1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_errortag.h"
14 #include "xfs_error.h"
15 #include "xfs_trans.h"
16 #include "xfs_trans_priv.h"
18 #include "xfs_log_priv.h"
19 #include "xfs_trace.h"
20 #include "xfs_sysfs.h"
22 #include "xfs_health.h"
24 kmem_zone_t
*xfs_log_ticket_zone
;
26 /* Local miscellaneous function prototypes */
30 struct xfs_buftarg
*log_target
,
31 xfs_daddr_t blk_offset
,
41 /* local state machine functions */
42 STATIC
void xlog_state_done_syncing(
43 struct xlog_in_core
*iclog
);
44 STATIC
void xlog_state_do_callback(
47 xlog_state_get_iclog_space(
50 struct xlog_in_core
**iclog
,
51 struct xlog_ticket
*ticket
,
61 struct xlog_in_core
*iclog
);
68 xlog_verify_grant_tail(
73 struct xlog_in_core
*iclog
,
78 struct xlog_in_core
*iclog
);
80 #define xlog_verify_dest_ptr(a,b)
81 #define xlog_verify_grant_tail(a)
82 #define xlog_verify_iclog(a,b,c)
83 #define xlog_verify_tail_lsn(a,b)
91 xfs_log_cover(struct xfs_mount
*);
99 int64_t head_val
= atomic64_read(head
);
105 xlog_crack_grant_head_val(head_val
, &cycle
, &space
);
109 space
+= log
->l_logsize
;
114 new = xlog_assign_grant_head_val(cycle
, space
);
115 head_val
= atomic64_cmpxchg(head
, old
, new);
116 } while (head_val
!= old
);
120 xlog_grant_add_space(
125 int64_t head_val
= atomic64_read(head
);
132 xlog_crack_grant_head_val(head_val
, &cycle
, &space
);
134 tmp
= log
->l_logsize
- space
;
143 new = xlog_assign_grant_head_val(cycle
, space
);
144 head_val
= atomic64_cmpxchg(head
, old
, new);
145 } while (head_val
!= old
);
149 xlog_grant_head_init(
150 struct xlog_grant_head
*head
)
152 xlog_assign_grant_head(&head
->grant
, 1, 0);
153 INIT_LIST_HEAD(&head
->waiters
);
154 spin_lock_init(&head
->lock
);
158 xlog_grant_head_wake_all(
159 struct xlog_grant_head
*head
)
161 struct xlog_ticket
*tic
;
163 spin_lock(&head
->lock
);
164 list_for_each_entry(tic
, &head
->waiters
, t_queue
)
165 wake_up_process(tic
->t_task
);
166 spin_unlock(&head
->lock
);
170 xlog_ticket_reservation(
172 struct xlog_grant_head
*head
,
173 struct xlog_ticket
*tic
)
175 if (head
== &log
->l_write_head
) {
176 ASSERT(tic
->t_flags
& XLOG_TIC_PERM_RESERV
);
177 return tic
->t_unit_res
;
179 if (tic
->t_flags
& XLOG_TIC_PERM_RESERV
)
180 return tic
->t_unit_res
* tic
->t_cnt
;
182 return tic
->t_unit_res
;
187 xlog_grant_head_wake(
189 struct xlog_grant_head
*head
,
192 struct xlog_ticket
*tic
;
194 bool woken_task
= false;
196 list_for_each_entry(tic
, &head
->waiters
, t_queue
) {
199 * There is a chance that the size of the CIL checkpoints in
200 * progress at the last AIL push target calculation resulted in
201 * limiting the target to the log head (l_last_sync_lsn) at the
202 * time. This may not reflect where the log head is now as the
203 * CIL checkpoints may have completed.
205 * Hence when we are woken here, it may be that the head of the
206 * log that has moved rather than the tail. As the tail didn't
207 * move, there still won't be space available for the
208 * reservation we require. However, if the AIL has already
209 * pushed to the target defined by the old log head location, we
210 * will hang here waiting for something else to update the AIL
213 * Therefore, if there isn't space to wake the first waiter on
214 * the grant head, we need to push the AIL again to ensure the
215 * target reflects both the current log tail and log head
216 * position before we wait for the tail to move again.
219 need_bytes
= xlog_ticket_reservation(log
, head
, tic
);
220 if (*free_bytes
< need_bytes
) {
222 xlog_grant_push_ail(log
, need_bytes
);
226 *free_bytes
-= need_bytes
;
227 trace_xfs_log_grant_wake_up(log
, tic
);
228 wake_up_process(tic
->t_task
);
236 xlog_grant_head_wait(
238 struct xlog_grant_head
*head
,
239 struct xlog_ticket
*tic
,
240 int need_bytes
) __releases(&head
->lock
)
241 __acquires(&head
->lock
)
243 list_add_tail(&tic
->t_queue
, &head
->waiters
);
246 if (xlog_is_shutdown(log
))
248 xlog_grant_push_ail(log
, need_bytes
);
250 __set_current_state(TASK_UNINTERRUPTIBLE
);
251 spin_unlock(&head
->lock
);
253 XFS_STATS_INC(log
->l_mp
, xs_sleep_logspace
);
255 trace_xfs_log_grant_sleep(log
, tic
);
257 trace_xfs_log_grant_wake(log
, tic
);
259 spin_lock(&head
->lock
);
260 if (xlog_is_shutdown(log
))
262 } while (xlog_space_left(log
, &head
->grant
) < need_bytes
);
264 list_del_init(&tic
->t_queue
);
267 list_del_init(&tic
->t_queue
);
272 * Atomically get the log space required for a log ticket.
274 * Once a ticket gets put onto head->waiters, it will only return after the
275 * needed reservation is satisfied.
277 * This function is structured so that it has a lock free fast path. This is
278 * necessary because every new transaction reservation will come through this
279 * path. Hence any lock will be globally hot if we take it unconditionally on
282 * As tickets are only ever moved on and off head->waiters under head->lock, we
283 * only need to take that lock if we are going to add the ticket to the queue
284 * and sleep. We can avoid taking the lock if the ticket was never added to
285 * head->waiters because the t_queue list head will be empty and we hold the
286 * only reference to it so it can safely be checked unlocked.
289 xlog_grant_head_check(
291 struct xlog_grant_head
*head
,
292 struct xlog_ticket
*tic
,
298 ASSERT(!xlog_in_recovery(log
));
301 * If there are other waiters on the queue then give them a chance at
302 * logspace before us. Wake up the first waiters, if we do not wake
303 * up all the waiters then go to sleep waiting for more free space,
304 * otherwise try to get some space for this transaction.
306 *need_bytes
= xlog_ticket_reservation(log
, head
, tic
);
307 free_bytes
= xlog_space_left(log
, &head
->grant
);
308 if (!list_empty_careful(&head
->waiters
)) {
309 spin_lock(&head
->lock
);
310 if (!xlog_grant_head_wake(log
, head
, &free_bytes
) ||
311 free_bytes
< *need_bytes
) {
312 error
= xlog_grant_head_wait(log
, head
, tic
,
315 spin_unlock(&head
->lock
);
316 } else if (free_bytes
< *need_bytes
) {
317 spin_lock(&head
->lock
);
318 error
= xlog_grant_head_wait(log
, head
, tic
, *need_bytes
);
319 spin_unlock(&head
->lock
);
326 xlog_tic_reset_res(xlog_ticket_t
*tic
)
329 tic
->t_res_arr_sum
= 0;
330 tic
->t_res_num_ophdrs
= 0;
334 xlog_tic_add_region(xlog_ticket_t
*tic
, uint len
, uint type
)
336 if (tic
->t_res_num
== XLOG_TIC_LEN_MAX
) {
337 /* add to overflow and start again */
338 tic
->t_res_o_flow
+= tic
->t_res_arr_sum
;
340 tic
->t_res_arr_sum
= 0;
343 tic
->t_res_arr
[tic
->t_res_num
].r_len
= len
;
344 tic
->t_res_arr
[tic
->t_res_num
].r_type
= type
;
345 tic
->t_res_arr_sum
+= len
;
351 struct xfs_mount
*mp
)
354 * Do not write to the log on norecovery mounts, if the data or log
355 * devices are read-only, or if the filesystem is shutdown. Read-only
356 * mounts allow internal writes for log recovery and unmount purposes,
357 * so don't restrict that case.
359 if (xfs_has_norecovery(mp
))
361 if (xfs_readonly_buftarg(mp
->m_ddev_targp
))
363 if (xfs_readonly_buftarg(mp
->m_log
->l_targ
))
365 if (xlog_is_shutdown(mp
->m_log
))
371 * Replenish the byte reservation required by moving the grant write head.
375 struct xfs_mount
*mp
,
376 struct xlog_ticket
*tic
)
378 struct xlog
*log
= mp
->m_log
;
382 if (xlog_is_shutdown(log
))
385 XFS_STATS_INC(mp
, xs_try_logspace
);
388 * This is a new transaction on the ticket, so we need to change the
389 * transaction ID so that the next transaction has a different TID in
390 * the log. Just add one to the existing tid so that we can see chains
391 * of rolling transactions in the log easily.
395 xlog_grant_push_ail(log
, tic
->t_unit_res
);
397 tic
->t_curr_res
= tic
->t_unit_res
;
398 xlog_tic_reset_res(tic
);
403 trace_xfs_log_regrant(log
, tic
);
405 error
= xlog_grant_head_check(log
, &log
->l_write_head
, tic
,
410 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, need_bytes
);
411 trace_xfs_log_regrant_exit(log
, tic
);
412 xlog_verify_grant_tail(log
);
417 * If we are failing, make sure the ticket doesn't have any current
418 * reservations. We don't want to add this back when the ticket/
419 * transaction gets cancelled.
422 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
427 * Reserve log space and return a ticket corresponding to the reservation.
429 * Each reservation is going to reserve extra space for a log record header.
430 * When writes happen to the on-disk log, we don't subtract the length of the
431 * log record header from any reservation. By wasting space in each
432 * reservation, we prevent over allocation problems.
436 struct xfs_mount
*mp
,
439 struct xlog_ticket
**ticp
,
443 struct xlog
*log
= mp
->m_log
;
444 struct xlog_ticket
*tic
;
448 ASSERT(client
== XFS_TRANSACTION
|| client
== XFS_LOG
);
450 if (xlog_is_shutdown(log
))
453 XFS_STATS_INC(mp
, xs_try_logspace
);
455 ASSERT(*ticp
== NULL
);
456 tic
= xlog_ticket_alloc(log
, unit_bytes
, cnt
, client
, permanent
);
459 xlog_grant_push_ail(log
, tic
->t_cnt
? tic
->t_unit_res
* tic
->t_cnt
462 trace_xfs_log_reserve(log
, tic
);
464 error
= xlog_grant_head_check(log
, &log
->l_reserve_head
, tic
,
469 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
, need_bytes
);
470 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, need_bytes
);
471 trace_xfs_log_reserve_exit(log
, tic
);
472 xlog_verify_grant_tail(log
);
477 * If we are failing, make sure the ticket doesn't have any current
478 * reservations. We don't want to add this back when the ticket/
479 * transaction gets cancelled.
482 tic
->t_cnt
= 0; /* ungrant will give back unit_res * t_cnt. */
487 * Run all the pending iclog callbacks and wake log force waiters and iclog
488 * space waiters so they can process the newly set shutdown state. We really
489 * don't care what order we process callbacks here because the log is shut down
490 * and so state cannot change on disk anymore.
492 * We avoid processing actively referenced iclogs so that we don't run callbacks
493 * while the iclog owner might still be preparing the iclog for IO submssion.
494 * These will be caught by xlog_state_iclog_release() and call this function
495 * again to process any callbacks that may have been added to that iclog.
498 xlog_state_shutdown_callbacks(
501 struct xlog_in_core
*iclog
;
504 spin_lock(&log
->l_icloglock
);
505 iclog
= log
->l_iclog
;
507 if (atomic_read(&iclog
->ic_refcnt
)) {
508 /* Reference holder will re-run iclog callbacks. */
511 list_splice_init(&iclog
->ic_callbacks
, &cb_list
);
512 wake_up_all(&iclog
->ic_write_wait
);
513 wake_up_all(&iclog
->ic_force_wait
);
514 } while ((iclog
= iclog
->ic_next
) != log
->l_iclog
);
516 wake_up_all(&log
->l_flush_wait
);
517 spin_unlock(&log
->l_icloglock
);
519 xlog_cil_process_committed(&cb_list
);
523 * Flush iclog to disk if this is the last reference to the given iclog and the
524 * it is in the WANT_SYNC state.
526 * If the caller passes in a non-zero @old_tail_lsn and the current log tail
527 * does not match, there may be metadata on disk that must be persisted before
528 * this iclog is written. To satisfy that requirement, set the
529 * XLOG_ICL_NEED_FLUSH flag as a condition for writing this iclog with the new
532 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
533 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
534 * written to stable storage, and implies that a commit record is contained
535 * within the iclog. We need to ensure that the log tail does not move beyond
536 * the tail that the first commit record in the iclog ordered against, otherwise
537 * correct recovery of that checkpoint becomes dependent on future operations
538 * performed on this iclog.
540 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
541 * current tail into iclog. Once the iclog tail is set, future operations must
542 * not modify it, otherwise they potentially violate ordering constraints for
543 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
544 * the iclog will get zeroed on activation of the iclog after sync, so we
545 * always capture the tail lsn on the iclog on the first NEED_FUA release
546 * regardless of the number of active reference counts on this iclog.
550 xlog_state_release_iclog(
552 struct xlog_in_core
*iclog
,
553 xfs_lsn_t old_tail_lsn
)
558 lockdep_assert_held(&log
->l_icloglock
);
560 trace_xlog_iclog_release(iclog
, _RET_IP_
);
562 * Grabbing the current log tail needs to be atomic w.r.t. the writing
563 * of the tail LSN into the iclog so we guarantee that the log tail does
564 * not move between deciding if a cache flush is required and writing
565 * the LSN into the iclog below.
567 if (old_tail_lsn
|| iclog
->ic_state
== XLOG_STATE_WANT_SYNC
) {
568 tail_lsn
= xlog_assign_tail_lsn(log
->l_mp
);
570 if (old_tail_lsn
&& tail_lsn
!= old_tail_lsn
)
571 iclog
->ic_flags
|= XLOG_ICL_NEED_FLUSH
;
573 if ((iclog
->ic_flags
& XLOG_ICL_NEED_FUA
) &&
574 !iclog
->ic_header
.h_tail_lsn
)
575 iclog
->ic_header
.h_tail_lsn
= cpu_to_be64(tail_lsn
);
578 last_ref
= atomic_dec_and_test(&iclog
->ic_refcnt
);
580 if (xlog_is_shutdown(log
)) {
582 * If there are no more references to this iclog, process the
583 * pending iclog callbacks that were waiting on the release of
587 spin_unlock(&log
->l_icloglock
);
588 xlog_state_shutdown_callbacks(log
);
589 spin_lock(&log
->l_icloglock
);
597 if (iclog
->ic_state
!= XLOG_STATE_WANT_SYNC
) {
598 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
);
602 iclog
->ic_state
= XLOG_STATE_SYNCING
;
603 if (!iclog
->ic_header
.h_tail_lsn
)
604 iclog
->ic_header
.h_tail_lsn
= cpu_to_be64(tail_lsn
);
605 xlog_verify_tail_lsn(log
, iclog
);
606 trace_xlog_iclog_syncing(iclog
, _RET_IP_
);
608 spin_unlock(&log
->l_icloglock
);
609 xlog_sync(log
, iclog
);
610 spin_lock(&log
->l_icloglock
);
615 * Mount a log filesystem
617 * mp - ubiquitous xfs mount point structure
618 * log_target - buftarg of on-disk log device
619 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
620 * num_bblocks - Number of BBSIZE blocks in on-disk log
622 * Return error or zero.
627 xfs_buftarg_t
*log_target
,
628 xfs_daddr_t blk_offset
,
632 bool fatal
= xfs_has_crc(mp
);
636 if (!xfs_has_norecovery(mp
)) {
637 xfs_notice(mp
, "Mounting V%d Filesystem",
638 XFS_SB_VERSION_NUM(&mp
->m_sb
));
641 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
642 XFS_SB_VERSION_NUM(&mp
->m_sb
));
643 ASSERT(xfs_is_readonly(mp
));
646 log
= xlog_alloc_log(mp
, log_target
, blk_offset
, num_bblks
);
648 error
= PTR_ERR(log
);
654 * Validate the given log space and drop a critical message via syslog
655 * if the log size is too small that would lead to some unexpected
656 * situations in transaction log space reservation stage.
658 * Note: we can't just reject the mount if the validation fails. This
659 * would mean that people would have to downgrade their kernel just to
660 * remedy the situation as there is no way to grow the log (short of
661 * black magic surgery with xfs_db).
663 * We can, however, reject mounts for CRC format filesystems, as the
664 * mkfs binary being used to make the filesystem should never create a
665 * filesystem with a log that is too small.
667 min_logfsbs
= xfs_log_calc_minimum_size(mp
);
669 if (mp
->m_sb
.sb_logblocks
< min_logfsbs
) {
671 "Log size %d blocks too small, minimum size is %d blocks",
672 mp
->m_sb
.sb_logblocks
, min_logfsbs
);
674 } else if (mp
->m_sb
.sb_logblocks
> XFS_MAX_LOG_BLOCKS
) {
676 "Log size %d blocks too large, maximum size is %lld blocks",
677 mp
->m_sb
.sb_logblocks
, XFS_MAX_LOG_BLOCKS
);
679 } else if (XFS_FSB_TO_B(mp
, mp
->m_sb
.sb_logblocks
) > XFS_MAX_LOG_BYTES
) {
681 "log size %lld bytes too large, maximum size is %lld bytes",
682 XFS_FSB_TO_B(mp
, mp
->m_sb
.sb_logblocks
),
685 } else if (mp
->m_sb
.sb_logsunit
> 1 &&
686 mp
->m_sb
.sb_logsunit
% mp
->m_sb
.sb_blocksize
) {
688 "log stripe unit %u bytes must be a multiple of block size",
689 mp
->m_sb
.sb_logsunit
);
695 * Log check errors are always fatal on v5; or whenever bad
696 * metadata leads to a crash.
699 xfs_crit(mp
, "AAIEEE! Log failed size checks. Abort!");
703 xfs_crit(mp
, "Log size out of supported range.");
705 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
709 * Initialize the AIL now we have a log.
711 error
= xfs_trans_ail_init(mp
);
713 xfs_warn(mp
, "AIL initialisation failed: error %d", error
);
716 log
->l_ailp
= mp
->m_ail
;
719 * skip log recovery on a norecovery mount. pretend it all
722 if (!xfs_has_norecovery(mp
)) {
724 * log recovery ignores readonly state and so we need to clear
725 * mount-based read only state so it can write to disk.
727 bool readonly
= test_and_clear_bit(XFS_OPSTATE_READONLY
,
729 error
= xlog_recover(log
);
731 set_bit(XFS_OPSTATE_READONLY
, &mp
->m_opstate
);
733 xfs_warn(mp
, "log mount/recovery failed: error %d",
735 xlog_recover_cancel(log
);
736 goto out_destroy_ail
;
740 error
= xfs_sysfs_init(&log
->l_kobj
, &xfs_log_ktype
, &mp
->m_kobj
,
743 goto out_destroy_ail
;
745 /* Normal transactions can now occur */
746 clear_bit(XLOG_ACTIVE_RECOVERY
, &log
->l_opstate
);
749 * Now the log has been fully initialised and we know were our
750 * space grant counters are, we can initialise the permanent ticket
751 * needed for delayed logging to work.
753 xlog_cil_init_post_recovery(log
);
758 xfs_trans_ail_destroy(mp
);
760 xlog_dealloc_log(log
);
766 * Finish the recovery of the file system. This is separate from the
767 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
768 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
771 * If we finish recovery successfully, start the background log work. If we are
772 * not doing recovery, then we have a RO filesystem and we don't need to start
776 xfs_log_mount_finish(
777 struct xfs_mount
*mp
)
779 struct xlog
*log
= mp
->m_log
;
783 if (xfs_has_norecovery(mp
)) {
784 ASSERT(xfs_is_readonly(mp
));
789 * log recovery ignores readonly state and so we need to clear
790 * mount-based read only state so it can write to disk.
792 readonly
= test_and_clear_bit(XFS_OPSTATE_READONLY
, &mp
->m_opstate
);
795 * During the second phase of log recovery, we need iget and
796 * iput to behave like they do for an active filesystem.
797 * xfs_fs_drop_inode needs to be able to prevent the deletion
798 * of inodes before we're done replaying log items on those
799 * inodes. Turn it off immediately after recovery finishes
800 * so that we don't leak the quota inodes if subsequent mount
803 * We let all inodes involved in redo item processing end up on
804 * the LRU instead of being evicted immediately so that if we do
805 * something to an unlinked inode, the irele won't cause
806 * premature truncation and freeing of the inode, which results
807 * in log recovery failure. We have to evict the unreferenced
808 * lru inodes after clearing SB_ACTIVE because we don't
809 * otherwise clean up the lru if there's a subsequent failure in
810 * xfs_mountfs, which leads to us leaking the inodes if nothing
811 * else (e.g. quotacheck) references the inodes before the
812 * mount failure occurs.
814 mp
->m_super
->s_flags
|= SB_ACTIVE
;
815 if (xlog_recovery_needed(log
))
816 error
= xlog_recover_finish(log
);
818 xfs_log_work_queue(mp
);
819 mp
->m_super
->s_flags
&= ~SB_ACTIVE
;
820 evict_inodes(mp
->m_super
);
823 * Drain the buffer LRU after log recovery. This is required for v4
824 * filesystems to avoid leaving around buffers with NULL verifier ops,
825 * but we do it unconditionally to make sure we're always in a clean
826 * cache state after mount.
828 * Don't push in the error case because the AIL may have pending intents
829 * that aren't removed until recovery is cancelled.
831 if (xlog_recovery_needed(log
)) {
833 xfs_log_force(mp
, XFS_LOG_SYNC
);
834 xfs_ail_push_all_sync(mp
->m_ail
);
836 xfs_notice(mp
, "Ending recovery (logdev: %s)",
837 mp
->m_logname
? mp
->m_logname
: "internal");
839 xfs_info(mp
, "Ending clean mount");
841 xfs_buftarg_drain(mp
->m_ddev_targp
);
843 clear_bit(XLOG_RECOVERY_NEEDED
, &log
->l_opstate
);
845 set_bit(XFS_OPSTATE_READONLY
, &mp
->m_opstate
);
847 /* Make sure the log is dead if we're returning failure. */
848 ASSERT(!error
|| xlog_is_shutdown(log
));
854 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
858 xfs_log_mount_cancel(
859 struct xfs_mount
*mp
)
861 xlog_recover_cancel(mp
->m_log
);
866 * Flush out the iclog to disk ensuring that device caches are flushed and
867 * the iclog hits stable storage before any completion waiters are woken.
871 struct xlog_in_core
*iclog
)
873 atomic_inc(&iclog
->ic_refcnt
);
874 iclog
->ic_flags
|= XLOG_ICL_NEED_FLUSH
| XLOG_ICL_NEED_FUA
;
875 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
)
876 xlog_state_switch_iclogs(iclog
->ic_log
, iclog
, 0);
877 return xlog_state_release_iclog(iclog
->ic_log
, iclog
, 0);
881 * Wait for the iclog and all prior iclogs to be written disk as required by the
882 * log force state machine. Waiting on ic_force_wait ensures iclog completions
883 * have been ordered and callbacks run before we are woken here, hence
884 * guaranteeing that all the iclogs up to this one are on stable storage.
888 struct xlog_in_core
*iclog
)
889 __releases(iclog
->ic_log
->l_icloglock
)
891 struct xlog
*log
= iclog
->ic_log
;
893 trace_xlog_iclog_wait_on(iclog
, _RET_IP_
);
894 if (!xlog_is_shutdown(log
) &&
895 iclog
->ic_state
!= XLOG_STATE_ACTIVE
&&
896 iclog
->ic_state
!= XLOG_STATE_DIRTY
) {
897 XFS_STATS_INC(log
->l_mp
, xs_log_force_sleep
);
898 xlog_wait(&iclog
->ic_force_wait
, &log
->l_icloglock
);
900 spin_unlock(&log
->l_icloglock
);
903 if (xlog_is_shutdown(log
))
909 * Write out an unmount record using the ticket provided. We have to account for
910 * the data space used in the unmount ticket as this write is not done from a
911 * transaction context that has already done the accounting for us.
914 xlog_write_unmount_record(
916 struct xlog_ticket
*ticket
)
918 struct xfs_unmount_log_format ulf
= {
919 .magic
= XLOG_UNMOUNT_TYPE
,
921 struct xfs_log_iovec reg
= {
923 .i_len
= sizeof(ulf
),
924 .i_type
= XLOG_REG_TYPE_UNMOUNT
,
926 struct xfs_log_vec vec
= {
931 /* account for space used by record data */
932 ticket
->t_curr_res
-= sizeof(ulf
);
934 return xlog_write(log
, NULL
, &vec
, ticket
, XLOG_UNMOUNT_TRANS
);
938 * Mark the filesystem clean by writing an unmount record to the head of the
945 struct xfs_mount
*mp
= log
->l_mp
;
946 struct xlog_in_core
*iclog
;
947 struct xlog_ticket
*tic
= NULL
;
950 error
= xfs_log_reserve(mp
, 600, 1, &tic
, XFS_LOG
, 0);
954 error
= xlog_write_unmount_record(log
, tic
);
956 * At this point, we're umounting anyway, so there's no point in
957 * transitioning log state to shutdown. Just continue...
961 xfs_alert(mp
, "%s: unmount record failed", __func__
);
963 spin_lock(&log
->l_icloglock
);
964 iclog
= log
->l_iclog
;
965 error
= xlog_force_iclog(iclog
);
966 xlog_wait_on_iclog(iclog
);
969 trace_xfs_log_umount_write(log
, tic
);
970 xfs_log_ticket_ungrant(log
, tic
);
975 xfs_log_unmount_verify_iclog(
978 struct xlog_in_core
*iclog
= log
->l_iclog
;
981 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
);
982 ASSERT(iclog
->ic_offset
== 0);
983 } while ((iclog
= iclog
->ic_next
) != log
->l_iclog
);
987 * Unmount record used to have a string "Unmount filesystem--" in the
988 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
989 * We just write the magic number now since that particular field isn't
990 * currently architecture converted and "Unmount" is a bit foo.
991 * As far as I know, there weren't any dependencies on the old behaviour.
994 xfs_log_unmount_write(
995 struct xfs_mount
*mp
)
997 struct xlog
*log
= mp
->m_log
;
999 if (!xfs_log_writable(mp
))
1002 xfs_log_force(mp
, XFS_LOG_SYNC
);
1004 if (xlog_is_shutdown(log
))
1008 * If we think the summary counters are bad, avoid writing the unmount
1009 * record to force log recovery at next mount, after which the summary
1010 * counters will be recalculated. Refer to xlog_check_unmount_rec for
1013 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp
, XFS_SICK_FS_COUNTERS
), mp
,
1014 XFS_ERRTAG_FORCE_SUMMARY_RECALC
)) {
1015 xfs_alert(mp
, "%s: will fix summary counters at next mount",
1020 xfs_log_unmount_verify_iclog(log
);
1021 xlog_unmount_write(log
);
1025 * Empty the log for unmount/freeze.
1027 * To do this, we first need to shut down the background log work so it is not
1028 * trying to cover the log as we clean up. We then need to unpin all objects in
1029 * the log so we can then flush them out. Once they have completed their IO and
1030 * run the callbacks removing themselves from the AIL, we can cover the log.
1034 struct xfs_mount
*mp
)
1037 * Clear log incompat features since we're quiescing the log. Report
1038 * failures, though it's not fatal to have a higher log feature
1039 * protection level than the log contents actually require.
1041 if (xfs_clear_incompat_log_features(mp
)) {
1044 error
= xfs_sync_sb(mp
, false);
1047 "Failed to clear log incompat features on quiesce");
1050 cancel_delayed_work_sync(&mp
->m_log
->l_work
);
1051 xfs_log_force(mp
, XFS_LOG_SYNC
);
1054 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1055 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1056 * xfs_buf_iowait() cannot be used because it was pushed with the
1057 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1058 * the IO to complete.
1060 xfs_ail_push_all_sync(mp
->m_ail
);
1061 xfs_buftarg_wait(mp
->m_ddev_targp
);
1062 xfs_buf_lock(mp
->m_sb_bp
);
1063 xfs_buf_unlock(mp
->m_sb_bp
);
1065 return xfs_log_cover(mp
);
1070 struct xfs_mount
*mp
)
1072 xfs_log_quiesce(mp
);
1073 xfs_log_unmount_write(mp
);
1077 * Shut down and release the AIL and Log.
1079 * During unmount, we need to ensure we flush all the dirty metadata objects
1080 * from the AIL so that the log is empty before we write the unmount record to
1081 * the log. Once this is done, we can tear down the AIL and the log.
1085 struct xfs_mount
*mp
)
1089 xfs_buftarg_drain(mp
->m_ddev_targp
);
1091 xfs_trans_ail_destroy(mp
);
1093 xfs_sysfs_del(&mp
->m_log
->l_kobj
);
1095 xlog_dealloc_log(mp
->m_log
);
1100 struct xfs_mount
*mp
,
1101 struct xfs_log_item
*item
,
1103 const struct xfs_item_ops
*ops
)
1105 item
->li_mountp
= mp
;
1106 item
->li_ailp
= mp
->m_ail
;
1107 item
->li_type
= type
;
1111 INIT_LIST_HEAD(&item
->li_ail
);
1112 INIT_LIST_HEAD(&item
->li_cil
);
1113 INIT_LIST_HEAD(&item
->li_bio_list
);
1114 INIT_LIST_HEAD(&item
->li_trans
);
1118 * Wake up processes waiting for log space after we have moved the log tail.
1122 struct xfs_mount
*mp
)
1124 struct xlog
*log
= mp
->m_log
;
1127 if (xlog_is_shutdown(log
))
1130 if (!list_empty_careful(&log
->l_write_head
.waiters
)) {
1131 ASSERT(!xlog_in_recovery(log
));
1133 spin_lock(&log
->l_write_head
.lock
);
1134 free_bytes
= xlog_space_left(log
, &log
->l_write_head
.grant
);
1135 xlog_grant_head_wake(log
, &log
->l_write_head
, &free_bytes
);
1136 spin_unlock(&log
->l_write_head
.lock
);
1139 if (!list_empty_careful(&log
->l_reserve_head
.waiters
)) {
1140 ASSERT(!xlog_in_recovery(log
));
1142 spin_lock(&log
->l_reserve_head
.lock
);
1143 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
1144 xlog_grant_head_wake(log
, &log
->l_reserve_head
, &free_bytes
);
1145 spin_unlock(&log
->l_reserve_head
.lock
);
1150 * Determine if we have a transaction that has gone to disk that needs to be
1151 * covered. To begin the transition to the idle state firstly the log needs to
1152 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1153 * we start attempting to cover the log.
1155 * Only if we are then in a state where covering is needed, the caller is
1156 * informed that dummy transactions are required to move the log into the idle
1159 * If there are any items in the AIl or CIL, then we do not want to attempt to
1160 * cover the log as we may be in a situation where there isn't log space
1161 * available to run a dummy transaction and this can lead to deadlocks when the
1162 * tail of the log is pinned by an item that is modified in the CIL. Hence
1163 * there's no point in running a dummy transaction at this point because we
1164 * can't start trying to idle the log until both the CIL and AIL are empty.
1167 xfs_log_need_covered(
1168 struct xfs_mount
*mp
)
1170 struct xlog
*log
= mp
->m_log
;
1171 bool needed
= false;
1173 if (!xlog_cil_empty(log
))
1176 spin_lock(&log
->l_icloglock
);
1177 switch (log
->l_covered_state
) {
1178 case XLOG_STATE_COVER_DONE
:
1179 case XLOG_STATE_COVER_DONE2
:
1180 case XLOG_STATE_COVER_IDLE
:
1182 case XLOG_STATE_COVER_NEED
:
1183 case XLOG_STATE_COVER_NEED2
:
1184 if (xfs_ail_min_lsn(log
->l_ailp
))
1186 if (!xlog_iclogs_empty(log
))
1190 if (log
->l_covered_state
== XLOG_STATE_COVER_NEED
)
1191 log
->l_covered_state
= XLOG_STATE_COVER_DONE
;
1193 log
->l_covered_state
= XLOG_STATE_COVER_DONE2
;
1199 spin_unlock(&log
->l_icloglock
);
1204 * Explicitly cover the log. This is similar to background log covering but
1205 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1206 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1207 * must all be empty.
1211 struct xfs_mount
*mp
)
1216 ASSERT((xlog_cil_empty(mp
->m_log
) && xlog_iclogs_empty(mp
->m_log
) &&
1217 !xfs_ail_min_lsn(mp
->m_log
->l_ailp
)) ||
1218 xlog_is_shutdown(mp
->m_log
));
1220 if (!xfs_log_writable(mp
))
1224 * xfs_log_need_covered() is not idempotent because it progresses the
1225 * state machine if the log requires covering. Therefore, we must call
1226 * this function once and use the result until we've issued an sb sync.
1227 * Do so first to make that abundantly clear.
1229 * Fall into the covering sequence if the log needs covering or the
1230 * mount has lazy superblock accounting to sync to disk. The sb sync
1231 * used for covering accumulates the in-core counters, so covering
1232 * handles this for us.
1234 need_covered
= xfs_log_need_covered(mp
);
1235 if (!need_covered
&& !xfs_has_lazysbcount(mp
))
1239 * To cover the log, commit the superblock twice (at most) in
1240 * independent checkpoints. The first serves as a reference for the
1241 * tail pointer. The sync transaction and AIL push empties the AIL and
1242 * updates the in-core tail to the LSN of the first checkpoint. The
1243 * second commit updates the on-disk tail with the in-core LSN,
1244 * covering the log. Push the AIL one more time to leave it empty, as
1248 error
= xfs_sync_sb(mp
, true);
1251 xfs_ail_push_all_sync(mp
->m_ail
);
1252 } while (xfs_log_need_covered(mp
));
1258 * We may be holding the log iclog lock upon entering this routine.
1261 xlog_assign_tail_lsn_locked(
1262 struct xfs_mount
*mp
)
1264 struct xlog
*log
= mp
->m_log
;
1265 struct xfs_log_item
*lip
;
1268 assert_spin_locked(&mp
->m_ail
->ail_lock
);
1271 * To make sure we always have a valid LSN for the log tail we keep
1272 * track of the last LSN which was committed in log->l_last_sync_lsn,
1273 * and use that when the AIL was empty.
1275 lip
= xfs_ail_min(mp
->m_ail
);
1277 tail_lsn
= lip
->li_lsn
;
1279 tail_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
1280 trace_xfs_log_assign_tail_lsn(log
, tail_lsn
);
1281 atomic64_set(&log
->l_tail_lsn
, tail_lsn
);
1286 xlog_assign_tail_lsn(
1287 struct xfs_mount
*mp
)
1291 spin_lock(&mp
->m_ail
->ail_lock
);
1292 tail_lsn
= xlog_assign_tail_lsn_locked(mp
);
1293 spin_unlock(&mp
->m_ail
->ail_lock
);
1299 * Return the space in the log between the tail and the head. The head
1300 * is passed in the cycle/bytes formal parms. In the special case where
1301 * the reserve head has wrapped passed the tail, this calculation is no
1302 * longer valid. In this case, just return 0 which means there is no space
1303 * in the log. This works for all places where this function is called
1304 * with the reserve head. Of course, if the write head were to ever
1305 * wrap the tail, we should blow up. Rather than catch this case here,
1306 * we depend on other ASSERTions in other parts of the code. XXXmiken
1308 * If reservation head is behind the tail, we have a problem. Warn about it,
1309 * but then treat it as if the log is empty.
1311 * If the log is shut down, the head and tail may be invalid or out of whack, so
1312 * shortcut invalidity asserts in this case so that we don't trigger them
1325 xlog_crack_grant_head(head
, &head_cycle
, &head_bytes
);
1326 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_bytes
);
1327 tail_bytes
= BBTOB(tail_bytes
);
1328 if (tail_cycle
== head_cycle
&& head_bytes
>= tail_bytes
)
1329 return log
->l_logsize
- (head_bytes
- tail_bytes
);
1330 if (tail_cycle
+ 1 < head_cycle
)
1333 /* Ignore potential inconsistency when shutdown. */
1334 if (xlog_is_shutdown(log
))
1335 return log
->l_logsize
;
1337 if (tail_cycle
< head_cycle
) {
1338 ASSERT(tail_cycle
== (head_cycle
- 1));
1339 return tail_bytes
- head_bytes
;
1343 * The reservation head is behind the tail. In this case we just want to
1344 * return the size of the log as the amount of space left.
1346 xfs_alert(log
->l_mp
, "xlog_space_left: head behind tail");
1347 xfs_alert(log
->l_mp
, " tail_cycle = %d, tail_bytes = %d",
1348 tail_cycle
, tail_bytes
);
1349 xfs_alert(log
->l_mp
, " GH cycle = %d, GH bytes = %d",
1350 head_cycle
, head_bytes
);
1352 return log
->l_logsize
;
1358 struct work_struct
*work
)
1360 struct xlog_in_core
*iclog
=
1361 container_of(work
, struct xlog_in_core
, ic_end_io_work
);
1362 struct xlog
*log
= iclog
->ic_log
;
1365 error
= blk_status_to_errno(iclog
->ic_bio
.bi_status
);
1367 /* treat writes with injected CRC errors as failed */
1368 if (iclog
->ic_fail_crc
)
1373 * Race to shutdown the filesystem if we see an error.
1375 if (XFS_TEST_ERROR(error
, log
->l_mp
, XFS_ERRTAG_IODONE_IOERR
)) {
1376 xfs_alert(log
->l_mp
, "log I/O error %d", error
);
1377 xfs_force_shutdown(log
->l_mp
, SHUTDOWN_LOG_IO_ERROR
);
1380 xlog_state_done_syncing(iclog
);
1381 bio_uninit(&iclog
->ic_bio
);
1384 * Drop the lock to signal that we are done. Nothing references the
1385 * iclog after this, so an unmount waiting on this lock can now tear it
1386 * down safely. As such, it is unsafe to reference the iclog after the
1387 * unlock as we could race with it being freed.
1389 up(&iclog
->ic_sema
);
1393 * Return size of each in-core log record buffer.
1395 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1397 * If the filesystem blocksize is too large, we may need to choose a
1398 * larger size since the directory code currently logs entire blocks.
1401 xlog_get_iclog_buffer_size(
1402 struct xfs_mount
*mp
,
1405 if (mp
->m_logbufs
<= 0)
1406 mp
->m_logbufs
= XLOG_MAX_ICLOGS
;
1407 if (mp
->m_logbsize
<= 0)
1408 mp
->m_logbsize
= XLOG_BIG_RECORD_BSIZE
;
1410 log
->l_iclog_bufs
= mp
->m_logbufs
;
1411 log
->l_iclog_size
= mp
->m_logbsize
;
1414 * # headers = size / 32k - one header holds cycles from 32k of data.
1416 log
->l_iclog_heads
=
1417 DIV_ROUND_UP(mp
->m_logbsize
, XLOG_HEADER_CYCLE_SIZE
);
1418 log
->l_iclog_hsize
= log
->l_iclog_heads
<< BBSHIFT
;
1423 struct xfs_mount
*mp
)
1425 queue_delayed_work(mp
->m_sync_workqueue
, &mp
->m_log
->l_work
,
1426 msecs_to_jiffies(xfs_syncd_centisecs
* 10));
1430 * Clear the log incompat flags if we have the opportunity.
1432 * This only happens if we're about to log the second dummy transaction as part
1433 * of covering the log and we can get the log incompat feature usage lock.
1436 xlog_clear_incompat(
1439 struct xfs_mount
*mp
= log
->l_mp
;
1441 if (!xfs_sb_has_incompat_log_feature(&mp
->m_sb
,
1442 XFS_SB_FEAT_INCOMPAT_LOG_ALL
))
1445 if (log
->l_covered_state
!= XLOG_STATE_COVER_DONE2
)
1448 if (!down_write_trylock(&log
->l_incompat_users
))
1451 xfs_clear_incompat_log_features(mp
);
1452 up_write(&log
->l_incompat_users
);
1456 * Every sync period we need to unpin all items in the AIL and push them to
1457 * disk. If there is nothing dirty, then we might need to cover the log to
1458 * indicate that the filesystem is idle.
1462 struct work_struct
*work
)
1464 struct xlog
*log
= container_of(to_delayed_work(work
),
1465 struct xlog
, l_work
);
1466 struct xfs_mount
*mp
= log
->l_mp
;
1468 /* dgc: errors ignored - not fatal and nowhere to report them */
1469 if (xfs_fs_writable(mp
, SB_FREEZE_WRITE
) && xfs_log_need_covered(mp
)) {
1471 * Dump a transaction into the log that contains no real change.
1472 * This is needed to stamp the current tail LSN into the log
1473 * during the covering operation.
1475 * We cannot use an inode here for this - that will push dirty
1476 * state back up into the VFS and then periodic inode flushing
1477 * will prevent log covering from making progress. Hence we
1478 * synchronously log the superblock instead to ensure the
1479 * superblock is immediately unpinned and can be written back.
1481 xlog_clear_incompat(log
);
1482 xfs_sync_sb(mp
, true);
1484 xfs_log_force(mp
, 0);
1486 /* start pushing all the metadata that is currently dirty */
1487 xfs_ail_push_all(mp
->m_ail
);
1489 /* queue us up again */
1490 xfs_log_work_queue(mp
);
1494 * This routine initializes some of the log structure for a given mount point.
1495 * Its primary purpose is to fill in enough, so recovery can occur. However,
1496 * some other stuff may be filled in too.
1498 STATIC
struct xlog
*
1500 struct xfs_mount
*mp
,
1501 struct xfs_buftarg
*log_target
,
1502 xfs_daddr_t blk_offset
,
1506 xlog_rec_header_t
*head
;
1507 xlog_in_core_t
**iclogp
;
1508 xlog_in_core_t
*iclog
, *prev_iclog
=NULL
;
1510 int error
= -ENOMEM
;
1513 log
= kmem_zalloc(sizeof(struct xlog
), KM_MAYFAIL
);
1515 xfs_warn(mp
, "Log allocation failed: No memory!");
1520 log
->l_targ
= log_target
;
1521 log
->l_logsize
= BBTOB(num_bblks
);
1522 log
->l_logBBstart
= blk_offset
;
1523 log
->l_logBBsize
= num_bblks
;
1524 log
->l_covered_state
= XLOG_STATE_COVER_IDLE
;
1525 set_bit(XLOG_ACTIVE_RECOVERY
, &log
->l_opstate
);
1526 INIT_DELAYED_WORK(&log
->l_work
, xfs_log_worker
);
1528 log
->l_prev_block
= -1;
1529 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1530 xlog_assign_atomic_lsn(&log
->l_tail_lsn
, 1, 0);
1531 xlog_assign_atomic_lsn(&log
->l_last_sync_lsn
, 1, 0);
1532 log
->l_curr_cycle
= 1; /* 0 is bad since this is initial value */
1534 if (xfs_has_logv2(mp
) && mp
->m_sb
.sb_logsunit
> 1)
1535 log
->l_iclog_roundoff
= mp
->m_sb
.sb_logsunit
;
1537 log
->l_iclog_roundoff
= BBSIZE
;
1539 xlog_grant_head_init(&log
->l_reserve_head
);
1540 xlog_grant_head_init(&log
->l_write_head
);
1542 error
= -EFSCORRUPTED
;
1543 if (xfs_has_sector(mp
)) {
1544 log2_size
= mp
->m_sb
.sb_logsectlog
;
1545 if (log2_size
< BBSHIFT
) {
1546 xfs_warn(mp
, "Log sector size too small (0x%x < 0x%x)",
1547 log2_size
, BBSHIFT
);
1551 log2_size
-= BBSHIFT
;
1552 if (log2_size
> mp
->m_sectbb_log
) {
1553 xfs_warn(mp
, "Log sector size too large (0x%x > 0x%x)",
1554 log2_size
, mp
->m_sectbb_log
);
1558 /* for larger sector sizes, must have v2 or external log */
1559 if (log2_size
&& log
->l_logBBstart
> 0 &&
1560 !xfs_has_logv2(mp
)) {
1562 "log sector size (0x%x) invalid for configuration.",
1567 log
->l_sectBBsize
= 1 << log2_size
;
1569 init_rwsem(&log
->l_incompat_users
);
1571 xlog_get_iclog_buffer_size(mp
, log
);
1573 spin_lock_init(&log
->l_icloglock
);
1574 init_waitqueue_head(&log
->l_flush_wait
);
1576 iclogp
= &log
->l_iclog
;
1578 * The amount of memory to allocate for the iclog structure is
1579 * rather funky due to the way the structure is defined. It is
1580 * done this way so that we can use different sizes for machines
1581 * with different amounts of memory. See the definition of
1582 * xlog_in_core_t in xfs_log_priv.h for details.
1584 ASSERT(log
->l_iclog_size
>= 4096);
1585 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
1586 size_t bvec_size
= howmany(log
->l_iclog_size
, PAGE_SIZE
) *
1587 sizeof(struct bio_vec
);
1589 iclog
= kmem_zalloc(sizeof(*iclog
) + bvec_size
, KM_MAYFAIL
);
1591 goto out_free_iclog
;
1594 iclog
->ic_prev
= prev_iclog
;
1597 iclog
->ic_data
= kvzalloc(log
->l_iclog_size
,
1598 GFP_KERNEL
| __GFP_RETRY_MAYFAIL
);
1599 if (!iclog
->ic_data
)
1600 goto out_free_iclog
;
1602 log
->l_iclog_bak
[i
] = &iclog
->ic_header
;
1604 head
= &iclog
->ic_header
;
1605 memset(head
, 0, sizeof(xlog_rec_header_t
));
1606 head
->h_magicno
= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
);
1607 head
->h_version
= cpu_to_be32(
1608 xfs_has_logv2(log
->l_mp
) ? 2 : 1);
1609 head
->h_size
= cpu_to_be32(log
->l_iclog_size
);
1611 head
->h_fmt
= cpu_to_be32(XLOG_FMT
);
1612 memcpy(&head
->h_fs_uuid
, &mp
->m_sb
.sb_uuid
, sizeof(uuid_t
));
1614 iclog
->ic_size
= log
->l_iclog_size
- log
->l_iclog_hsize
;
1615 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
1616 iclog
->ic_log
= log
;
1617 atomic_set(&iclog
->ic_refcnt
, 0);
1618 INIT_LIST_HEAD(&iclog
->ic_callbacks
);
1619 iclog
->ic_datap
= (char *)iclog
->ic_data
+ log
->l_iclog_hsize
;
1621 init_waitqueue_head(&iclog
->ic_force_wait
);
1622 init_waitqueue_head(&iclog
->ic_write_wait
);
1623 INIT_WORK(&iclog
->ic_end_io_work
, xlog_ioend_work
);
1624 sema_init(&iclog
->ic_sema
, 1);
1626 iclogp
= &iclog
->ic_next
;
1628 *iclogp
= log
->l_iclog
; /* complete ring */
1629 log
->l_iclog
->ic_prev
= prev_iclog
; /* re-write 1st prev ptr */
1631 log
->l_ioend_workqueue
= alloc_workqueue("xfs-log/%s",
1632 XFS_WQFLAGS(WQ_FREEZABLE
| WQ_MEM_RECLAIM
|
1634 0, mp
->m_super
->s_id
);
1635 if (!log
->l_ioend_workqueue
)
1636 goto out_free_iclog
;
1638 error
= xlog_cil_init(log
);
1640 goto out_destroy_workqueue
;
1643 out_destroy_workqueue
:
1644 destroy_workqueue(log
->l_ioend_workqueue
);
1646 for (iclog
= log
->l_iclog
; iclog
; iclog
= prev_iclog
) {
1647 prev_iclog
= iclog
->ic_next
;
1648 kmem_free(iclog
->ic_data
);
1650 if (prev_iclog
== log
->l_iclog
)
1656 return ERR_PTR(error
);
1657 } /* xlog_alloc_log */
1660 * Compute the LSN that we'd need to push the log tail towards in order to have
1661 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1662 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1663 * log free space already meets all three thresholds, this function returns
1667 xlog_grant_push_threshold(
1671 xfs_lsn_t threshold_lsn
= 0;
1672 xfs_lsn_t last_sync_lsn
;
1675 int threshold_block
;
1676 int threshold_cycle
;
1679 ASSERT(BTOBB(need_bytes
) < log
->l_logBBsize
);
1681 free_bytes
= xlog_space_left(log
, &log
->l_reserve_head
.grant
);
1682 free_blocks
= BTOBBT(free_bytes
);
1685 * Set the threshold for the minimum number of free blocks in the
1686 * log to the maximum of what the caller needs, one quarter of the
1687 * log, and 256 blocks.
1689 free_threshold
= BTOBB(need_bytes
);
1690 free_threshold
= max(free_threshold
, (log
->l_logBBsize
>> 2));
1691 free_threshold
= max(free_threshold
, 256);
1692 if (free_blocks
>= free_threshold
)
1693 return NULLCOMMITLSN
;
1695 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &threshold_cycle
,
1697 threshold_block
+= free_threshold
;
1698 if (threshold_block
>= log
->l_logBBsize
) {
1699 threshold_block
-= log
->l_logBBsize
;
1700 threshold_cycle
+= 1;
1702 threshold_lsn
= xlog_assign_lsn(threshold_cycle
,
1705 * Don't pass in an lsn greater than the lsn of the last
1706 * log record known to be on disk. Use a snapshot of the last sync lsn
1707 * so that it doesn't change between the compare and the set.
1709 last_sync_lsn
= atomic64_read(&log
->l_last_sync_lsn
);
1710 if (XFS_LSN_CMP(threshold_lsn
, last_sync_lsn
) > 0)
1711 threshold_lsn
= last_sync_lsn
;
1713 return threshold_lsn
;
1717 * Push the tail of the log if we need to do so to maintain the free log space
1718 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1719 * policy which pushes on an lsn which is further along in the log once we
1720 * reach the high water mark. In this manner, we would be creating a low water
1724 xlog_grant_push_ail(
1728 xfs_lsn_t threshold_lsn
;
1730 threshold_lsn
= xlog_grant_push_threshold(log
, need_bytes
);
1731 if (threshold_lsn
== NULLCOMMITLSN
|| xlog_is_shutdown(log
))
1735 * Get the transaction layer to kick the dirty buffers out to
1736 * disk asynchronously. No point in trying to do this if
1737 * the filesystem is shutting down.
1739 xfs_ail_push(log
->l_ailp
, threshold_lsn
);
1743 * Stamp cycle number in every block
1748 struct xlog_in_core
*iclog
,
1752 int size
= iclog
->ic_offset
+ roundoff
;
1756 cycle_lsn
= CYCLE_LSN_DISK(iclog
->ic_header
.h_lsn
);
1758 dp
= iclog
->ic_datap
;
1759 for (i
= 0; i
< BTOBB(size
); i
++) {
1760 if (i
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
))
1762 iclog
->ic_header
.h_cycle_data
[i
] = *(__be32
*)dp
;
1763 *(__be32
*)dp
= cycle_lsn
;
1767 if (xfs_has_logv2(log
->l_mp
)) {
1768 xlog_in_core_2_t
*xhdr
= iclog
->ic_data
;
1770 for ( ; i
< BTOBB(size
); i
++) {
1771 j
= i
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
1772 k
= i
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
1773 xhdr
[j
].hic_xheader
.xh_cycle_data
[k
] = *(__be32
*)dp
;
1774 *(__be32
*)dp
= cycle_lsn
;
1778 for (i
= 1; i
< log
->l_iclog_heads
; i
++)
1779 xhdr
[i
].hic_xheader
.xh_cycle
= cycle_lsn
;
1784 * Calculate the checksum for a log buffer.
1786 * This is a little more complicated than it should be because the various
1787 * headers and the actual data are non-contiguous.
1792 struct xlog_rec_header
*rhead
,
1798 /* first generate the crc for the record header ... */
1799 crc
= xfs_start_cksum_update((char *)rhead
,
1800 sizeof(struct xlog_rec_header
),
1801 offsetof(struct xlog_rec_header
, h_crc
));
1803 /* ... then for additional cycle data for v2 logs ... */
1804 if (xfs_has_logv2(log
->l_mp
)) {
1805 union xlog_in_core2
*xhdr
= (union xlog_in_core2
*)rhead
;
1809 xheads
= DIV_ROUND_UP(size
, XLOG_HEADER_CYCLE_SIZE
);
1811 for (i
= 1; i
< xheads
; i
++) {
1812 crc
= crc32c(crc
, &xhdr
[i
].hic_xheader
,
1813 sizeof(struct xlog_rec_ext_header
));
1817 /* ... and finally for the payload */
1818 crc
= crc32c(crc
, dp
, size
);
1820 return xfs_end_cksum(crc
);
1827 struct xlog_in_core
*iclog
= bio
->bi_private
;
1829 queue_work(iclog
->ic_log
->l_ioend_workqueue
,
1830 &iclog
->ic_end_io_work
);
1834 xlog_map_iclog_data(
1840 struct page
*page
= kmem_to_page(data
);
1841 unsigned int off
= offset_in_page(data
);
1842 size_t len
= min_t(size_t, count
, PAGE_SIZE
- off
);
1844 if (bio_add_page(bio
, page
, len
, off
) != len
)
1857 struct xlog_in_core
*iclog
,
1861 ASSERT(bno
< log
->l_logBBsize
);
1862 trace_xlog_iclog_write(iclog
, _RET_IP_
);
1865 * We lock the iclogbufs here so that we can serialise against I/O
1866 * completion during unmount. We might be processing a shutdown
1867 * triggered during unmount, and that can occur asynchronously to the
1868 * unmount thread, and hence we need to ensure that completes before
1869 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1870 * across the log IO to archieve that.
1872 down(&iclog
->ic_sema
);
1873 if (xlog_is_shutdown(log
)) {
1875 * It would seem logical to return EIO here, but we rely on
1876 * the log state machine to propagate I/O errors instead of
1877 * doing it here. We kick of the state machine and unlock
1878 * the buffer manually, the code needs to be kept in sync
1879 * with the I/O completion path.
1881 xlog_state_done_syncing(iclog
);
1882 up(&iclog
->ic_sema
);
1886 bio_init(&iclog
->ic_bio
, iclog
->ic_bvec
, howmany(count
, PAGE_SIZE
));
1887 bio_set_dev(&iclog
->ic_bio
, log
->l_targ
->bt_bdev
);
1888 iclog
->ic_bio
.bi_iter
.bi_sector
= log
->l_logBBstart
+ bno
;
1889 iclog
->ic_bio
.bi_end_io
= xlog_bio_end_io
;
1890 iclog
->ic_bio
.bi_private
= iclog
;
1893 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1894 * IOs coming immediately after this one. This prevents the block layer
1895 * writeback throttle from throttling log writes behind background
1896 * metadata writeback and causing priority inversions.
1898 iclog
->ic_bio
.bi_opf
= REQ_OP_WRITE
| REQ_META
| REQ_SYNC
| REQ_IDLE
;
1899 if (iclog
->ic_flags
& XLOG_ICL_NEED_FLUSH
) {
1900 iclog
->ic_bio
.bi_opf
|= REQ_PREFLUSH
;
1902 * For external log devices, we also need to flush the data
1903 * device cache first to ensure all metadata writeback covered
1904 * by the LSN in this iclog is on stable storage. This is slow,
1905 * but it *must* complete before we issue the external log IO.
1907 if (log
->l_targ
!= log
->l_mp
->m_ddev_targp
)
1908 blkdev_issue_flush(log
->l_mp
->m_ddev_targp
->bt_bdev
);
1910 if (iclog
->ic_flags
& XLOG_ICL_NEED_FUA
)
1911 iclog
->ic_bio
.bi_opf
|= REQ_FUA
;
1913 iclog
->ic_flags
&= ~(XLOG_ICL_NEED_FLUSH
| XLOG_ICL_NEED_FUA
);
1915 if (xlog_map_iclog_data(&iclog
->ic_bio
, iclog
->ic_data
, count
)) {
1916 xfs_force_shutdown(log
->l_mp
, SHUTDOWN_LOG_IO_ERROR
);
1919 if (is_vmalloc_addr(iclog
->ic_data
))
1920 flush_kernel_vmap_range(iclog
->ic_data
, count
);
1923 * If this log buffer would straddle the end of the log we will have
1924 * to split it up into two bios, so that we can continue at the start.
1926 if (bno
+ BTOBB(count
) > log
->l_logBBsize
) {
1929 split
= bio_split(&iclog
->ic_bio
, log
->l_logBBsize
- bno
,
1930 GFP_NOIO
, &fs_bio_set
);
1931 bio_chain(split
, &iclog
->ic_bio
);
1934 /* restart at logical offset zero for the remainder */
1935 iclog
->ic_bio
.bi_iter
.bi_sector
= log
->l_logBBstart
;
1938 submit_bio(&iclog
->ic_bio
);
1942 * We need to bump cycle number for the part of the iclog that is
1943 * written to the start of the log. Watch out for the header magic
1944 * number case, though.
1953 unsigned int split_offset
= BBTOB(log
->l_logBBsize
- bno
);
1956 for (i
= split_offset
; i
< count
; i
+= BBSIZE
) {
1957 uint32_t cycle
= get_unaligned_be32(data
+ i
);
1959 if (++cycle
== XLOG_HEADER_MAGIC_NUM
)
1961 put_unaligned_be32(cycle
, data
+ i
);
1966 xlog_calc_iclog_size(
1968 struct xlog_in_core
*iclog
,
1971 uint32_t count_init
, count
;
1973 /* Add for LR header */
1974 count_init
= log
->l_iclog_hsize
+ iclog
->ic_offset
;
1975 count
= roundup(count_init
, log
->l_iclog_roundoff
);
1977 *roundoff
= count
- count_init
;
1979 ASSERT(count
>= count_init
);
1980 ASSERT(*roundoff
< log
->l_iclog_roundoff
);
1985 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1986 * fashion. Previously, we should have moved the current iclog
1987 * ptr in the log to point to the next available iclog. This allows further
1988 * write to continue while this code syncs out an iclog ready to go.
1989 * Before an in-core log can be written out, the data section must be scanned
1990 * to save away the 1st word of each BBSIZE block into the header. We replace
1991 * it with the current cycle count. Each BBSIZE block is tagged with the
1992 * cycle count because there in an implicit assumption that drives will
1993 * guarantee that entire 512 byte blocks get written at once. In other words,
1994 * we can't have part of a 512 byte block written and part not written. By
1995 * tagging each block, we will know which blocks are valid when recovering
1996 * after an unclean shutdown.
1998 * This routine is single threaded on the iclog. No other thread can be in
1999 * this routine with the same iclog. Changing contents of iclog can there-
2000 * fore be done without grabbing the state machine lock. Updating the global
2001 * log will require grabbing the lock though.
2003 * The entire log manager uses a logical block numbering scheme. Only
2004 * xlog_write_iclog knows about the fact that the log may not start with
2005 * block zero on a given device.
2010 struct xlog_in_core
*iclog
)
2012 unsigned int count
; /* byte count of bwrite */
2013 unsigned int roundoff
; /* roundoff to BB or stripe */
2017 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
2018 trace_xlog_iclog_sync(iclog
, _RET_IP_
);
2020 count
= xlog_calc_iclog_size(log
, iclog
, &roundoff
);
2022 /* move grant heads by roundoff in sync */
2023 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
, roundoff
);
2024 xlog_grant_add_space(log
, &log
->l_write_head
.grant
, roundoff
);
2026 /* put cycle number in every block */
2027 xlog_pack_data(log
, iclog
, roundoff
);
2029 /* real byte length */
2030 size
= iclog
->ic_offset
;
2031 if (xfs_has_logv2(log
->l_mp
))
2033 iclog
->ic_header
.h_len
= cpu_to_be32(size
);
2035 XFS_STATS_INC(log
->l_mp
, xs_log_writes
);
2036 XFS_STATS_ADD(log
->l_mp
, xs_log_blocks
, BTOBB(count
));
2038 bno
= BLOCK_LSN(be64_to_cpu(iclog
->ic_header
.h_lsn
));
2040 /* Do we need to split this write into 2 parts? */
2041 if (bno
+ BTOBB(count
) > log
->l_logBBsize
)
2042 xlog_split_iclog(log
, &iclog
->ic_header
, bno
, count
);
2044 /* calculcate the checksum */
2045 iclog
->ic_header
.h_crc
= xlog_cksum(log
, &iclog
->ic_header
,
2046 iclog
->ic_datap
, size
);
2048 * Intentionally corrupt the log record CRC based on the error injection
2049 * frequency, if defined. This facilitates testing log recovery in the
2050 * event of torn writes. Hence, set the IOABORT state to abort the log
2051 * write on I/O completion and shutdown the fs. The subsequent mount
2052 * detects the bad CRC and attempts to recover.
2055 if (XFS_TEST_ERROR(false, log
->l_mp
, XFS_ERRTAG_LOG_BAD_CRC
)) {
2056 iclog
->ic_header
.h_crc
&= cpu_to_le32(0xAAAAAAAA);
2057 iclog
->ic_fail_crc
= true;
2059 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2060 be64_to_cpu(iclog
->ic_header
.h_lsn
));
2063 xlog_verify_iclog(log
, iclog
, count
);
2064 xlog_write_iclog(log
, iclog
, bno
, count
);
2068 * Deallocate a log structure
2074 xlog_in_core_t
*iclog
, *next_iclog
;
2077 xlog_cil_destroy(log
);
2080 * Cycle all the iclogbuf locks to make sure all log IO completion
2081 * is done before we tear down these buffers.
2083 iclog
= log
->l_iclog
;
2084 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
2085 down(&iclog
->ic_sema
);
2086 up(&iclog
->ic_sema
);
2087 iclog
= iclog
->ic_next
;
2090 iclog
= log
->l_iclog
;
2091 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
2092 next_iclog
= iclog
->ic_next
;
2093 kmem_free(iclog
->ic_data
);
2098 log
->l_mp
->m_log
= NULL
;
2099 destroy_workqueue(log
->l_ioend_workqueue
);
2104 * Update counters atomically now that memcpy is done.
2107 xlog_state_finish_copy(
2109 struct xlog_in_core
*iclog
,
2113 lockdep_assert_held(&log
->l_icloglock
);
2115 be32_add_cpu(&iclog
->ic_header
.h_num_logops
, record_cnt
);
2116 iclog
->ic_offset
+= copy_bytes
;
2120 * print out info relating to regions written which consume
2125 struct xfs_mount
*mp
,
2126 struct xlog_ticket
*ticket
)
2129 uint ophdr_spc
= ticket
->t_res_num_ophdrs
* (uint
)sizeof(xlog_op_header_t
);
2131 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2132 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2133 static char *res_type_str
[] = {
2134 REG_TYPE_STR(BFORMAT
, "bformat"),
2135 REG_TYPE_STR(BCHUNK
, "bchunk"),
2136 REG_TYPE_STR(EFI_FORMAT
, "efi_format"),
2137 REG_TYPE_STR(EFD_FORMAT
, "efd_format"),
2138 REG_TYPE_STR(IFORMAT
, "iformat"),
2139 REG_TYPE_STR(ICORE
, "icore"),
2140 REG_TYPE_STR(IEXT
, "iext"),
2141 REG_TYPE_STR(IBROOT
, "ibroot"),
2142 REG_TYPE_STR(ILOCAL
, "ilocal"),
2143 REG_TYPE_STR(IATTR_EXT
, "iattr_ext"),
2144 REG_TYPE_STR(IATTR_BROOT
, "iattr_broot"),
2145 REG_TYPE_STR(IATTR_LOCAL
, "iattr_local"),
2146 REG_TYPE_STR(QFORMAT
, "qformat"),
2147 REG_TYPE_STR(DQUOT
, "dquot"),
2148 REG_TYPE_STR(QUOTAOFF
, "quotaoff"),
2149 REG_TYPE_STR(LRHEADER
, "LR header"),
2150 REG_TYPE_STR(UNMOUNT
, "unmount"),
2151 REG_TYPE_STR(COMMIT
, "commit"),
2152 REG_TYPE_STR(TRANSHDR
, "trans header"),
2153 REG_TYPE_STR(ICREATE
, "inode create"),
2154 REG_TYPE_STR(RUI_FORMAT
, "rui_format"),
2155 REG_TYPE_STR(RUD_FORMAT
, "rud_format"),
2156 REG_TYPE_STR(CUI_FORMAT
, "cui_format"),
2157 REG_TYPE_STR(CUD_FORMAT
, "cud_format"),
2158 REG_TYPE_STR(BUI_FORMAT
, "bui_format"),
2159 REG_TYPE_STR(BUD_FORMAT
, "bud_format"),
2161 BUILD_BUG_ON(ARRAY_SIZE(res_type_str
) != XLOG_REG_TYPE_MAX
+ 1);
2164 xfs_warn(mp
, "ticket reservation summary:");
2165 xfs_warn(mp
, " unit res = %d bytes",
2166 ticket
->t_unit_res
);
2167 xfs_warn(mp
, " current res = %d bytes",
2168 ticket
->t_curr_res
);
2169 xfs_warn(mp
, " total reg = %u bytes (o/flow = %u bytes)",
2170 ticket
->t_res_arr_sum
, ticket
->t_res_o_flow
);
2171 xfs_warn(mp
, " ophdrs = %u (ophdr space = %u bytes)",
2172 ticket
->t_res_num_ophdrs
, ophdr_spc
);
2173 xfs_warn(mp
, " ophdr + reg = %u bytes",
2174 ticket
->t_res_arr_sum
+ ticket
->t_res_o_flow
+ ophdr_spc
);
2175 xfs_warn(mp
, " num regions = %u",
2178 for (i
= 0; i
< ticket
->t_res_num
; i
++) {
2179 uint r_type
= ticket
->t_res_arr
[i
].r_type
;
2180 xfs_warn(mp
, "region[%u]: %s - %u bytes", i
,
2181 ((r_type
<= 0 || r_type
> XLOG_REG_TYPE_MAX
) ?
2182 "bad-rtype" : res_type_str
[r_type
]),
2183 ticket
->t_res_arr
[i
].r_len
);
2188 * Print a summary of the transaction.
2192 struct xfs_trans
*tp
)
2194 struct xfs_mount
*mp
= tp
->t_mountp
;
2195 struct xfs_log_item
*lip
;
2197 /* dump core transaction and ticket info */
2198 xfs_warn(mp
, "transaction summary:");
2199 xfs_warn(mp
, " log res = %d", tp
->t_log_res
);
2200 xfs_warn(mp
, " log count = %d", tp
->t_log_count
);
2201 xfs_warn(mp
, " flags = 0x%x", tp
->t_flags
);
2203 xlog_print_tic_res(mp
, tp
->t_ticket
);
2205 /* dump each log item */
2206 list_for_each_entry(lip
, &tp
->t_items
, li_trans
) {
2207 struct xfs_log_vec
*lv
= lip
->li_lv
;
2208 struct xfs_log_iovec
*vec
;
2211 xfs_warn(mp
, "log item: ");
2212 xfs_warn(mp
, " type = 0x%x", lip
->li_type
);
2213 xfs_warn(mp
, " flags = 0x%lx", lip
->li_flags
);
2216 xfs_warn(mp
, " niovecs = %d", lv
->lv_niovecs
);
2217 xfs_warn(mp
, " size = %d", lv
->lv_size
);
2218 xfs_warn(mp
, " bytes = %d", lv
->lv_bytes
);
2219 xfs_warn(mp
, " buf len = %d", lv
->lv_buf_len
);
2221 /* dump each iovec for the log item */
2222 vec
= lv
->lv_iovecp
;
2223 for (i
= 0; i
< lv
->lv_niovecs
; i
++) {
2224 int dumplen
= min(vec
->i_len
, 32);
2226 xfs_warn(mp
, " iovec[%d]", i
);
2227 xfs_warn(mp
, " type = 0x%x", vec
->i_type
);
2228 xfs_warn(mp
, " len = %d", vec
->i_len
);
2229 xfs_warn(mp
, " first %d bytes of iovec[%d]:", dumplen
, i
);
2230 xfs_hex_dump(vec
->i_addr
, dumplen
);
2238 * Calculate the potential space needed by the log vector. We may need a start
2239 * record, and each region gets its own struct xlog_op_header and may need to be
2240 * double word aligned.
2243 xlog_write_calc_vec_length(
2244 struct xlog_ticket
*ticket
,
2245 struct xfs_log_vec
*log_vector
,
2248 struct xfs_log_vec
*lv
;
2253 if (optype
& XLOG_START_TRANS
)
2256 for (lv
= log_vector
; lv
; lv
= lv
->lv_next
) {
2257 /* we don't write ordered log vectors */
2258 if (lv
->lv_buf_len
== XFS_LOG_VEC_ORDERED
)
2261 headers
+= lv
->lv_niovecs
;
2263 for (i
= 0; i
< lv
->lv_niovecs
; i
++) {
2264 struct xfs_log_iovec
*vecp
= &lv
->lv_iovecp
[i
];
2267 xlog_tic_add_region(ticket
, vecp
->i_len
, vecp
->i_type
);
2271 ticket
->t_res_num_ophdrs
+= headers
;
2272 len
+= headers
* sizeof(struct xlog_op_header
);
2278 xlog_write_start_rec(
2279 struct xlog_op_header
*ophdr
,
2280 struct xlog_ticket
*ticket
)
2282 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
2283 ophdr
->oh_clientid
= ticket
->t_clientid
;
2285 ophdr
->oh_flags
= XLOG_START_TRANS
;
2289 static xlog_op_header_t
*
2290 xlog_write_setup_ophdr(
2292 struct xlog_op_header
*ophdr
,
2293 struct xlog_ticket
*ticket
,
2296 ophdr
->oh_tid
= cpu_to_be32(ticket
->t_tid
);
2297 ophdr
->oh_clientid
= ticket
->t_clientid
;
2300 /* are we copying a commit or unmount record? */
2301 ophdr
->oh_flags
= flags
;
2304 * We've seen logs corrupted with bad transaction client ids. This
2305 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2306 * and shut down the filesystem.
2308 switch (ophdr
->oh_clientid
) {
2309 case XFS_TRANSACTION
:
2315 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT
,
2316 ophdr
->oh_clientid
, ticket
);
2324 * Set up the parameters of the region copy into the log. This has
2325 * to handle region write split across multiple log buffers - this
2326 * state is kept external to this function so that this code can
2327 * be written in an obvious, self documenting manner.
2330 xlog_write_setup_copy(
2331 struct xlog_ticket
*ticket
,
2332 struct xlog_op_header
*ophdr
,
2333 int space_available
,
2337 int *last_was_partial_copy
,
2338 int *bytes_consumed
)
2342 still_to_copy
= space_required
- *bytes_consumed
;
2343 *copy_off
= *bytes_consumed
;
2345 if (still_to_copy
<= space_available
) {
2346 /* write of region completes here */
2347 *copy_len
= still_to_copy
;
2348 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
2349 if (*last_was_partial_copy
)
2350 ophdr
->oh_flags
|= (XLOG_END_TRANS
|XLOG_WAS_CONT_TRANS
);
2351 *last_was_partial_copy
= 0;
2352 *bytes_consumed
= 0;
2356 /* partial write of region, needs extra log op header reservation */
2357 *copy_len
= space_available
;
2358 ophdr
->oh_len
= cpu_to_be32(*copy_len
);
2359 ophdr
->oh_flags
|= XLOG_CONTINUE_TRANS
;
2360 if (*last_was_partial_copy
)
2361 ophdr
->oh_flags
|= XLOG_WAS_CONT_TRANS
;
2362 *bytes_consumed
+= *copy_len
;
2363 (*last_was_partial_copy
)++;
2365 /* account for new log op header */
2366 ticket
->t_curr_res
-= sizeof(struct xlog_op_header
);
2367 ticket
->t_res_num_ophdrs
++;
2369 return sizeof(struct xlog_op_header
);
2373 xlog_write_copy_finish(
2375 struct xlog_in_core
*iclog
,
2380 int *partial_copy_len
,
2385 if (*partial_copy
) {
2387 * This iclog has already been marked WANT_SYNC by
2388 * xlog_state_get_iclog_space.
2390 spin_lock(&log
->l_icloglock
);
2391 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
2398 *partial_copy_len
= 0;
2400 if (iclog
->ic_size
- log_offset
> sizeof(xlog_op_header_t
))
2403 /* no more space in this iclog - push it. */
2404 spin_lock(&log
->l_icloglock
);
2405 xlog_state_finish_copy(log
, iclog
, *record_cnt
, *data_cnt
);
2409 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
)
2410 xlog_state_switch_iclogs(log
, iclog
, 0);
2412 ASSERT(iclog
->ic_state
== XLOG_STATE_WANT_SYNC
||
2413 xlog_is_shutdown(log
));
2415 error
= xlog_state_release_iclog(log
, iclog
, 0);
2416 spin_unlock(&log
->l_icloglock
);
2421 * Write some region out to in-core log
2423 * This will be called when writing externally provided regions or when
2424 * writing out a commit record for a given transaction.
2426 * General algorithm:
2427 * 1. Find total length of this write. This may include adding to the
2428 * lengths passed in.
2429 * 2. Check whether we violate the tickets reservation.
2430 * 3. While writing to this iclog
2431 * A. Reserve as much space in this iclog as can get
2432 * B. If this is first write, save away start lsn
2433 * C. While writing this region:
2434 * 1. If first write of transaction, write start record
2435 * 2. Write log operation header (header per region)
2436 * 3. Find out if we can fit entire region into this iclog
2437 * 4. Potentially, verify destination memcpy ptr
2438 * 5. Memcpy (partial) region
2439 * 6. If partial copy, release iclog; otherwise, continue
2440 * copying more regions into current iclog
2441 * 4. Mark want sync bit (in simulation mode)
2442 * 5. Release iclog for potential flush to on-disk log.
2445 * 1. Panic if reservation is overrun. This should never happen since
2446 * reservation amounts are generated internal to the filesystem.
2448 * 1. Tickets are single threaded data structures.
2449 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2450 * syncing routine. When a single log_write region needs to span
2451 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2452 * on all log operation writes which don't contain the end of the
2453 * region. The XLOG_END_TRANS bit is used for the in-core log
2454 * operation which contains the end of the continued log_write region.
2455 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2456 * we don't really know exactly how much space will be used. As a result,
2457 * we don't update ic_offset until the end when we know exactly how many
2458 * bytes have been written out.
2463 struct xfs_cil_ctx
*ctx
,
2464 struct xfs_log_vec
*log_vector
,
2465 struct xlog_ticket
*ticket
,
2468 struct xlog_in_core
*iclog
= NULL
;
2469 struct xfs_log_vec
*lv
= log_vector
;
2470 struct xfs_log_iovec
*vecp
= lv
->lv_iovecp
;
2473 int partial_copy
= 0;
2474 int partial_copy_len
= 0;
2481 * If this is a commit or unmount transaction, we don't need a start
2482 * record to be written. We do, however, have to account for the
2483 * commit or unmount header that gets written. Hence we always have
2484 * to account for an extra xlog_op_header here.
2486 ticket
->t_curr_res
-= sizeof(struct xlog_op_header
);
2487 if (ticket
->t_curr_res
< 0) {
2488 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
2489 "ctx ticket reservation ran out. Need to up reservation");
2490 xlog_print_tic_res(log
->l_mp
, ticket
);
2491 xfs_force_shutdown(log
->l_mp
, SHUTDOWN_LOG_IO_ERROR
);
2494 len
= xlog_write_calc_vec_length(ticket
, log_vector
, optype
);
2495 while (lv
&& (!lv
->lv_niovecs
|| index
< lv
->lv_niovecs
)) {
2499 error
= xlog_state_get_iclog_space(log
, len
, &iclog
, ticket
,
2500 &contwr
, &log_offset
);
2504 ASSERT(log_offset
<= iclog
->ic_size
- 1);
2505 ptr
= iclog
->ic_datap
+ log_offset
;
2508 * If we have a context pointer, pass it the first iclog we are
2509 * writing to so it can record state needed for iclog write
2513 xlog_cil_set_ctx_write_state(ctx
, iclog
);
2518 * This loop writes out as many regions as can fit in the amount
2519 * of space which was allocated by xlog_state_get_iclog_space().
2521 while (lv
&& (!lv
->lv_niovecs
|| index
< lv
->lv_niovecs
)) {
2522 struct xfs_log_iovec
*reg
;
2523 struct xlog_op_header
*ophdr
;
2526 bool ordered
= false;
2527 bool wrote_start_rec
= false;
2529 /* ordered log vectors have no regions to write */
2530 if (lv
->lv_buf_len
== XFS_LOG_VEC_ORDERED
) {
2531 ASSERT(lv
->lv_niovecs
== 0);
2537 ASSERT(reg
->i_len
% sizeof(int32_t) == 0);
2538 ASSERT((unsigned long)ptr
% sizeof(int32_t) == 0);
2541 * Before we start formatting log vectors, we need to
2542 * write a start record. Only do this for the first
2543 * iclog we write to.
2545 if (optype
& XLOG_START_TRANS
) {
2546 xlog_write_start_rec(ptr
, ticket
);
2547 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2548 sizeof(struct xlog_op_header
));
2549 optype
&= ~XLOG_START_TRANS
;
2550 wrote_start_rec
= true;
2553 ophdr
= xlog_write_setup_ophdr(log
, ptr
, ticket
, optype
);
2557 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2558 sizeof(struct xlog_op_header
));
2560 len
+= xlog_write_setup_copy(ticket
, ophdr
,
2561 iclog
->ic_size
-log_offset
,
2563 ©_off
, ©_len
,
2566 xlog_verify_dest_ptr(log
, ptr
);
2571 * Unmount records just log an opheader, so can have
2572 * empty payloads with no data region to copy. Hence we
2573 * only copy the payload if the vector says it has data
2576 ASSERT(copy_len
>= 0);
2578 memcpy(ptr
, reg
->i_addr
+ copy_off
, copy_len
);
2579 xlog_write_adv_cnt(&ptr
, &len
, &log_offset
,
2582 copy_len
+= sizeof(struct xlog_op_header
);
2584 if (wrote_start_rec
) {
2585 copy_len
+= sizeof(struct xlog_op_header
);
2588 data_cnt
+= contwr
? copy_len
: 0;
2590 error
= xlog_write_copy_finish(log
, iclog
, optype
,
2591 &record_cnt
, &data_cnt
,
2599 * if we had a partial copy, we need to get more iclog
2600 * space but we don't want to increment the region
2601 * index because there is still more is this region to
2604 * If we completed writing this region, and we flushed
2605 * the iclog (indicated by resetting of the record
2606 * count), then we also need to get more log space. If
2607 * this was the last record, though, we are done and
2613 if (++index
== lv
->lv_niovecs
) {
2618 vecp
= lv
->lv_iovecp
;
2620 if (record_cnt
== 0 && !ordered
) {
2630 spin_lock(&log
->l_icloglock
);
2631 xlog_state_finish_copy(log
, iclog
, record_cnt
, data_cnt
);
2632 error
= xlog_state_release_iclog(log
, iclog
, 0);
2633 spin_unlock(&log
->l_icloglock
);
2639 xlog_state_activate_iclog(
2640 struct xlog_in_core
*iclog
,
2641 int *iclogs_changed
)
2643 ASSERT(list_empty_careful(&iclog
->ic_callbacks
));
2644 trace_xlog_iclog_activate(iclog
, _RET_IP_
);
2647 * If the number of ops in this iclog indicate it just contains the
2648 * dummy transaction, we can change state into IDLE (the second time
2649 * around). Otherwise we should change the state into NEED a dummy.
2650 * We don't need to cover the dummy.
2652 if (*iclogs_changed
== 0 &&
2653 iclog
->ic_header
.h_num_logops
== cpu_to_be32(XLOG_COVER_OPS
)) {
2654 *iclogs_changed
= 1;
2657 * We have two dirty iclogs so start over. This could also be
2658 * num of ops indicating this is not the dummy going out.
2660 *iclogs_changed
= 2;
2663 iclog
->ic_state
= XLOG_STATE_ACTIVE
;
2664 iclog
->ic_offset
= 0;
2665 iclog
->ic_header
.h_num_logops
= 0;
2666 memset(iclog
->ic_header
.h_cycle_data
, 0,
2667 sizeof(iclog
->ic_header
.h_cycle_data
));
2668 iclog
->ic_header
.h_lsn
= 0;
2669 iclog
->ic_header
.h_tail_lsn
= 0;
2673 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2674 * ACTIVE after iclog I/O has completed.
2677 xlog_state_activate_iclogs(
2679 int *iclogs_changed
)
2681 struct xlog_in_core
*iclog
= log
->l_iclog
;
2684 if (iclog
->ic_state
== XLOG_STATE_DIRTY
)
2685 xlog_state_activate_iclog(iclog
, iclogs_changed
);
2687 * The ordering of marking iclogs ACTIVE must be maintained, so
2688 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2690 else if (iclog
->ic_state
!= XLOG_STATE_ACTIVE
)
2692 } while ((iclog
= iclog
->ic_next
) != log
->l_iclog
);
2701 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2702 * wrote the first covering record (DONE). We go to IDLE if we just
2703 * wrote the second covering record (DONE2) and remain in IDLE until a
2704 * non-covering write occurs.
2706 switch (prev_state
) {
2707 case XLOG_STATE_COVER_IDLE
:
2708 if (iclogs_changed
== 1)
2709 return XLOG_STATE_COVER_IDLE
;
2711 case XLOG_STATE_COVER_NEED
:
2712 case XLOG_STATE_COVER_NEED2
:
2714 case XLOG_STATE_COVER_DONE
:
2715 if (iclogs_changed
== 1)
2716 return XLOG_STATE_COVER_NEED2
;
2718 case XLOG_STATE_COVER_DONE2
:
2719 if (iclogs_changed
== 1)
2720 return XLOG_STATE_COVER_IDLE
;
2726 return XLOG_STATE_COVER_NEED
;
2730 xlog_state_clean_iclog(
2732 struct xlog_in_core
*dirty_iclog
)
2734 int iclogs_changed
= 0;
2736 trace_xlog_iclog_clean(dirty_iclog
, _RET_IP_
);
2738 dirty_iclog
->ic_state
= XLOG_STATE_DIRTY
;
2740 xlog_state_activate_iclogs(log
, &iclogs_changed
);
2741 wake_up_all(&dirty_iclog
->ic_force_wait
);
2743 if (iclogs_changed
) {
2744 log
->l_covered_state
= xlog_covered_state(log
->l_covered_state
,
2750 xlog_get_lowest_lsn(
2753 struct xlog_in_core
*iclog
= log
->l_iclog
;
2754 xfs_lsn_t lowest_lsn
= 0, lsn
;
2757 if (iclog
->ic_state
== XLOG_STATE_ACTIVE
||
2758 iclog
->ic_state
== XLOG_STATE_DIRTY
)
2761 lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
2762 if ((lsn
&& !lowest_lsn
) || XFS_LSN_CMP(lsn
, lowest_lsn
) < 0)
2764 } while ((iclog
= iclog
->ic_next
) != log
->l_iclog
);
2770 * Completion of a iclog IO does not imply that a transaction has completed, as
2771 * transactions can be large enough to span many iclogs. We cannot change the
2772 * tail of the log half way through a transaction as this may be the only
2773 * transaction in the log and moving the tail to point to the middle of it
2774 * will prevent recovery from finding the start of the transaction. Hence we
2775 * should only update the last_sync_lsn if this iclog contains transaction
2776 * completion callbacks on it.
2778 * We have to do this before we drop the icloglock to ensure we are the only one
2779 * that can update it.
2781 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2782 * the reservation grant head pushing. This is due to the fact that the push
2783 * target is bound by the current last_sync_lsn value. Hence if we have a large
2784 * amount of log space bound up in this committing transaction then the
2785 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2786 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2787 * should push the AIL to ensure the push target (and hence the grant head) is
2788 * no longer bound by the old log head location and can move forwards and make
2792 xlog_state_set_callback(
2794 struct xlog_in_core
*iclog
,
2795 xfs_lsn_t header_lsn
)
2797 trace_xlog_iclog_callback(iclog
, _RET_IP_
);
2798 iclog
->ic_state
= XLOG_STATE_CALLBACK
;
2800 ASSERT(XFS_LSN_CMP(atomic64_read(&log
->l_last_sync_lsn
),
2803 if (list_empty_careful(&iclog
->ic_callbacks
))
2806 atomic64_set(&log
->l_last_sync_lsn
, header_lsn
);
2807 xlog_grant_push_ail(log
, 0);
2811 * Return true if we need to stop processing, false to continue to the next
2812 * iclog. The caller will need to run callbacks if the iclog is returned in the
2813 * XLOG_STATE_CALLBACK state.
2816 xlog_state_iodone_process_iclog(
2818 struct xlog_in_core
*iclog
)
2820 xfs_lsn_t lowest_lsn
;
2821 xfs_lsn_t header_lsn
;
2823 switch (iclog
->ic_state
) {
2824 case XLOG_STATE_ACTIVE
:
2825 case XLOG_STATE_DIRTY
:
2827 * Skip all iclogs in the ACTIVE & DIRTY states:
2830 case XLOG_STATE_DONE_SYNC
:
2832 * Now that we have an iclog that is in the DONE_SYNC state, do
2833 * one more check here to see if we have chased our tail around.
2834 * If this is not the lowest lsn iclog, then we will leave it
2835 * for another completion to process.
2837 header_lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
2838 lowest_lsn
= xlog_get_lowest_lsn(log
);
2839 if (lowest_lsn
&& XFS_LSN_CMP(lowest_lsn
, header_lsn
) < 0)
2841 xlog_state_set_callback(log
, iclog
, header_lsn
);
2845 * Can only perform callbacks in order. Since this iclog is not
2846 * in the DONE_SYNC state, we skip the rest and just try to
2854 * Loop over all the iclogs, running attached callbacks on them. Return true if
2855 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2856 * to handle transient shutdown state here at all because
2857 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2858 * cleanup of the callbacks.
2861 xlog_state_do_iclog_callbacks(
2863 __releases(&log
->l_icloglock
)
2864 __acquires(&log
->l_icloglock
)
2866 struct xlog_in_core
*first_iclog
= log
->l_iclog
;
2867 struct xlog_in_core
*iclog
= first_iclog
;
2868 bool ran_callback
= false;
2873 if (xlog_state_iodone_process_iclog(log
, iclog
))
2875 if (iclog
->ic_state
!= XLOG_STATE_CALLBACK
) {
2876 iclog
= iclog
->ic_next
;
2879 list_splice_init(&iclog
->ic_callbacks
, &cb_list
);
2880 spin_unlock(&log
->l_icloglock
);
2882 trace_xlog_iclog_callbacks_start(iclog
, _RET_IP_
);
2883 xlog_cil_process_committed(&cb_list
);
2884 trace_xlog_iclog_callbacks_done(iclog
, _RET_IP_
);
2885 ran_callback
= true;
2887 spin_lock(&log
->l_icloglock
);
2888 xlog_state_clean_iclog(log
, iclog
);
2889 iclog
= iclog
->ic_next
;
2890 } while (iclog
!= first_iclog
);
2892 return ran_callback
;
2897 * Loop running iclog completion callbacks until there are no more iclogs in a
2898 * state that can run callbacks.
2901 xlog_state_do_callback(
2907 spin_lock(&log
->l_icloglock
);
2908 while (xlog_state_do_iclog_callbacks(log
)) {
2909 if (xlog_is_shutdown(log
))
2912 if (++repeats
> 5000) {
2913 flushcnt
+= repeats
;
2916 "%s: possible infinite loop (%d iterations)",
2917 __func__
, flushcnt
);
2921 if (log
->l_iclog
->ic_state
== XLOG_STATE_ACTIVE
)
2922 wake_up_all(&log
->l_flush_wait
);
2924 spin_unlock(&log
->l_icloglock
);
2929 * Finish transitioning this iclog to the dirty state.
2931 * Callbacks could take time, so they are done outside the scope of the
2932 * global state machine log lock.
2935 xlog_state_done_syncing(
2936 struct xlog_in_core
*iclog
)
2938 struct xlog
*log
= iclog
->ic_log
;
2940 spin_lock(&log
->l_icloglock
);
2941 ASSERT(atomic_read(&iclog
->ic_refcnt
) == 0);
2942 trace_xlog_iclog_sync_done(iclog
, _RET_IP_
);
2945 * If we got an error, either on the first buffer, or in the case of
2946 * split log writes, on the second, we shut down the file system and
2947 * no iclogs should ever be attempted to be written to disk again.
2949 if (!xlog_is_shutdown(log
)) {
2950 ASSERT(iclog
->ic_state
== XLOG_STATE_SYNCING
);
2951 iclog
->ic_state
= XLOG_STATE_DONE_SYNC
;
2955 * Someone could be sleeping prior to writing out the next
2956 * iclog buffer, we wake them all, one will get to do the
2957 * I/O, the others get to wait for the result.
2959 wake_up_all(&iclog
->ic_write_wait
);
2960 spin_unlock(&log
->l_icloglock
);
2961 xlog_state_do_callback(log
);
2965 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2966 * sleep. We wait on the flush queue on the head iclog as that should be
2967 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2968 * we will wait here and all new writes will sleep until a sync completes.
2970 * The in-core logs are used in a circular fashion. They are not used
2971 * out-of-order even when an iclog past the head is free.
2974 * * log_offset where xlog_write() can start writing into the in-core
2976 * * in-core log pointer to which xlog_write() should write.
2977 * * boolean indicating this is a continued write to an in-core log.
2978 * If this is the last write, then the in-core log's offset field
2979 * needs to be incremented, depending on the amount of data which
2983 xlog_state_get_iclog_space(
2986 struct xlog_in_core
**iclogp
,
2987 struct xlog_ticket
*ticket
,
2988 int *continued_write
,
2992 xlog_rec_header_t
*head
;
2993 xlog_in_core_t
*iclog
;
2996 spin_lock(&log
->l_icloglock
);
2997 if (xlog_is_shutdown(log
)) {
2998 spin_unlock(&log
->l_icloglock
);
3002 iclog
= log
->l_iclog
;
3003 if (iclog
->ic_state
!= XLOG_STATE_ACTIVE
) {
3004 XFS_STATS_INC(log
->l_mp
, xs_log_noiclogs
);
3006 /* Wait for log writes to have flushed */
3007 xlog_wait(&log
->l_flush_wait
, &log
->l_icloglock
);
3011 head
= &iclog
->ic_header
;
3013 atomic_inc(&iclog
->ic_refcnt
); /* prevents sync */
3014 log_offset
= iclog
->ic_offset
;
3016 trace_xlog_iclog_get_space(iclog
, _RET_IP_
);
3018 /* On the 1st write to an iclog, figure out lsn. This works
3019 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3020 * committing to. If the offset is set, that's how many blocks
3023 if (log_offset
== 0) {
3024 ticket
->t_curr_res
-= log
->l_iclog_hsize
;
3025 xlog_tic_add_region(ticket
,
3027 XLOG_REG_TYPE_LRHEADER
);
3028 head
->h_cycle
= cpu_to_be32(log
->l_curr_cycle
);
3029 head
->h_lsn
= cpu_to_be64(
3030 xlog_assign_lsn(log
->l_curr_cycle
, log
->l_curr_block
));
3031 ASSERT(log
->l_curr_block
>= 0);
3034 /* If there is enough room to write everything, then do it. Otherwise,
3035 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3036 * bit is on, so this will get flushed out. Don't update ic_offset
3037 * until you know exactly how many bytes get copied. Therefore, wait
3038 * until later to update ic_offset.
3040 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3041 * can fit into remaining data section.
3043 if (iclog
->ic_size
- iclog
->ic_offset
< 2*sizeof(xlog_op_header_t
)) {
3046 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
3049 * If we are the only one writing to this iclog, sync it to
3050 * disk. We need to do an atomic compare and decrement here to
3051 * avoid racing with concurrent atomic_dec_and_lock() calls in
3052 * xlog_state_release_iclog() when there is more than one
3053 * reference to the iclog.
3055 if (!atomic_add_unless(&iclog
->ic_refcnt
, -1, 1))
3056 error
= xlog_state_release_iclog(log
, iclog
, 0);
3057 spin_unlock(&log
->l_icloglock
);
3063 /* Do we have enough room to write the full amount in the remainder
3064 * of this iclog? Or must we continue a write on the next iclog and
3065 * mark this iclog as completely taken? In the case where we switch
3066 * iclogs (to mark it taken), this particular iclog will release/sync
3067 * to disk in xlog_write().
3069 if (len
<= iclog
->ic_size
- iclog
->ic_offset
) {
3070 *continued_write
= 0;
3071 iclog
->ic_offset
+= len
;
3073 *continued_write
= 1;
3074 xlog_state_switch_iclogs(log
, iclog
, iclog
->ic_size
);
3078 ASSERT(iclog
->ic_offset
<= iclog
->ic_size
);
3079 spin_unlock(&log
->l_icloglock
);
3081 *logoffsetp
= log_offset
;
3086 * The first cnt-1 times a ticket goes through here we don't need to move the
3087 * grant write head because the permanent reservation has reserved cnt times the
3088 * unit amount. Release part of current permanent unit reservation and reset
3089 * current reservation to be one units worth. Also move grant reservation head
3093 xfs_log_ticket_regrant(
3095 struct xlog_ticket
*ticket
)
3097 trace_xfs_log_ticket_regrant(log
, ticket
);
3099 if (ticket
->t_cnt
> 0)
3102 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
,
3103 ticket
->t_curr_res
);
3104 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
,
3105 ticket
->t_curr_res
);
3106 ticket
->t_curr_res
= ticket
->t_unit_res
;
3107 xlog_tic_reset_res(ticket
);
3109 trace_xfs_log_ticket_regrant_sub(log
, ticket
);
3111 /* just return if we still have some of the pre-reserved space */
3112 if (!ticket
->t_cnt
) {
3113 xlog_grant_add_space(log
, &log
->l_reserve_head
.grant
,
3114 ticket
->t_unit_res
);
3115 trace_xfs_log_ticket_regrant_exit(log
, ticket
);
3117 ticket
->t_curr_res
= ticket
->t_unit_res
;
3118 xlog_tic_reset_res(ticket
);
3121 xfs_log_ticket_put(ticket
);
3125 * Give back the space left from a reservation.
3127 * All the information we need to make a correct determination of space left
3128 * is present. For non-permanent reservations, things are quite easy. The
3129 * count should have been decremented to zero. We only need to deal with the
3130 * space remaining in the current reservation part of the ticket. If the
3131 * ticket contains a permanent reservation, there may be left over space which
3132 * needs to be released. A count of N means that N-1 refills of the current
3133 * reservation can be done before we need to ask for more space. The first
3134 * one goes to fill up the first current reservation. Once we run out of
3135 * space, the count will stay at zero and the only space remaining will be
3136 * in the current reservation field.
3139 xfs_log_ticket_ungrant(
3141 struct xlog_ticket
*ticket
)
3145 trace_xfs_log_ticket_ungrant(log
, ticket
);
3147 if (ticket
->t_cnt
> 0)
3150 trace_xfs_log_ticket_ungrant_sub(log
, ticket
);
3153 * If this is a permanent reservation ticket, we may be able to free
3154 * up more space based on the remaining count.
3156 bytes
= ticket
->t_curr_res
;
3157 if (ticket
->t_cnt
> 0) {
3158 ASSERT(ticket
->t_flags
& XLOG_TIC_PERM_RESERV
);
3159 bytes
+= ticket
->t_unit_res
*ticket
->t_cnt
;
3162 xlog_grant_sub_space(log
, &log
->l_reserve_head
.grant
, bytes
);
3163 xlog_grant_sub_space(log
, &log
->l_write_head
.grant
, bytes
);
3165 trace_xfs_log_ticket_ungrant_exit(log
, ticket
);
3167 xfs_log_space_wake(log
->l_mp
);
3168 xfs_log_ticket_put(ticket
);
3172 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3173 * the current iclog pointer to the next iclog in the ring.
3176 xlog_state_switch_iclogs(
3178 struct xlog_in_core
*iclog
,
3181 ASSERT(iclog
->ic_state
== XLOG_STATE_ACTIVE
);
3182 assert_spin_locked(&log
->l_icloglock
);
3183 trace_xlog_iclog_switch(iclog
, _RET_IP_
);
3186 eventual_size
= iclog
->ic_offset
;
3187 iclog
->ic_state
= XLOG_STATE_WANT_SYNC
;
3188 iclog
->ic_header
.h_prev_block
= cpu_to_be32(log
->l_prev_block
);
3189 log
->l_prev_block
= log
->l_curr_block
;
3190 log
->l_prev_cycle
= log
->l_curr_cycle
;
3192 /* roll log?: ic_offset changed later */
3193 log
->l_curr_block
+= BTOBB(eventual_size
)+BTOBB(log
->l_iclog_hsize
);
3195 /* Round up to next log-sunit */
3196 if (log
->l_iclog_roundoff
> BBSIZE
) {
3197 uint32_t sunit_bb
= BTOBB(log
->l_iclog_roundoff
);
3198 log
->l_curr_block
= roundup(log
->l_curr_block
, sunit_bb
);
3201 if (log
->l_curr_block
>= log
->l_logBBsize
) {
3203 * Rewind the current block before the cycle is bumped to make
3204 * sure that the combined LSN never transiently moves forward
3205 * when the log wraps to the next cycle. This is to support the
3206 * unlocked sample of these fields from xlog_valid_lsn(). Most
3207 * other cases should acquire l_icloglock.
3209 log
->l_curr_block
-= log
->l_logBBsize
;
3210 ASSERT(log
->l_curr_block
>= 0);
3212 log
->l_curr_cycle
++;
3213 if (log
->l_curr_cycle
== XLOG_HEADER_MAGIC_NUM
)
3214 log
->l_curr_cycle
++;
3216 ASSERT(iclog
== log
->l_iclog
);
3217 log
->l_iclog
= iclog
->ic_next
;
3221 * Force the iclog to disk and check if the iclog has been completed before
3222 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3223 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3224 * If completion has already occurred, tell the caller so that it can avoid an
3225 * unnecessary wait on the iclog.
3228 xlog_force_and_check_iclog(
3229 struct xlog_in_core
*iclog
,
3232 xfs_lsn_t lsn
= be64_to_cpu(iclog
->ic_header
.h_lsn
);
3236 error
= xlog_force_iclog(iclog
);
3241 * If the iclog has already been completed and reused the header LSN
3242 * will have been rewritten by completion
3244 if (be64_to_cpu(iclog
->ic_header
.h_lsn
) != lsn
)
3250 * Write out all data in the in-core log as of this exact moment in time.
3252 * Data may be written to the in-core log during this call. However,
3253 * we don't guarantee this data will be written out. A change from past
3254 * implementation means this routine will *not* write out zero length LRs.
3256 * Basically, we try and perform an intelligent scan of the in-core logs.
3257 * If we determine there is no flushable data, we just return. There is no
3258 * flushable data if:
3260 * 1. the current iclog is active and has no data; the previous iclog
3261 * is in the active or dirty state.
3262 * 2. the current iclog is drity, and the previous iclog is in the
3263 * active or dirty state.
3267 * 1. the current iclog is not in the active nor dirty state.
3268 * 2. the current iclog dirty, and the previous iclog is not in the
3269 * active nor dirty state.
3270 * 3. the current iclog is active, and there is another thread writing
3271 * to this particular iclog.
3272 * 4. a) the current iclog is active and has no other writers
3273 * b) when we return from flushing out this iclog, it is still
3274 * not in the active nor dirty state.
3278 struct xfs_mount
*mp
,
3281 struct xlog
*log
= mp
->m_log
;
3282 struct xlog_in_core
*iclog
;
3284 XFS_STATS_INC(mp
, xs_log_force
);
3285 trace_xfs_log_force(mp
, 0, _RET_IP_
);
3287 xlog_cil_force(log
);
3289 spin_lock(&log
->l_icloglock
);
3290 if (xlog_is_shutdown(log
))
3293 iclog
= log
->l_iclog
;
3294 trace_xlog_iclog_force(iclog
, _RET_IP_
);
3296 if (iclog
->ic_state
== XLOG_STATE_DIRTY
||
3297 (iclog
->ic_state
== XLOG_STATE_ACTIVE
&&
3298 atomic_read(&iclog
->ic_refcnt
) == 0 && iclog
->ic_offset
== 0)) {
3300 * If the head is dirty or (active and empty), then we need to
3301 * look at the previous iclog.
3303 * If the previous iclog is active or dirty we are done. There
3304 * is nothing to sync out. Otherwise, we attach ourselves to the
3305 * previous iclog and go to sleep.
3307 iclog
= iclog
->ic_prev
;
3308 } else if (iclog
->ic_state
== XLOG_STATE_ACTIVE
) {
3309 if (atomic_read(&iclog
->ic_refcnt
) == 0) {
3310 /* We have exclusive access to this iclog. */
3313 if (xlog_force_and_check_iclog(iclog
, &completed
))
3320 * Someone else is still writing to this iclog, so we
3321 * need to ensure that when they release the iclog it
3322 * gets synced immediately as we may be waiting on it.
3324 xlog_state_switch_iclogs(log
, iclog
, 0);
3329 * The iclog we are about to wait on may contain the checkpoint pushed
3330 * by the above xlog_cil_force() call, but it may not have been pushed
3331 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3332 * are flushed when this iclog is written.
3334 if (iclog
->ic_state
== XLOG_STATE_WANT_SYNC
)
3335 iclog
->ic_flags
|= XLOG_ICL_NEED_FLUSH
| XLOG_ICL_NEED_FUA
;
3337 if (flags
& XFS_LOG_SYNC
)
3338 return xlog_wait_on_iclog(iclog
);
3340 spin_unlock(&log
->l_icloglock
);
3343 spin_unlock(&log
->l_icloglock
);
3348 * Force the log to a specific LSN.
3350 * If an iclog with that lsn can be found:
3351 * If it is in the DIRTY state, just return.
3352 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3353 * state and go to sleep or return.
3354 * If it is in any other state, go to sleep or return.
3356 * Synchronous forces are implemented with a wait queue. All callers trying
3357 * to force a given lsn to disk must wait on the queue attached to the
3358 * specific in-core log. When given in-core log finally completes its write
3359 * to disk, that thread will wake up all threads waiting on the queue.
3369 struct xlog_in_core
*iclog
;
3372 spin_lock(&log
->l_icloglock
);
3373 if (xlog_is_shutdown(log
))
3376 iclog
= log
->l_iclog
;
3377 while (be64_to_cpu(iclog
->ic_header
.h_lsn
) != lsn
) {
3378 trace_xlog_iclog_force_lsn(iclog
, _RET_IP_
);
3379 iclog
= iclog
->ic_next
;
3380 if (iclog
== log
->l_iclog
)
3384 switch (iclog
->ic_state
) {
3385 case XLOG_STATE_ACTIVE
:
3387 * We sleep here if we haven't already slept (e.g. this is the
3388 * first time we've looked at the correct iclog buf) and the
3389 * buffer before us is going to be sync'ed. The reason for this
3390 * is that if we are doing sync transactions here, by waiting
3391 * for the previous I/O to complete, we can allow a few more
3392 * transactions into this iclog before we close it down.
3394 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3395 * refcnt so we can release the log (which drops the ref count).
3396 * The state switch keeps new transaction commits from using
3397 * this buffer. When the current commits finish writing into
3398 * the buffer, the refcount will drop to zero and the buffer
3401 if (!already_slept
&&
3402 (iclog
->ic_prev
->ic_state
== XLOG_STATE_WANT_SYNC
||
3403 iclog
->ic_prev
->ic_state
== XLOG_STATE_SYNCING
)) {
3404 xlog_wait(&iclog
->ic_prev
->ic_write_wait
,
3408 if (xlog_force_and_check_iclog(iclog
, &completed
))
3415 case XLOG_STATE_WANT_SYNC
:
3417 * This iclog may contain the checkpoint pushed by the
3418 * xlog_cil_force_seq() call, but there are other writers still
3419 * accessing it so it hasn't been pushed to disk yet. Like the
3420 * ACTIVE case above, we need to make sure caches are flushed
3421 * when this iclog is written.
3423 iclog
->ic_flags
|= XLOG_ICL_NEED_FLUSH
| XLOG_ICL_NEED_FUA
;
3427 * The entire checkpoint was written by the CIL force and is on
3428 * its way to disk already. It will be stable when it
3429 * completes, so we don't need to manipulate caches here at all.
3430 * We just need to wait for completion if necessary.
3435 if (flags
& XFS_LOG_SYNC
)
3436 return xlog_wait_on_iclog(iclog
);
3438 spin_unlock(&log
->l_icloglock
);
3441 spin_unlock(&log
->l_icloglock
);
3446 * Force the log to a specific checkpoint sequence.
3448 * First force the CIL so that all the required changes have been flushed to the
3449 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3450 * the iclog that needs to be flushed to stable storage. If the caller needs
3451 * a synchronous log force, we will wait on the iclog with the LSN returned by
3452 * xlog_cil_force_seq() to be completed.
3456 struct xfs_mount
*mp
,
3461 struct xlog
*log
= mp
->m_log
;
3466 XFS_STATS_INC(mp
, xs_log_force
);
3467 trace_xfs_log_force(mp
, seq
, _RET_IP_
);
3469 lsn
= xlog_cil_force_seq(log
, seq
);
3470 if (lsn
== NULLCOMMITLSN
)
3473 ret
= xlog_force_lsn(log
, lsn
, flags
, log_flushed
, false);
3474 if (ret
== -EAGAIN
) {
3475 XFS_STATS_INC(mp
, xs_log_force_sleep
);
3476 ret
= xlog_force_lsn(log
, lsn
, flags
, log_flushed
, true);
3482 * Free a used ticket when its refcount falls to zero.
3486 xlog_ticket_t
*ticket
)
3488 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3489 if (atomic_dec_and_test(&ticket
->t_ref
))
3490 kmem_cache_free(xfs_log_ticket_zone
, ticket
);
3495 xlog_ticket_t
*ticket
)
3497 ASSERT(atomic_read(&ticket
->t_ref
) > 0);
3498 atomic_inc(&ticket
->t_ref
);
3503 * Figure out the total log space unit (in bytes) that would be
3504 * required for a log ticket.
3515 * Permanent reservations have up to 'cnt'-1 active log operations
3516 * in the log. A unit in this case is the amount of space for one
3517 * of these log operations. Normal reservations have a cnt of 1
3518 * and their unit amount is the total amount of space required.
3520 * The following lines of code account for non-transaction data
3521 * which occupy space in the on-disk log.
3523 * Normal form of a transaction is:
3524 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3525 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3527 * We need to account for all the leadup data and trailer data
3528 * around the transaction data.
3529 * And then we need to account for the worst case in terms of using
3531 * The worst case will happen if:
3532 * - the placement of the transaction happens to be such that the
3533 * roundoff is at its maximum
3534 * - the transaction data is synced before the commit record is synced
3535 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3536 * Therefore the commit record is in its own Log Record.
3537 * This can happen as the commit record is called with its
3538 * own region to xlog_write().
3539 * This then means that in the worst case, roundoff can happen for
3540 * the commit-rec as well.
3541 * The commit-rec is smaller than padding in this scenario and so it is
3542 * not added separately.
3545 /* for trans header */
3546 unit_bytes
+= sizeof(xlog_op_header_t
);
3547 unit_bytes
+= sizeof(xfs_trans_header_t
);
3550 unit_bytes
+= sizeof(xlog_op_header_t
);
3553 * for LR headers - the space for data in an iclog is the size minus
3554 * the space used for the headers. If we use the iclog size, then we
3555 * undercalculate the number of headers required.
3557 * Furthermore - the addition of op headers for split-recs might
3558 * increase the space required enough to require more log and op
3559 * headers, so take that into account too.
3561 * IMPORTANT: This reservation makes the assumption that if this
3562 * transaction is the first in an iclog and hence has the LR headers
3563 * accounted to it, then the remaining space in the iclog is
3564 * exclusively for this transaction. i.e. if the transaction is larger
3565 * than the iclog, it will be the only thing in that iclog.
3566 * Fundamentally, this means we must pass the entire log vector to
3567 * xlog_write to guarantee this.
3569 iclog_space
= log
->l_iclog_size
- log
->l_iclog_hsize
;
3570 num_headers
= howmany(unit_bytes
, iclog_space
);
3572 /* for split-recs - ophdrs added when data split over LRs */
3573 unit_bytes
+= sizeof(xlog_op_header_t
) * num_headers
;
3575 /* add extra header reservations if we overrun */
3576 while (!num_headers
||
3577 howmany(unit_bytes
, iclog_space
) > num_headers
) {
3578 unit_bytes
+= sizeof(xlog_op_header_t
);
3581 unit_bytes
+= log
->l_iclog_hsize
* num_headers
;
3583 /* for commit-rec LR header - note: padding will subsume the ophdr */
3584 unit_bytes
+= log
->l_iclog_hsize
;
3586 /* roundoff padding for transaction data and one for commit record */
3587 unit_bytes
+= 2 * log
->l_iclog_roundoff
;
3593 xfs_log_calc_unit_res(
3594 struct xfs_mount
*mp
,
3597 return xlog_calc_unit_res(mp
->m_log
, unit_bytes
);
3601 * Allocate and initialise a new log ticket.
3603 struct xlog_ticket
*
3611 struct xlog_ticket
*tic
;
3614 tic
= kmem_cache_zalloc(xfs_log_ticket_zone
, GFP_NOFS
| __GFP_NOFAIL
);
3616 unit_res
= xlog_calc_unit_res(log
, unit_bytes
);
3618 atomic_set(&tic
->t_ref
, 1);
3619 tic
->t_task
= current
;
3620 INIT_LIST_HEAD(&tic
->t_queue
);
3621 tic
->t_unit_res
= unit_res
;
3622 tic
->t_curr_res
= unit_res
;
3625 tic
->t_tid
= prandom_u32();
3626 tic
->t_clientid
= client
;
3628 tic
->t_flags
|= XLOG_TIC_PERM_RESERV
;
3630 xlog_tic_reset_res(tic
);
3637 * Make sure that the destination ptr is within the valid data region of
3638 * one of the iclogs. This uses backup pointers stored in a different
3639 * part of the log in case we trash the log structure.
3642 xlog_verify_dest_ptr(
3649 for (i
= 0; i
< log
->l_iclog_bufs
; i
++) {
3650 if (ptr
>= log
->l_iclog_bak
[i
] &&
3651 ptr
<= log
->l_iclog_bak
[i
] + log
->l_iclog_size
)
3656 xfs_emerg(log
->l_mp
, "%s: invalid ptr", __func__
);
3660 * Check to make sure the grant write head didn't just over lap the tail. If
3661 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3662 * the cycles differ by exactly one and check the byte count.
3664 * This check is run unlocked, so can give false positives. Rather than assert
3665 * on failures, use a warn-once flag and a panic tag to allow the admin to
3666 * determine if they want to panic the machine when such an error occurs. For
3667 * debug kernels this will have the same effect as using an assert but, unlinke
3668 * an assert, it can be turned off at runtime.
3671 xlog_verify_grant_tail(
3674 int tail_cycle
, tail_blocks
;
3677 xlog_crack_grant_head(&log
->l_write_head
.grant
, &cycle
, &space
);
3678 xlog_crack_atomic_lsn(&log
->l_tail_lsn
, &tail_cycle
, &tail_blocks
);
3679 if (tail_cycle
!= cycle
) {
3680 if (cycle
- 1 != tail_cycle
&&
3681 !test_and_set_bit(XLOG_TAIL_WARN
, &log
->l_opstate
)) {
3682 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3683 "%s: cycle - 1 != tail_cycle", __func__
);
3686 if (space
> BBTOB(tail_blocks
) &&
3687 !test_and_set_bit(XLOG_TAIL_WARN
, &log
->l_opstate
)) {
3688 xfs_alert_tag(log
->l_mp
, XFS_PTAG_LOGRES
,
3689 "%s: space > BBTOB(tail_blocks)", __func__
);
3694 /* check if it will fit */
3696 xlog_verify_tail_lsn(
3698 struct xlog_in_core
*iclog
)
3700 xfs_lsn_t tail_lsn
= be64_to_cpu(iclog
->ic_header
.h_tail_lsn
);
3703 if (CYCLE_LSN(tail_lsn
) == log
->l_prev_cycle
) {
3705 log
->l_logBBsize
- (log
->l_prev_block
- BLOCK_LSN(tail_lsn
));
3706 if (blocks
< BTOBB(iclog
->ic_offset
)+BTOBB(log
->l_iclog_hsize
))
3707 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3709 ASSERT(CYCLE_LSN(tail_lsn
)+1 == log
->l_prev_cycle
);
3711 if (BLOCK_LSN(tail_lsn
) == log
->l_prev_block
)
3712 xfs_emerg(log
->l_mp
, "%s: tail wrapped", __func__
);
3714 blocks
= BLOCK_LSN(tail_lsn
) - log
->l_prev_block
;
3715 if (blocks
< BTOBB(iclog
->ic_offset
) + 1)
3716 xfs_emerg(log
->l_mp
, "%s: ran out of log space", __func__
);
3721 * Perform a number of checks on the iclog before writing to disk.
3723 * 1. Make sure the iclogs are still circular
3724 * 2. Make sure we have a good magic number
3725 * 3. Make sure we don't have magic numbers in the data
3726 * 4. Check fields of each log operation header for:
3727 * A. Valid client identifier
3728 * B. tid ptr value falls in valid ptr space (user space code)
3729 * C. Length in log record header is correct according to the
3730 * individual operation headers within record.
3731 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3732 * log, check the preceding blocks of the physical log to make sure all
3733 * the cycle numbers agree with the current cycle number.
3738 struct xlog_in_core
*iclog
,
3741 xlog_op_header_t
*ophead
;
3742 xlog_in_core_t
*icptr
;
3743 xlog_in_core_2_t
*xhdr
;
3744 void *base_ptr
, *ptr
, *p
;
3745 ptrdiff_t field_offset
;
3747 int len
, i
, j
, k
, op_len
;
3750 /* check validity of iclog pointers */
3751 spin_lock(&log
->l_icloglock
);
3752 icptr
= log
->l_iclog
;
3753 for (i
= 0; i
< log
->l_iclog_bufs
; i
++, icptr
= icptr
->ic_next
)
3756 if (icptr
!= log
->l_iclog
)
3757 xfs_emerg(log
->l_mp
, "%s: corrupt iclog ring", __func__
);
3758 spin_unlock(&log
->l_icloglock
);
3760 /* check log magic numbers */
3761 if (iclog
->ic_header
.h_magicno
!= cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3762 xfs_emerg(log
->l_mp
, "%s: invalid magic num", __func__
);
3764 base_ptr
= ptr
= &iclog
->ic_header
;
3765 p
= &iclog
->ic_header
;
3766 for (ptr
+= BBSIZE
; ptr
< base_ptr
+ count
; ptr
+= BBSIZE
) {
3767 if (*(__be32
*)ptr
== cpu_to_be32(XLOG_HEADER_MAGIC_NUM
))
3768 xfs_emerg(log
->l_mp
, "%s: unexpected magic num",
3773 len
= be32_to_cpu(iclog
->ic_header
.h_num_logops
);
3774 base_ptr
= ptr
= iclog
->ic_datap
;
3776 xhdr
= iclog
->ic_data
;
3777 for (i
= 0; i
< len
; i
++) {
3780 /* clientid is only 1 byte */
3781 p
= &ophead
->oh_clientid
;
3782 field_offset
= p
- base_ptr
;
3783 if (field_offset
& 0x1ff) {
3784 clientid
= ophead
->oh_clientid
;
3786 idx
= BTOBBT((char *)&ophead
->oh_clientid
- iclog
->ic_datap
);
3787 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3788 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3789 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3790 clientid
= xlog_get_client_id(
3791 xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3793 clientid
= xlog_get_client_id(
3794 iclog
->ic_header
.h_cycle_data
[idx
]);
3797 if (clientid
!= XFS_TRANSACTION
&& clientid
!= XFS_LOG
)
3799 "%s: invalid clientid %d op "PTR_FMT
" offset 0x%lx",
3800 __func__
, clientid
, ophead
,
3801 (unsigned long)field_offset
);
3804 p
= &ophead
->oh_len
;
3805 field_offset
= p
- base_ptr
;
3806 if (field_offset
& 0x1ff) {
3807 op_len
= be32_to_cpu(ophead
->oh_len
);
3809 idx
= BTOBBT((uintptr_t)&ophead
->oh_len
-
3810 (uintptr_t)iclog
->ic_datap
);
3811 if (idx
>= (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
)) {
3812 j
= idx
/ (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3813 k
= idx
% (XLOG_HEADER_CYCLE_SIZE
/ BBSIZE
);
3814 op_len
= be32_to_cpu(xhdr
[j
].hic_xheader
.xh_cycle_data
[k
]);
3816 op_len
= be32_to_cpu(iclog
->ic_header
.h_cycle_data
[idx
]);
3819 ptr
+= sizeof(xlog_op_header_t
) + op_len
;
3825 * Perform a forced shutdown on the log. This should be called once and once
3826 * only by the high level filesystem shutdown code to shut the log subsystem
3829 * Our main objectives here are to make sure that:
3830 * a. if the shutdown was not due to a log IO error, flush the logs to
3831 * disk. Anything modified after this is ignored.
3832 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3833 * parties to find out. Nothing new gets queued after this is done.
3834 * c. Tasks sleeping on log reservations, pinned objects and
3835 * other resources get woken up.
3837 * Return true if the shutdown cause was a log IO error and we actually shut the
3841 xlog_force_shutdown(
3845 bool log_error
= (shutdown_flags
& SHUTDOWN_LOG_IO_ERROR
);
3848 * If this happens during log recovery then we aren't using the runtime
3849 * log mechanisms yet so there's nothing to shut down.
3851 if (!log
|| xlog_in_recovery(log
))
3854 ASSERT(!xlog_is_shutdown(log
));
3857 * Flush all the completed transactions to disk before marking the log
3858 * being shut down. We need to do this first as shutting down the log
3859 * before the force will prevent the log force from flushing the iclogs
3862 * Re-entry due to a log IO error shutdown during the log force is
3863 * prevented by the atomicity of higher level shutdown code.
3866 xfs_log_force(log
->l_mp
, XFS_LOG_SYNC
);
3869 * Atomically set the shutdown state. If the shutdown state is already
3870 * set, there someone else is performing the shutdown and so we are done
3871 * here. This should never happen because we should only ever get called
3872 * once by the first shutdown caller.
3874 * Much of the log state machine transitions assume that shutdown state
3875 * cannot change once they hold the log->l_icloglock. Hence we need to
3876 * hold that lock here, even though we use the atomic test_and_set_bit()
3877 * operation to set the shutdown state.
3879 spin_lock(&log
->l_icloglock
);
3880 if (test_and_set_bit(XLOG_IO_ERROR
, &log
->l_opstate
)) {
3881 spin_unlock(&log
->l_icloglock
);
3885 spin_unlock(&log
->l_icloglock
);
3888 * We don't want anybody waiting for log reservations after this. That
3889 * means we have to wake up everybody queued up on reserveq as well as
3890 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3891 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3892 * action is protected by the grant locks.
3894 xlog_grant_head_wake_all(&log
->l_reserve_head
);
3895 xlog_grant_head_wake_all(&log
->l_write_head
);
3898 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3899 * as if the log writes were completed. The abort handling in the log
3900 * item committed callback functions will do this again under lock to
3903 spin_lock(&log
->l_cilp
->xc_push_lock
);
3904 wake_up_all(&log
->l_cilp
->xc_start_wait
);
3905 wake_up_all(&log
->l_cilp
->xc_commit_wait
);
3906 spin_unlock(&log
->l_cilp
->xc_push_lock
);
3907 xlog_state_shutdown_callbacks(log
);
3916 xlog_in_core_t
*iclog
;
3918 iclog
= log
->l_iclog
;
3920 /* endianness does not matter here, zero is zero in
3923 if (iclog
->ic_header
.h_num_logops
)
3925 iclog
= iclog
->ic_next
;
3926 } while (iclog
!= log
->l_iclog
);
3931 * Verify that an LSN stamped into a piece of metadata is valid. This is
3932 * intended for use in read verifiers on v5 superblocks.
3936 struct xfs_mount
*mp
,
3939 struct xlog
*log
= mp
->m_log
;
3943 * norecovery mode skips mount-time log processing and unconditionally
3944 * resets the in-core LSN. We can't validate in this mode, but
3945 * modifications are not allowed anyways so just return true.
3947 if (xfs_has_norecovery(mp
))
3951 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3952 * handled by recovery and thus safe to ignore here.
3954 if (lsn
== NULLCOMMITLSN
)
3957 valid
= xlog_valid_lsn(mp
->m_log
, lsn
);
3959 /* warn the user about what's gone wrong before verifier failure */
3961 spin_lock(&log
->l_icloglock
);
3963 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3964 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3965 CYCLE_LSN(lsn
), BLOCK_LSN(lsn
),
3966 log
->l_curr_cycle
, log
->l_curr_block
);
3967 spin_unlock(&log
->l_icloglock
);
3974 * Notify the log that we're about to start using a feature that is protected
3975 * by a log incompat feature flag. This will prevent log covering from
3976 * clearing those flags.
3979 xlog_use_incompat_feat(
3982 down_read(&log
->l_incompat_users
);
3985 /* Notify the log that we've finished using log incompat features. */
3987 xlog_drop_incompat_feat(
3990 up_read(&log
->l_incompat_users
);