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_types.h"
22 #include "xfs_trans.h"
25 #include "xfs_mount.h"
26 #include "xfs_error.h"
27 #include "xfs_log_priv.h"
28 #include "xfs_buf_item.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_alloc_btree.h"
31 #include "xfs_ialloc_btree.h"
32 #include "xfs_log_recover.h"
33 #include "xfs_trans_priv.h"
34 #include "xfs_dinode.h"
35 #include "xfs_inode.h"
36 #include "xfs_trace.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
,
260 list_add_tail(&tic
->t_queue
, &head
->waiters
);
263 if (XLOG_FORCED_SHUTDOWN(log
))
265 xlog_grant_push_ail(log
, need_bytes
);
267 __set_current_state(TASK_UNINTERRUPTIBLE
);
268 spin_unlock(&head
->lock
);
270 XFS_STATS_INC(xs_sleep_logspace
);
272 trace_xfs_log_grant_sleep(log
, tic
);
274 trace_xfs_log_grant_wake(log
, tic
);
276 spin_lock(&head
->lock
);
277 if (XLOG_FORCED_SHUTDOWN(log
))
279 } while (xlog_space_left(log
, &head
->grant
) < need_bytes
);
281 list_del_init(&tic
->t_queue
);
284 list_del_init(&tic
->t_queue
);
285 return XFS_ERROR(EIO
);
289 * Atomically get the log space required for a log ticket.
291 * Once a ticket gets put onto head->waiters, it will only return after the
292 * needed reservation is satisfied.
294 * This function is structured so that it has a lock free fast path. This is
295 * necessary because every new transaction reservation will come through this
296 * path. Hence any lock will be globally hot if we take it unconditionally on
299 * As tickets are only ever moved on and off head->waiters under head->lock, we
300 * only need to take that lock if we are going to add the ticket to the queue
301 * and sleep. We can avoid taking the lock if the ticket was never added to
302 * head->waiters because the t_queue list head will be empty and we hold the
303 * only reference to it so it can safely be checked unlocked.
306 xlog_grant_head_check(
308 struct xlog_grant_head
*head
,
309 struct xlog_ticket
*tic
,
315 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
318 * If there are other waiters on the queue then give them a chance at
319 * logspace before us. Wake up the first waiters, if we do not wake
320 * up all the waiters then go to sleep waiting for more free space,
321 * otherwise try to get some space for this transaction.
323 *need_bytes
= xlog_ticket_reservation(log
, head
, tic
);
324 free_bytes
= xlog_space_left(log
, &head
->grant
);
325 if (!list_empty_careful(&head
->waiters
)) {
326 spin_lock(&head
->lock
);
327 if (!xlog_grant_head_wake(log
, head
, &free_bytes
) ||
328 free_bytes
< *need_bytes
) {
329 error
= xlog_grant_head_wait(log
, head
, tic
,
332 spin_unlock(&head
->lock
);
333 } else if (free_bytes
< *need_bytes
) {
334 spin_lock(&head
->lock
);
335 error
= xlog_grant_head_wait(log
, head
, tic
, *need_bytes
);
336 spin_unlock(&head
->lock
);
343 xlog_tic_reset_res(xlog_ticket_t
*tic
)
346 tic
->t_res_arr_sum
= 0;
347 tic
->t_res_num_ophdrs
= 0;
351 xlog_tic_add_region(xlog_ticket_t
*tic
, uint len
, uint type
)
353 if (tic
->t_res_num
== XLOG_TIC_LEN_MAX
) {
354 /* add to overflow and start again */
355 tic
->t_res_o_flow
+= tic
->t_res_arr_sum
;
357 tic
->t_res_arr_sum
= 0;
360 tic
->t_res_arr
[tic
->t_res_num
].r_len
= len
;
361 tic
->t_res_arr
[tic
->t_res_num
].r_type
= type
;
362 tic
->t_res_arr_sum
+= len
;
367 * Replenish the byte reservation required by moving the grant write head.
371 struct xfs_mount
*mp
,
372 struct xlog_ticket
*tic
)
374 struct xlog
*log
= mp
->m_log
;
378 if (XLOG_FORCED_SHUTDOWN(log
))
379 return XFS_ERROR(EIO
);
381 XFS_STATS_INC(xs_try_logspace
);
384 * This is a new transaction on the ticket, so we need to change the
385 * transaction ID so that the next transaction has a different TID in
386 * the log. Just add one to the existing tid so that we can see chains
387 * of rolling transactions in the log easily.
391 xlog_grant_push_ail(log
, tic
->t_unit_res
);
393 tic
->t_curr_res
= tic
->t_unit_res
;
394 xlog_tic_reset_res(tic
);
399 trace_xfs_log_regrant(log
, tic
);
401 error
= xlog_grant_head_check(log
, &log
->l_write_head
, tic
,
406 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, need_bytes
);
407 trace_xfs_log_regrant_exit(log
, tic
);
408 xlog_verify_grant_tail(log
);
413 * If we are failing, make sure the ticket doesn't have any current
414 * reservations. We don't want to add this back when the ticket/
415 * transaction gets cancelled.
418 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
423 * Reserve log space and return a ticket corresponding the reservation.
425 * Each reservation is going to reserve extra space for a log record header.
426 * When writes happen to the on-disk log, we don't subtract the length of the
427 * log record header from any reservation. By wasting space in each
428 * reservation, we prevent over allocation problems.
432 struct xfs_mount
*mp
,
435 struct xlog_ticket
**ticp
,
440 struct xlog
*log
= mp
->m_log
;
441 struct xlog_ticket
*tic
;
445 ASSERT(client
== XFS_TRANSACTION
|| client
== XFS_LOG
);
447 if (XLOG_FORCED_SHUTDOWN(log
))
448 return XFS_ERROR(EIO
);
450 XFS_STATS_INC(xs_try_logspace
);
452 ASSERT(*ticp
== NULL
);
453 tic
= xlog_ticket_alloc(log
, unit_bytes
, cnt
, client
, permanent
,
454 KM_SLEEP
| KM_MAYFAIL
);
456 return XFS_ERROR(ENOMEM
);
458 tic
->t_trans_type
= t_type
;
461 xlog_grant_push_ail(log
, tic
->t_unit_res
* tic
->t_cnt
);
463 trace_xfs_log_reserve(log
, tic
);
465 error
= xlog_grant_head_check(log
, &log
->l_reserve_head
, tic
,
470 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
, need_bytes
);
471 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, need_bytes
);
472 trace_xfs_log_reserve_exit(log
, tic
);
473 xlog_verify_grant_tail(log
);
478 * If we are failing, make sure the ticket doesn't have any current
479 * reservations. We don't want to add this back when the ticket/
480 * transaction gets cancelled.
483 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
491 * 1. currblock field gets updated at startup and after in-core logs
492 * marked as with WANT_SYNC.
496 * This routine is called when a user of a log manager ticket is done with
497 * the reservation. If the ticket was ever used, then a commit record for
498 * the associated transaction is written out as a log operation header with
499 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
500 * a given ticket. If the ticket was one with a permanent reservation, then
501 * a few operations are done differently. Permanent reservation tickets by
502 * default don't release the reservation. They just commit the current
503 * transaction with the belief that the reservation is still needed. A flag
504 * must be passed in before permanent reservations are actually released.
505 * When these type of tickets are not released, they need to be set into
506 * the inited state again. By doing this, a start record will be written
507 * out when the next write occurs.
511 struct xfs_mount
*mp
,
512 struct xlog_ticket
*ticket
,
513 struct xlog_in_core
**iclog
,
516 struct xlog
*log
= mp
->m_log
;
519 if (XLOG_FORCED_SHUTDOWN(log
) ||
521 * If nothing was ever written, don't write out commit record.
522 * If we get an error, just continue and give back the log ticket.
524 (((ticket
->t_flags
& XLOG_TIC_INITED
) == 0) &&
525 (xlog_commit_record(log
, ticket
, iclog
, &lsn
)))) {
526 lsn
= (xfs_lsn_t
) -1;
527 if (ticket
->t_flags
& XLOG_TIC_PERM_RESERV
) {
528 flags
|= XFS_LOG_REL_PERM_RESERV
;
533 if ((ticket
->t_flags
& XLOG_TIC_PERM_RESERV
) == 0 ||
534 (flags
& XFS_LOG_REL_PERM_RESERV
)) {
535 trace_xfs_log_done_nonperm(log
, ticket
);
538 * Release ticket if not permanent reservation or a specific
539 * request has been made to release a permanent reservation.
541 xlog_ungrant_log_space(log
, ticket
);
542 xfs_log_ticket_put(ticket
);
544 trace_xfs_log_done_perm(log
, ticket
);
546 xlog_regrant_reserve_log_space(log
, ticket
);
547 /* If this ticket was a permanent reservation and we aren't
548 * trying to release it, reset the inited flags; so next time
549 * we write, a start record will be written out.
551 ticket
->t_flags
|= XLOG_TIC_INITED
;
558 * Attaches a new iclog I/O completion callback routine during
559 * transaction commit. If the log is in error state, a non-zero
560 * return code is handed back and the caller is responsible for
561 * executing the callback at an appropriate time.
565 struct xfs_mount
*mp
,
566 struct xlog_in_core
*iclog
,
567 xfs_log_callback_t
*cb
)
571 spin_lock(&iclog
->ic_callback_lock
);
572 abortflg
= (iclog
->ic_state
& XLOG_STATE_IOERROR
);
574 ASSERT_ALWAYS((iclog
->ic_state
== XLOG_STATE_ACTIVE
) ||
575 (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
));
577 *(iclog
->ic_callback_tail
) = cb
;
578 iclog
->ic_callback_tail
= &(cb
->cb_next
);
580 spin_unlock(&iclog
->ic_callback_lock
);
585 xfs_log_release_iclog(
586 struct xfs_mount
*mp
,
587 struct xlog_in_core
*iclog
)
589 if (xlog_state_release_iclog(mp
->m_log
, iclog
)) {
590 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
598 * Mount a log filesystem
600 * mp - ubiquitous xfs mount point structure
601 * log_target - buftarg of on-disk log device
602 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
603 * num_bblocks - Number of BBSIZE blocks in on-disk log
605 * Return error or zero.
610 xfs_buftarg_t
*log_target
,
611 xfs_daddr_t blk_offset
,
616 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
))
617 xfs_notice(mp
, "Mounting Filesystem");
620 "Mounting filesystem in no-recovery mode. Filesystem will be inconsistent.");
621 ASSERT(mp
->m_flags
& XFS_MOUNT_RDONLY
);
624 mp
->m_log
= xlog_alloc_log(mp
, log_target
, blk_offset
, num_bblks
);
625 if (IS_ERR(mp
->m_log
)) {
626 error
= -PTR_ERR(mp
->m_log
);
631 * Initialize the AIL now we have a log.
633 error
= xfs_trans_ail_init(mp
);
635 xfs_warn(mp
, "AIL initialisation failed: error %d", error
);
638 mp
->m_log
->l_ailp
= mp
->m_ail
;
641 * skip log recovery on a norecovery mount. pretend it all
644 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
)) {
645 int readonly
= (mp
->m_flags
& XFS_MOUNT_RDONLY
);
648 mp
->m_flags
&= ~XFS_MOUNT_RDONLY
;
650 error
= xlog_recover(mp
->m_log
);
653 mp
->m_flags
|= XFS_MOUNT_RDONLY
;
655 xfs_warn(mp
, "log mount/recovery failed: error %d",
657 goto out_destroy_ail
;
661 /* Normal transactions can now occur */
662 mp
->m_log
->l_flags
&= ~XLOG_ACTIVE_RECOVERY
;
665 * Now the log has been fully initialised and we know were our
666 * space grant counters are, we can initialise the permanent ticket
667 * needed for delayed logging to work.
669 xlog_cil_init_post_recovery(mp
->m_log
);
674 xfs_trans_ail_destroy(mp
);
676 xlog_dealloc_log(mp
->m_log
);
682 * Finish the recovery of the file system. This is separate from
683 * the xfs_log_mount() call, because it depends on the code in
684 * xfs_mountfs() to read in the root and real-time bitmap inodes
685 * between calling xfs_log_mount() and here.
687 * mp - ubiquitous xfs mount point structure
690 xfs_log_mount_finish(xfs_mount_t
*mp
)
694 if (!(mp
->m_flags
& XFS_MOUNT_NORECOVERY
))
695 error
= xlog_recover_finish(mp
->m_log
);
698 ASSERT(mp
->m_flags
& XFS_MOUNT_RDONLY
);
705 * Final log writes as part of unmount.
707 * Mark the filesystem clean as unmount happens. Note that during relocation
708 * this routine needs to be executed as part of source-bag while the
709 * deallocation must not be done until source-end.
713 * Unmount record used to have a string "Unmount filesystem--" in the
714 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
715 * We just write the magic number now since that particular field isn't
716 * currently architecture converted and "nUmount" is a bit foo.
717 * As far as I know, there weren't any dependencies on the old behaviour.
721 xfs_log_unmount_write(xfs_mount_t
*mp
)
723 struct xlog
*log
= mp
->m_log
;
724 xlog_in_core_t
*iclog
;
726 xlog_in_core_t
*first_iclog
;
728 xlog_ticket_t
*tic
= NULL
;
733 * Don't write out unmount record on read-only mounts.
734 * Or, if we are doing a forced umount (typically because of IO errors).
736 if (mp
->m_flags
& XFS_MOUNT_RDONLY
)
739 error
= _xfs_log_force(mp
, XFS_LOG_SYNC
, NULL
);
740 ASSERT(error
|| !(XLOG_FORCED_SHUTDOWN(log
)));
743 first_iclog
= iclog
= log
->l_iclog
;
745 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
746 ASSERT(iclog
->ic_state
& XLOG_STATE_ACTIVE
);
747 ASSERT(iclog
->ic_offset
== 0);
749 iclog
= iclog
->ic_next
;
750 } while (iclog
!= first_iclog
);
752 if (! (XLOG_FORCED_SHUTDOWN(log
))) {
753 error
= xfs_log_reserve(mp
, 600, 1, &tic
,
754 XFS_LOG
, 0, XLOG_UNMOUNT_REC_TYPE
);
756 /* the data section must be 32 bit size aligned */
760 __uint32_t pad2
; /* may as well make it 64 bits */
762 .magic
= XLOG_UNMOUNT_TYPE
,
764 struct xfs_log_iovec reg
= {
766 .i_len
= sizeof(magic
),
767 .i_type
= XLOG_REG_TYPE_UNMOUNT
,
769 struct xfs_log_vec vec
= {
774 /* remove inited flag, and account for space used */
776 tic
->t_curr_res
-= sizeof(magic
);
777 error
= xlog_write(log
, &vec
, tic
, &lsn
,
778 NULL
, XLOG_UNMOUNT_TRANS
);
780 * At this point, we're umounting anyway,
781 * so there's no point in transitioning log state
782 * to IOERROR. Just continue...
787 xfs_alert(mp
, "%s: unmount record failed", __func__
);
790 spin_lock(&log
->l_icloglock
);
791 iclog
= log
->l_iclog
;
792 atomic_inc(&iclog
->ic_refcnt
);
793 xlog_state_want_sync(log
, iclog
);
794 spin_unlock(&log
->l_icloglock
);
795 error
= xlog_state_release_iclog(log
, iclog
);
797 spin_lock(&log
->l_icloglock
);
798 if (!(iclog
->ic_state
== XLOG_STATE_ACTIVE
||
799 iclog
->ic_state
== XLOG_STATE_DIRTY
)) {
800 if (!XLOG_FORCED_SHUTDOWN(log
)) {
801 xlog_wait(&iclog
->ic_force_wait
,
804 spin_unlock(&log
->l_icloglock
);
807 spin_unlock(&log
->l_icloglock
);
810 trace_xfs_log_umount_write(log
, tic
);
811 xlog_ungrant_log_space(log
, tic
);
812 xfs_log_ticket_put(tic
);
816 * We're already in forced_shutdown mode, couldn't
817 * even attempt to write out the unmount transaction.
819 * Go through the motions of sync'ing and releasing
820 * the iclog, even though no I/O will actually happen,
821 * we need to wait for other log I/Os that may already
822 * be in progress. Do this as a separate section of
823 * code so we'll know if we ever get stuck here that
824 * we're in this odd situation of trying to unmount
825 * a file system that went into forced_shutdown as
826 * the result of an unmount..
828 spin_lock(&log
->l_icloglock
);
829 iclog
= log
->l_iclog
;
830 atomic_inc(&iclog
->ic_refcnt
);
832 xlog_state_want_sync(log
, iclog
);
833 spin_unlock(&log
->l_icloglock
);
834 error
= xlog_state_release_iclog(log
, iclog
);
836 spin_lock(&log
->l_icloglock
);
838 if ( ! ( iclog
->ic_state
== XLOG_STATE_ACTIVE
839 || iclog
->ic_state
== XLOG_STATE_DIRTY
840 || iclog
->ic_state
== XLOG_STATE_IOERROR
) ) {
842 xlog_wait(&iclog
->ic_force_wait
,
845 spin_unlock(&log
->l_icloglock
);
850 } /* xfs_log_unmount_write */
853 * Deallocate log structures for unmount/relocation.
855 * We need to stop the aild from running before we destroy
856 * and deallocate the log as the aild references the log.
859 xfs_log_unmount(xfs_mount_t
*mp
)
861 cancel_delayed_work_sync(&mp
->m_sync_work
);
862 xfs_trans_ail_destroy(mp
);
863 xlog_dealloc_log(mp
->m_log
);
868 struct xfs_mount
*mp
,
869 struct xfs_log_item
*item
,
871 const struct xfs_item_ops
*ops
)
873 item
->li_mountp
= mp
;
874 item
->li_ailp
= mp
->m_ail
;
875 item
->li_type
= type
;
879 INIT_LIST_HEAD(&item
->li_ail
);
880 INIT_LIST_HEAD(&item
->li_cil
);
884 * Wake up processes waiting for log space after we have moved the log tail.
888 struct xfs_mount
*mp
)
890 struct xlog
*log
= mp
->m_log
;
893 if (XLOG_FORCED_SHUTDOWN(log
))
896 if (!list_empty_careful(&log
->l_write_head
.waiters
)) {
897 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
899 spin_lock(&log
->l_write_head
.lock
);
900 free_bytes
= xlog_space_left(log
, &log
->l_write_head
.grant
);
901 xlog_grant_head_wake(log
, &log
->l_write_head
, &free_bytes
);
902 spin_unlock(&log
->l_write_head
.lock
);
905 if (!list_empty_careful(&log
->l_reserve_head
.waiters
)) {
906 ASSERT(!(log
->l_flags
& XLOG_ACTIVE_RECOVERY
));
908 spin_lock(&log
->l_reserve_head
.lock
);
909 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
910 xlog_grant_head_wake(log
, &log
->l_reserve_head
, &free_bytes
);
911 spin_unlock(&log
->l_reserve_head
.lock
);
916 * Determine if we have a transaction that has gone to disk
917 * that needs to be covered. To begin the transition to the idle state
918 * firstly the log needs to be idle (no AIL and nothing in the iclogs).
919 * If we are then in a state where covering is needed, the caller is informed
920 * that dummy transactions are required to move the log into the idle state.
922 * Because this is called as part of the sync process, we should also indicate
923 * that dummy transactions should be issued in anything but the covered or
924 * idle states. This ensures that the log tail is accurately reflected in
925 * the log at the end of the sync, hence if a crash occurrs avoids replay
926 * of transactions where the metadata is already on disk.
929 xfs_log_need_covered(xfs_mount_t
*mp
)
932 struct xlog
*log
= mp
->m_log
;
934 if (!xfs_fs_writable(mp
))
937 spin_lock(&log
->l_icloglock
);
938 switch (log
->l_covered_state
) {
939 case XLOG_STATE_COVER_DONE
:
940 case XLOG_STATE_COVER_DONE2
:
941 case XLOG_STATE_COVER_IDLE
:
943 case XLOG_STATE_COVER_NEED
:
944 case XLOG_STATE_COVER_NEED2
:
945 if (!xfs_ail_min_lsn(log
->l_ailp
) &&
946 xlog_iclogs_empty(log
)) {
947 if (log
->l_covered_state
== XLOG_STATE_COVER_NEED
)
948 log
->l_covered_state
= XLOG_STATE_COVER_DONE
;
950 log
->l_covered_state
= XLOG_STATE_COVER_DONE2
;
957 spin_unlock(&log
->l_icloglock
);
962 * We may be holding the log iclog lock upon entering this routine.
965 xlog_assign_tail_lsn_locked(
966 struct xfs_mount
*mp
)
968 struct xlog
*log
= mp
->m_log
;
969 struct xfs_log_item
*lip
;
972 assert_spin_locked(&mp
->m_ail
->xa_lock
);
975 * To make sure we always have a valid LSN for the log tail we keep
976 * track of the last LSN which was committed in log->l_last_sync_lsn,
977 * and use that when the AIL was empty.
979 lip
= xfs_ail_min(mp
->m_ail
);
981 tail_lsn
= lip
->li_lsn
;
983 tail_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
984 atomic64_set(&log
->l_tail_lsn
, tail_lsn
);
989 xlog_assign_tail_lsn(
990 struct xfs_mount
*mp
)
994 spin_lock(&mp
->m_ail
->xa_lock
);
995 tail_lsn
= xlog_assign_tail_lsn_locked(mp
);
996 spin_unlock(&mp
->m_ail
->xa_lock
);
1002 * Return the space in the log between the tail and the head. The head
1003 * is passed in the cycle/bytes formal parms. In the special case where
1004 * the reserve head has wrapped passed the tail, this calculation is no
1005 * longer valid. In this case, just return 0 which means there is no space
1006 * in the log. This works for all places where this function is called
1007 * with the reserve head. Of course, if the write head were to ever
1008 * wrap the tail, we should blow up. Rather than catch this case here,
1009 * we depend on other ASSERTions in other parts of the code. XXXmiken
1011 * This code also handles the case where the reservation head is behind
1012 * the tail. The details of this case are described below, but the end
1013 * result is that we return the size of the log as the amount of space left.
1026 xlog_crack_grant_head(head
, &head_cycle
, &head_bytes
);
1027 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_bytes
);
1028 tail_bytes
= BBTOB(tail_bytes
);
1029 if (tail_cycle
== head_cycle
&& head_bytes
>= tail_bytes
)
1030 free_bytes
= log
->l_logsize
- (head_bytes
- tail_bytes
);
1031 else if (tail_cycle
+ 1 < head_cycle
)
1033 else if (tail_cycle
< head_cycle
) {
1034 ASSERT(tail_cycle
== (head_cycle
- 1));
1035 free_bytes
= tail_bytes
- head_bytes
;
1038 * The reservation head is behind the tail.
1039 * In this case we just want to return the size of the
1040 * log as the amount of space left.
1042 xfs_alert(log
->l_mp
,
1043 "xlog_space_left: head behind tail\n"
1044 " tail_cycle = %d, tail_bytes = %d\n"
1045 " GH cycle = %d, GH bytes = %d",
1046 tail_cycle
, tail_bytes
, head_cycle
, head_bytes
);
1048 free_bytes
= log
->l_logsize
;
1055 * Log function which is called when an io completes.
1057 * The log manager needs its own routine, in order to control what
1058 * happens with the buffer after the write completes.
1061 xlog_iodone(xfs_buf_t
*bp
)
1063 struct xlog_in_core
*iclog
= bp
->b_fspriv
;
1064 struct xlog
*l
= iclog
->ic_log
;
1068 * Race to shutdown the filesystem if we see an error.
1070 if (XFS_TEST_ERROR((xfs_buf_geterror(bp
)), l
->l_mp
,
1071 XFS_ERRTAG_IODONE_IOERR
, XFS_RANDOM_IODONE_IOERR
)) {
1072 xfs_buf_ioerror_alert(bp
, __func__
);
1074 xfs_force_shutdown(l
->l_mp
, SHUTDOWN_LOG_IO_ERROR
);
1076 * This flag will be propagated to the trans-committed
1077 * callback routines to let them know that the log-commit
1080 aborted
= XFS_LI_ABORTED
;
1081 } else if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
1082 aborted
= XFS_LI_ABORTED
;
1085 /* log I/O is always issued ASYNC */
1086 ASSERT(XFS_BUF_ISASYNC(bp
));
1087 xlog_state_done_syncing(iclog
, aborted
);
1089 * do not reference the buffer (bp) here as we could race
1090 * with it being freed after writing the unmount record to the
1097 * Return size of each in-core log record buffer.
1099 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1101 * If the filesystem blocksize is too large, we may need to choose a
1102 * larger size since the directory code currently logs entire blocks.
1106 xlog_get_iclog_buffer_size(
1107 struct xfs_mount
*mp
,
1113 if (mp
->m_logbufs
<= 0)
1114 log
->l_iclog_bufs
= XLOG_MAX_ICLOGS
;
1116 log
->l_iclog_bufs
= mp
->m_logbufs
;
1119 * Buffer size passed in from mount system call.
1121 if (mp
->m_logbsize
> 0) {
1122 size
= log
->l_iclog_size
= mp
->m_logbsize
;
1123 log
->l_iclog_size_log
= 0;
1125 log
->l_iclog_size_log
++;
1129 if (xfs_sb_version_haslogv2(&mp
->m_sb
)) {
1130 /* # headers = size / 32k
1131 * one header holds cycles from 32k of data
1134 xhdrs
= mp
->m_logbsize
/ XLOG_HEADER_CYCLE_SIZE
;
1135 if (mp
->m_logbsize
% XLOG_HEADER_CYCLE_SIZE
)
1137 log
->l_iclog_hsize
= xhdrs
<< BBSHIFT
;
1138 log
->l_iclog_heads
= xhdrs
;
1140 ASSERT(mp
->m_logbsize
<= XLOG_BIG_RECORD_BSIZE
);
1141 log
->l_iclog_hsize
= BBSIZE
;
1142 log
->l_iclog_heads
= 1;
1147 /* All machines use 32kB buffers by default. */
1148 log
->l_iclog_size
= XLOG_BIG_RECORD_BSIZE
;
1149 log
->l_iclog_size_log
= XLOG_BIG_RECORD_BSHIFT
;
1151 /* the default log size is 16k or 32k which is one header sector */
1152 log
->l_iclog_hsize
= BBSIZE
;
1153 log
->l_iclog_heads
= 1;
1156 /* are we being asked to make the sizes selected above visible? */
1157 if (mp
->m_logbufs
== 0)
1158 mp
->m_logbufs
= log
->l_iclog_bufs
;
1159 if (mp
->m_logbsize
== 0)
1160 mp
->m_logbsize
= log
->l_iclog_size
;
1161 } /* xlog_get_iclog_buffer_size */
1165 * This routine initializes some of the log structure for a given mount point.
1166 * Its primary purpose is to fill in enough, so recovery can occur. However,
1167 * some other stuff may be filled in too.
1169 STATIC
struct xlog
*
1171 struct xfs_mount
*mp
,
1172 struct xfs_buftarg
*log_target
,
1173 xfs_daddr_t blk_offset
,
1177 xlog_rec_header_t
*head
;
1178 xlog_in_core_t
**iclogp
;
1179 xlog_in_core_t
*iclog
, *prev_iclog
=NULL
;
1185 log
= kmem_zalloc(sizeof(struct xlog
), KM_MAYFAIL
);
1187 xfs_warn(mp
, "Log allocation failed: No memory!");
1192 log
->l_targ
= log_target
;
1193 log
->l_logsize
= BBTOB(num_bblks
);
1194 log
->l_logBBstart
= blk_offset
;
1195 log
->l_logBBsize
= num_bblks
;
1196 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
1197 log
->l_flags
|= XLOG_ACTIVE_RECOVERY
;
1199 log
->l_prev_block
= -1;
1200 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1201 xlog_assign_atomic_lsn(&log
->l_tail_lsn
, 1, 0);
1202 xlog_assign_atomic_lsn(&log
->l_last_sync_lsn
, 1, 0);
1203 log
->l_curr_cycle
= 1; /* 0 is bad since this is initial value */
1205 xlog_grant_head_init(&log
->l_reserve_head
);
1206 xlog_grant_head_init(&log
->l_write_head
);
1208 error
= EFSCORRUPTED
;
1209 if (xfs_sb_version_hassector(&mp
->m_sb
)) {
1210 log2_size
= mp
->m_sb
.sb_logsectlog
;
1211 if (log2_size
< BBSHIFT
) {
1212 xfs_warn(mp
, "Log sector size too small (0x%x < 0x%x)",
1213 log2_size
, BBSHIFT
);
1217 log2_size
-= BBSHIFT
;
1218 if (log2_size
> mp
->m_sectbb_log
) {
1219 xfs_warn(mp
, "Log sector size too large (0x%x > 0x%x)",
1220 log2_size
, mp
->m_sectbb_log
);
1224 /* for larger sector sizes, must have v2 or external log */
1225 if (log2_size
&& log
->l_logBBstart
> 0 &&
1226 !xfs_sb_version_haslogv2(&mp
->m_sb
)) {
1228 "log sector size (0x%x) invalid for configuration.",
1233 log
->l_sectBBsize
= 1 << log2_size
;
1235 xlog_get_iclog_buffer_size(mp
, log
);
1238 bp
= xfs_buf_alloc(mp
->m_logdev_targp
, 0, BTOBB(log
->l_iclog_size
), 0);
1241 bp
->b_iodone
= xlog_iodone
;
1242 ASSERT(xfs_buf_islocked(bp
));
1245 spin_lock_init(&log
->l_icloglock
);
1246 init_waitqueue_head(&log
->l_flush_wait
);
1248 iclogp
= &log
->l_iclog
;
1250 * The amount of memory to allocate for the iclog structure is
1251 * rather funky due to the way the structure is defined. It is
1252 * done this way so that we can use different sizes for machines
1253 * with different amounts of memory. See the definition of
1254 * xlog_in_core_t in xfs_log_priv.h for details.
1256 ASSERT(log
->l_iclog_size
>= 4096);
1257 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
1258 *iclogp
= kmem_zalloc(sizeof(xlog_in_core_t
), KM_MAYFAIL
);
1260 goto out_free_iclog
;
1263 iclog
->ic_prev
= prev_iclog
;
1266 bp
= xfs_buf_get_uncached(mp
->m_logdev_targp
,
1267 BTOBB(log
->l_iclog_size
), 0);
1269 goto out_free_iclog
;
1271 bp
->b_iodone
= xlog_iodone
;
1273 iclog
->ic_data
= bp
->b_addr
;
1275 log
->l_iclog_bak
[i
] = (xfs_caddr_t
)&(iclog
->ic_header
);
1277 head
= &iclog
->ic_header
;
1278 memset(head
, 0, sizeof(xlog_rec_header_t
));
1279 head
->h_magicno
= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
);
1280 head
->h_version
= cpu_to_be32(
1281 xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) ? 2 : 1);
1282 head
->h_size
= cpu_to_be32(log
->l_iclog_size
);
1284 head
->h_fmt
= cpu_to_be32(XLOG_FMT
);
1285 memcpy(&head
->h_fs_uuid
, &mp
->m_sb
.sb_uuid
, sizeof(uuid_t
));
1287 iclog
->ic_size
= BBTOB(bp
->b_length
) - log
->l_iclog_hsize
;
1288 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
1289 iclog
->ic_log
= log
;
1290 atomic_set(&iclog
->ic_refcnt
, 0);
1291 spin_lock_init(&iclog
->ic_callback_lock
);
1292 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
1293 iclog
->ic_datap
= (char *)iclog
->ic_data
+ log
->l_iclog_hsize
;
1295 ASSERT(xfs_buf_islocked(iclog
->ic_bp
));
1296 init_waitqueue_head(&iclog
->ic_force_wait
);
1297 init_waitqueue_head(&iclog
->ic_write_wait
);
1299 iclogp
= &iclog
->ic_next
;
1301 *iclogp
= log
->l_iclog
; /* complete ring */
1302 log
->l_iclog
->ic_prev
= prev_iclog
; /* re-write 1st prev ptr */
1304 error
= xlog_cil_init(log
);
1306 goto out_free_iclog
;
1310 for (iclog
= log
->l_iclog
; iclog
; iclog
= prev_iclog
) {
1311 prev_iclog
= iclog
->ic_next
;
1313 xfs_buf_free(iclog
->ic_bp
);
1316 spinlock_destroy(&log
->l_icloglock
);
1317 xfs_buf_free(log
->l_xbuf
);
1321 return ERR_PTR(-error
);
1322 } /* xlog_alloc_log */
1326 * Write out the commit record of a transaction associated with the given
1327 * ticket. Return the lsn of the commit record.
1332 struct xlog_ticket
*ticket
,
1333 struct xlog_in_core
**iclog
,
1334 xfs_lsn_t
*commitlsnp
)
1336 struct xfs_mount
*mp
= log
->l_mp
;
1338 struct xfs_log_iovec reg
= {
1341 .i_type
= XLOG_REG_TYPE_COMMIT
,
1343 struct xfs_log_vec vec
= {
1348 ASSERT_ALWAYS(iclog
);
1349 error
= xlog_write(log
, &vec
, ticket
, commitlsnp
, iclog
,
1352 xfs_force_shutdown(mp
, SHUTDOWN_LOG_IO_ERROR
);
1357 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1358 * log space. This code pushes on the lsn which would supposedly free up
1359 * the 25% which we want to leave free. We may need to adopt a policy which
1360 * pushes on an lsn which is further along in the log once we reach the high
1361 * water mark. In this manner, we would be creating a low water mark.
1364 xlog_grant_push_ail(
1368 xfs_lsn_t threshold_lsn
= 0;
1369 xfs_lsn_t last_sync_lsn
;
1372 int threshold_block
;
1373 int threshold_cycle
;
1376 ASSERT(BTOBB(need_bytes
) < log
->l_logBBsize
);
1378 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
1379 free_blocks
= BTOBBT(free_bytes
);
1382 * Set the threshold for the minimum number of free blocks in the
1383 * log to the maximum of what the caller needs, one quarter of the
1384 * log, and 256 blocks.
1386 free_threshold
= BTOBB(need_bytes
);
1387 free_threshold
= MAX(free_threshold
, (log
->l_logBBsize
>> 2));
1388 free_threshold
= MAX(free_threshold
, 256);
1389 if (free_blocks
>= free_threshold
)
1392 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &threshold_cycle
,
1394 threshold_block
+= free_threshold
;
1395 if (threshold_block
>= log
->l_logBBsize
) {
1396 threshold_block
-= log
->l_logBBsize
;
1397 threshold_cycle
+= 1;
1399 threshold_lsn
= xlog_assign_lsn(threshold_cycle
,
1402 * Don't pass in an lsn greater than the lsn of the last
1403 * log record known to be on disk. Use a snapshot of the last sync lsn
1404 * so that it doesn't change between the compare and the set.
1406 last_sync_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
1407 if (XFS_LSN_CMP(threshold_lsn
, last_sync_lsn
) > 0)
1408 threshold_lsn
= last_sync_lsn
;
1411 * Get the transaction layer to kick the dirty buffers out to
1412 * disk asynchronously. No point in trying to do this if
1413 * the filesystem is shutting down.
1415 if (!XLOG_FORCED_SHUTDOWN(log
))
1416 xfs_ail_push(log
->l_ailp
, threshold_lsn
);
1420 * The bdstrat callback function for log bufs. This gives us a central
1421 * place to trap bufs in case we get hit by a log I/O error and need to
1422 * shutdown. Actually, in practice, even when we didn't get a log error,
1423 * we transition the iclogs to IOERROR state *after* flushing all existing
1424 * iclogs to disk. This is because we don't want anymore new transactions to be
1425 * started or completed afterwards.
1431 struct xlog_in_core
*iclog
= bp
->b_fspriv
;
1433 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
1434 xfs_buf_ioerror(bp
, EIO
);
1436 xfs_buf_ioend(bp
, 0);
1438 * It would seem logical to return EIO here, but we rely on
1439 * the log state machine to propagate I/O errors instead of
1445 xfs_buf_iorequest(bp
);
1450 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1451 * fashion. Previously, we should have moved the current iclog
1452 * ptr in the log to point to the next available iclog. This allows further
1453 * write to continue while this code syncs out an iclog ready to go.
1454 * Before an in-core log can be written out, the data section must be scanned
1455 * to save away the 1st word of each BBSIZE block into the header. We replace
1456 * it with the current cycle count. Each BBSIZE block is tagged with the
1457 * cycle count because there in an implicit assumption that drives will
1458 * guarantee that entire 512 byte blocks get written at once. In other words,
1459 * we can't have part of a 512 byte block written and part not written. By
1460 * tagging each block, we will know which blocks are valid when recovering
1461 * after an unclean shutdown.
1463 * This routine is single threaded on the iclog. No other thread can be in
1464 * this routine with the same iclog. Changing contents of iclog can there-
1465 * fore be done without grabbing the state machine lock. Updating the global
1466 * log will require grabbing the lock though.
1468 * The entire log manager uses a logical block numbering scheme. Only
1469 * log_sync (and then only bwrite()) know about the fact that the log may
1470 * not start with block zero on a given device. The log block start offset
1471 * is added immediately before calling bwrite().
1477 struct xlog_in_core
*iclog
)
1479 xfs_caddr_t dptr
; /* pointer to byte sized element */
1482 uint count
; /* byte count of bwrite */
1483 uint count_init
; /* initial count before roundup */
1484 int roundoff
; /* roundoff to BB or stripe */
1485 int split
= 0; /* split write into two regions */
1487 int v2
= xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
);
1489 XFS_STATS_INC(xs_log_writes
);
1490 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
1492 /* Add for LR header */
1493 count_init
= log
->l_iclog_hsize
+ iclog
->ic_offset
;
1495 /* Round out the log write size */
1496 if (v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1) {
1497 /* we have a v2 stripe unit to use */
1498 count
= XLOG_LSUNITTOB(log
, XLOG_BTOLSUNIT(log
, count_init
));
1500 count
= BBTOB(BTOBB(count_init
));
1502 roundoff
= count
- count_init
;
1503 ASSERT(roundoff
>= 0);
1504 ASSERT((v2
&& log
->l_mp
->m_sb
.sb_logsunit
> 1 &&
1505 roundoff
< log
->l_mp
->m_sb
.sb_logsunit
)
1507 (log
->l_mp
->m_sb
.sb_logsunit
<= 1 &&
1508 roundoff
< BBTOB(1)));
1510 /* move grant heads by roundoff in sync */
1511 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
, roundoff
);
1512 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, roundoff
);
1514 /* put cycle number in every block */
1515 xlog_pack_data(log
, iclog
, roundoff
);
1517 /* real byte length */
1519 iclog
->ic_header
.h_len
=
1520 cpu_to_be32(iclog
->ic_offset
+ roundoff
);
1522 iclog
->ic_header
.h_len
=
1523 cpu_to_be32(iclog
->ic_offset
);
1527 XFS_BUF_SET_ADDR(bp
, BLOCK_LSN(be64_to_cpu(iclog
->ic_header
.h_lsn
)));
1529 XFS_STATS_ADD(xs_log_blocks
, BTOBB(count
));
1531 /* Do we need to split this write into 2 parts? */
1532 if (XFS_BUF_ADDR(bp
) + BTOBB(count
) > log
->l_logBBsize
) {
1533 split
= count
- (BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
)));
1534 count
= BBTOB(log
->l_logBBsize
- XFS_BUF_ADDR(bp
));
1535 iclog
->ic_bwritecnt
= 2; /* split into 2 writes */
1537 iclog
->ic_bwritecnt
= 1;
1539 bp
->b_io_length
= BTOBB(count
);
1540 bp
->b_fspriv
= iclog
;
1541 XFS_BUF_ZEROFLAGS(bp
);
1543 bp
->b_flags
|= XBF_SYNCIO
;
1545 if (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
) {
1546 bp
->b_flags
|= XBF_FUA
;
1549 * Flush the data device before flushing the log to make
1550 * sure all meta data written back from the AIL actually made
1551 * it to disk before stamping the new log tail LSN into the
1552 * log buffer. For an external log we need to issue the
1553 * flush explicitly, and unfortunately synchronously here;
1554 * for an internal log we can simply use the block layer
1555 * state machine for preflushes.
1557 if (log
->l_mp
->m_logdev_targp
!= log
->l_mp
->m_ddev_targp
)
1558 xfs_blkdev_issue_flush(log
->l_mp
->m_ddev_targp
);
1560 bp
->b_flags
|= XBF_FLUSH
;
1563 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1564 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1566 xlog_verify_iclog(log
, iclog
, count
, B_TRUE
);
1568 /* account for log which doesn't start at block #0 */
1569 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1571 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1576 error
= xlog_bdstrat(bp
);
1578 xfs_buf_ioerror_alert(bp
, "xlog_sync");
1582 bp
= iclog
->ic_log
->l_xbuf
;
1583 XFS_BUF_SET_ADDR(bp
, 0); /* logical 0 */
1584 xfs_buf_associate_memory(bp
,
1585 (char *)&iclog
->ic_header
+ count
, split
);
1586 bp
->b_fspriv
= iclog
;
1587 XFS_BUF_ZEROFLAGS(bp
);
1589 bp
->b_flags
|= XBF_SYNCIO
;
1590 if (log
->l_mp
->m_flags
& XFS_MOUNT_BARRIER
)
1591 bp
->b_flags
|= XBF_FUA
;
1594 * Bump the cycle numbers at the start of each block
1595 * since this part of the buffer is at the start of
1596 * a new cycle. Watch out for the header magic number
1599 for (i
= 0; i
< split
; i
+= BBSIZE
) {
1600 be32_add_cpu((__be32
*)dptr
, 1);
1601 if (be32_to_cpu(*(__be32
*)dptr
) == XLOG_HEADER_MAGIC_NUM
)
1602 be32_add_cpu((__be32
*)dptr
, 1);
1606 ASSERT(XFS_BUF_ADDR(bp
) <= log
->l_logBBsize
-1);
1607 ASSERT(XFS_BUF_ADDR(bp
) + BTOBB(count
) <= log
->l_logBBsize
);
1609 /* account for internal log which doesn't start at block #0 */
1610 XFS_BUF_SET_ADDR(bp
, XFS_BUF_ADDR(bp
) + log
->l_logBBstart
);
1612 error
= xlog_bdstrat(bp
);
1614 xfs_buf_ioerror_alert(bp
, "xlog_sync (split)");
1623 * Deallocate a log structure
1629 xlog_in_core_t
*iclog
, *next_iclog
;
1632 xlog_cil_destroy(log
);
1635 * always need to ensure that the extra buffer does not point to memory
1636 * owned by another log buffer before we free it.
1638 xfs_buf_set_empty(log
->l_xbuf
, BTOBB(log
->l_iclog_size
));
1639 xfs_buf_free(log
->l_xbuf
);
1641 iclog
= log
->l_iclog
;
1642 for (i
=0; i
<log
->l_iclog_bufs
; i
++) {
1643 xfs_buf_free(iclog
->ic_bp
);
1644 next_iclog
= iclog
->ic_next
;
1648 spinlock_destroy(&log
->l_icloglock
);
1650 log
->l_mp
->m_log
= NULL
;
1652 } /* xlog_dealloc_log */
1655 * Update counters atomically now that memcpy is done.
1659 xlog_state_finish_copy(
1661 struct xlog_in_core
*iclog
,
1665 spin_lock(&log
->l_icloglock
);
1667 be32_add_cpu(&iclog
->ic_header
.h_num_logops
, record_cnt
);
1668 iclog
->ic_offset
+= copy_bytes
;
1670 spin_unlock(&log
->l_icloglock
);
1671 } /* xlog_state_finish_copy */
1677 * print out info relating to regions written which consume
1682 struct xfs_mount
*mp
,
1683 struct xlog_ticket
*ticket
)
1686 uint ophdr_spc
= ticket
->t_res_num_ophdrs
* (uint
)sizeof(xlog_op_header_t
);
1688 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1689 static char *res_type_str
[XLOG_REG_TYPE_MAX
] = {
1710 static char *trans_type_str
[XFS_TRANS_TYPE_MAX
] = {
1754 "xlog_write: reservation summary:\n"
1755 " trans type = %s (%u)\n"
1756 " unit res = %d bytes\n"
1757 " current res = %d bytes\n"
1758 " total reg = %u bytes (o/flow = %u bytes)\n"
1759 " ophdrs = %u (ophdr space = %u bytes)\n"
1760 " ophdr + reg = %u bytes\n"
1761 " num regions = %u\n",
1762 ((ticket
->t_trans_type
<= 0 ||
1763 ticket
->t_trans_type
> XFS_TRANS_TYPE_MAX
) ?
1764 "bad-trans-type" : trans_type_str
[ticket
->t_trans_type
-1]),
1765 ticket
->t_trans_type
,
1768 ticket
->t_res_arr_sum
, ticket
->t_res_o_flow
,
1769 ticket
->t_res_num_ophdrs
, ophdr_spc
,
1770 ticket
->t_res_arr_sum
+
1771 ticket
->t_res_o_flow
+ ophdr_spc
,
1774 for (i
= 0; i
< ticket
->t_res_num
; i
++) {
1775 uint r_type
= ticket
->t_res_arr
[i
].r_type
;
1776 xfs_warn(mp
, "region[%u]: %s - %u bytes\n", i
,
1777 ((r_type
<= 0 || r_type
> XLOG_REG_TYPE_MAX
) ?
1778 "bad-rtype" : res_type_str
[r_type
-1]),
1779 ticket
->t_res_arr
[i
].r_len
);
1782 xfs_alert_tag(mp
, XFS_PTAG_LOGRES
,
1783 "xlog_write: reservation ran out. Need to up reservation");
1784 xfs_force_shutdown(mp
, SHUTDOWN_CORRUPT_INCORE
);
1788 * Calculate the potential space needed by the log vector. Each region gets
1789 * its own xlog_op_header_t and may need to be double word aligned.
1792 xlog_write_calc_vec_length(
1793 struct xlog_ticket
*ticket
,
1794 struct xfs_log_vec
*log_vector
)
1796 struct xfs_log_vec
*lv
;
1801 /* acct for start rec of xact */
1802 if (ticket
->t_flags
& XLOG_TIC_INITED
)
1805 for (lv
= log_vector
; lv
; lv
= lv
->lv_next
) {
1806 headers
+= lv
->lv_niovecs
;
1808 for (i
= 0; i
< lv
->lv_niovecs
; i
++) {
1809 struct xfs_log_iovec
*vecp
= &lv
->lv_iovecp
[i
];
1812 xlog_tic_add_region(ticket
, vecp
->i_len
, vecp
->i_type
);
1816 ticket
->t_res_num_ophdrs
+= headers
;
1817 len
+= headers
* sizeof(struct xlog_op_header
);
1823 * If first write for transaction, insert start record We can't be trying to
1824 * commit if we are inited. We can't have any "partial_copy" if we are inited.
1827 xlog_write_start_rec(
1828 struct xlog_op_header
*ophdr
,
1829 struct xlog_ticket
*ticket
)
1831 if (!(ticket
->t_flags
& XLOG_TIC_INITED
))
1834 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
1835 ophdr
->oh_clientid
= ticket
->t_clientid
;
1837 ophdr
->oh_flags
= XLOG_START_TRANS
;
1840 ticket
->t_flags
&= ~XLOG_TIC_INITED
;
1842 return sizeof(struct xlog_op_header
);
1845 static xlog_op_header_t
*
1846 xlog_write_setup_ophdr(
1848 struct xlog_op_header
*ophdr
,
1849 struct xlog_ticket
*ticket
,
1852 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
1853 ophdr
->oh_clientid
= ticket
->t_clientid
;
1856 /* are we copying a commit or unmount record? */
1857 ophdr
->oh_flags
= flags
;
1860 * We've seen logs corrupted with bad transaction client ids. This
1861 * makes sure that XFS doesn't generate them on. Turn this into an EIO
1862 * and shut down the filesystem.
1864 switch (ophdr
->oh_clientid
) {
1865 case XFS_TRANSACTION
:
1871 "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1872 ophdr
->oh_clientid
, ticket
);
1880 * Set up the parameters of the region copy into the log. This has
1881 * to handle region write split across multiple log buffers - this
1882 * state is kept external to this function so that this code can
1883 * can be written in an obvious, self documenting manner.
1886 xlog_write_setup_copy(
1887 struct xlog_ticket
*ticket
,
1888 struct xlog_op_header
*ophdr
,
1889 int space_available
,
1893 int *last_was_partial_copy
,
1894 int *bytes_consumed
)
1898 still_to_copy
= space_required
- *bytes_consumed
;
1899 *copy_off
= *bytes_consumed
;
1901 if (still_to_copy
<= space_available
) {
1902 /* write of region completes here */
1903 *copy_len
= still_to_copy
;
1904 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
1905 if (*last_was_partial_copy
)
1906 ophdr
->oh_flags
|= (XLOG_END_TRANS
|XLOG_WAS_CONT_TRANS
);
1907 *last_was_partial_copy
= 0;
1908 *bytes_consumed
= 0;
1912 /* partial write of region, needs extra log op header reservation */
1913 *copy_len
= space_available
;
1914 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
1915 ophdr
->oh_flags
|= XLOG_CONTINUE_TRANS
;
1916 if (*last_was_partial_copy
)
1917 ophdr
->oh_flags
|= XLOG_WAS_CONT_TRANS
;
1918 *bytes_consumed
+= *copy_len
;
1919 (*last_was_partial_copy
)++;
1921 /* account for new log op header */
1922 ticket
->t_curr_res
-= sizeof(struct xlog_op_header
);
1923 ticket
->t_res_num_ophdrs
++;
1925 return sizeof(struct xlog_op_header
);
1929 xlog_write_copy_finish(
1931 struct xlog_in_core
*iclog
,
1936 int *partial_copy_len
,
1938 struct xlog_in_core
**commit_iclog
)
1940 if (*partial_copy
) {
1942 * This iclog has already been marked WANT_SYNC by
1943 * xlog_state_get_iclog_space.
1945 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
1948 return xlog_state_release_iclog(log
, iclog
);
1952 *partial_copy_len
= 0;
1954 if (iclog
->ic_size
- log_offset
<= sizeof(xlog_op_header_t
)) {
1955 /* no more space in this iclog - push it. */
1956 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
1960 spin_lock(&log
->l_icloglock
);
1961 xlog_state_want_sync(log
, iclog
);
1962 spin_unlock(&log
->l_icloglock
);
1965 return xlog_state_release_iclog(log
, iclog
);
1966 ASSERT(flags
& XLOG_COMMIT_TRANS
);
1967 *commit_iclog
= iclog
;
1974 * Write some region out to in-core log
1976 * This will be called when writing externally provided regions or when
1977 * writing out a commit record for a given transaction.
1979 * General algorithm:
1980 * 1. Find total length of this write. This may include adding to the
1981 * lengths passed in.
1982 * 2. Check whether we violate the tickets reservation.
1983 * 3. While writing to this iclog
1984 * A. Reserve as much space in this iclog as can get
1985 * B. If this is first write, save away start lsn
1986 * C. While writing this region:
1987 * 1. If first write of transaction, write start record
1988 * 2. Write log operation header (header per region)
1989 * 3. Find out if we can fit entire region into this iclog
1990 * 4. Potentially, verify destination memcpy ptr
1991 * 5. Memcpy (partial) region
1992 * 6. If partial copy, release iclog; otherwise, continue
1993 * copying more regions into current iclog
1994 * 4. Mark want sync bit (in simulation mode)
1995 * 5. Release iclog for potential flush to on-disk log.
1998 * 1. Panic if reservation is overrun. This should never happen since
1999 * reservation amounts are generated internal to the filesystem.
2001 * 1. Tickets are single threaded data structures.
2002 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2003 * syncing routine. When a single log_write region needs to span
2004 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2005 * on all log operation writes which don't contain the end of the
2006 * region. The XLOG_END_TRANS bit is used for the in-core log
2007 * operation which contains the end of the continued log_write region.
2008 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2009 * we don't really know exactly how much space will be used. As a result,
2010 * we don't update ic_offset until the end when we know exactly how many
2011 * bytes have been written out.
2016 struct xfs_log_vec
*log_vector
,
2017 struct xlog_ticket
*ticket
,
2018 xfs_lsn_t
*start_lsn
,
2019 struct xlog_in_core
**commit_iclog
,
2022 struct xlog_in_core
*iclog
= NULL
;
2023 struct xfs_log_iovec
*vecp
;
2024 struct xfs_log_vec
*lv
;
2027 int partial_copy
= 0;
2028 int partial_copy_len
= 0;
2036 len
= xlog_write_calc_vec_length(ticket
, log_vector
);
2039 * Region headers and bytes are already accounted for.
2040 * We only need to take into account start records and
2041 * split regions in this function.
2043 if (ticket
->t_flags
& XLOG_TIC_INITED
)
2044 ticket
->t_curr_res
-= sizeof(xlog_op_header_t
);
2047 * Commit record headers need to be accounted for. These
2048 * come in as separate writes so are easy to detect.
2050 if (flags
& (XLOG_COMMIT_TRANS
| XLOG_UNMOUNT_TRANS
))
2051 ticket
->t_curr_res
-= sizeof(xlog_op_header_t
);
2053 if (ticket
->t_curr_res
< 0)
2054 xlog_print_tic_res(log
->l_mp
, ticket
);
2058 vecp
= lv
->lv_iovecp
;
2059 while (lv
&& index
< lv
->lv_niovecs
) {
2063 error
= xlog_state_get_iclog_space(log
, len
, &iclog
, ticket
,
2064 &contwr
, &log_offset
);
2068 ASSERT(log_offset
<= iclog
->ic_size
- 1);
2069 ptr
= iclog
->ic_datap
+ log_offset
;
2071 /* start_lsn is the first lsn written to. That's all we need. */
2073 *start_lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
2076 * This loop writes out as many regions as can fit in the amount
2077 * of space which was allocated by xlog_state_get_iclog_space().
2079 while (lv
&& index
< lv
->lv_niovecs
) {
2080 struct xfs_log_iovec
*reg
= &vecp
[index
];
2081 struct xlog_op_header
*ophdr
;
2086 ASSERT(reg
->i_len
% sizeof(__int32_t
) == 0);
2087 ASSERT((unsigned long)ptr
% sizeof(__int32_t
) == 0);
2089 start_rec_copy
= xlog_write_start_rec(ptr
, ticket
);
2090 if (start_rec_copy
) {
2092 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2096 ophdr
= xlog_write_setup_ophdr(log
, ptr
, ticket
, flags
);
2098 return XFS_ERROR(EIO
);
2100 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2101 sizeof(struct xlog_op_header
));
2103 len
+= xlog_write_setup_copy(ticket
, ophdr
,
2104 iclog
->ic_size
-log_offset
,
2106 ©_off
, ©_len
,
2109 xlog_verify_dest_ptr(log
, ptr
);
2112 ASSERT(copy_len
>= 0);
2113 memcpy(ptr
, reg
->i_addr
+ copy_off
, copy_len
);
2114 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
, copy_len
);
2116 copy_len
+= start_rec_copy
+ sizeof(xlog_op_header_t
);
2118 data_cnt
+= contwr
? copy_len
: 0;
2120 error
= xlog_write_copy_finish(log
, iclog
, flags
,
2121 &record_cnt
, &data_cnt
,
2130 * if we had a partial copy, we need to get more iclog
2131 * space but we don't want to increment the region
2132 * index because there is still more is this region to
2135 * If we completed writing this region, and we flushed
2136 * the iclog (indicated by resetting of the record
2137 * count), then we also need to get more log space. If
2138 * this was the last record, though, we are done and
2144 if (++index
== lv
->lv_niovecs
) {
2148 vecp
= lv
->lv_iovecp
;
2150 if (record_cnt
== 0) {
2160 xlog_state_finish_copy(log
, iclog
, record_cnt
, data_cnt
);
2162 return xlog_state_release_iclog(log
, iclog
);
2164 ASSERT(flags
& XLOG_COMMIT_TRANS
);
2165 *commit_iclog
= iclog
;
2170 /*****************************************************************************
2172 * State Machine functions
2174 *****************************************************************************
2177 /* Clean iclogs starting from the head. This ordering must be
2178 * maintained, so an iclog doesn't become ACTIVE beyond one that
2179 * is SYNCING. This is also required to maintain the notion that we use
2180 * a ordered wait queue to hold off would be writers to the log when every
2181 * iclog is trying to sync to disk.
2183 * State Change: DIRTY -> ACTIVE
2186 xlog_state_clean_log(
2189 xlog_in_core_t
*iclog
;
2192 iclog
= log
->l_iclog
;
2194 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
2195 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
2196 iclog
->ic_offset
= 0;
2197 ASSERT(iclog
->ic_callback
== NULL
);
2199 * If the number of ops in this iclog indicate it just
2200 * contains the dummy transaction, we can
2201 * change state into IDLE (the second time around).
2202 * Otherwise we should change the state into
2204 * We don't need to cover the dummy.
2207 (be32_to_cpu(iclog
->ic_header
.h_num_logops
) ==
2212 * We have two dirty iclogs so start over
2213 * This could also be num of ops indicates
2214 * this is not the dummy going out.
2218 iclog
->ic_header
.h_num_logops
= 0;
2219 memset(iclog
->ic_header
.h_cycle_data
, 0,
2220 sizeof(iclog
->ic_header
.h_cycle_data
));
2221 iclog
->ic_header
.h_lsn
= 0;
2222 } else if (iclog
->ic_state
== XLOG_STATE_ACTIVE
)
2225 break; /* stop cleaning */
2226 iclog
= iclog
->ic_next
;
2227 } while (iclog
!= log
->l_iclog
);
2229 /* log is locked when we are called */
2231 * Change state for the dummy log recording.
2232 * We usually go to NEED. But we go to NEED2 if the changed indicates
2233 * we are done writing the dummy record.
2234 * If we are done with the second dummy recored (DONE2), then
2238 switch (log
->l_covered_state
) {
2239 case XLOG_STATE_COVER_IDLE
:
2240 case XLOG_STATE_COVER_NEED
:
2241 case XLOG_STATE_COVER_NEED2
:
2242 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2245 case XLOG_STATE_COVER_DONE
:
2247 log
->l_covered_state
= XLOG_STATE_COVER_NEED2
;
2249 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2252 case XLOG_STATE_COVER_DONE2
:
2254 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
2256 log
->l_covered_state
= XLOG_STATE_COVER_NEED
;
2263 } /* xlog_state_clean_log */
2266 xlog_get_lowest_lsn(
2269 xlog_in_core_t
*lsn_log
;
2270 xfs_lsn_t lowest_lsn
, lsn
;
2272 lsn_log
= log
->l_iclog
;
2275 if (!(lsn_log
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
))) {
2276 lsn
= be64_to_cpu(lsn_log
->ic_header
.h_lsn
);
2277 if ((lsn
&& !lowest_lsn
) ||
2278 (XFS_LSN_CMP(lsn
, lowest_lsn
) < 0)) {
2282 lsn_log
= lsn_log
->ic_next
;
2283 } while (lsn_log
!= log
->l_iclog
);
2289 xlog_state_do_callback(
2292 struct xlog_in_core
*ciclog
)
2294 xlog_in_core_t
*iclog
;
2295 xlog_in_core_t
*first_iclog
; /* used to know when we've
2296 * processed all iclogs once */
2297 xfs_log_callback_t
*cb
, *cb_next
;
2299 xfs_lsn_t lowest_lsn
;
2300 int ioerrors
; /* counter: iclogs with errors */
2301 int loopdidcallbacks
; /* flag: inner loop did callbacks*/
2302 int funcdidcallbacks
; /* flag: function did callbacks */
2303 int repeats
; /* for issuing console warnings if
2304 * looping too many times */
2307 spin_lock(&log
->l_icloglock
);
2308 first_iclog
= iclog
= log
->l_iclog
;
2310 funcdidcallbacks
= 0;
2315 * Scan all iclogs starting with the one pointed to by the
2316 * log. Reset this starting point each time the log is
2317 * unlocked (during callbacks).
2319 * Keep looping through iclogs until one full pass is made
2320 * without running any callbacks.
2322 first_iclog
= log
->l_iclog
;
2323 iclog
= log
->l_iclog
;
2324 loopdidcallbacks
= 0;
2329 /* skip all iclogs in the ACTIVE & DIRTY states */
2330 if (iclog
->ic_state
&
2331 (XLOG_STATE_ACTIVE
|XLOG_STATE_DIRTY
)) {
2332 iclog
= iclog
->ic_next
;
2337 * Between marking a filesystem SHUTDOWN and stopping
2338 * the log, we do flush all iclogs to disk (if there
2339 * wasn't a log I/O error). So, we do want things to
2340 * go smoothly in case of just a SHUTDOWN w/o a
2343 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
2345 * Can only perform callbacks in order. Since
2346 * this iclog is not in the DONE_SYNC/
2347 * DO_CALLBACK state, we skip the rest and
2348 * just try to clean up. If we set our iclog
2349 * to DO_CALLBACK, we will not process it when
2350 * we retry since a previous iclog is in the
2351 * CALLBACK and the state cannot change since
2352 * we are holding the l_icloglock.
2354 if (!(iclog
->ic_state
&
2355 (XLOG_STATE_DONE_SYNC
|
2356 XLOG_STATE_DO_CALLBACK
))) {
2357 if (ciclog
&& (ciclog
->ic_state
==
2358 XLOG_STATE_DONE_SYNC
)) {
2359 ciclog
->ic_state
= XLOG_STATE_DO_CALLBACK
;
2364 * We now have an iclog that is in either the
2365 * DO_CALLBACK or DONE_SYNC states. The other
2366 * states (WANT_SYNC, SYNCING, or CALLBACK were
2367 * caught by the above if and are going to
2368 * clean (i.e. we aren't doing their callbacks)
2373 * We will do one more check here to see if we
2374 * have chased our tail around.
2377 lowest_lsn
= xlog_get_lowest_lsn(log
);
2379 XFS_LSN_CMP(lowest_lsn
,
2380 be64_to_cpu(iclog
->ic_header
.h_lsn
)) < 0) {
2381 iclog
= iclog
->ic_next
;
2382 continue; /* Leave this iclog for
2386 iclog
->ic_state
= XLOG_STATE_CALLBACK
;
2390 * Completion of a iclog IO does not imply that
2391 * a transaction has completed, as transactions
2392 * can be large enough to span many iclogs. We
2393 * cannot change the tail of the log half way
2394 * through a transaction as this may be the only
2395 * transaction in the log and moving th etail to
2396 * point to the middle of it will prevent
2397 * recovery from finding the start of the
2398 * transaction. Hence we should only update the
2399 * last_sync_lsn if this iclog contains
2400 * transaction completion callbacks on it.
2402 * We have to do this before we drop the
2403 * icloglock to ensure we are the only one that
2406 ASSERT(XFS_LSN_CMP(atomic64_read(&log
->l_last_sync_lsn
),
2407 be64_to_cpu(iclog
->ic_header
.h_lsn
)) <= 0);
2408 if (iclog
->ic_callback
)
2409 atomic64_set(&log
->l_last_sync_lsn
,
2410 be64_to_cpu(iclog
->ic_header
.h_lsn
));
2415 spin_unlock(&log
->l_icloglock
);
2418 * Keep processing entries in the callback list until
2419 * we come around and it is empty. We need to
2420 * atomically see that the list is empty and change the
2421 * state to DIRTY so that we don't miss any more
2422 * callbacks being added.
2424 spin_lock(&iclog
->ic_callback_lock
);
2425 cb
= iclog
->ic_callback
;
2427 iclog
->ic_callback_tail
= &(iclog
->ic_callback
);
2428 iclog
->ic_callback
= NULL
;
2429 spin_unlock(&iclog
->ic_callback_lock
);
2431 /* perform callbacks in the order given */
2432 for (; cb
; cb
= cb_next
) {
2433 cb_next
= cb
->cb_next
;
2434 cb
->cb_func(cb
->cb_arg
, aborted
);
2436 spin_lock(&iclog
->ic_callback_lock
);
2437 cb
= iclog
->ic_callback
;
2443 spin_lock(&log
->l_icloglock
);
2444 ASSERT(iclog
->ic_callback
== NULL
);
2445 spin_unlock(&iclog
->ic_callback_lock
);
2446 if (!(iclog
->ic_state
& XLOG_STATE_IOERROR
))
2447 iclog
->ic_state
= XLOG_STATE_DIRTY
;
2450 * Transition from DIRTY to ACTIVE if applicable.
2451 * NOP if STATE_IOERROR.
2453 xlog_state_clean_log(log
);
2455 /* wake up threads waiting in xfs_log_force() */
2456 wake_up_all(&iclog
->ic_force_wait
);
2458 iclog
= iclog
->ic_next
;
2459 } while (first_iclog
!= iclog
);
2461 if (repeats
> 5000) {
2462 flushcnt
+= repeats
;
2465 "%s: possible infinite loop (%d iterations)",
2466 __func__
, flushcnt
);
2468 } while (!ioerrors
&& loopdidcallbacks
);
2471 * make one last gasp attempt to see if iclogs are being left in
2475 if (funcdidcallbacks
) {
2476 first_iclog
= iclog
= log
->l_iclog
;
2478 ASSERT(iclog
->ic_state
!= XLOG_STATE_DO_CALLBACK
);
2480 * Terminate the loop if iclogs are found in states
2481 * which will cause other threads to clean up iclogs.
2483 * SYNCING - i/o completion will go through logs
2484 * DONE_SYNC - interrupt thread should be waiting for
2486 * IOERROR - give up hope all ye who enter here
2488 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
||
2489 iclog
->ic_state
== XLOG_STATE_SYNCING
||
2490 iclog
->ic_state
== XLOG_STATE_DONE_SYNC
||
2491 iclog
->ic_state
== XLOG_STATE_IOERROR
)
2493 iclog
= iclog
->ic_next
;
2494 } while (first_iclog
!= iclog
);
2498 if (log
->l_iclog
->ic_state
& (XLOG_STATE_ACTIVE
|XLOG_STATE_IOERROR
))
2500 spin_unlock(&log
->l_icloglock
);
2503 wake_up_all(&log
->l_flush_wait
);
2508 * Finish transitioning this iclog to the dirty state.
2510 * Make sure that we completely execute this routine only when this is
2511 * the last call to the iclog. There is a good chance that iclog flushes,
2512 * when we reach the end of the physical log, get turned into 2 separate
2513 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2514 * routine. By using the reference count bwritecnt, we guarantee that only
2515 * the second completion goes through.
2517 * Callbacks could take time, so they are done outside the scope of the
2518 * global state machine log lock.
2521 xlog_state_done_syncing(
2522 xlog_in_core_t
*iclog
,
2525 struct xlog
*log
= iclog
->ic_log
;
2527 spin_lock(&log
->l_icloglock
);
2529 ASSERT(iclog
->ic_state
== XLOG_STATE_SYNCING
||
2530 iclog
->ic_state
== XLOG_STATE_IOERROR
);
2531 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
2532 ASSERT(iclog
->ic_bwritecnt
== 1 || iclog
->ic_bwritecnt
== 2);
2536 * If we got an error, either on the first buffer, or in the case of
2537 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2538 * and none should ever be attempted to be written to disk
2541 if (iclog
->ic_state
!= XLOG_STATE_IOERROR
) {
2542 if (--iclog
->ic_bwritecnt
== 1) {
2543 spin_unlock(&log
->l_icloglock
);
2546 iclog
->ic_state
= XLOG_STATE_DONE_SYNC
;
2550 * Someone could be sleeping prior to writing out the next
2551 * iclog buffer, we wake them all, one will get to do the
2552 * I/O, the others get to wait for the result.
2554 wake_up_all(&iclog
->ic_write_wait
);
2555 spin_unlock(&log
->l_icloglock
);
2556 xlog_state_do_callback(log
, aborted
, iclog
); /* also cleans log */
2557 } /* xlog_state_done_syncing */
2561 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2562 * sleep. We wait on the flush queue on the head iclog as that should be
2563 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2564 * we will wait here and all new writes will sleep until a sync completes.
2566 * The in-core logs are used in a circular fashion. They are not used
2567 * out-of-order even when an iclog past the head is free.
2570 * * log_offset where xlog_write() can start writing into the in-core
2572 * * in-core log pointer to which xlog_write() should write.
2573 * * boolean indicating this is a continued write to an in-core log.
2574 * If this is the last write, then the in-core log's offset field
2575 * needs to be incremented, depending on the amount of data which
2579 xlog_state_get_iclog_space(
2582 struct xlog_in_core
**iclogp
,
2583 struct xlog_ticket
*ticket
,
2584 int *continued_write
,
2588 xlog_rec_header_t
*head
;
2589 xlog_in_core_t
*iclog
;
2593 spin_lock(&log
->l_icloglock
);
2594 if (XLOG_FORCED_SHUTDOWN(log
)) {
2595 spin_unlock(&log
->l_icloglock
);
2596 return XFS_ERROR(EIO
);
2599 iclog
= log
->l_iclog
;
2600 if (iclog
->ic_state
!= XLOG_STATE_ACTIVE
) {
2601 XFS_STATS_INC(xs_log_noiclogs
);
2603 /* Wait for log writes to have flushed */
2604 xlog_wait(&log
->l_flush_wait
, &log
->l_icloglock
);
2608 head
= &iclog
->ic_header
;
2610 atomic_inc(&iclog
->ic_refcnt
); /* prevents sync */
2611 log_offset
= iclog
->ic_offset
;
2613 /* On the 1st write to an iclog, figure out lsn. This works
2614 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2615 * committing to. If the offset is set, that's how many blocks
2618 if (log_offset
== 0) {
2619 ticket
->t_curr_res
-= log
->l_iclog_hsize
;
2620 xlog_tic_add_region(ticket
,
2622 XLOG_REG_TYPE_LRHEADER
);
2623 head
->h_cycle
= cpu_to_be32(log
->l_curr_cycle
);
2624 head
->h_lsn
= cpu_to_be64(
2625 xlog_assign_lsn(log
->l_curr_cycle
, log
->l_curr_block
));
2626 ASSERT(log
->l_curr_block
>= 0);
2629 /* If there is enough room to write everything, then do it. Otherwise,
2630 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2631 * bit is on, so this will get flushed out. Don't update ic_offset
2632 * until you know exactly how many bytes get copied. Therefore, wait
2633 * until later to update ic_offset.
2635 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2636 * can fit into remaining data section.
2638 if (iclog
->ic_size
- iclog
->ic_offset
< 2*sizeof(xlog_op_header_t
)) {
2639 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2642 * If I'm the only one writing to this iclog, sync it to disk.
2643 * We need to do an atomic compare and decrement here to avoid
2644 * racing with concurrent atomic_dec_and_lock() calls in
2645 * xlog_state_release_iclog() when there is more than one
2646 * reference to the iclog.
2648 if (!atomic_add_unless(&iclog
->ic_refcnt
, -1, 1)) {
2649 /* we are the only one */
2650 spin_unlock(&log
->l_icloglock
);
2651 error
= xlog_state_release_iclog(log
, iclog
);
2655 spin_unlock(&log
->l_icloglock
);
2660 /* Do we have enough room to write the full amount in the remainder
2661 * of this iclog? Or must we continue a write on the next iclog and
2662 * mark this iclog as completely taken? In the case where we switch
2663 * iclogs (to mark it taken), this particular iclog will release/sync
2664 * to disk in xlog_write().
2666 if (len
<= iclog
->ic_size
- iclog
->ic_offset
) {
2667 *continued_write
= 0;
2668 iclog
->ic_offset
+= len
;
2670 *continued_write
= 1;
2671 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
2675 ASSERT(iclog
->ic_offset
<= iclog
->ic_size
);
2676 spin_unlock(&log
->l_icloglock
);
2678 *logoffsetp
= log_offset
;
2680 } /* xlog_state_get_iclog_space */
2682 /* The first cnt-1 times through here we don't need to
2683 * move the grant write head because the permanent
2684 * reservation has reserved cnt times the unit amount.
2685 * Release part of current permanent unit reservation and
2686 * reset current reservation to be one units worth. Also
2687 * move grant reservation head forward.
2690 xlog_regrant_reserve_log_space(
2692 struct xlog_ticket
*ticket
)
2694 trace_xfs_log_regrant_reserve_enter(log
, ticket
);
2696 if (ticket
->t_cnt
> 0)
2699 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
,
2700 ticket
->t_curr_res
);
2701 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
,
2702 ticket
->t_curr_res
);
2703 ticket
->t_curr_res
= ticket
->t_unit_res
;
2704 xlog_tic_reset_res(ticket
);
2706 trace_xfs_log_regrant_reserve_sub(log
, ticket
);
2708 /* just return if we still have some of the pre-reserved space */
2709 if (ticket
->t_cnt
> 0)
2712 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
,
2713 ticket
->t_unit_res
);
2715 trace_xfs_log_regrant_reserve_exit(log
, ticket
);
2717 ticket
->t_curr_res
= ticket
->t_unit_res
;
2718 xlog_tic_reset_res(ticket
);
2719 } /* xlog_regrant_reserve_log_space */
2723 * Give back the space left from a reservation.
2725 * All the information we need to make a correct determination of space left
2726 * is present. For non-permanent reservations, things are quite easy. The
2727 * count should have been decremented to zero. We only need to deal with the
2728 * space remaining in the current reservation part of the ticket. If the
2729 * ticket contains a permanent reservation, there may be left over space which
2730 * needs to be released. A count of N means that N-1 refills of the current
2731 * reservation can be done before we need to ask for more space. The first
2732 * one goes to fill up the first current reservation. Once we run out of
2733 * space, the count will stay at zero and the only space remaining will be
2734 * in the current reservation field.
2737 xlog_ungrant_log_space(
2739 struct xlog_ticket
*ticket
)
2743 if (ticket
->t_cnt
> 0)
2746 trace_xfs_log_ungrant_enter(log
, ticket
);
2747 trace_xfs_log_ungrant_sub(log
, ticket
);
2750 * If this is a permanent reservation ticket, we may be able to free
2751 * up more space based on the remaining count.
2753 bytes
= ticket
->t_curr_res
;
2754 if (ticket
->t_cnt
> 0) {
2755 ASSERT(ticket
->t_flags
& XLOG_TIC_PERM_RESERV
);
2756 bytes
+= ticket
->t_unit_res
*ticket
->t_cnt
;
2759 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
, bytes
);
2760 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
, bytes
);
2762 trace_xfs_log_ungrant_exit(log
, ticket
);
2764 xfs_log_space_wake(log
->l_mp
);
2768 * Flush iclog to disk if this is the last reference to the given iclog and
2769 * the WANT_SYNC bit is set.
2771 * When this function is entered, the iclog is not necessarily in the
2772 * WANT_SYNC state. It may be sitting around waiting to get filled.
2777 xlog_state_release_iclog(
2779 struct xlog_in_core
*iclog
)
2781 int sync
= 0; /* do we sync? */
2783 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
2784 return XFS_ERROR(EIO
);
2786 ASSERT(atomic_read(&iclog
->ic_refcnt
) > 0);
2787 if (!atomic_dec_and_lock(&iclog
->ic_refcnt
, &log
->l_icloglock
))
2790 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
2791 spin_unlock(&log
->l_icloglock
);
2792 return XFS_ERROR(EIO
);
2794 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2795 iclog
->ic_state
== XLOG_STATE_WANT_SYNC
);
2797 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
) {
2798 /* update tail before writing to iclog */
2799 xfs_lsn_t tail_lsn
= xlog_assign_tail_lsn(log
->l_mp
);
2801 iclog
->ic_state
= XLOG_STATE_SYNCING
;
2802 iclog
->ic_header
.h_tail_lsn
= cpu_to_be64(tail_lsn
);
2803 xlog_verify_tail_lsn(log
, iclog
, tail_lsn
);
2804 /* cycle incremented when incrementing curr_block */
2806 spin_unlock(&log
->l_icloglock
);
2809 * We let the log lock go, so it's possible that we hit a log I/O
2810 * error or some other SHUTDOWN condition that marks the iclog
2811 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2812 * this iclog has consistent data, so we ignore IOERROR
2813 * flags after this point.
2816 return xlog_sync(log
, iclog
);
2818 } /* xlog_state_release_iclog */
2822 * This routine will mark the current iclog in the ring as WANT_SYNC
2823 * and move the current iclog pointer to the next iclog in the ring.
2824 * When this routine is called from xlog_state_get_iclog_space(), the
2825 * exact size of the iclog has not yet been determined. All we know is
2826 * that every data block. We have run out of space in this log record.
2829 xlog_state_switch_iclogs(
2831 struct xlog_in_core
*iclog
,
2834 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
);
2836 eventual_size
= iclog
->ic_offset
;
2837 iclog
->ic_state
= XLOG_STATE_WANT_SYNC
;
2838 iclog
->ic_header
.h_prev_block
= cpu_to_be32(log
->l_prev_block
);
2839 log
->l_prev_block
= log
->l_curr_block
;
2840 log
->l_prev_cycle
= log
->l_curr_cycle
;
2842 /* roll log?: ic_offset changed later */
2843 log
->l_curr_block
+= BTOBB(eventual_size
)+BTOBB(log
->l_iclog_hsize
);
2845 /* Round up to next log-sunit */
2846 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) &&
2847 log
->l_mp
->m_sb
.sb_logsunit
> 1) {
2848 __uint32_t sunit_bb
= BTOBB(log
->l_mp
->m_sb
.sb_logsunit
);
2849 log
->l_curr_block
= roundup(log
->l_curr_block
, sunit_bb
);
2852 if (log
->l_curr_block
>= log
->l_logBBsize
) {
2853 log
->l_curr_cycle
++;
2854 if (log
->l_curr_cycle
== XLOG_HEADER_MAGIC_NUM
)
2855 log
->l_curr_cycle
++;
2856 log
->l_curr_block
-= log
->l_logBBsize
;
2857 ASSERT(log
->l_curr_block
>= 0);
2859 ASSERT(iclog
== log
->l_iclog
);
2860 log
->l_iclog
= iclog
->ic_next
;
2861 } /* xlog_state_switch_iclogs */
2864 * Write out all data in the in-core log as of this exact moment in time.
2866 * Data may be written to the in-core log during this call. However,
2867 * we don't guarantee this data will be written out. A change from past
2868 * implementation means this routine will *not* write out zero length LRs.
2870 * Basically, we try and perform an intelligent scan of the in-core logs.
2871 * If we determine there is no flushable data, we just return. There is no
2872 * flushable data if:
2874 * 1. the current iclog is active and has no data; the previous iclog
2875 * is in the active or dirty state.
2876 * 2. the current iclog is drity, and the previous iclog is in the
2877 * active or dirty state.
2881 * 1. the current iclog is not in the active nor dirty state.
2882 * 2. the current iclog dirty, and the previous iclog is not in the
2883 * active nor dirty state.
2884 * 3. the current iclog is active, and there is another thread writing
2885 * to this particular iclog.
2886 * 4. a) the current iclog is active and has no other writers
2887 * b) when we return from flushing out this iclog, it is still
2888 * not in the active nor dirty state.
2892 struct xfs_mount
*mp
,
2896 struct xlog
*log
= mp
->m_log
;
2897 struct xlog_in_core
*iclog
;
2900 XFS_STATS_INC(xs_log_force
);
2902 xlog_cil_force(log
);
2904 spin_lock(&log
->l_icloglock
);
2906 iclog
= log
->l_iclog
;
2907 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
2908 spin_unlock(&log
->l_icloglock
);
2909 return XFS_ERROR(EIO
);
2912 /* If the head iclog is not active nor dirty, we just attach
2913 * ourselves to the head and go to sleep.
2915 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2916 iclog
->ic_state
== XLOG_STATE_DIRTY
) {
2918 * If the head is dirty or (active and empty), then
2919 * we need to look at the previous iclog. If the previous
2920 * iclog is active or dirty we are done. There is nothing
2921 * to sync out. Otherwise, we attach ourselves to the
2922 * previous iclog and go to sleep.
2924 if (iclog
->ic_state
== XLOG_STATE_DIRTY
||
2925 (atomic_read(&iclog
->ic_refcnt
) == 0
2926 && iclog
->ic_offset
== 0)) {
2927 iclog
= iclog
->ic_prev
;
2928 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2929 iclog
->ic_state
== XLOG_STATE_DIRTY
)
2934 if (atomic_read(&iclog
->ic_refcnt
) == 0) {
2935 /* We are the only one with access to this
2936 * iclog. Flush it out now. There should
2937 * be a roundoff of zero to show that someone
2938 * has already taken care of the roundoff from
2939 * the previous sync.
2941 atomic_inc(&iclog
->ic_refcnt
);
2942 lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
2943 xlog_state_switch_iclogs(log
, iclog
, 0);
2944 spin_unlock(&log
->l_icloglock
);
2946 if (xlog_state_release_iclog(log
, iclog
))
2947 return XFS_ERROR(EIO
);
2951 spin_lock(&log
->l_icloglock
);
2952 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) == lsn
&&
2953 iclog
->ic_state
!= XLOG_STATE_DIRTY
)
2958 /* Someone else is writing to this iclog.
2959 * Use its call to flush out the data. However,
2960 * the other thread may not force out this LR,
2961 * so we mark it WANT_SYNC.
2963 xlog_state_switch_iclogs(log
, iclog
, 0);
2969 /* By the time we come around again, the iclog could've been filled
2970 * which would give it another lsn. If we have a new lsn, just
2971 * return because the relevant data has been flushed.
2974 if (flags
& XFS_LOG_SYNC
) {
2976 * We must check if we're shutting down here, before
2977 * we wait, while we're holding the l_icloglock.
2978 * Then we check again after waking up, in case our
2979 * sleep was disturbed by a bad news.
2981 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
2982 spin_unlock(&log
->l_icloglock
);
2983 return XFS_ERROR(EIO
);
2985 XFS_STATS_INC(xs_log_force_sleep
);
2986 xlog_wait(&iclog
->ic_force_wait
, &log
->l_icloglock
);
2988 * No need to grab the log lock here since we're
2989 * only deciding whether or not to return EIO
2990 * and the memory read should be atomic.
2992 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
2993 return XFS_ERROR(EIO
);
2999 spin_unlock(&log
->l_icloglock
);
3005 * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3006 * about errors or whether the log was flushed or not. This is the normal
3007 * interface to use when trying to unpin items or move the log forward.
3016 trace_xfs_log_force(mp
, 0);
3017 error
= _xfs_log_force(mp
, flags
, NULL
);
3019 xfs_warn(mp
, "%s: error %d returned.", __func__
, error
);
3023 * Force the in-core log to disk for a specific LSN.
3025 * Find in-core log with lsn.
3026 * If it is in the DIRTY state, just return.
3027 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3028 * state and go to sleep or return.
3029 * If it is in any other state, go to sleep or return.
3031 * Synchronous forces are implemented with a signal variable. All callers
3032 * to force a given lsn to disk will wait on a the sv attached to the
3033 * specific in-core log. When given in-core log finally completes its
3034 * write to disk, that thread will wake up all threads waiting on the
3039 struct xfs_mount
*mp
,
3044 struct xlog
*log
= mp
->m_log
;
3045 struct xlog_in_core
*iclog
;
3046 int already_slept
= 0;
3050 XFS_STATS_INC(xs_log_force
);
3052 lsn
= xlog_cil_force_lsn(log
, lsn
);
3053 if (lsn
== NULLCOMMITLSN
)
3057 spin_lock(&log
->l_icloglock
);
3058 iclog
= log
->l_iclog
;
3059 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3060 spin_unlock(&log
->l_icloglock
);
3061 return XFS_ERROR(EIO
);
3065 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) != lsn
) {
3066 iclog
= iclog
->ic_next
;
3070 if (iclog
->ic_state
== XLOG_STATE_DIRTY
) {
3071 spin_unlock(&log
->l_icloglock
);
3075 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3077 * We sleep here if we haven't already slept (e.g.
3078 * this is the first time we've looked at the correct
3079 * iclog buf) and the buffer before us is going to
3080 * be sync'ed. The reason for this is that if we
3081 * are doing sync transactions here, by waiting for
3082 * the previous I/O to complete, we can allow a few
3083 * more transactions into this iclog before we close
3086 * Otherwise, we mark the buffer WANT_SYNC, and bump
3087 * up the refcnt so we can release the log (which
3088 * drops the ref count). The state switch keeps new
3089 * transaction commits from using this buffer. When
3090 * the current commits finish writing into the buffer,
3091 * the refcount will drop to zero and the buffer will
3094 if (!already_slept
&&
3095 (iclog
->ic_prev
->ic_state
&
3096 (XLOG_STATE_WANT_SYNC
| XLOG_STATE_SYNCING
))) {
3097 ASSERT(!(iclog
->ic_state
& XLOG_STATE_IOERROR
));
3099 XFS_STATS_INC(xs_log_force_sleep
);
3101 xlog_wait(&iclog
->ic_prev
->ic_write_wait
,
3108 atomic_inc(&iclog
->ic_refcnt
);
3109 xlog_state_switch_iclogs(log
, iclog
, 0);
3110 spin_unlock(&log
->l_icloglock
);
3111 if (xlog_state_release_iclog(log
, iclog
))
3112 return XFS_ERROR(EIO
);
3115 spin_lock(&log
->l_icloglock
);
3118 if ((flags
& XFS_LOG_SYNC
) && /* sleep */
3120 (XLOG_STATE_ACTIVE
| XLOG_STATE_DIRTY
))) {
3122 * Don't wait on completion if we know that we've
3123 * gotten a log write error.
3125 if (iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3126 spin_unlock(&log
->l_icloglock
);
3127 return XFS_ERROR(EIO
);
3129 XFS_STATS_INC(xs_log_force_sleep
);
3130 xlog_wait(&iclog
->ic_force_wait
, &log
->l_icloglock
);
3132 * No need to grab the log lock here since we're
3133 * only deciding whether or not to return EIO
3134 * and the memory read should be atomic.
3136 if (iclog
->ic_state
& XLOG_STATE_IOERROR
)
3137 return XFS_ERROR(EIO
);
3141 } else { /* just return */
3142 spin_unlock(&log
->l_icloglock
);
3146 } while (iclog
!= log
->l_iclog
);
3148 spin_unlock(&log
->l_icloglock
);
3153 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3154 * about errors or whether the log was flushed or not. This is the normal
3155 * interface to use when trying to unpin items or move the log forward.
3165 trace_xfs_log_force(mp
, lsn
);
3166 error
= _xfs_log_force_lsn(mp
, lsn
, flags
, NULL
);
3168 xfs_warn(mp
, "%s: error %d returned.", __func__
, error
);
3172 * Called when we want to mark the current iclog as being ready to sync to
3176 xlog_state_want_sync(
3178 struct xlog_in_core
*iclog
)
3180 assert_spin_locked(&log
->l_icloglock
);
3182 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3183 xlog_state_switch_iclogs(log
, iclog
, 0);
3185 ASSERT(iclog
->ic_state
&
3186 (XLOG_STATE_WANT_SYNC
|XLOG_STATE_IOERROR
));
3191 /*****************************************************************************
3195 *****************************************************************************
3199 * Free a used ticket when its refcount falls to zero.
3203 xlog_ticket_t
*ticket
)
3205 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3206 if (atomic_dec_and_test(&ticket
->t_ref
))
3207 kmem_zone_free(xfs_log_ticket_zone
, ticket
);
3212 xlog_ticket_t
*ticket
)
3214 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3215 atomic_inc(&ticket
->t_ref
);
3220 * Allocate and initialise a new log ticket.
3222 struct xlog_ticket
*
3229 xfs_km_flags_t alloc_flags
)
3231 struct xlog_ticket
*tic
;
3235 tic
= kmem_zone_zalloc(xfs_log_ticket_zone
, alloc_flags
);
3240 * Permanent reservations have up to 'cnt'-1 active log operations
3241 * in the log. A unit in this case is the amount of space for one
3242 * of these log operations. Normal reservations have a cnt of 1
3243 * and their unit amount is the total amount of space required.
3245 * The following lines of code account for non-transaction data
3246 * which occupy space in the on-disk log.
3248 * Normal form of a transaction is:
3249 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3250 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3252 * We need to account for all the leadup data and trailer data
3253 * around the transaction data.
3254 * And then we need to account for the worst case in terms of using
3256 * The worst case will happen if:
3257 * - the placement of the transaction happens to be such that the
3258 * roundoff is at its maximum
3259 * - the transaction data is synced before the commit record is synced
3260 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3261 * Therefore the commit record is in its own Log Record.
3262 * This can happen as the commit record is called with its
3263 * own region to xlog_write().
3264 * This then means that in the worst case, roundoff can happen for
3265 * the commit-rec as well.
3266 * The commit-rec is smaller than padding in this scenario and so it is
3267 * not added separately.
3270 /* for trans header */
3271 unit_bytes
+= sizeof(xlog_op_header_t
);
3272 unit_bytes
+= sizeof(xfs_trans_header_t
);
3275 unit_bytes
+= sizeof(xlog_op_header_t
);
3278 * for LR headers - the space for data in an iclog is the size minus
3279 * the space used for the headers. If we use the iclog size, then we
3280 * undercalculate the number of headers required.
3282 * Furthermore - the addition of op headers for split-recs might
3283 * increase the space required enough to require more log and op
3284 * headers, so take that into account too.
3286 * IMPORTANT: This reservation makes the assumption that if this
3287 * transaction is the first in an iclog and hence has the LR headers
3288 * accounted to it, then the remaining space in the iclog is
3289 * exclusively for this transaction. i.e. if the transaction is larger
3290 * than the iclog, it will be the only thing in that iclog.
3291 * Fundamentally, this means we must pass the entire log vector to
3292 * xlog_write to guarantee this.
3294 iclog_space
= log
->l_iclog_size
- log
->l_iclog_hsize
;
3295 num_headers
= howmany(unit_bytes
, iclog_space
);
3297 /* for split-recs - ophdrs added when data split over LRs */
3298 unit_bytes
+= sizeof(xlog_op_header_t
) * num_headers
;
3300 /* add extra header reservations if we overrun */
3301 while (!num_headers
||
3302 howmany(unit_bytes
, iclog_space
) > num_headers
) {
3303 unit_bytes
+= sizeof(xlog_op_header_t
);
3306 unit_bytes
+= log
->l_iclog_hsize
* num_headers
;
3308 /* for commit-rec LR header - note: padding will subsume the ophdr */
3309 unit_bytes
+= log
->l_iclog_hsize
;
3311 /* for roundoff padding for transaction data and one for commit record */
3312 if (xfs_sb_version_haslogv2(&log
->l_mp
->m_sb
) &&
3313 log
->l_mp
->m_sb
.sb_logsunit
> 1) {
3314 /* log su roundoff */
3315 unit_bytes
+= 2*log
->l_mp
->m_sb
.sb_logsunit
;
3318 unit_bytes
+= 2*BBSIZE
;
3321 atomic_set(&tic
->t_ref
, 1);
3322 tic
->t_task
= current
;
3323 INIT_LIST_HEAD(&tic
->t_queue
);
3324 tic
->t_unit_res
= unit_bytes
;
3325 tic
->t_curr_res
= unit_bytes
;
3328 tic
->t_tid
= random32();
3329 tic
->t_clientid
= client
;
3330 tic
->t_flags
= XLOG_TIC_INITED
;
3331 tic
->t_trans_type
= 0;
3333 tic
->t_flags
|= XLOG_TIC_PERM_RESERV
;
3335 xlog_tic_reset_res(tic
);
3341 /******************************************************************************
3343 * Log debug routines
3345 ******************************************************************************
3349 * Make sure that the destination ptr is within the valid data region of
3350 * one of the iclogs. This uses backup pointers stored in a different
3351 * part of the log in case we trash the log structure.
3354 xlog_verify_dest_ptr(
3361 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
3362 if (ptr
>= log
->l_iclog_bak
[i
] &&
3363 ptr
<= log
->l_iclog_bak
[i
] + log
->l_iclog_size
)
3368 xfs_emerg(log
->l_mp
, "%s: invalid ptr", __func__
);
3372 * Check to make sure the grant write head didn't just over lap the tail. If
3373 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3374 * the cycles differ by exactly one and check the byte count.
3376 * This check is run unlocked, so can give false positives. Rather than assert
3377 * on failures, use a warn-once flag and a panic tag to allow the admin to
3378 * determine if they want to panic the machine when such an error occurs. For
3379 * debug kernels this will have the same effect as using an assert but, unlinke
3380 * an assert, it can be turned off at runtime.
3383 xlog_verify_grant_tail(
3386 int tail_cycle
, tail_blocks
;
3389 xlog_crack_grant_head(&log
->l_write_head
.grant
, &cycle
, &space
);
3390 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_blocks
);
3391 if (tail_cycle
!= cycle
) {
3392 if (cycle
- 1 != tail_cycle
&&
3393 !(log
->l_flags
& XLOG_TAIL_WARN
)) {
3394 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3395 "%s: cycle - 1 != tail_cycle", __func__
);
3396 log
->l_flags
|= XLOG_TAIL_WARN
;
3399 if (space
> BBTOB(tail_blocks
) &&
3400 !(log
->l_flags
& XLOG_TAIL_WARN
)) {
3401 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3402 "%s: space > BBTOB(tail_blocks)", __func__
);
3403 log
->l_flags
|= XLOG_TAIL_WARN
;
3408 /* check if it will fit */
3410 xlog_verify_tail_lsn(
3412 struct xlog_in_core
*iclog
,
3417 if (CYCLE_LSN(tail_lsn
) == log
->l_prev_cycle
) {
3419 log
->l_logBBsize
- (log
->l_prev_block
- BLOCK_LSN(tail_lsn
));
3420 if (blocks
< BTOBB(iclog
->ic_offset
)+BTOBB(log
->l_iclog_hsize
))
3421 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3423 ASSERT(CYCLE_LSN(tail_lsn
)+1 == log
->l_prev_cycle
);
3425 if (BLOCK_LSN(tail_lsn
) == log
->l_prev_block
)
3426 xfs_emerg(log
->l_mp
, "%s: tail wrapped", __func__
);
3428 blocks
= BLOCK_LSN(tail_lsn
) - log
->l_prev_block
;
3429 if (blocks
< BTOBB(iclog
->ic_offset
) + 1)
3430 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3432 } /* xlog_verify_tail_lsn */
3435 * Perform a number of checks on the iclog before writing to disk.
3437 * 1. Make sure the iclogs are still circular
3438 * 2. Make sure we have a good magic number
3439 * 3. Make sure we don't have magic numbers in the data
3440 * 4. Check fields of each log operation header for:
3441 * A. Valid client identifier
3442 * B. tid ptr value falls in valid ptr space (user space code)
3443 * C. Length in log record header is correct according to the
3444 * individual operation headers within record.
3445 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3446 * log, check the preceding blocks of the physical log to make sure all
3447 * the cycle numbers agree with the current cycle number.
3452 struct xlog_in_core
*iclog
,
3456 xlog_op_header_t
*ophead
;
3457 xlog_in_core_t
*icptr
;
3458 xlog_in_core_2_t
*xhdr
;
3460 xfs_caddr_t base_ptr
;
3461 __psint_t field_offset
;
3463 int len
, i
, j
, k
, op_len
;
3466 /* check validity of iclog pointers */
3467 spin_lock(&log
->l_icloglock
);
3468 icptr
= log
->l_iclog
;
3469 for (i
=0; i
< log
->l_iclog_bufs
; i
++) {
3471 xfs_emerg(log
->l_mp
, "%s: invalid ptr", __func__
);
3472 icptr
= icptr
->ic_next
;
3474 if (icptr
!= log
->l_iclog
)
3475 xfs_emerg(log
->l_mp
, "%s: corrupt iclog ring", __func__
);
3476 spin_unlock(&log
->l_icloglock
);
3478 /* check log magic numbers */
3479 if (iclog
->ic_header
.h_magicno
!= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3480 xfs_emerg(log
->l_mp
, "%s: invalid magic num", __func__
);
3482 ptr
= (xfs_caddr_t
) &iclog
->ic_header
;
3483 for (ptr
+= BBSIZE
; ptr
< ((xfs_caddr_t
)&iclog
->ic_header
) + count
;
3485 if (*(__be32
*)ptr
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3486 xfs_emerg(log
->l_mp
, "%s: unexpected magic num",
3491 len
= be32_to_cpu(iclog
->ic_header
.h_num_logops
);
3492 ptr
= iclog
->ic_datap
;
3494 ophead
= (xlog_op_header_t
*)ptr
;
3495 xhdr
= iclog
->ic_data
;
3496 for (i
= 0; i
< len
; i
++) {
3497 ophead
= (xlog_op_header_t
*)ptr
;
3499 /* clientid is only 1 byte */
3500 field_offset
= (__psint_t
)
3501 ((xfs_caddr_t
)&(ophead
->oh_clientid
) - base_ptr
);
3502 if (syncing
== B_FALSE
|| (field_offset
& 0x1ff)) {
3503 clientid
= ophead
->oh_clientid
;
3505 idx
= BTOBBT((xfs_caddr_t
)&(ophead
->oh_clientid
) - iclog
->ic_datap
);
3506 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3507 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3508 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3509 clientid
= xlog_get_client_id(
3510 xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3512 clientid
= xlog_get_client_id(
3513 iclog
->ic_header
.h_cycle_data
[idx
]);
3516 if (clientid
!= XFS_TRANSACTION
&& clientid
!= XFS_LOG
)
3518 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3519 __func__
, clientid
, ophead
,
3520 (unsigned long)field_offset
);
3523 field_offset
= (__psint_t
)
3524 ((xfs_caddr_t
)&(ophead
->oh_len
) - base_ptr
);
3525 if (syncing
== B_FALSE
|| (field_offset
& 0x1ff)) {
3526 op_len
= be32_to_cpu(ophead
->oh_len
);
3528 idx
= BTOBBT((__psint_t
)&ophead
->oh_len
-
3529 (__psint_t
)iclog
->ic_datap
);
3530 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3531 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3532 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3533 op_len
= be32_to_cpu(xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3535 op_len
= be32_to_cpu(iclog
->ic_header
.h_cycle_data
[idx
]);
3538 ptr
+= sizeof(xlog_op_header_t
) + op_len
;
3540 } /* xlog_verify_iclog */
3544 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3550 xlog_in_core_t
*iclog
, *ic
;
3552 iclog
= log
->l_iclog
;
3553 if (! (iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
3555 * Mark all the incore logs IOERROR.
3556 * From now on, no log flushes will result.
3560 ic
->ic_state
= XLOG_STATE_IOERROR
;
3562 } while (ic
!= iclog
);
3566 * Return non-zero, if state transition has already happened.
3572 * This is called from xfs_force_shutdown, when we're forcibly
3573 * shutting down the filesystem, typically because of an IO error.
3574 * Our main objectives here are to make sure that:
3575 * a. the filesystem gets marked 'SHUTDOWN' for all interested
3576 * parties to find out, 'atomically'.
3577 * b. those who're sleeping on log reservations, pinned objects and
3578 * other resources get woken up, and be told the bad news.
3579 * c. nothing new gets queued up after (a) and (b) are done.
3580 * d. if !logerror, flush the iclogs to disk, then seal them off
3583 * Note: for delayed logging the !logerror case needs to flush the regions
3584 * held in memory out to the iclogs before flushing them to disk. This needs
3585 * to be done before the log is marked as shutdown, otherwise the flush to the
3589 xfs_log_force_umount(
3590 struct xfs_mount
*mp
,
3599 * If this happens during log recovery, don't worry about
3600 * locking; the log isn't open for business yet.
3603 log
->l_flags
& XLOG_ACTIVE_RECOVERY
) {
3604 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3606 XFS_BUF_DONE(mp
->m_sb_bp
);
3611 * Somebody could've already done the hard work for us.
3612 * No need to get locks for this.
3614 if (logerror
&& log
->l_iclog
->ic_state
& XLOG_STATE_IOERROR
) {
3615 ASSERT(XLOG_FORCED_SHUTDOWN(log
));
3621 * Flush the in memory commit item list before marking the log as
3622 * being shut down. We need to do it in this order to ensure all the
3623 * completed transactions are flushed to disk with the xfs_log_force()
3627 xlog_cil_force(log
);
3630 * mark the filesystem and the as in a shutdown state and wake
3631 * everybody up to tell them the bad news.
3633 spin_lock(&log
->l_icloglock
);
3634 mp
->m_flags
|= XFS_MOUNT_FS_SHUTDOWN
;
3636 XFS_BUF_DONE(mp
->m_sb_bp
);
3639 * This flag is sort of redundant because of the mount flag, but
3640 * it's good to maintain the separation between the log and the rest
3643 log
->l_flags
|= XLOG_IO_ERROR
;
3646 * If we hit a log error, we want to mark all the iclogs IOERROR
3647 * while we're still holding the loglock.
3650 retval
= xlog_state_ioerror(log
);
3651 spin_unlock(&log
->l_icloglock
);
3654 * We don't want anybody waiting for log reservations after this. That
3655 * means we have to wake up everybody queued up on reserveq as well as
3656 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3657 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3658 * action is protected by the grant locks.
3660 xlog_grant_head_wake_all(&log
->l_reserve_head
);
3661 xlog_grant_head_wake_all(&log
->l_write_head
);
3663 if (!(log
->l_iclog
->ic_state
& XLOG_STATE_IOERROR
)) {
3666 * Force the incore logs to disk before shutting the
3667 * log down completely.
3669 _xfs_log_force(mp
, XFS_LOG_SYNC
, NULL
);
3671 spin_lock(&log
->l_icloglock
);
3672 retval
= xlog_state_ioerror(log
);
3673 spin_unlock(&log
->l_icloglock
);
3676 * Wake up everybody waiting on xfs_log_force.
3677 * Callback all log item committed functions as if the
3678 * log writes were completed.
3680 xlog_state_do_callback(log
, XFS_LI_ABORTED
, NULL
);
3682 #ifdef XFSERRORDEBUG
3684 xlog_in_core_t
*iclog
;
3686 spin_lock(&log
->l_icloglock
);
3687 iclog
= log
->l_iclog
;
3689 ASSERT(iclog
->ic_callback
== 0);
3690 iclog
= iclog
->ic_next
;
3691 } while (iclog
!= log
->l_iclog
);
3692 spin_unlock(&log
->l_icloglock
);
3695 /* return non-zero if log IOERROR transition had already happened */
3703 xlog_in_core_t
*iclog
;
3705 iclog
= log
->l_iclog
;
3707 /* endianness does not matter here, zero is zero in
3710 if (iclog
->ic_header
.h_num_logops
)
3712 iclog
= iclog
->ic_next
;
3713 } while (iclog
!= log
->l_iclog
);