567 * - the transaction data is synced before the commit record is synced
3568 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3569 * Therefore the commit record is in its own Log Record.
3570 * This can happen as the commit record is called with its
3571 * own region to xlog_write().
3572 * This then means that in the worst case, roundoff can happen for
3573 * the commit-rec as well.
3574 * The commit-rec is smaller than padding in this scenario and so it is
3575 * not added separately.
3578 /* for trans header */
3579 unit_bytes
+= sizeof(xlog_op_header_t
);
3580 unit_bytes
+= sizeof(xfs_trans_header_t
);
3583 unit_bytes
+= sizeof(xlog_op_header_t
);
3586 * for LR headers - the space for data in an iclog is the size minus
3587 * the space used for the headers. If we use the iclog size, then we
3588 * undercalculate the number of headers required.
3590 * Furthermore - the addition of op headers for split-recs might
3591 * increase the space required enough to require more log and op
3592 * headers, so take that into account too.
3594 * IMPORTANT: This reservation makes the assumption that if this
3595 * transaction is the first in an iclog and hence has the LR headers
3596 * accounted to it, then the remaining space in the iclog is
3597 * exclusively for this transaction. i.e. if the transaction is larger
3598 * than the iclog, it will be the only thing in that iclog.
3599 * Fundamentally, this means we must pass the entire log vector to
3600 * xlog_write to guarantee this.
3602 iclog_space
= log
->l_iclog_size
- log
->l_iclog_hsize
;
3603 num_headers
= howmany(unit_bytes
, iclog_space
);
3605 /* for split-recs - ophdrs added when data split over LRs */
3606 unit_bytes
+= sizeof(xlog_op_header_t
) * num_headers
;
3608 /* add extra header reservations if we overrun */
3609 while (!num_headers
||
3610 howmany(unit_bytes
, iclog_space
) > num_headers
) {
3611 unit_bytes
+= sizeof(xlog_op_header_t
);
3614 unit_bytes
+= log
->l_iclog_hsize
* num_headers
;
3616 /* for commit-rec LR header - note: padding will subsume the ophdr */
3617 unit_bytes
+= log
->l_iclog_hsize
;
3619 /* for roundoff padding for transaction data and one for commit record */
3620 if (xfs_sb_version_haslogv2(&mp
->m_sb
) && mp
->m_sb
.sb_logsunit
> 1) {
3621 /* log su roundoff */
3622 unit_bytes
+= 2 * mp
->m_sb
.sb_logsunit
;
3625 unit_bytes
+= 2 * BBSIZE
;
3632 * Allocate and initialise a new log ticket.
3634 struct xlog_ticket
*
3641 xfs_km_flags_t alloc_flags
)
3643 struct xlog_ticket
*tic
;
3646 tic
= kmem_zone_zalloc(xfs_log_ticket_zone
, alloc_flags
);
3650 unit_res
= xfs_log_calc_unit_res(log
->l_mp
, unit_bytes
);
3652 atomic_set(&tic
->t_ref
, 1);
3653 tic
->t_task
= current
;
3654 INIT_LIST_HEAD(&tic
->t_queue
);
3655 tic
->t_unit_res
= unit_res
;
3656 tic
->t_curr_res
= unit_res
;
3659 tic
->t_tid
= prandom_u32();
3660 tic
->t_clientid
= client
;
3661 tic
->t_flags
= XLOG_TIC_INITED
;
3662 tic
->t_trans_type
= 0;
3664 tic
->t_flags
|= XLOG_TIC_PERM_RESERV
;
3666 xlog_tic_reset_res(tic
);
3672 /******************************************************************************
3674 * Log debug routines
3676 ******************************************************************************
3680 * Make sure that the destination ptr is within the valid data region of
3681 * one of the iclogs. This uses backup pointers stored in a different
3682 * part of the log in case we trash the log structure.
3685 xlog_verify_dest_ptr(
3692 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
3693 if (ptr
>= log
->l_iclog_bak
[i
] &&
3694 ptr
<= log
->l_iclog_bak
[i
] + log
->l_iclog_size
)
3699 xfs_emerg(log
->l_mp
, "%s: invalid ptr", __func__
);
3703 * Check to make sure the grant write head didn't just over lap the tail. If
3704 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3705 * the cycles differ by exactly one and check the byte count.
3707 * This check is run unlocked, so can give false positives. Rather than assert
3708 * on failures, use a warn-once flag and a panic tag to allow the admin to
3709 * determine if they want to panic the machine when such an error occurs. For
3710 * debug kernels this will have the same effect as using an assert but, unlinke
3711 * an assert, it can be turned off at runtime.
3714 xlog_verify_grant_tail(
3717 int tail_cycle
, tail_blocks
;
3720 xlog_crack_grant_head(&log
->l_write_head
.grant
, &cycle
, &space
);
3721 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_blocks
);
3722 if (tail_cycle
!= cycle
) {
3723 if (cycle
- 1 != tail_cycle
&&
3724 !(log
->l_flags
& XLOG_TAIL_WARN
)) {
3725 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3726 "%s: cycle - 1 != tail_cycle", __func__
);
3727 log
->l_flags
|= XLOG_TAIL_WARN
;
3730 if (space
> BBTOB(tail_blocks
) &&
3731 !(log
->l_flags
& XLOG_TAIL_WARN
)) {
3732 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3733 "%s: space > BBTOB(tail_blocks)", __func__
);
3734 log
->l_flags
|= XLOG_TAIL_WARN
;
3739 /* check if it will fit */
3741 xlog_verify_tail_lsn(
3743 struct xlog_in_core
*iclog
,
3748 if (CYCLE_LSN(tail_lsn
) == log
->l_prev_cycle
) {
3750 log
->l_logBBsize
- (log
->l_prev_block
- BLOCK_LSN(tail_lsn
));
3751 if (blocks
< BTOBB(iclog
->ic_offset
)+BTOBB(log
->l_iclog_hsize
))
3752 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3754 ASSERT(CYCLE_LSN(tail_lsn
)+1 == log
->l_prev_cycle
);
3756 if (BLOCK_LSN(tail_lsn
) == log
->l_prev_block
)
3757 xfs_emerg(log
->l_mp
, "%s: tail wrapped", __func__
);
3759 blocks
= BLOCK_LSN(tail_lsn
) - log
->l_prev_block
;
3760 if (blocks
< BTOBB(iclog
->ic_offset
) + 1)
3761 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3763 } /* xlog_verify_tail_lsn */
3766 * Perform a number of checks on the iclog before writing to disk.
3768 * 1. Make sure the iclogs are still circular
3769 * 2. Make sure we have a good magic number
3770 * 3. Make sure we don't have magic numbers in the data
3771 * 4. Check fields of each log operation header for:
3772 * A. Valid client identifier
3773 * B. tid ptr value falls in valid ptr space (user space code)
3774 * C. Length in log record header is correct according to the
3775 * individual operation headers within record.
3776 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3777 * log, check the preceding blocks of the physical log to make sure all
3778 * the cycle numbers agree with the current cycle number.
3783 struct xlog_in_core
*iclog
,
3787 xlog_op_header_t
*ophead
;
3788 xlog_in_core_t
*icptr
;
3789 xlog_in_core_2_t
*xhdr
;
3790 void *base_ptr
, *ptr
, *p
;
3791 ptrdiff_t field_offset
;
3793 int len
, i
, j
, k
, op_len
;
3796 /* check validity of iclog pointers */
3797 spin_lock(&log
->l_icloglock
);
3798 icptr
= log
->l_iclog
;
3799 for (i
= 0; i
< log
->l_iclog_bufs
; i
++, icptr
= icptr
->ic_next
)
3802 if (icptr
!= log
->l_iclog
)
3803 xfs_emerg(log
->l_mp
, "%s: corrupt iclog ring", __func__
);
3804 spin_unlock(&log
->l_icloglock
);
3806 /* check log magic numbers */
3807 if (iclog
->ic_header
.h_magicno
!= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3808 xfs_emerg(log
->l_mp
, "%s: invalid magic num", __func__
);
3810 base_ptr
= ptr
= &iclog
->ic_header
;
3811 p
= &iclog
->ic_header
;
3812 for (ptr
+= BBSIZE
; ptr
< base_ptr
+ count
; ptr
+= BBSIZE
) {
3813 if (*(__be32
*)ptr
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3814 xfs_emerg(log
->l_mp
, "%s: unexpected magic num",
3819 len
= be32_to_cpu(iclog
->ic_header
.h_num_logops
);
3820 base_ptr
= ptr
= iclog
->ic_datap
;
3822 xhdr
= iclog
->ic_data
;
3823 for (i
= 0; i
< len
; i
++) {
3826 /* clientid is only 1 byte */
3827 p
= &ophead
->oh_clientid
;
3828 field_offset
= p
- base_ptr
;
3829 if (!syncing
|| (field_offset
& 0x1ff)) {
3830 clientid
= ophead
->oh_clientid
;
3832 idx
= BTOBBT((char *)&ophead
->oh_clientid
- iclog
->ic_datap
);
3833 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3834 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3835 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3836 clientid
= xlog_get_client_id(
3837 xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3839 clientid
= xlog_get_client_id(
3840 iclog
->ic_header
.h_cycle_data
[idx
]);
3843 if (clientid
!= XFS_TRANSACTION
&& clientid
!= XFS_LOG
)
3845 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3846 __func__
, clientid
, ophead
,
3847 (unsigned long)field_offset
);
3850 p
= &ophead
->oh_len
;
3851 field_offset
= p
- base_ptr
;
3852 if (!syncing
|| (field_offset
& 0x1ff)) {
3853 op_len
= be32_to_cpu(ophead
->oh_len
);
3855 idx
= BTOBBT((uintptr_t)&ophead
->oh_len
-
3856 (uintptr_t)iclog
->ic_datap
);
3857 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3858 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3859 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3860 op_len
= be32_to_cpu(xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3862 op_len
= be32_to_cpu(iclog
->ic_header
.h_cycle_data
[idx
]);
3865 ptr
+= sizeof(xlog_op_header_t
) + op_len
;
3867 } /* xlog_verify_iclog */
3871 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3877 xlog_in_core_t
*iclog
, *ic
;
3879 iclog
= log
->l_iclog
;
3880 if (! (iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
3882 * Mark all the incore logs IOERROR.
3883 * From now on, no log flushes will result.
3887 ic
->ic_state
= XLOG_STATE_IOERROR
;
3889 } while (ic
!= iclog
);
3893 * Return non-zero, if state transition has already happened.
3899 * This is called from xfs_force_shutdown, when we're forcibly
3900 * shutting down the filesystem, typically because of an IO error.
3901 * Our main objectives here are to make sure that:
3902 * a. if !logerror, flush the logs to disk. Anything modified
3903 * after this is ignored.
3904 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3905 * parties to find out, 'atomically'.
3906 * c. those who're sleeping on log reservations, pinned objects and
3907 * other resources get woken up, and be told the bad news.
3908 * d. nothing new gets queued up after (b) and (c) are done.
3910 * Note: for the !logerror case we need to flush the regions held in memory out
3911 * to disk first. This needs to be done before the log is marked as shutdown,
3912 * otherwise the iclog writes will fail.
3915 xfs_log_force_umount(
3916 struct xfs_mount
*mp
,
3925 * If this happens during log recovery, don't worry about
3926 * locking; the log isn't open for business yet.
3929 log
->l_flags
& XLOG_ACTIVE_RECOVERY
) {
3930 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3932 XFS_BUF_DONE(mp
->m_sb_bp
);
3937 * Somebody could've already done the hard work for us.
3938 * No need to get locks for this.
3940 if (logerror
&& log
->l_iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3941 ASSERT(XLOG_FORCED_SHUTDOWN(log
));
3946 * Flush all the completed transactions to disk before marking the log
3947 * being shut down. We need to do it in this order to ensure that
3948 * completed operations are safely on disk before we shut down, and that
3949 * we don't have to issue any buffer IO after the shutdown flags are set
3950 * to guarantee this.
3953 _xfs_log_force(mp
, XFS_LOG_SYNC
, NULL
);
3956 * mark the filesystem and the as in a shutdown state and wake
3957 * everybody up to tell them the bad news.
3959 spin_lock(&log
->l_icloglock
);
3960 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3962 XFS_BUF_DONE(mp
->m_sb_bp
);
3965 * Mark the log and the iclogs with IO error flags to prevent any
3966 * further log IO from being issued or completed.
3968 log
->l_flags
|= XLOG_IO_ERROR
;
3969 retval
= xlog_state_ioerror(log
);
3970 spin_unlock(&log
->l_icloglock
);
3973 * We don't want anybody waiting for log reservations after this. That
3974 * means we have to wake up everybody queued up on reserveq as well as
3975 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3976 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3977 * action is protected by the grant locks.
3979 xlog_grant_head_wake_all(&log
->l_reserve_head
);
3980 xlog_grant_head_wake_all(&log
->l_write_head
);
3983 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3984 * as if the log writes were completed. The abort handling in the log
3985 * item committed callback functions will do this again under lock to
3988 wake_up_all(&log
->l_cilp
->xc_commit_wait
);
3989 xlog_state_do_callback(log
, XFS_LI_ABORTED
, NULL
);
3991 #ifdef XFSERRORDEBUG
3993 xlog_in_core_t
*iclog
;
3995 spin_lock(&log
->l_icloglock
);
3996 iclog
= log
->l_iclog
;
3998 ASSERT(iclog
->ic_callback
== 0);
3999 iclog
= iclog
->ic_next
;
4000 } while (iclog
!= log
->l_iclog
);
4001 spin_unlock(&log
->l_icloglock
);
4004 /* return non-zero if log IOERROR transition had already happened */
4012 xlog_in_core_t
*iclog
;
4014 iclog
= log
->l_iclog
;
4016 /* endianness does not matter here, zero is zero in
4019 if (iclog
->ic_header
.h_num_logops
)
4021 iclog
= iclog
->ic_next
;
4022 } while (iclog
!= log
->l_iclog
);