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 struct kmem_cache *xfs_log_ticket_cache;
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,
60 struct xlog_in_core *iclog,
61 struct xlog_ticket *ticket);
64 xlog_verify_grant_tail(
69 struct xlog_in_core *iclog,
74 struct xlog_in_core *iclog);
76 #define xlog_verify_grant_tail(a)
77 #define xlog_verify_iclog(a,b,c)
78 #define xlog_verify_tail_lsn(a,b)
86 xfs_log_cover(struct xfs_mount *);
89 * We need to make sure the buffer pointer returned is naturally aligned for the
90 * biggest basic data type we put into it. We have already accounted for this
91 * padding when sizing the buffer.
93 * However, this padding does not get written into the log, and hence we have to
94 * track the space used by the log vectors separately to prevent log space hangs
95 * due to inaccurate accounting (i.e. a leak) of the used log space through the
98 * We also add space for the xlog_op_header that describes this region in the
99 * log. This prepends the data region we return to the caller to copy their data
100 * into, so do all the static initialisation of the ophdr now. Because the ophdr
101 * is not 8 byte aligned, we have to be careful to ensure that we align the
102 * start of the buffer such that the region we return to the call is 8 byte
103 * aligned and packed against the tail of the ophdr.
107 struct xfs_log_vec *lv,
108 struct xfs_log_iovec **vecp,
111 struct xfs_log_iovec *vec = *vecp;
112 struct xlog_op_header *oph;
117 ASSERT(vec - lv->lv_iovecp < lv->lv_niovecs);
120 vec = &lv->lv_iovecp[0];
123 len = lv->lv_buf_len + sizeof(struct xlog_op_header);
124 if (!IS_ALIGNED(len, sizeof(uint64_t))) {
125 lv->lv_buf_len = round_up(len, sizeof(uint64_t)) -
126 sizeof(struct xlog_op_header);
130 vec->i_addr = lv->lv_buf + lv->lv_buf_len;
133 oph->oh_clientid = XFS_TRANSACTION;
137 buf = vec->i_addr + sizeof(struct xlog_op_header);
138 ASSERT(IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)));
145 xlog_grant_sub_space(
150 int64_t head_val = atomic64_read(head);
156 xlog_crack_grant_head_val(head_val, &cycle, &space);
160 space += log->l_logsize;
165 new = xlog_assign_grant_head_val(cycle, space);
166 head_val = atomic64_cmpxchg(head, old, new);
167 } while (head_val != old);
171 xlog_grant_add_space(
176 int64_t head_val = atomic64_read(head);
183 xlog_crack_grant_head_val(head_val, &cycle, &space);
185 tmp = log->l_logsize - space;
194 new = xlog_assign_grant_head_val(cycle, space);
195 head_val = atomic64_cmpxchg(head, old, new);
196 } while (head_val != old);
200 xlog_grant_head_init(
201 struct xlog_grant_head *head)
203 xlog_assign_grant_head(&head->grant, 1, 0);
204 INIT_LIST_HEAD(&head->waiters);
205 spin_lock_init(&head->lock);
209 xlog_grant_head_wake_all(
210 struct xlog_grant_head *head)
212 struct xlog_ticket *tic;
214 spin_lock(&head->lock);
215 list_for_each_entry(tic, &head->waiters, t_queue)
216 wake_up_process(tic->t_task);
217 spin_unlock(&head->lock);
221 xlog_ticket_reservation(
223 struct xlog_grant_head *head,
224 struct xlog_ticket *tic)
226 if (head == &log->l_write_head) {
227 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
228 return tic->t_unit_res;
231 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
232 return tic->t_unit_res * tic->t_cnt;
234 return tic->t_unit_res;
238 xlog_grant_head_wake(
240 struct xlog_grant_head *head,
243 struct xlog_ticket *tic;
245 bool woken_task = false;
247 list_for_each_entry(tic, &head->waiters, t_queue) {
250 * There is a chance that the size of the CIL checkpoints in
251 * progress at the last AIL push target calculation resulted in
252 * limiting the target to the log head (l_last_sync_lsn) at the
253 * time. This may not reflect where the log head is now as the
254 * CIL checkpoints may have completed.
256 * Hence when we are woken here, it may be that the head of the
257 * log that has moved rather than the tail. As the tail didn't
258 * move, there still won't be space available for the
259 * reservation we require. However, if the AIL has already
260 * pushed to the target defined by the old log head location, we
261 * will hang here waiting for something else to update the AIL
264 * Therefore, if there isn't space to wake the first waiter on
265 * the grant head, we need to push the AIL again to ensure the
266 * target reflects both the current log tail and log head
267 * position before we wait for the tail to move again.
270 need_bytes = xlog_ticket_reservation(log, head, tic);
271 if (*free_bytes < need_bytes) {
273 xlog_grant_push_ail(log, need_bytes);
277 *free_bytes -= need_bytes;
278 trace_xfs_log_grant_wake_up(log, tic);
279 wake_up_process(tic->t_task);
287 xlog_grant_head_wait(
289 struct xlog_grant_head *head,
290 struct xlog_ticket *tic,
291 int need_bytes) __releases(&head->lock)
292 __acquires(&head->lock)
294 list_add_tail(&tic->t_queue, &head->waiters);
297 if (xlog_is_shutdown(log))
299 xlog_grant_push_ail(log, need_bytes);
301 __set_current_state(TASK_UNINTERRUPTIBLE);
302 spin_unlock(&head->lock);
304 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
306 trace_xfs_log_grant_sleep(log, tic);
308 trace_xfs_log_grant_wake(log, tic);
310 spin_lock(&head->lock);
311 if (xlog_is_shutdown(log))
313 } while (xlog_space_left(log, &head->grant) < need_bytes);
315 list_del_init(&tic->t_queue);
318 list_del_init(&tic->t_queue);
323 * Atomically get the log space required for a log ticket.
325 * Once a ticket gets put onto head->waiters, it will only return after the
326 * needed reservation is satisfied.
328 * This function is structured so that it has a lock free fast path. This is
329 * necessary because every new transaction reservation will come through this
330 * path. Hence any lock will be globally hot if we take it unconditionally on
333 * As tickets are only ever moved on and off head->waiters under head->lock, we
334 * only need to take that lock if we are going to add the ticket to the queue
335 * and sleep. We can avoid taking the lock if the ticket was never added to
336 * head->waiters because the t_queue list head will be empty and we hold the
337 * only reference to it so it can safely be checked unlocked.
340 xlog_grant_head_check(
342 struct xlog_grant_head *head,
343 struct xlog_ticket *tic,
349 ASSERT(!xlog_in_recovery(log));
352 * If there are other waiters on the queue then give them a chance at
353 * logspace before us. Wake up the first waiters, if we do not wake
354 * up all the waiters then go to sleep waiting for more free space,
355 * otherwise try to get some space for this transaction.
357 *need_bytes = xlog_ticket_reservation(log, head, tic);
358 free_bytes = xlog_space_left(log, &head->grant);
359 if (!list_empty_careful(&head->waiters)) {
360 spin_lock(&head->lock);
361 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
362 free_bytes < *need_bytes) {
363 error = xlog_grant_head_wait(log, head, tic,
366 spin_unlock(&head->lock);
367 } else if (free_bytes < *need_bytes) {
368 spin_lock(&head->lock);
369 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
370 spin_unlock(&head->lock);
378 struct xfs_mount *mp)
381 * Do not write to the log on norecovery mounts, if the data or log
382 * devices are read-only, or if the filesystem is shutdown. Read-only
383 * mounts allow internal writes for log recovery and unmount purposes,
384 * so don't restrict that case.
386 if (xfs_has_norecovery(mp))
388 if (xfs_readonly_buftarg(mp->m_ddev_targp))
390 if (xfs_readonly_buftarg(mp->m_log->l_targ))
392 if (xlog_is_shutdown(mp->m_log))
398 * Replenish the byte reservation required by moving the grant write head.
402 struct xfs_mount *mp,
403 struct xlog_ticket *tic)
405 struct xlog *log = mp->m_log;
409 if (xlog_is_shutdown(log))
412 XFS_STATS_INC(mp, xs_try_logspace);
415 * This is a new transaction on the ticket, so we need to change the
416 * transaction ID so that the next transaction has a different TID in
417 * the log. Just add one to the existing tid so that we can see chains
418 * of rolling transactions in the log easily.
422 xlog_grant_push_ail(log, tic->t_unit_res);
424 tic->t_curr_res = tic->t_unit_res;
428 trace_xfs_log_regrant(log, tic);
430 error = xlog_grant_head_check(log, &log->l_write_head, tic,
435 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
436 trace_xfs_log_regrant_exit(log, tic);
437 xlog_verify_grant_tail(log);
442 * If we are failing, make sure the ticket doesn't have any current
443 * reservations. We don't want to add this back when the ticket/
444 * transaction gets cancelled.
447 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
452 * Reserve log space and return a ticket corresponding to the reservation.
454 * Each reservation is going to reserve extra space for a log record header.
455 * When writes happen to the on-disk log, we don't subtract the length of the
456 * log record header from any reservation. By wasting space in each
457 * reservation, we prevent over allocation problems.
461 struct xfs_mount *mp,
464 struct xlog_ticket **ticp,
467 struct xlog *log = mp->m_log;
468 struct xlog_ticket *tic;
472 if (xlog_is_shutdown(log))
475 XFS_STATS_INC(mp, xs_try_logspace);
477 ASSERT(*ticp == NULL);
478 tic = xlog_ticket_alloc(log, unit_bytes, cnt, permanent);
481 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
484 trace_xfs_log_reserve(log, tic);
486 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
491 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
492 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
493 trace_xfs_log_reserve_exit(log, tic);
494 xlog_verify_grant_tail(log);
499 * If we are failing, make sure the ticket doesn't have any current
500 * reservations. We don't want to add this back when the ticket/
501 * transaction gets cancelled.
504 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
509 * Run all the pending iclog callbacks and wake log force waiters and iclog
510 * space waiters so they can process the newly set shutdown state. We really
511 * don't care what order we process callbacks here because the log is shut down
512 * and so state cannot change on disk anymore. However, we cannot wake waiters
513 * until the callbacks have been processed because we may be in unmount and
514 * we must ensure that all AIL operations the callbacks perform have completed
515 * before we tear down the AIL.
517 * We avoid processing actively referenced iclogs so that we don't run callbacks
518 * while the iclog owner might still be preparing the iclog for IO submssion.
519 * These will be caught by xlog_state_iclog_release() and call this function
520 * again to process any callbacks that may have been added to that iclog.
523 xlog_state_shutdown_callbacks(
526 struct xlog_in_core *iclog;
529 iclog = log->l_iclog;
531 if (atomic_read(&iclog->ic_refcnt)) {
532 /* Reference holder will re-run iclog callbacks. */
535 list_splice_init(&iclog->ic_callbacks, &cb_list);
536 spin_unlock(&log->l_icloglock);
538 xlog_cil_process_committed(&cb_list);
540 spin_lock(&log->l_icloglock);
541 wake_up_all(&iclog->ic_write_wait);
542 wake_up_all(&iclog->ic_force_wait);
543 } while ((iclog = iclog->ic_next) != log->l_iclog);
545 wake_up_all(&log->l_flush_wait);
549 * Flush iclog to disk if this is the last reference to the given iclog and the
550 * it is in the WANT_SYNC state.
552 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
553 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
554 * written to stable storage, and implies that a commit record is contained
555 * within the iclog. We need to ensure that the log tail does not move beyond
556 * the tail that the first commit record in the iclog ordered against, otherwise
557 * correct recovery of that checkpoint becomes dependent on future operations
558 * performed on this iclog.
560 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
561 * current tail into iclog. Once the iclog tail is set, future operations must
562 * not modify it, otherwise they potentially violate ordering constraints for
563 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
564 * the iclog will get zeroed on activation of the iclog after sync, so we
565 * always capture the tail lsn on the iclog on the first NEED_FUA release
566 * regardless of the number of active reference counts on this iclog.
569 xlog_state_release_iclog(
571 struct xlog_in_core *iclog,
572 struct xlog_ticket *ticket)
577 lockdep_assert_held(&log->l_icloglock);
579 trace_xlog_iclog_release(iclog, _RET_IP_);
581 * Grabbing the current log tail needs to be atomic w.r.t. the writing
582 * of the tail LSN into the iclog so we guarantee that the log tail does
583 * not move between the first time we know that the iclog needs to be
584 * made stable and when we eventually submit it.
586 if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
587 (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
588 !iclog->ic_header.h_tail_lsn) {
589 tail_lsn = xlog_assign_tail_lsn(log->l_mp);
590 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
593 last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
595 if (xlog_is_shutdown(log)) {
597 * If there are no more references to this iclog, process the
598 * pending iclog callbacks that were waiting on the release of
602 xlog_state_shutdown_callbacks(log);
609 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
610 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
614 iclog->ic_state = XLOG_STATE_SYNCING;
615 xlog_verify_tail_lsn(log, iclog);
616 trace_xlog_iclog_syncing(iclog, _RET_IP_);
618 spin_unlock(&log->l_icloglock);
619 xlog_sync(log, iclog, ticket);
620 spin_lock(&log->l_icloglock);
625 * Mount a log filesystem
627 * mp - ubiquitous xfs mount point structure
628 * log_target - buftarg of on-disk log device
629 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
630 * num_bblocks - Number of BBSIZE blocks in on-disk log
632 * Return error or zero.
637 xfs_buftarg_t *log_target,
638 xfs_daddr_t blk_offset,
642 bool fatal = xfs_has_crc(mp);
646 if (!xfs_has_norecovery(mp)) {
647 xfs_notice(mp, "Mounting V%d Filesystem %pU",
648 XFS_SB_VERSION_NUM(&mp->m_sb),
652 "Mounting V%d filesystem %pU in no-recovery mode. Filesystem will be inconsistent.",
653 XFS_SB_VERSION_NUM(&mp->m_sb),
655 ASSERT(xfs_is_readonly(mp));
658 log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
660 error = PTR_ERR(log);
666 * Validate the given log space and drop a critical message via syslog
667 * if the log size is too small that would lead to some unexpected
668 * situations in transaction log space reservation stage.
670 * Note: we can't just reject the mount if the validation fails. This
671 * would mean that people would have to downgrade their kernel just to
672 * remedy the situation as there is no way to grow the log (short of
673 * black magic surgery with xfs_db).
675 * We can, however, reject mounts for CRC format filesystems, as the
676 * mkfs binary being used to make the filesystem should never create a
677 * filesystem with a log that is too small.
679 min_logfsbs = xfs_log_calc_minimum_size(mp);
681 if (mp->m_sb.sb_logblocks < min_logfsbs) {
683 "Log size %d blocks too small, minimum size is %d blocks",
684 mp->m_sb.sb_logblocks, min_logfsbs);
686 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
688 "Log size %d blocks too large, maximum size is %lld blocks",
689 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
691 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
693 "log size %lld bytes too large, maximum size is %lld bytes",
694 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
697 } else if (mp->m_sb.sb_logsunit > 1 &&
698 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
700 "log stripe unit %u bytes must be a multiple of block size",
701 mp->m_sb.sb_logsunit);
707 * Log check errors are always fatal on v5; or whenever bad
708 * metadata leads to a crash.
711 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
715 xfs_crit(mp, "Log size out of supported range.");
717 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
721 * Initialize the AIL now we have a log.
723 error = xfs_trans_ail_init(mp);
725 xfs_warn(mp, "AIL initialisation failed: error %d", error);
728 log->l_ailp = mp->m_ail;
731 * skip log recovery on a norecovery mount. pretend it all
734 if (!xfs_has_norecovery(mp)) {
736 * log recovery ignores readonly state and so we need to clear
737 * mount-based read only state so it can write to disk.
739 bool readonly = test_and_clear_bit(XFS_OPSTATE_READONLY,
741 error = xlog_recover(log);
743 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
745 xfs_warn(mp, "log mount/recovery failed: error %d",
747 xlog_recover_cancel(log);
748 goto out_destroy_ail;
752 error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
755 goto out_destroy_ail;
757 /* Normal transactions can now occur */
758 clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
761 * Now the log has been fully initialised and we know were our
762 * space grant counters are, we can initialise the permanent ticket
763 * needed for delayed logging to work.
765 xlog_cil_init_post_recovery(log);
770 xfs_trans_ail_destroy(mp);
772 xlog_dealloc_log(log);
778 * Finish the recovery of the file system. This is separate from the
779 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
780 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
783 * If we finish recovery successfully, start the background log work. If we are
784 * not doing recovery, then we have a RO filesystem and we don't need to start
788 xfs_log_mount_finish(
789 struct xfs_mount *mp)
791 struct xlog *log = mp->m_log;
795 if (xfs_has_norecovery(mp)) {
796 ASSERT(xfs_is_readonly(mp));
801 * log recovery ignores readonly state and so we need to clear
802 * mount-based read only state so it can write to disk.
804 readonly = test_and_clear_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
807 * During the second phase of log recovery, we need iget and
808 * iput to behave like they do for an active filesystem.
809 * xfs_fs_drop_inode needs to be able to prevent the deletion
810 * of inodes before we're done replaying log items on those
811 * inodes. Turn it off immediately after recovery finishes
812 * so that we don't leak the quota inodes if subsequent mount
815 * We let all inodes involved in redo item processing end up on
816 * the LRU instead of being evicted immediately so that if we do
817 * something to an unlinked inode, the irele won't cause
818 * premature truncation and freeing of the inode, which results
819 * in log recovery failure. We have to evict the unreferenced
820 * lru inodes after clearing SB_ACTIVE because we don't
821 * otherwise clean up the lru if there's a subsequent failure in
822 * xfs_mountfs, which leads to us leaking the inodes if nothing
823 * else (e.g. quotacheck) references the inodes before the
824 * mount failure occurs.
826 mp->m_super->s_flags |= SB_ACTIVE;
827 xfs_log_work_queue(mp);
828 if (xlog_recovery_needed(log))
829 error = xlog_recover_finish(log);
830 mp->m_super->s_flags &= ~SB_ACTIVE;
831 evict_inodes(mp->m_super);
834 * Drain the buffer LRU after log recovery. This is required for v4
835 * filesystems to avoid leaving around buffers with NULL verifier ops,
836 * but we do it unconditionally to make sure we're always in a clean
837 * cache state after mount.
839 * Don't push in the error case because the AIL may have pending intents
840 * that aren't removed until recovery is cancelled.
842 if (xlog_recovery_needed(log)) {
844 xfs_log_force(mp, XFS_LOG_SYNC);
845 xfs_ail_push_all_sync(mp->m_ail);
847 xfs_notice(mp, "Ending recovery (logdev: %s)",
848 mp->m_logname ? mp->m_logname : "internal");
850 xfs_info(mp, "Ending clean mount");
852 xfs_buftarg_drain(mp->m_ddev_targp);
854 clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
856 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
858 /* Make sure the log is dead if we're returning failure. */
859 ASSERT(!error || xlog_is_shutdown(log));
865 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
869 xfs_log_mount_cancel(
870 struct xfs_mount *mp)
872 xlog_recover_cancel(mp->m_log);
877 * Flush out the iclog to disk ensuring that device caches are flushed and
878 * the iclog hits stable storage before any completion waiters are woken.
882 struct xlog_in_core *iclog)
884 atomic_inc(&iclog->ic_refcnt);
885 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
886 if (iclog->ic_state == XLOG_STATE_ACTIVE)
887 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
888 return xlog_state_release_iclog(iclog->ic_log, iclog, NULL);
892 * Cycle all the iclogbuf locks to make sure all log IO completion
893 * is done before we tear down these buffers.
896 xlog_wait_iclog_completion(struct xlog *log)
899 struct xlog_in_core *iclog = log->l_iclog;
901 for (i = 0; i < log->l_iclog_bufs; i++) {
902 down(&iclog->ic_sema);
904 iclog = iclog->ic_next;
909 * Wait for the iclog and all prior iclogs to be written disk as required by the
910 * log force state machine. Waiting on ic_force_wait ensures iclog completions
911 * have been ordered and callbacks run before we are woken here, hence
912 * guaranteeing that all the iclogs up to this one are on stable storage.
916 struct xlog_in_core *iclog)
917 __releases(iclog->ic_log->l_icloglock)
919 struct xlog *log = iclog->ic_log;
921 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
922 if (!xlog_is_shutdown(log) &&
923 iclog->ic_state != XLOG_STATE_ACTIVE &&
924 iclog->ic_state != XLOG_STATE_DIRTY) {
925 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
926 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
928 spin_unlock(&log->l_icloglock);
931 if (xlog_is_shutdown(log))
937 * Write out an unmount record using the ticket provided. We have to account for
938 * the data space used in the unmount ticket as this write is not done from a
939 * transaction context that has already done the accounting for us.
942 xlog_write_unmount_record(
944 struct xlog_ticket *ticket)
947 struct xlog_op_header ophdr;
948 struct xfs_unmount_log_format ulf;
951 .oh_clientid = XFS_LOG,
952 .oh_tid = cpu_to_be32(ticket->t_tid),
953 .oh_flags = XLOG_UNMOUNT_TRANS,
956 .magic = XLOG_UNMOUNT_TYPE,
959 struct xfs_log_iovec reg = {
960 .i_addr = &unmount_rec,
961 .i_len = sizeof(unmount_rec),
962 .i_type = XLOG_REG_TYPE_UNMOUNT,
964 struct xfs_log_vec vec = {
969 list_add(&vec.lv_list, &lv_chain);
971 BUILD_BUG_ON((sizeof(struct xlog_op_header) +
972 sizeof(struct xfs_unmount_log_format)) !=
973 sizeof(unmount_rec));
975 /* account for space used by record data */
976 ticket->t_curr_res -= sizeof(unmount_rec);
978 return xlog_write(log, NULL, &lv_chain, ticket, reg.i_len);
982 * Mark the filesystem clean by writing an unmount record to the head of the
989 struct xfs_mount *mp = log->l_mp;
990 struct xlog_in_core *iclog;
991 struct xlog_ticket *tic = NULL;
994 error = xfs_log_reserve(mp, 600, 1, &tic, 0);
998 error = xlog_write_unmount_record(log, tic);
1000 * At this point, we're umounting anyway, so there's no point in
1001 * transitioning log state to shutdown. Just continue...
1005 xfs_alert(mp, "%s: unmount record failed", __func__);
1007 spin_lock(&log->l_icloglock);
1008 iclog = log->l_iclog;
1009 error = xlog_force_iclog(iclog);
1010 xlog_wait_on_iclog(iclog);
1013 trace_xfs_log_umount_write(log, tic);
1014 xfs_log_ticket_ungrant(log, tic);
1019 xfs_log_unmount_verify_iclog(
1022 struct xlog_in_core *iclog = log->l_iclog;
1025 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
1026 ASSERT(iclog->ic_offset == 0);
1027 } while ((iclog = iclog->ic_next) != log->l_iclog);
1031 * Unmount record used to have a string "Unmount filesystem--" in the
1032 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
1033 * We just write the magic number now since that particular field isn't
1034 * currently architecture converted and "Unmount" is a bit foo.
1035 * As far as I know, there weren't any dependencies on the old behaviour.
1038 xfs_log_unmount_write(
1039 struct xfs_mount *mp)
1041 struct xlog *log = mp->m_log;
1043 if (!xfs_log_writable(mp))
1046 xfs_log_force(mp, XFS_LOG_SYNC);
1048 if (xlog_is_shutdown(log))
1052 * If we think the summary counters are bad, avoid writing the unmount
1053 * record to force log recovery at next mount, after which the summary
1054 * counters will be recalculated. Refer to xlog_check_unmount_rec for
1057 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1058 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1059 xfs_alert(mp, "%s: will fix summary counters at next mount",
1064 xfs_log_unmount_verify_iclog(log);
1065 xlog_unmount_write(log);
1069 * Empty the log for unmount/freeze.
1071 * To do this, we first need to shut down the background log work so it is not
1072 * trying to cover the log as we clean up. We then need to unpin all objects in
1073 * the log so we can then flush them out. Once they have completed their IO and
1074 * run the callbacks removing themselves from the AIL, we can cover the log.
1078 struct xfs_mount *mp)
1081 * Clear log incompat features since we're quiescing the log. Report
1082 * failures, though it's not fatal to have a higher log feature
1083 * protection level than the log contents actually require.
1085 if (xfs_clear_incompat_log_features(mp)) {
1088 error = xfs_sync_sb(mp, false);
1091 "Failed to clear log incompat features on quiesce");
1094 cancel_delayed_work_sync(&mp->m_log->l_work);
1095 xfs_log_force(mp, XFS_LOG_SYNC);
1098 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1099 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1100 * xfs_buf_iowait() cannot be used because it was pushed with the
1101 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1102 * the IO to complete.
1104 xfs_ail_push_all_sync(mp->m_ail);
1105 xfs_buftarg_wait(mp->m_ddev_targp);
1106 xfs_buf_lock(mp->m_sb_bp);
1107 xfs_buf_unlock(mp->m_sb_bp);
1109 return xfs_log_cover(mp);
1114 struct xfs_mount *mp)
1116 xfs_log_quiesce(mp);
1117 xfs_log_unmount_write(mp);
1121 * Shut down and release the AIL and Log.
1123 * During unmount, we need to ensure we flush all the dirty metadata objects
1124 * from the AIL so that the log is empty before we write the unmount record to
1125 * the log. Once this is done, we can tear down the AIL and the log.
1129 struct xfs_mount *mp)
1134 * If shutdown has come from iclog IO context, the log
1135 * cleaning will have been skipped and so we need to wait
1136 * for the iclog to complete shutdown processing before we
1137 * tear anything down.
1139 xlog_wait_iclog_completion(mp->m_log);
1141 xfs_buftarg_drain(mp->m_ddev_targp);
1143 xfs_trans_ail_destroy(mp);
1145 xfs_sysfs_del(&mp->m_log->l_kobj);
1147 xlog_dealloc_log(mp->m_log);
1152 struct xfs_mount *mp,
1153 struct xfs_log_item *item,
1155 const struct xfs_item_ops *ops)
1157 item->li_log = mp->m_log;
1158 item->li_ailp = mp->m_ail;
1159 item->li_type = type;
1163 INIT_LIST_HEAD(&item->li_ail);
1164 INIT_LIST_HEAD(&item->li_cil);
1165 INIT_LIST_HEAD(&item->li_bio_list);
1166 INIT_LIST_HEAD(&item->li_trans);
1170 * Wake up processes waiting for log space after we have moved the log tail.
1174 struct xfs_mount *mp)
1176 struct xlog *log = mp->m_log;
1179 if (xlog_is_shutdown(log))
1182 if (!list_empty_careful(&log->l_write_head.waiters)) {
1183 ASSERT(!xlog_in_recovery(log));
1185 spin_lock(&log->l_write_head.lock);
1186 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1187 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1188 spin_unlock(&log->l_write_head.lock);
1191 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1192 ASSERT(!xlog_in_recovery(log));
1194 spin_lock(&log->l_reserve_head.lock);
1195 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1196 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1197 spin_unlock(&log->l_reserve_head.lock);
1202 * Determine if we have a transaction that has gone to disk that needs to be
1203 * covered. To begin the transition to the idle state firstly the log needs to
1204 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1205 * we start attempting to cover the log.
1207 * Only if we are then in a state where covering is needed, the caller is
1208 * informed that dummy transactions are required to move the log into the idle
1211 * If there are any items in the AIl or CIL, then we do not want to attempt to
1212 * cover the log as we may be in a situation where there isn't log space
1213 * available to run a dummy transaction and this can lead to deadlocks when the
1214 * tail of the log is pinned by an item that is modified in the CIL. Hence
1215 * there's no point in running a dummy transaction at this point because we
1216 * can't start trying to idle the log until both the CIL and AIL are empty.
1219 xfs_log_need_covered(
1220 struct xfs_mount *mp)
1222 struct xlog *log = mp->m_log;
1223 bool needed = false;
1225 if (!xlog_cil_empty(log))
1228 spin_lock(&log->l_icloglock);
1229 switch (log->l_covered_state) {
1230 case XLOG_STATE_COVER_DONE:
1231 case XLOG_STATE_COVER_DONE2:
1232 case XLOG_STATE_COVER_IDLE:
1234 case XLOG_STATE_COVER_NEED:
1235 case XLOG_STATE_COVER_NEED2:
1236 if (xfs_ail_min_lsn(log->l_ailp))
1238 if (!xlog_iclogs_empty(log))
1242 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1243 log->l_covered_state = XLOG_STATE_COVER_DONE;
1245 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1251 spin_unlock(&log->l_icloglock);
1256 * Explicitly cover the log. This is similar to background log covering but
1257 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1258 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1259 * must all be empty.
1263 struct xfs_mount *mp)
1268 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1269 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1270 xlog_is_shutdown(mp->m_log));
1272 if (!xfs_log_writable(mp))
1276 * xfs_log_need_covered() is not idempotent because it progresses the
1277 * state machine if the log requires covering. Therefore, we must call
1278 * this function once and use the result until we've issued an sb sync.
1279 * Do so first to make that abundantly clear.
1281 * Fall into the covering sequence if the log needs covering or the
1282 * mount has lazy superblock accounting to sync to disk. The sb sync
1283 * used for covering accumulates the in-core counters, so covering
1284 * handles this for us.
1286 need_covered = xfs_log_need_covered(mp);
1287 if (!need_covered && !xfs_has_lazysbcount(mp))
1291 * To cover the log, commit the superblock twice (at most) in
1292 * independent checkpoints. The first serves as a reference for the
1293 * tail pointer. The sync transaction and AIL push empties the AIL and
1294 * updates the in-core tail to the LSN of the first checkpoint. The
1295 * second commit updates the on-disk tail with the in-core LSN,
1296 * covering the log. Push the AIL one more time to leave it empty, as
1300 error = xfs_sync_sb(mp, true);
1303 xfs_ail_push_all_sync(mp->m_ail);
1304 } while (xfs_log_need_covered(mp));
1310 * We may be holding the log iclog lock upon entering this routine.
1313 xlog_assign_tail_lsn_locked(
1314 struct xfs_mount *mp)
1316 struct xlog *log = mp->m_log;
1317 struct xfs_log_item *lip;
1320 assert_spin_locked(&mp->m_ail->ail_lock);
1323 * To make sure we always have a valid LSN for the log tail we keep
1324 * track of the last LSN which was committed in log->l_last_sync_lsn,
1325 * and use that when the AIL was empty.
1327 lip = xfs_ail_min(mp->m_ail);
1329 tail_lsn = lip->li_lsn;
1331 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1332 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1333 atomic64_set(&log->l_tail_lsn, tail_lsn);
1338 xlog_assign_tail_lsn(
1339 struct xfs_mount *mp)
1343 spin_lock(&mp->m_ail->ail_lock);
1344 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1345 spin_unlock(&mp->m_ail->ail_lock);
1351 * Return the space in the log between the tail and the head. The head
1352 * is passed in the cycle/bytes formal parms. In the special case where
1353 * the reserve head has wrapped passed the tail, this calculation is no
1354 * longer valid. In this case, just return 0 which means there is no space
1355 * in the log. This works for all places where this function is called
1356 * with the reserve head. Of course, if the write head were to ever
1357 * wrap the tail, we should blow up. Rather than catch this case here,
1358 * we depend on other ASSERTions in other parts of the code. XXXmiken
1360 * If reservation head is behind the tail, we have a problem. Warn about it,
1361 * but then treat it as if the log is empty.
1363 * If the log is shut down, the head and tail may be invalid or out of whack, so
1364 * shortcut invalidity asserts in this case so that we don't trigger them
1377 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1378 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1379 tail_bytes = BBTOB(tail_bytes);
1380 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1381 return log->l_logsize - (head_bytes - tail_bytes);
1382 if (tail_cycle + 1 < head_cycle)
1385 /* Ignore potential inconsistency when shutdown. */
1386 if (xlog_is_shutdown(log))
1387 return log->l_logsize;
1389 if (tail_cycle < head_cycle) {
1390 ASSERT(tail_cycle == (head_cycle - 1));
1391 return tail_bytes - head_bytes;
1395 * The reservation head is behind the tail. In this case we just want to
1396 * return the size of the log as the amount of space left.
1398 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1399 xfs_alert(log->l_mp, " tail_cycle = %d, tail_bytes = %d",
1400 tail_cycle, tail_bytes);
1401 xfs_alert(log->l_mp, " GH cycle = %d, GH bytes = %d",
1402 head_cycle, head_bytes);
1404 return log->l_logsize;
1410 struct work_struct *work)
1412 struct xlog_in_core *iclog =
1413 container_of(work, struct xlog_in_core, ic_end_io_work);
1414 struct xlog *log = iclog->ic_log;
1417 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1419 /* treat writes with injected CRC errors as failed */
1420 if (iclog->ic_fail_crc)
1425 * Race to shutdown the filesystem if we see an error.
1427 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1428 xfs_alert(log->l_mp, "log I/O error %d", error);
1429 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1432 xlog_state_done_syncing(iclog);
1433 bio_uninit(&iclog->ic_bio);
1436 * Drop the lock to signal that we are done. Nothing references the
1437 * iclog after this, so an unmount waiting on this lock can now tear it
1438 * down safely. As such, it is unsafe to reference the iclog after the
1439 * unlock as we could race with it being freed.
1441 up(&iclog->ic_sema);
1445 * Return size of each in-core log record buffer.
1447 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1449 * If the filesystem blocksize is too large, we may need to choose a
1450 * larger size since the directory code currently logs entire blocks.
1453 xlog_get_iclog_buffer_size(
1454 struct xfs_mount *mp,
1457 if (mp->m_logbufs <= 0)
1458 mp->m_logbufs = XLOG_MAX_ICLOGS;
1459 if (mp->m_logbsize <= 0)
1460 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1462 log->l_iclog_bufs = mp->m_logbufs;
1463 log->l_iclog_size = mp->m_logbsize;
1466 * # headers = size / 32k - one header holds cycles from 32k of data.
1468 log->l_iclog_heads =
1469 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1470 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1475 struct xfs_mount *mp)
1477 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1478 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1482 * Clear the log incompat flags if we have the opportunity.
1484 * This only happens if we're about to log the second dummy transaction as part
1485 * of covering the log and we can get the log incompat feature usage lock.
1488 xlog_clear_incompat(
1491 struct xfs_mount *mp = log->l_mp;
1493 if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1494 XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1497 if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1500 if (!down_write_trylock(&log->l_incompat_users))
1503 xfs_clear_incompat_log_features(mp);
1504 up_write(&log->l_incompat_users);
1508 * Every sync period we need to unpin all items in the AIL and push them to
1509 * disk. If there is nothing dirty, then we might need to cover the log to
1510 * indicate that the filesystem is idle.
1514 struct work_struct *work)
1516 struct xlog *log = container_of(to_delayed_work(work),
1517 struct xlog, l_work);
1518 struct xfs_mount *mp = log->l_mp;
1520 /* dgc: errors ignored - not fatal and nowhere to report them */
1521 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1523 * Dump a transaction into the log that contains no real change.
1524 * This is needed to stamp the current tail LSN into the log
1525 * during the covering operation.
1527 * We cannot use an inode here for this - that will push dirty
1528 * state back up into the VFS and then periodic inode flushing
1529 * will prevent log covering from making progress. Hence we
1530 * synchronously log the superblock instead to ensure the
1531 * superblock is immediately unpinned and can be written back.
1533 xlog_clear_incompat(log);
1534 xfs_sync_sb(mp, true);
1536 xfs_log_force(mp, 0);
1538 /* start pushing all the metadata that is currently dirty */
1539 xfs_ail_push_all(mp->m_ail);
1541 /* queue us up again */
1542 xfs_log_work_queue(mp);
1546 * This routine initializes some of the log structure for a given mount point.
1547 * Its primary purpose is to fill in enough, so recovery can occur. However,
1548 * some other stuff may be filled in too.
1550 STATIC struct xlog *
1552 struct xfs_mount *mp,
1553 struct xfs_buftarg *log_target,
1554 xfs_daddr_t blk_offset,
1558 xlog_rec_header_t *head;
1559 xlog_in_core_t **iclogp;
1560 xlog_in_core_t *iclog, *prev_iclog=NULL;
1562 int error = -ENOMEM;
1565 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1567 xfs_warn(mp, "Log allocation failed: No memory!");
1572 log->l_targ = log_target;
1573 log->l_logsize = BBTOB(num_bblks);
1574 log->l_logBBstart = blk_offset;
1575 log->l_logBBsize = num_bblks;
1576 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1577 set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1578 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1580 log->l_prev_block = -1;
1581 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1582 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1583 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1584 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1586 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1587 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1589 log->l_iclog_roundoff = BBSIZE;
1591 xlog_grant_head_init(&log->l_reserve_head);
1592 xlog_grant_head_init(&log->l_write_head);
1594 error = -EFSCORRUPTED;
1595 if (xfs_has_sector(mp)) {
1596 log2_size = mp->m_sb.sb_logsectlog;
1597 if (log2_size < BBSHIFT) {
1598 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1599 log2_size, BBSHIFT);
1603 log2_size -= BBSHIFT;
1604 if (log2_size > mp->m_sectbb_log) {
1605 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1606 log2_size, mp->m_sectbb_log);
1610 /* for larger sector sizes, must have v2 or external log */
1611 if (log2_size && log->l_logBBstart > 0 &&
1612 !xfs_has_logv2(mp)) {
1614 "log sector size (0x%x) invalid for configuration.",
1619 log->l_sectBBsize = 1 << log2_size;
1621 init_rwsem(&log->l_incompat_users);
1623 xlog_get_iclog_buffer_size(mp, log);
1625 spin_lock_init(&log->l_icloglock);
1626 init_waitqueue_head(&log->l_flush_wait);
1628 iclogp = &log->l_iclog;
1630 * The amount of memory to allocate for the iclog structure is
1631 * rather funky due to the way the structure is defined. It is
1632 * done this way so that we can use different sizes for machines
1633 * with different amounts of memory. See the definition of
1634 * xlog_in_core_t in xfs_log_priv.h for details.
1636 ASSERT(log->l_iclog_size >= 4096);
1637 for (i = 0; i < log->l_iclog_bufs; i++) {
1638 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1639 sizeof(struct bio_vec);
1641 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1643 goto out_free_iclog;
1646 iclog->ic_prev = prev_iclog;
1649 iclog->ic_data = kvzalloc(log->l_iclog_size,
1650 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1651 if (!iclog->ic_data)
1652 goto out_free_iclog;
1653 head = &iclog->ic_header;
1654 memset(head, 0, sizeof(xlog_rec_header_t));
1655 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1656 head->h_version = cpu_to_be32(
1657 xfs_has_logv2(log->l_mp) ? 2 : 1);
1658 head->h_size = cpu_to_be32(log->l_iclog_size);
1660 head->h_fmt = cpu_to_be32(XLOG_FMT);
1661 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1663 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1664 iclog->ic_state = XLOG_STATE_ACTIVE;
1665 iclog->ic_log = log;
1666 atomic_set(&iclog->ic_refcnt, 0);
1667 INIT_LIST_HEAD(&iclog->ic_callbacks);
1668 iclog->ic_datap = (void *)iclog->ic_data + log->l_iclog_hsize;
1670 init_waitqueue_head(&iclog->ic_force_wait);
1671 init_waitqueue_head(&iclog->ic_write_wait);
1672 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1673 sema_init(&iclog->ic_sema, 1);
1675 iclogp = &iclog->ic_next;
1677 *iclogp = log->l_iclog; /* complete ring */
1678 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1680 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1681 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1683 0, mp->m_super->s_id);
1684 if (!log->l_ioend_workqueue)
1685 goto out_free_iclog;
1687 error = xlog_cil_init(log);
1689 goto out_destroy_workqueue;
1692 out_destroy_workqueue:
1693 destroy_workqueue(log->l_ioend_workqueue);
1695 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1696 prev_iclog = iclog->ic_next;
1697 kmem_free(iclog->ic_data);
1699 if (prev_iclog == log->l_iclog)
1705 return ERR_PTR(error);
1706 } /* xlog_alloc_log */
1709 * Compute the LSN that we'd need to push the log tail towards in order to have
1710 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1711 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1712 * log free space already meets all three thresholds, this function returns
1716 xlog_grant_push_threshold(
1720 xfs_lsn_t threshold_lsn = 0;
1721 xfs_lsn_t last_sync_lsn;
1724 int threshold_block;
1725 int threshold_cycle;
1728 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1730 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1731 free_blocks = BTOBBT(free_bytes);
1734 * Set the threshold for the minimum number of free blocks in the
1735 * log to the maximum of what the caller needs, one quarter of the
1736 * log, and 256 blocks.
1738 free_threshold = BTOBB(need_bytes);
1739 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1740 free_threshold = max(free_threshold, 256);
1741 if (free_blocks >= free_threshold)
1742 return NULLCOMMITLSN;
1744 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1746 threshold_block += free_threshold;
1747 if (threshold_block >= log->l_logBBsize) {
1748 threshold_block -= log->l_logBBsize;
1749 threshold_cycle += 1;
1751 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1754 * Don't pass in an lsn greater than the lsn of the last
1755 * log record known to be on disk. Use a snapshot of the last sync lsn
1756 * so that it doesn't change between the compare and the set.
1758 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1759 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1760 threshold_lsn = last_sync_lsn;
1762 return threshold_lsn;
1766 * Push the tail of the log if we need to do so to maintain the free log space
1767 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1768 * policy which pushes on an lsn which is further along in the log once we
1769 * reach the high water mark. In this manner, we would be creating a low water
1773 xlog_grant_push_ail(
1777 xfs_lsn_t threshold_lsn;
1779 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1780 if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1784 * Get the transaction layer to kick the dirty buffers out to
1785 * disk asynchronously. No point in trying to do this if
1786 * the filesystem is shutting down.
1788 xfs_ail_push(log->l_ailp, threshold_lsn);
1792 * Stamp cycle number in every block
1797 struct xlog_in_core *iclog,
1801 int size = iclog->ic_offset + roundoff;
1805 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1807 dp = iclog->ic_datap;
1808 for (i = 0; i < BTOBB(size); i++) {
1809 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1811 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1812 *(__be32 *)dp = cycle_lsn;
1816 if (xfs_has_logv2(log->l_mp)) {
1817 xlog_in_core_2_t *xhdr = iclog->ic_data;
1819 for ( ; i < BTOBB(size); i++) {
1820 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1821 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1822 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1823 *(__be32 *)dp = cycle_lsn;
1827 for (i = 1; i < log->l_iclog_heads; i++)
1828 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1833 * Calculate the checksum for a log buffer.
1835 * This is a little more complicated than it should be because the various
1836 * headers and the actual data are non-contiguous.
1841 struct xlog_rec_header *rhead,
1847 /* first generate the crc for the record header ... */
1848 crc = xfs_start_cksum_update((char *)rhead,
1849 sizeof(struct xlog_rec_header),
1850 offsetof(struct xlog_rec_header, h_crc));
1852 /* ... then for additional cycle data for v2 logs ... */
1853 if (xfs_has_logv2(log->l_mp)) {
1854 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1858 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1860 for (i = 1; i < xheads; i++) {
1861 crc = crc32c(crc, &xhdr[i].hic_xheader,
1862 sizeof(struct xlog_rec_ext_header));
1866 /* ... and finally for the payload */
1867 crc = crc32c(crc, dp, size);
1869 return xfs_end_cksum(crc);
1876 struct xlog_in_core *iclog = bio->bi_private;
1878 queue_work(iclog->ic_log->l_ioend_workqueue,
1879 &iclog->ic_end_io_work);
1883 xlog_map_iclog_data(
1889 struct page *page = kmem_to_page(data);
1890 unsigned int off = offset_in_page(data);
1891 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1893 if (bio_add_page(bio, page, len, off) != len)
1906 struct xlog_in_core *iclog,
1910 ASSERT(bno < log->l_logBBsize);
1911 trace_xlog_iclog_write(iclog, _RET_IP_);
1914 * We lock the iclogbufs here so that we can serialise against I/O
1915 * completion during unmount. We might be processing a shutdown
1916 * triggered during unmount, and that can occur asynchronously to the
1917 * unmount thread, and hence we need to ensure that completes before
1918 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1919 * across the log IO to archieve that.
1921 down(&iclog->ic_sema);
1922 if (xlog_is_shutdown(log)) {
1924 * It would seem logical to return EIO here, but we rely on
1925 * the log state machine to propagate I/O errors instead of
1926 * doing it here. We kick of the state machine and unlock
1927 * the buffer manually, the code needs to be kept in sync
1928 * with the I/O completion path.
1930 xlog_state_done_syncing(iclog);
1931 up(&iclog->ic_sema);
1936 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1937 * IOs coming immediately after this one. This prevents the block layer
1938 * writeback throttle from throttling log writes behind background
1939 * metadata writeback and causing priority inversions.
1941 bio_init(&iclog->ic_bio, log->l_targ->bt_bdev, iclog->ic_bvec,
1942 howmany(count, PAGE_SIZE),
1943 REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE);
1944 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1945 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1946 iclog->ic_bio.bi_private = iclog;
1948 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1949 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1951 * For external log devices, we also need to flush the data
1952 * device cache first to ensure all metadata writeback covered
1953 * by the LSN in this iclog is on stable storage. This is slow,
1954 * but it *must* complete before we issue the external log IO.
1956 * If the flush fails, we cannot conclude that past metadata
1957 * writeback from the log succeeded. Repeating the flush is
1958 * not possible, hence we must shut down with log IO error to
1959 * avoid shutdown re-entering this path and erroring out again.
1961 if (log->l_targ != log->l_mp->m_ddev_targp &&
1962 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev)) {
1963 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1967 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1968 iclog->ic_bio.bi_opf |= REQ_FUA;
1970 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1972 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1973 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
1976 if (is_vmalloc_addr(iclog->ic_data))
1977 flush_kernel_vmap_range(iclog->ic_data, count);
1980 * If this log buffer would straddle the end of the log we will have
1981 * to split it up into two bios, so that we can continue at the start.
1983 if (bno + BTOBB(count) > log->l_logBBsize) {
1986 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1987 GFP_NOIO, &fs_bio_set);
1988 bio_chain(split, &iclog->ic_bio);
1991 /* restart at logical offset zero for the remainder */
1992 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1995 submit_bio(&iclog->ic_bio);
1999 * We need to bump cycle number for the part of the iclog that is
2000 * written to the start of the log. Watch out for the header magic
2001 * number case, though.
2010 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
2013 for (i = split_offset; i < count; i += BBSIZE) {
2014 uint32_t cycle = get_unaligned_be32(data + i);
2016 if (++cycle == XLOG_HEADER_MAGIC_NUM)
2018 put_unaligned_be32(cycle, data + i);
2023 xlog_calc_iclog_size(
2025 struct xlog_in_core *iclog,
2028 uint32_t count_init, count;
2030 /* Add for LR header */
2031 count_init = log->l_iclog_hsize + iclog->ic_offset;
2032 count = roundup(count_init, log->l_iclog_roundoff);
2034 *roundoff = count - count_init;
2036 ASSERT(count >= count_init);
2037 ASSERT(*roundoff < log->l_iclog_roundoff);
2042 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
2043 * fashion. Previously, we should have moved the current iclog
2044 * ptr in the log to point to the next available iclog. This allows further
2045 * write to continue while this code syncs out an iclog ready to go.
2046 * Before an in-core log can be written out, the data section must be scanned
2047 * to save away the 1st word of each BBSIZE block into the header. We replace
2048 * it with the current cycle count. Each BBSIZE block is tagged with the
2049 * cycle count because there in an implicit assumption that drives will
2050 * guarantee that entire 512 byte blocks get written at once. In other words,
2051 * we can't have part of a 512 byte block written and part not written. By
2052 * tagging each block, we will know which blocks are valid when recovering
2053 * after an unclean shutdown.
2055 * This routine is single threaded on the iclog. No other thread can be in
2056 * this routine with the same iclog. Changing contents of iclog can there-
2057 * fore be done without grabbing the state machine lock. Updating the global
2058 * log will require grabbing the lock though.
2060 * The entire log manager uses a logical block numbering scheme. Only
2061 * xlog_write_iclog knows about the fact that the log may not start with
2062 * block zero on a given device.
2067 struct xlog_in_core *iclog,
2068 struct xlog_ticket *ticket)
2070 unsigned int count; /* byte count of bwrite */
2071 unsigned int roundoff; /* roundoff to BB or stripe */
2075 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2076 trace_xlog_iclog_sync(iclog, _RET_IP_);
2078 count = xlog_calc_iclog_size(log, iclog, &roundoff);
2081 * If we have a ticket, account for the roundoff via the ticket
2082 * reservation to avoid touching the hot grant heads needlessly.
2083 * Otherwise, we have to move grant heads directly.
2086 ticket->t_curr_res -= roundoff;
2088 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2089 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2092 /* put cycle number in every block */
2093 xlog_pack_data(log, iclog, roundoff);
2095 /* real byte length */
2096 size = iclog->ic_offset;
2097 if (xfs_has_logv2(log->l_mp))
2099 iclog->ic_header.h_len = cpu_to_be32(size);
2101 XFS_STATS_INC(log->l_mp, xs_log_writes);
2102 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2104 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2106 /* Do we need to split this write into 2 parts? */
2107 if (bno + BTOBB(count) > log->l_logBBsize)
2108 xlog_split_iclog(log, &iclog->ic_header, bno, count);
2110 /* calculcate the checksum */
2111 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2112 iclog->ic_datap, size);
2114 * Intentionally corrupt the log record CRC based on the error injection
2115 * frequency, if defined. This facilitates testing log recovery in the
2116 * event of torn writes. Hence, set the IOABORT state to abort the log
2117 * write on I/O completion and shutdown the fs. The subsequent mount
2118 * detects the bad CRC and attempts to recover.
2121 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2122 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2123 iclog->ic_fail_crc = true;
2125 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2126 be64_to_cpu(iclog->ic_header.h_lsn));
2129 xlog_verify_iclog(log, iclog, count);
2130 xlog_write_iclog(log, iclog, bno, count);
2134 * Deallocate a log structure
2140 xlog_in_core_t *iclog, *next_iclog;
2144 * Destroy the CIL after waiting for iclog IO completion because an
2145 * iclog EIO error will try to shut down the log, which accesses the
2146 * CIL to wake up the waiters.
2148 xlog_cil_destroy(log);
2150 iclog = log->l_iclog;
2151 for (i = 0; i < log->l_iclog_bufs; i++) {
2152 next_iclog = iclog->ic_next;
2153 kmem_free(iclog->ic_data);
2158 log->l_mp->m_log = NULL;
2159 destroy_workqueue(log->l_ioend_workqueue);
2164 * Update counters atomically now that memcpy is done.
2167 xlog_state_finish_copy(
2169 struct xlog_in_core *iclog,
2173 lockdep_assert_held(&log->l_icloglock);
2175 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2176 iclog->ic_offset += copy_bytes;
2180 * print out info relating to regions written which consume
2185 struct xfs_mount *mp,
2186 struct xlog_ticket *ticket)
2188 xfs_warn(mp, "ticket reservation summary:");
2189 xfs_warn(mp, " unit res = %d bytes", ticket->t_unit_res);
2190 xfs_warn(mp, " current res = %d bytes", ticket->t_curr_res);
2191 xfs_warn(mp, " original count = %d", ticket->t_ocnt);
2192 xfs_warn(mp, " remaining count = %d", ticket->t_cnt);
2196 * Print a summary of the transaction.
2200 struct xfs_trans *tp)
2202 struct xfs_mount *mp = tp->t_mountp;
2203 struct xfs_log_item *lip;
2205 /* dump core transaction and ticket info */
2206 xfs_warn(mp, "transaction summary:");
2207 xfs_warn(mp, " log res = %d", tp->t_log_res);
2208 xfs_warn(mp, " log count = %d", tp->t_log_count);
2209 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2211 xlog_print_tic_res(mp, tp->t_ticket);
2213 /* dump each log item */
2214 list_for_each_entry(lip, &tp->t_items, li_trans) {
2215 struct xfs_log_vec *lv = lip->li_lv;
2216 struct xfs_log_iovec *vec;
2219 xfs_warn(mp, "log item: ");
2220 xfs_warn(mp, " type = 0x%x", lip->li_type);
2221 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2224 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2225 xfs_warn(mp, " size = %d", lv->lv_size);
2226 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2227 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2229 /* dump each iovec for the log item */
2230 vec = lv->lv_iovecp;
2231 for (i = 0; i < lv->lv_niovecs; i++) {
2232 int dumplen = min(vec->i_len, 32);
2234 xfs_warn(mp, " iovec[%d]", i);
2235 xfs_warn(mp, " type = 0x%x", vec->i_type);
2236 xfs_warn(mp, " len = %d", vec->i_len);
2237 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2238 xfs_hex_dump(vec->i_addr, dumplen);
2247 struct xlog_in_core *iclog,
2248 uint32_t *log_offset,
2252 uint32_t *record_cnt,
2255 ASSERT(*log_offset < iclog->ic_log->l_iclog_size);
2256 ASSERT(*log_offset % sizeof(int32_t) == 0);
2257 ASSERT(write_len % sizeof(int32_t) == 0);
2259 memcpy(iclog->ic_datap + *log_offset, data, write_len);
2260 *log_offset += write_len;
2261 *bytes_left -= write_len;
2263 *data_cnt += write_len;
2267 * Write log vectors into a single iclog which is guaranteed by the caller
2268 * to have enough space to write the entire log vector into.
2272 struct xfs_log_vec *lv,
2273 struct xlog_ticket *ticket,
2274 struct xlog_in_core *iclog,
2275 uint32_t *log_offset,
2277 uint32_t *record_cnt,
2282 ASSERT(*log_offset + *len <= iclog->ic_size ||
2283 iclog->ic_state == XLOG_STATE_WANT_SYNC);
2286 * Ordered log vectors have no regions to write so this
2287 * loop will naturally skip them.
2289 for (index = 0; index < lv->lv_niovecs; index++) {
2290 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2291 struct xlog_op_header *ophdr = reg->i_addr;
2293 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2294 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2295 reg->i_len, len, record_cnt, data_cnt);
2300 xlog_write_get_more_iclog_space(
2301 struct xlog_ticket *ticket,
2302 struct xlog_in_core **iclogp,
2303 uint32_t *log_offset,
2305 uint32_t *record_cnt,
2308 struct xlog_in_core *iclog = *iclogp;
2309 struct xlog *log = iclog->ic_log;
2312 spin_lock(&log->l_icloglock);
2313 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC);
2314 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2315 error = xlog_state_release_iclog(log, iclog, ticket);
2316 spin_unlock(&log->l_icloglock);
2320 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2331 * Write log vectors into a single iclog which is smaller than the current chain
2332 * length. We write until we cannot fit a full record into the remaining space
2333 * and then stop. We return the log vector that is to be written that cannot
2334 * wholly fit in the iclog.
2338 struct xfs_log_vec *lv,
2339 struct xlog_ticket *ticket,
2340 struct xlog_in_core **iclogp,
2341 uint32_t *log_offset,
2343 uint32_t *record_cnt,
2346 struct xlog_in_core *iclog = *iclogp;
2347 struct xlog_op_header *ophdr;
2352 /* walk the logvec, copying until we run out of space in the iclog */
2353 for (index = 0; index < lv->lv_niovecs; index++) {
2354 struct xfs_log_iovec *reg = &lv->lv_iovecp[index];
2355 uint32_t reg_offset = 0;
2358 * The first region of a continuation must have a non-zero
2359 * length otherwise log recovery will just skip over it and
2360 * start recovering from the next opheader it finds. Because we
2361 * mark the next opheader as a continuation, recovery will then
2362 * incorrectly add the continuation to the previous region and
2363 * that breaks stuff.
2365 * Hence if there isn't space for region data after the
2366 * opheader, then we need to start afresh with a new iclog.
2368 if (iclog->ic_size - *log_offset <=
2369 sizeof(struct xlog_op_header)) {
2370 error = xlog_write_get_more_iclog_space(ticket,
2371 &iclog, log_offset, *len, record_cnt,
2377 ophdr = reg->i_addr;
2378 rlen = min_t(uint32_t, reg->i_len, iclog->ic_size - *log_offset);
2380 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2381 ophdr->oh_len = cpu_to_be32(rlen - sizeof(struct xlog_op_header));
2382 if (rlen != reg->i_len)
2383 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2385 xlog_write_iovec(iclog, log_offset, reg->i_addr,
2386 rlen, len, record_cnt, data_cnt);
2388 /* If we wrote the whole region, move to the next. */
2389 if (rlen == reg->i_len)
2393 * We now have a partially written iovec, but it can span
2394 * multiple iclogs so we loop here. First we release the iclog
2395 * we currently have, then we get a new iclog and add a new
2396 * opheader. Then we continue copying from where we were until
2397 * we either complete the iovec or fill the iclog. If we
2398 * complete the iovec, then we increment the index and go right
2399 * back to the top of the outer loop. if we fill the iclog, we
2400 * run the inner loop again.
2402 * This is complicated by the tail of a region using all the
2403 * space in an iclog and hence requiring us to release the iclog
2404 * and get a new one before returning to the outer loop. We must
2405 * always guarantee that we exit this inner loop with at least
2406 * space for log transaction opheaders left in the current
2407 * iclog, hence we cannot just terminate the loop at the end
2408 * of the of the continuation. So we loop while there is no
2409 * space left in the current iclog, and check for the end of the
2410 * continuation after getting a new iclog.
2414 * Ensure we include the continuation opheader in the
2415 * space we need in the new iclog by adding that size
2416 * to the length we require. This continuation opheader
2417 * needs to be accounted to the ticket as the space it
2418 * consumes hasn't been accounted to the lv we are
2421 error = xlog_write_get_more_iclog_space(ticket,
2423 *len + sizeof(struct xlog_op_header),
2424 record_cnt, data_cnt);
2428 ophdr = iclog->ic_datap + *log_offset;
2429 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2430 ophdr->oh_clientid = XFS_TRANSACTION;
2432 ophdr->oh_flags = XLOG_WAS_CONT_TRANS;
2434 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2435 *log_offset += sizeof(struct xlog_op_header);
2436 *data_cnt += sizeof(struct xlog_op_header);
2439 * If rlen fits in the iclog, then end the region
2440 * continuation. Otherwise we're going around again.
2443 rlen = reg->i_len - reg_offset;
2444 if (rlen <= iclog->ic_size - *log_offset)
2445 ophdr->oh_flags |= XLOG_END_TRANS;
2447 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2449 rlen = min_t(uint32_t, rlen, iclog->ic_size - *log_offset);
2450 ophdr->oh_len = cpu_to_be32(rlen);
2452 xlog_write_iovec(iclog, log_offset,
2453 reg->i_addr + reg_offset,
2454 rlen, len, record_cnt, data_cnt);
2456 } while (ophdr->oh_flags & XLOG_CONTINUE_TRANS);
2460 * No more iovecs remain in this logvec so return the next log vec to
2461 * the caller so it can go back to fast path copying.
2468 * Write some region out to in-core log
2470 * This will be called when writing externally provided regions or when
2471 * writing out a commit record for a given transaction.
2473 * General algorithm:
2474 * 1. Find total length of this write. This may include adding to the
2475 * lengths passed in.
2476 * 2. Check whether we violate the tickets reservation.
2477 * 3. While writing to this iclog
2478 * A. Reserve as much space in this iclog as can get
2479 * B. If this is first write, save away start lsn
2480 * C. While writing this region:
2481 * 1. If first write of transaction, write start record
2482 * 2. Write log operation header (header per region)
2483 * 3. Find out if we can fit entire region into this iclog
2484 * 4. Potentially, verify destination memcpy ptr
2485 * 5. Memcpy (partial) region
2486 * 6. If partial copy, release iclog; otherwise, continue
2487 * copying more regions into current iclog
2488 * 4. Mark want sync bit (in simulation mode)
2489 * 5. Release iclog for potential flush to on-disk log.
2492 * 1. Panic if reservation is overrun. This should never happen since
2493 * reservation amounts are generated internal to the filesystem.
2495 * 1. Tickets are single threaded data structures.
2496 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2497 * syncing routine. When a single log_write region needs to span
2498 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2499 * on all log operation writes which don't contain the end of the
2500 * region. The XLOG_END_TRANS bit is used for the in-core log
2501 * operation which contains the end of the continued log_write region.
2502 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2503 * we don't really know exactly how much space will be used. As a result,
2504 * we don't update ic_offset until the end when we know exactly how many
2505 * bytes have been written out.
2510 struct xfs_cil_ctx *ctx,
2511 struct list_head *lv_chain,
2512 struct xlog_ticket *ticket,
2516 struct xlog_in_core *iclog = NULL;
2517 struct xfs_log_vec *lv;
2518 uint32_t record_cnt = 0;
2519 uint32_t data_cnt = 0;
2523 if (ticket->t_curr_res < 0) {
2524 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2525 "ctx ticket reservation ran out. Need to up reservation");
2526 xlog_print_tic_res(log->l_mp, ticket);
2527 xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
2530 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2535 ASSERT(log_offset <= iclog->ic_size - 1);
2538 * If we have a context pointer, pass it the first iclog we are
2539 * writing to so it can record state needed for iclog write
2543 xlog_cil_set_ctx_write_state(ctx, iclog);
2545 list_for_each_entry(lv, lv_chain, lv_list) {
2547 * If the entire log vec does not fit in the iclog, punt it to
2548 * the partial copy loop which can handle this case.
2550 if (lv->lv_niovecs &&
2551 lv->lv_bytes > iclog->ic_size - log_offset) {
2552 error = xlog_write_partial(lv, ticket, &iclog,
2553 &log_offset, &len, &record_cnt,
2557 * We have no iclog to release, so just return
2558 * the error immediately.
2563 xlog_write_full(lv, ticket, iclog, &log_offset,
2564 &len, &record_cnt, &data_cnt);
2570 * We've already been guaranteed that the last writes will fit inside
2571 * the current iclog, and hence it will already have the space used by
2572 * those writes accounted to it. Hence we do not need to update the
2573 * iclog with the number of bytes written here.
2575 spin_lock(&log->l_icloglock);
2576 xlog_state_finish_copy(log, iclog, record_cnt, 0);
2577 error = xlog_state_release_iclog(log, iclog, ticket);
2578 spin_unlock(&log->l_icloglock);
2584 xlog_state_activate_iclog(
2585 struct xlog_in_core *iclog,
2586 int *iclogs_changed)
2588 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2589 trace_xlog_iclog_activate(iclog, _RET_IP_);
2592 * If the number of ops in this iclog indicate it just contains the
2593 * dummy transaction, we can change state into IDLE (the second time
2594 * around). Otherwise we should change the state into NEED a dummy.
2595 * We don't need to cover the dummy.
2597 if (*iclogs_changed == 0 &&
2598 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2599 *iclogs_changed = 1;
2602 * We have two dirty iclogs so start over. This could also be
2603 * num of ops indicating this is not the dummy going out.
2605 *iclogs_changed = 2;
2608 iclog->ic_state = XLOG_STATE_ACTIVE;
2609 iclog->ic_offset = 0;
2610 iclog->ic_header.h_num_logops = 0;
2611 memset(iclog->ic_header.h_cycle_data, 0,
2612 sizeof(iclog->ic_header.h_cycle_data));
2613 iclog->ic_header.h_lsn = 0;
2614 iclog->ic_header.h_tail_lsn = 0;
2618 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2619 * ACTIVE after iclog I/O has completed.
2622 xlog_state_activate_iclogs(
2624 int *iclogs_changed)
2626 struct xlog_in_core *iclog = log->l_iclog;
2629 if (iclog->ic_state == XLOG_STATE_DIRTY)
2630 xlog_state_activate_iclog(iclog, iclogs_changed);
2632 * The ordering of marking iclogs ACTIVE must be maintained, so
2633 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2635 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2637 } while ((iclog = iclog->ic_next) != log->l_iclog);
2646 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2647 * wrote the first covering record (DONE). We go to IDLE if we just
2648 * wrote the second covering record (DONE2) and remain in IDLE until a
2649 * non-covering write occurs.
2651 switch (prev_state) {
2652 case XLOG_STATE_COVER_IDLE:
2653 if (iclogs_changed == 1)
2654 return XLOG_STATE_COVER_IDLE;
2656 case XLOG_STATE_COVER_NEED:
2657 case XLOG_STATE_COVER_NEED2:
2659 case XLOG_STATE_COVER_DONE:
2660 if (iclogs_changed == 1)
2661 return XLOG_STATE_COVER_NEED2;
2663 case XLOG_STATE_COVER_DONE2:
2664 if (iclogs_changed == 1)
2665 return XLOG_STATE_COVER_IDLE;
2671 return XLOG_STATE_COVER_NEED;
2675 xlog_state_clean_iclog(
2677 struct xlog_in_core *dirty_iclog)
2679 int iclogs_changed = 0;
2681 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2683 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2685 xlog_state_activate_iclogs(log, &iclogs_changed);
2686 wake_up_all(&dirty_iclog->ic_force_wait);
2688 if (iclogs_changed) {
2689 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2695 xlog_get_lowest_lsn(
2698 struct xlog_in_core *iclog = log->l_iclog;
2699 xfs_lsn_t lowest_lsn = 0, lsn;
2702 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2703 iclog->ic_state == XLOG_STATE_DIRTY)
2706 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2707 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2709 } while ((iclog = iclog->ic_next) != log->l_iclog);
2715 * Completion of a iclog IO does not imply that a transaction has completed, as
2716 * transactions can be large enough to span many iclogs. We cannot change the
2717 * tail of the log half way through a transaction as this may be the only
2718 * transaction in the log and moving the tail to point to the middle of it
2719 * will prevent recovery from finding the start of the transaction. Hence we
2720 * should only update the last_sync_lsn if this iclog contains transaction
2721 * completion callbacks on it.
2723 * We have to do this before we drop the icloglock to ensure we are the only one
2724 * that can update it.
2726 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2727 * the reservation grant head pushing. This is due to the fact that the push
2728 * target is bound by the current last_sync_lsn value. Hence if we have a large
2729 * amount of log space bound up in this committing transaction then the
2730 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2731 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2732 * should push the AIL to ensure the push target (and hence the grant head) is
2733 * no longer bound by the old log head location and can move forwards and make
2737 xlog_state_set_callback(
2739 struct xlog_in_core *iclog,
2740 xfs_lsn_t header_lsn)
2742 trace_xlog_iclog_callback(iclog, _RET_IP_);
2743 iclog->ic_state = XLOG_STATE_CALLBACK;
2745 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2748 if (list_empty_careful(&iclog->ic_callbacks))
2751 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2752 xlog_grant_push_ail(log, 0);
2756 * Return true if we need to stop processing, false to continue to the next
2757 * iclog. The caller will need to run callbacks if the iclog is returned in the
2758 * XLOG_STATE_CALLBACK state.
2761 xlog_state_iodone_process_iclog(
2763 struct xlog_in_core *iclog)
2765 xfs_lsn_t lowest_lsn;
2766 xfs_lsn_t header_lsn;
2768 switch (iclog->ic_state) {
2769 case XLOG_STATE_ACTIVE:
2770 case XLOG_STATE_DIRTY:
2772 * Skip all iclogs in the ACTIVE & DIRTY states:
2775 case XLOG_STATE_DONE_SYNC:
2777 * Now that we have an iclog that is in the DONE_SYNC state, do
2778 * one more check here to see if we have chased our tail around.
2779 * If this is not the lowest lsn iclog, then we will leave it
2780 * for another completion to process.
2782 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2783 lowest_lsn = xlog_get_lowest_lsn(log);
2784 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2786 xlog_state_set_callback(log, iclog, header_lsn);
2790 * Can only perform callbacks in order. Since this iclog is not
2791 * in the DONE_SYNC state, we skip the rest and just try to
2799 * Loop over all the iclogs, running attached callbacks on them. Return true if
2800 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2801 * to handle transient shutdown state here at all because
2802 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2803 * cleanup of the callbacks.
2806 xlog_state_do_iclog_callbacks(
2808 __releases(&log->l_icloglock)
2809 __acquires(&log->l_icloglock)
2811 struct xlog_in_core *first_iclog = log->l_iclog;
2812 struct xlog_in_core *iclog = first_iclog;
2813 bool ran_callback = false;
2818 if (xlog_state_iodone_process_iclog(log, iclog))
2820 if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2821 iclog = iclog->ic_next;
2824 list_splice_init(&iclog->ic_callbacks, &cb_list);
2825 spin_unlock(&log->l_icloglock);
2827 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2828 xlog_cil_process_committed(&cb_list);
2829 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2830 ran_callback = true;
2832 spin_lock(&log->l_icloglock);
2833 xlog_state_clean_iclog(log, iclog);
2834 iclog = iclog->ic_next;
2835 } while (iclog != first_iclog);
2837 return ran_callback;
2842 * Loop running iclog completion callbacks until there are no more iclogs in a
2843 * state that can run callbacks.
2846 xlog_state_do_callback(
2852 spin_lock(&log->l_icloglock);
2853 while (xlog_state_do_iclog_callbacks(log)) {
2854 if (xlog_is_shutdown(log))
2857 if (++repeats > 5000) {
2858 flushcnt += repeats;
2861 "%s: possible infinite loop (%d iterations)",
2862 __func__, flushcnt);
2866 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2867 wake_up_all(&log->l_flush_wait);
2869 spin_unlock(&log->l_icloglock);
2874 * Finish transitioning this iclog to the dirty state.
2876 * Callbacks could take time, so they are done outside the scope of the
2877 * global state machine log lock.
2880 xlog_state_done_syncing(
2881 struct xlog_in_core *iclog)
2883 struct xlog *log = iclog->ic_log;
2885 spin_lock(&log->l_icloglock);
2886 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2887 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2890 * If we got an error, either on the first buffer, or in the case of
2891 * split log writes, on the second, we shut down the file system and
2892 * no iclogs should ever be attempted to be written to disk again.
2894 if (!xlog_is_shutdown(log)) {
2895 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2896 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2900 * Someone could be sleeping prior to writing out the next
2901 * iclog buffer, we wake them all, one will get to do the
2902 * I/O, the others get to wait for the result.
2904 wake_up_all(&iclog->ic_write_wait);
2905 spin_unlock(&log->l_icloglock);
2906 xlog_state_do_callback(log);
2910 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2911 * sleep. We wait on the flush queue on the head iclog as that should be
2912 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2913 * we will wait here and all new writes will sleep until a sync completes.
2915 * The in-core logs are used in a circular fashion. They are not used
2916 * out-of-order even when an iclog past the head is free.
2919 * * log_offset where xlog_write() can start writing into the in-core
2921 * * in-core log pointer to which xlog_write() should write.
2922 * * boolean indicating this is a continued write to an in-core log.
2923 * If this is the last write, then the in-core log's offset field
2924 * needs to be incremented, depending on the amount of data which
2928 xlog_state_get_iclog_space(
2931 struct xlog_in_core **iclogp,
2932 struct xlog_ticket *ticket,
2936 xlog_rec_header_t *head;
2937 xlog_in_core_t *iclog;
2940 spin_lock(&log->l_icloglock);
2941 if (xlog_is_shutdown(log)) {
2942 spin_unlock(&log->l_icloglock);
2946 iclog = log->l_iclog;
2947 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2948 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2950 /* Wait for log writes to have flushed */
2951 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2955 head = &iclog->ic_header;
2957 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2958 log_offset = iclog->ic_offset;
2960 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2962 /* On the 1st write to an iclog, figure out lsn. This works
2963 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2964 * committing to. If the offset is set, that's how many blocks
2967 if (log_offset == 0) {
2968 ticket->t_curr_res -= log->l_iclog_hsize;
2969 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2970 head->h_lsn = cpu_to_be64(
2971 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2972 ASSERT(log->l_curr_block >= 0);
2975 /* If there is enough room to write everything, then do it. Otherwise,
2976 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2977 * bit is on, so this will get flushed out. Don't update ic_offset
2978 * until you know exactly how many bytes get copied. Therefore, wait
2979 * until later to update ic_offset.
2981 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2982 * can fit into remaining data section.
2984 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2987 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2990 * If we are the only one writing to this iclog, sync it to
2991 * disk. We need to do an atomic compare and decrement here to
2992 * avoid racing with concurrent atomic_dec_and_lock() calls in
2993 * xlog_state_release_iclog() when there is more than one
2994 * reference to the iclog.
2996 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2997 error = xlog_state_release_iclog(log, iclog, ticket);
2998 spin_unlock(&log->l_icloglock);
3004 /* Do we have enough room to write the full amount in the remainder
3005 * of this iclog? Or must we continue a write on the next iclog and
3006 * mark this iclog as completely taken? In the case where we switch
3007 * iclogs (to mark it taken), this particular iclog will release/sync
3008 * to disk in xlog_write().
3010 if (len <= iclog->ic_size - iclog->ic_offset)
3011 iclog->ic_offset += len;
3013 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3016 ASSERT(iclog->ic_offset <= iclog->ic_size);
3017 spin_unlock(&log->l_icloglock);
3019 *logoffsetp = log_offset;
3024 * The first cnt-1 times a ticket goes through here we don't need to move the
3025 * grant write head because the permanent reservation has reserved cnt times the
3026 * unit amount. Release part of current permanent unit reservation and reset
3027 * current reservation to be one units worth. Also move grant reservation head
3031 xfs_log_ticket_regrant(
3033 struct xlog_ticket *ticket)
3035 trace_xfs_log_ticket_regrant(log, ticket);
3037 if (ticket->t_cnt > 0)
3040 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3041 ticket->t_curr_res);
3042 xlog_grant_sub_space(log, &log->l_write_head.grant,
3043 ticket->t_curr_res);
3044 ticket->t_curr_res = ticket->t_unit_res;
3046 trace_xfs_log_ticket_regrant_sub(log, ticket);
3048 /* just return if we still have some of the pre-reserved space */
3049 if (!ticket->t_cnt) {
3050 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3051 ticket->t_unit_res);
3052 trace_xfs_log_ticket_regrant_exit(log, ticket);
3054 ticket->t_curr_res = ticket->t_unit_res;
3057 xfs_log_ticket_put(ticket);
3061 * Give back the space left from a reservation.
3063 * All the information we need to make a correct determination of space left
3064 * is present. For non-permanent reservations, things are quite easy. The
3065 * count should have been decremented to zero. We only need to deal with the
3066 * space remaining in the current reservation part of the ticket. If the
3067 * ticket contains a permanent reservation, there may be left over space which
3068 * needs to be released. A count of N means that N-1 refills of the current
3069 * reservation can be done before we need to ask for more space. The first
3070 * one goes to fill up the first current reservation. Once we run out of
3071 * space, the count will stay at zero and the only space remaining will be
3072 * in the current reservation field.
3075 xfs_log_ticket_ungrant(
3077 struct xlog_ticket *ticket)
3081 trace_xfs_log_ticket_ungrant(log, ticket);
3083 if (ticket->t_cnt > 0)
3086 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3089 * If this is a permanent reservation ticket, we may be able to free
3090 * up more space based on the remaining count.
3092 bytes = ticket->t_curr_res;
3093 if (ticket->t_cnt > 0) {
3094 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3095 bytes += ticket->t_unit_res*ticket->t_cnt;
3098 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3099 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3101 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3103 xfs_log_space_wake(log->l_mp);
3104 xfs_log_ticket_put(ticket);
3108 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3109 * the current iclog pointer to the next iclog in the ring.
3112 xlog_state_switch_iclogs(
3114 struct xlog_in_core *iclog,
3117 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3118 assert_spin_locked(&log->l_icloglock);
3119 trace_xlog_iclog_switch(iclog, _RET_IP_);
3122 eventual_size = iclog->ic_offset;
3123 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3124 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3125 log->l_prev_block = log->l_curr_block;
3126 log->l_prev_cycle = log->l_curr_cycle;
3128 /* roll log?: ic_offset changed later */
3129 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3131 /* Round up to next log-sunit */
3132 if (log->l_iclog_roundoff > BBSIZE) {
3133 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3134 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3137 if (log->l_curr_block >= log->l_logBBsize) {
3139 * Rewind the current block before the cycle is bumped to make
3140 * sure that the combined LSN never transiently moves forward
3141 * when the log wraps to the next cycle. This is to support the
3142 * unlocked sample of these fields from xlog_valid_lsn(). Most
3143 * other cases should acquire l_icloglock.
3145 log->l_curr_block -= log->l_logBBsize;
3146 ASSERT(log->l_curr_block >= 0);
3148 log->l_curr_cycle++;
3149 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3150 log->l_curr_cycle++;
3152 ASSERT(iclog == log->l_iclog);
3153 log->l_iclog = iclog->ic_next;
3157 * Force the iclog to disk and check if the iclog has been completed before
3158 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3159 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3160 * If completion has already occurred, tell the caller so that it can avoid an
3161 * unnecessary wait on the iclog.
3164 xlog_force_and_check_iclog(
3165 struct xlog_in_core *iclog,
3168 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3172 error = xlog_force_iclog(iclog);
3177 * If the iclog has already been completed and reused the header LSN
3178 * will have been rewritten by completion
3180 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3186 * Write out all data in the in-core log as of this exact moment in time.
3188 * Data may be written to the in-core log during this call. However,
3189 * we don't guarantee this data will be written out. A change from past
3190 * implementation means this routine will *not* write out zero length LRs.
3192 * Basically, we try and perform an intelligent scan of the in-core logs.
3193 * If we determine there is no flushable data, we just return. There is no
3194 * flushable data if:
3196 * 1. the current iclog is active and has no data; the previous iclog
3197 * is in the active or dirty state.
3198 * 2. the current iclog is drity, and the previous iclog is in the
3199 * active or dirty state.
3203 * 1. the current iclog is not in the active nor dirty state.
3204 * 2. the current iclog dirty, and the previous iclog is not in the
3205 * active nor dirty state.
3206 * 3. the current iclog is active, and there is another thread writing
3207 * to this particular iclog.
3208 * 4. a) the current iclog is active and has no other writers
3209 * b) when we return from flushing out this iclog, it is still
3210 * not in the active nor dirty state.
3214 struct xfs_mount *mp,
3217 struct xlog *log = mp->m_log;
3218 struct xlog_in_core *iclog;
3220 XFS_STATS_INC(mp, xs_log_force);
3221 trace_xfs_log_force(mp, 0, _RET_IP_);
3223 xlog_cil_force(log);
3225 spin_lock(&log->l_icloglock);
3226 if (xlog_is_shutdown(log))
3229 iclog = log->l_iclog;
3230 trace_xlog_iclog_force(iclog, _RET_IP_);
3232 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3233 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3234 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3236 * If the head is dirty or (active and empty), then we need to
3237 * look at the previous iclog.
3239 * If the previous iclog is active or dirty we are done. There
3240 * is nothing to sync out. Otherwise, we attach ourselves to the
3241 * previous iclog and go to sleep.
3243 iclog = iclog->ic_prev;
3244 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3245 if (atomic_read(&iclog->ic_refcnt) == 0) {
3246 /* We have exclusive access to this iclog. */
3249 if (xlog_force_and_check_iclog(iclog, &completed))
3256 * Someone else is still writing to this iclog, so we
3257 * need to ensure that when they release the iclog it
3258 * gets synced immediately as we may be waiting on it.
3260 xlog_state_switch_iclogs(log, iclog, 0);
3265 * The iclog we are about to wait on may contain the checkpoint pushed
3266 * by the above xlog_cil_force() call, but it may not have been pushed
3267 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3268 * are flushed when this iclog is written.
3270 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3271 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3273 if (flags & XFS_LOG_SYNC)
3274 return xlog_wait_on_iclog(iclog);
3276 spin_unlock(&log->l_icloglock);
3279 spin_unlock(&log->l_icloglock);
3284 * Force the log to a specific LSN.
3286 * If an iclog with that lsn can be found:
3287 * If it is in the DIRTY state, just return.
3288 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3289 * state and go to sleep or return.
3290 * If it is in any other state, go to sleep or return.
3292 * Synchronous forces are implemented with a wait queue. All callers trying
3293 * to force a given lsn to disk must wait on the queue attached to the
3294 * specific in-core log. When given in-core log finally completes its write
3295 * to disk, that thread will wake up all threads waiting on the queue.
3305 struct xlog_in_core *iclog;
3308 spin_lock(&log->l_icloglock);
3309 if (xlog_is_shutdown(log))
3312 iclog = log->l_iclog;
3313 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3314 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3315 iclog = iclog->ic_next;
3316 if (iclog == log->l_iclog)
3320 switch (iclog->ic_state) {
3321 case XLOG_STATE_ACTIVE:
3323 * We sleep here if we haven't already slept (e.g. this is the
3324 * first time we've looked at the correct iclog buf) and the
3325 * buffer before us is going to be sync'ed. The reason for this
3326 * is that if we are doing sync transactions here, by waiting
3327 * for the previous I/O to complete, we can allow a few more
3328 * transactions into this iclog before we close it down.
3330 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3331 * refcnt so we can release the log (which drops the ref count).
3332 * The state switch keeps new transaction commits from using
3333 * this buffer. When the current commits finish writing into
3334 * the buffer, the refcount will drop to zero and the buffer
3337 if (!already_slept &&
3338 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3339 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3340 xlog_wait(&iclog->ic_prev->ic_write_wait,
3344 if (xlog_force_and_check_iclog(iclog, &completed))
3351 case XLOG_STATE_WANT_SYNC:
3353 * This iclog may contain the checkpoint pushed by the
3354 * xlog_cil_force_seq() call, but there are other writers still
3355 * accessing it so it hasn't been pushed to disk yet. Like the
3356 * ACTIVE case above, we need to make sure caches are flushed
3357 * when this iclog is written.
3359 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3363 * The entire checkpoint was written by the CIL force and is on
3364 * its way to disk already. It will be stable when it
3365 * completes, so we don't need to manipulate caches here at all.
3366 * We just need to wait for completion if necessary.
3371 if (flags & XFS_LOG_SYNC)
3372 return xlog_wait_on_iclog(iclog);
3374 spin_unlock(&log->l_icloglock);
3377 spin_unlock(&log->l_icloglock);
3382 * Force the log to a specific checkpoint sequence.
3384 * First force the CIL so that all the required changes have been flushed to the
3385 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3386 * the iclog that needs to be flushed to stable storage. If the caller needs
3387 * a synchronous log force, we will wait on the iclog with the LSN returned by
3388 * xlog_cil_force_seq() to be completed.
3392 struct xfs_mount *mp,
3397 struct xlog *log = mp->m_log;
3402 XFS_STATS_INC(mp, xs_log_force);
3403 trace_xfs_log_force(mp, seq, _RET_IP_);
3405 lsn = xlog_cil_force_seq(log, seq);
3406 if (lsn == NULLCOMMITLSN)
3409 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3410 if (ret == -EAGAIN) {
3411 XFS_STATS_INC(mp, xs_log_force_sleep);
3412 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3418 * Free a used ticket when its refcount falls to zero.
3422 xlog_ticket_t *ticket)
3424 ASSERT(atomic_read(&ticket->t_ref) > 0);
3425 if (atomic_dec_and_test(&ticket->t_ref))
3426 kmem_cache_free(xfs_log_ticket_cache, ticket);
3431 xlog_ticket_t *ticket)
3433 ASSERT(atomic_read(&ticket->t_ref) > 0);
3434 atomic_inc(&ticket->t_ref);
3439 * Figure out the total log space unit (in bytes) that would be
3440 * required for a log ticket.
3452 * Permanent reservations have up to 'cnt'-1 active log operations
3453 * in the log. A unit in this case is the amount of space for one
3454 * of these log operations. Normal reservations have a cnt of 1
3455 * and their unit amount is the total amount of space required.
3457 * The following lines of code account for non-transaction data
3458 * which occupy space in the on-disk log.
3460 * Normal form of a transaction is:
3461 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3462 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3464 * We need to account for all the leadup data and trailer data
3465 * around the transaction data.
3466 * And then we need to account for the worst case in terms of using
3468 * The worst case will happen if:
3469 * - the placement of the transaction happens to be such that the
3470 * roundoff is at its maximum
3471 * - the transaction data is synced before the commit record is synced
3472 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3473 * Therefore the commit record is in its own Log Record.
3474 * This can happen as the commit record is called with its
3475 * own region to xlog_write().
3476 * This then means that in the worst case, roundoff can happen for
3477 * the commit-rec as well.
3478 * The commit-rec is smaller than padding in this scenario and so it is
3479 * not added separately.
3482 /* for trans header */
3483 unit_bytes += sizeof(xlog_op_header_t);
3484 unit_bytes += sizeof(xfs_trans_header_t);
3487 unit_bytes += sizeof(xlog_op_header_t);
3490 * for LR headers - the space for data in an iclog is the size minus
3491 * the space used for the headers. If we use the iclog size, then we
3492 * undercalculate the number of headers required.
3494 * Furthermore - the addition of op headers for split-recs might
3495 * increase the space required enough to require more log and op
3496 * headers, so take that into account too.
3498 * IMPORTANT: This reservation makes the assumption that if this
3499 * transaction is the first in an iclog and hence has the LR headers
3500 * accounted to it, then the remaining space in the iclog is
3501 * exclusively for this transaction. i.e. if the transaction is larger
3502 * than the iclog, it will be the only thing in that iclog.
3503 * Fundamentally, this means we must pass the entire log vector to
3504 * xlog_write to guarantee this.
3506 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3507 num_headers = howmany(unit_bytes, iclog_space);
3509 /* for split-recs - ophdrs added when data split over LRs */
3510 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3512 /* add extra header reservations if we overrun */
3513 while (!num_headers ||
3514 howmany(unit_bytes, iclog_space) > num_headers) {
3515 unit_bytes += sizeof(xlog_op_header_t);
3518 unit_bytes += log->l_iclog_hsize * num_headers;
3520 /* for commit-rec LR header - note: padding will subsume the ophdr */
3521 unit_bytes += log->l_iclog_hsize;
3523 /* roundoff padding for transaction data and one for commit record */
3524 unit_bytes += 2 * log->l_iclog_roundoff;
3527 *niclogs = num_headers;
3532 xfs_log_calc_unit_res(
3533 struct xfs_mount *mp,
3536 return xlog_calc_unit_res(mp->m_log, unit_bytes, NULL);
3540 * Allocate and initialise a new log ticket.
3542 struct xlog_ticket *
3549 struct xlog_ticket *tic;
3552 tic = kmem_cache_zalloc(xfs_log_ticket_cache, GFP_NOFS | __GFP_NOFAIL);
3554 unit_res = xlog_calc_unit_res(log, unit_bytes, &tic->t_iclog_hdrs);
3556 atomic_set(&tic->t_ref, 1);
3557 tic->t_task = current;
3558 INIT_LIST_HEAD(&tic->t_queue);
3559 tic->t_unit_res = unit_res;
3560 tic->t_curr_res = unit_res;
3563 tic->t_tid = get_random_u32();
3565 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3572 * Check to make sure the grant write head didn't just over lap the tail. If
3573 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3574 * the cycles differ by exactly one and check the byte count.
3576 * This check is run unlocked, so can give false positives. Rather than assert
3577 * on failures, use a warn-once flag and a panic tag to allow the admin to
3578 * determine if they want to panic the machine when such an error occurs. For
3579 * debug kernels this will have the same effect as using an assert but, unlinke
3580 * an assert, it can be turned off at runtime.
3583 xlog_verify_grant_tail(
3586 int tail_cycle, tail_blocks;
3589 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3590 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3591 if (tail_cycle != cycle) {
3592 if (cycle - 1 != tail_cycle &&
3593 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3594 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3595 "%s: cycle - 1 != tail_cycle", __func__);
3598 if (space > BBTOB(tail_blocks) &&
3599 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3600 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3601 "%s: space > BBTOB(tail_blocks)", __func__);
3606 /* check if it will fit */
3608 xlog_verify_tail_lsn(
3610 struct xlog_in_core *iclog)
3612 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3615 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3617 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3618 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3619 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3621 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3623 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3624 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3626 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3627 if (blocks < BTOBB(iclog->ic_offset) + 1)
3628 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3633 * Perform a number of checks on the iclog before writing to disk.
3635 * 1. Make sure the iclogs are still circular
3636 * 2. Make sure we have a good magic number
3637 * 3. Make sure we don't have magic numbers in the data
3638 * 4. Check fields of each log operation header for:
3639 * A. Valid client identifier
3640 * B. tid ptr value falls in valid ptr space (user space code)
3641 * C. Length in log record header is correct according to the
3642 * individual operation headers within record.
3643 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3644 * log, check the preceding blocks of the physical log to make sure all
3645 * the cycle numbers agree with the current cycle number.
3650 struct xlog_in_core *iclog,
3653 xlog_op_header_t *ophead;
3654 xlog_in_core_t *icptr;
3655 xlog_in_core_2_t *xhdr;
3656 void *base_ptr, *ptr, *p;
3657 ptrdiff_t field_offset;
3659 int len, i, j, k, op_len;
3662 /* check validity of iclog pointers */
3663 spin_lock(&log->l_icloglock);
3664 icptr = log->l_iclog;
3665 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3668 if (icptr != log->l_iclog)
3669 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3670 spin_unlock(&log->l_icloglock);
3672 /* check log magic numbers */
3673 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3674 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3676 base_ptr = ptr = &iclog->ic_header;
3677 p = &iclog->ic_header;
3678 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3679 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3680 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3685 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3686 base_ptr = ptr = iclog->ic_datap;
3688 xhdr = iclog->ic_data;
3689 for (i = 0; i < len; i++) {
3692 /* clientid is only 1 byte */
3693 p = &ophead->oh_clientid;
3694 field_offset = p - base_ptr;
3695 if (field_offset & 0x1ff) {
3696 clientid = ophead->oh_clientid;
3698 idx = BTOBBT((void *)&ophead->oh_clientid - iclog->ic_datap);
3699 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3700 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3701 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3702 clientid = xlog_get_client_id(
3703 xhdr[j].hic_xheader.xh_cycle_data[k]);
3705 clientid = xlog_get_client_id(
3706 iclog->ic_header.h_cycle_data[idx]);
3709 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) {
3711 "%s: op %d invalid clientid %d op "PTR_FMT" offset 0x%lx",
3712 __func__, i, clientid, ophead,
3713 (unsigned long)field_offset);
3717 p = &ophead->oh_len;
3718 field_offset = p - base_ptr;
3719 if (field_offset & 0x1ff) {
3720 op_len = be32_to_cpu(ophead->oh_len);
3722 idx = BTOBBT((void *)&ophead->oh_len - iclog->ic_datap);
3723 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3724 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3725 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3726 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3728 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3731 ptr += sizeof(xlog_op_header_t) + op_len;
3737 * Perform a forced shutdown on the log.
3739 * This can be called from low level log code to trigger a shutdown, or from the
3740 * high level mount shutdown code when the mount shuts down.
3742 * Our main objectives here are to make sure that:
3743 * a. if the shutdown was not due to a log IO error, flush the logs to
3744 * disk. Anything modified after this is ignored.
3745 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3746 * parties to find out. Nothing new gets queued after this is done.
3747 * c. Tasks sleeping on log reservations, pinned objects and
3748 * other resources get woken up.
3749 * d. The mount is also marked as shut down so that log triggered shutdowns
3750 * still behave the same as if they called xfs_forced_shutdown().
3752 * Return true if the shutdown cause was a log IO error and we actually shut the
3756 xlog_force_shutdown(
3758 uint32_t shutdown_flags)
3760 bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3766 * Flush all the completed transactions to disk before marking the log
3767 * being shut down. We need to do this first as shutting down the log
3768 * before the force will prevent the log force from flushing the iclogs
3771 * When we are in recovery, there are no transactions to flush, and
3772 * we don't want to touch the log because we don't want to perturb the
3773 * current head/tail for future recovery attempts. Hence we need to
3774 * avoid a log force in this case.
3776 * If we are shutting down due to a log IO error, then we must avoid
3777 * trying to write the log as that may just result in more IO errors and
3778 * an endless shutdown/force loop.
3780 if (!log_error && !xlog_in_recovery(log))
3781 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3784 * Atomically set the shutdown state. If the shutdown state is already
3785 * set, there someone else is performing the shutdown and so we are done
3786 * here. This should never happen because we should only ever get called
3787 * once by the first shutdown caller.
3789 * Much of the log state machine transitions assume that shutdown state
3790 * cannot change once they hold the log->l_icloglock. Hence we need to
3791 * hold that lock here, even though we use the atomic test_and_set_bit()
3792 * operation to set the shutdown state.
3794 spin_lock(&log->l_icloglock);
3795 if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3796 spin_unlock(&log->l_icloglock);
3799 spin_unlock(&log->l_icloglock);
3802 * If this log shutdown also sets the mount shutdown state, issue a
3803 * shutdown warning message.
3805 if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &log->l_mp->m_opstate)) {
3806 xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
3807 "Filesystem has been shut down due to log error (0x%x).",
3809 xfs_alert(log->l_mp,
3810 "Please unmount the filesystem and rectify the problem(s).");
3811 if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
3816 * We don't want anybody waiting for log reservations after this. That
3817 * means we have to wake up everybody queued up on reserveq as well as
3818 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3819 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3820 * action is protected by the grant locks.
3822 xlog_grant_head_wake_all(&log->l_reserve_head);
3823 xlog_grant_head_wake_all(&log->l_write_head);
3826 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3827 * as if the log writes were completed. The abort handling in the log
3828 * item committed callback functions will do this again under lock to
3831 spin_lock(&log->l_cilp->xc_push_lock);
3832 wake_up_all(&log->l_cilp->xc_start_wait);
3833 wake_up_all(&log->l_cilp->xc_commit_wait);
3834 spin_unlock(&log->l_cilp->xc_push_lock);
3836 spin_lock(&log->l_icloglock);
3837 xlog_state_shutdown_callbacks(log);
3838 spin_unlock(&log->l_icloglock);
3840 wake_up_var(&log->l_opstate);
3848 xlog_in_core_t *iclog;
3850 iclog = log->l_iclog;
3852 /* endianness does not matter here, zero is zero in
3855 if (iclog->ic_header.h_num_logops)
3857 iclog = iclog->ic_next;
3858 } while (iclog != log->l_iclog);
3863 * Verify that an LSN stamped into a piece of metadata is valid. This is
3864 * intended for use in read verifiers on v5 superblocks.
3868 struct xfs_mount *mp,
3871 struct xlog *log = mp->m_log;
3875 * norecovery mode skips mount-time log processing and unconditionally
3876 * resets the in-core LSN. We can't validate in this mode, but
3877 * modifications are not allowed anyways so just return true.
3879 if (xfs_has_norecovery(mp))
3883 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3884 * handled by recovery and thus safe to ignore here.
3886 if (lsn == NULLCOMMITLSN)
3889 valid = xlog_valid_lsn(mp->m_log, lsn);
3891 /* warn the user about what's gone wrong before verifier failure */
3893 spin_lock(&log->l_icloglock);
3895 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3896 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3897 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3898 log->l_curr_cycle, log->l_curr_block);
3899 spin_unlock(&log->l_icloglock);
3906 * Notify the log that we're about to start using a feature that is protected
3907 * by a log incompat feature flag. This will prevent log covering from
3908 * clearing those flags.
3911 xlog_use_incompat_feat(
3914 down_read(&log->l_incompat_users);
3917 /* Notify the log that we've finished using log incompat features. */
3919 xlog_drop_incompat_feat(
3922 up_read(&log->l_incompat_users);