1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_errortag.h"
14 #include "xfs_error.h"
15 #include "xfs_trans.h"
16 #include "xfs_trans_priv.h"
18 #include "xfs_log_priv.h"
19 #include "xfs_trace.h"
20 #include "xfs_sysfs.h"
22 #include "xfs_health.h"
24 kmem_zone_t *xfs_log_ticket_zone;
26 /* Local miscellaneous function prototypes */
30 struct xfs_buftarg *log_target,
31 xfs_daddr_t blk_offset,
41 /* local state machine functions */
42 STATIC void xlog_state_done_syncing(
43 struct xlog_in_core *iclog);
45 xlog_state_get_iclog_space(
48 struct xlog_in_core **iclog,
49 struct xlog_ticket *ticket,
53 xlog_state_switch_iclogs(
55 struct xlog_in_core *iclog,
64 struct xlog_in_core *iclog);
71 xlog_verify_grant_tail(
76 struct xlog_in_core *iclog,
81 struct xlog_in_core *iclog);
83 #define xlog_verify_dest_ptr(a,b)
84 #define xlog_verify_grant_tail(a)
85 #define xlog_verify_iclog(a,b,c)
86 #define xlog_verify_tail_lsn(a,b)
94 xfs_log_cover(struct xfs_mount *);
102 int64_t head_val = atomic64_read(head);
108 xlog_crack_grant_head_val(head_val, &cycle, &space);
112 space += log->l_logsize;
117 new = xlog_assign_grant_head_val(cycle, space);
118 head_val = atomic64_cmpxchg(head, old, new);
119 } while (head_val != old);
123 xlog_grant_add_space(
128 int64_t head_val = atomic64_read(head);
135 xlog_crack_grant_head_val(head_val, &cycle, &space);
137 tmp = log->l_logsize - space;
146 new = xlog_assign_grant_head_val(cycle, space);
147 head_val = atomic64_cmpxchg(head, old, new);
148 } while (head_val != old);
152 xlog_grant_head_init(
153 struct xlog_grant_head *head)
155 xlog_assign_grant_head(&head->grant, 1, 0);
156 INIT_LIST_HEAD(&head->waiters);
157 spin_lock_init(&head->lock);
161 xlog_grant_head_wake_all(
162 struct xlog_grant_head *head)
164 struct xlog_ticket *tic;
166 spin_lock(&head->lock);
167 list_for_each_entry(tic, &head->waiters, t_queue)
168 wake_up_process(tic->t_task);
169 spin_unlock(&head->lock);
173 xlog_ticket_reservation(
175 struct xlog_grant_head *head,
176 struct xlog_ticket *tic)
178 if (head == &log->l_write_head) {
179 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
180 return tic->t_unit_res;
182 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
183 return tic->t_unit_res * tic->t_cnt;
185 return tic->t_unit_res;
190 xlog_grant_head_wake(
192 struct xlog_grant_head *head,
195 struct xlog_ticket *tic;
197 bool woken_task = false;
199 list_for_each_entry(tic, &head->waiters, t_queue) {
202 * There is a chance that the size of the CIL checkpoints in
203 * progress at the last AIL push target calculation resulted in
204 * limiting the target to the log head (l_last_sync_lsn) at the
205 * time. This may not reflect where the log head is now as the
206 * CIL checkpoints may have completed.
208 * Hence when we are woken here, it may be that the head of the
209 * log that has moved rather than the tail. As the tail didn't
210 * move, there still won't be space available for the
211 * reservation we require. However, if the AIL has already
212 * pushed to the target defined by the old log head location, we
213 * will hang here waiting for something else to update the AIL
216 * Therefore, if there isn't space to wake the first waiter on
217 * the grant head, we need to push the AIL again to ensure the
218 * target reflects both the current log tail and log head
219 * position before we wait for the tail to move again.
222 need_bytes = xlog_ticket_reservation(log, head, tic);
223 if (*free_bytes < need_bytes) {
225 xlog_grant_push_ail(log, need_bytes);
229 *free_bytes -= need_bytes;
230 trace_xfs_log_grant_wake_up(log, tic);
231 wake_up_process(tic->t_task);
239 xlog_grant_head_wait(
241 struct xlog_grant_head *head,
242 struct xlog_ticket *tic,
243 int need_bytes) __releases(&head->lock)
244 __acquires(&head->lock)
246 list_add_tail(&tic->t_queue, &head->waiters);
249 if (XLOG_FORCED_SHUTDOWN(log))
251 xlog_grant_push_ail(log, need_bytes);
253 __set_current_state(TASK_UNINTERRUPTIBLE);
254 spin_unlock(&head->lock);
256 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
258 trace_xfs_log_grant_sleep(log, tic);
260 trace_xfs_log_grant_wake(log, tic);
262 spin_lock(&head->lock);
263 if (XLOG_FORCED_SHUTDOWN(log))
265 } while (xlog_space_left(log, &head->grant) < need_bytes);
267 list_del_init(&tic->t_queue);
270 list_del_init(&tic->t_queue);
275 * Atomically get the log space required for a log ticket.
277 * Once a ticket gets put onto head->waiters, it will only return after the
278 * needed reservation is satisfied.
280 * This function is structured so that it has a lock free fast path. This is
281 * necessary because every new transaction reservation will come through this
282 * path. Hence any lock will be globally hot if we take it unconditionally on
285 * As tickets are only ever moved on and off head->waiters under head->lock, we
286 * only need to take that lock if we are going to add the ticket to the queue
287 * and sleep. We can avoid taking the lock if the ticket was never added to
288 * head->waiters because the t_queue list head will be empty and we hold the
289 * only reference to it so it can safely be checked unlocked.
292 xlog_grant_head_check(
294 struct xlog_grant_head *head,
295 struct xlog_ticket *tic,
301 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
304 * If there are other waiters on the queue then give them a chance at
305 * logspace before us. Wake up the first waiters, if we do not wake
306 * up all the waiters then go to sleep waiting for more free space,
307 * otherwise try to get some space for this transaction.
309 *need_bytes = xlog_ticket_reservation(log, head, tic);
310 free_bytes = xlog_space_left(log, &head->grant);
311 if (!list_empty_careful(&head->waiters)) {
312 spin_lock(&head->lock);
313 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
314 free_bytes < *need_bytes) {
315 error = xlog_grant_head_wait(log, head, tic,
318 spin_unlock(&head->lock);
319 } else if (free_bytes < *need_bytes) {
320 spin_lock(&head->lock);
321 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
322 spin_unlock(&head->lock);
329 xlog_tic_reset_res(xlog_ticket_t *tic)
332 tic->t_res_arr_sum = 0;
333 tic->t_res_num_ophdrs = 0;
337 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
339 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
340 /* add to overflow and start again */
341 tic->t_res_o_flow += tic->t_res_arr_sum;
343 tic->t_res_arr_sum = 0;
346 tic->t_res_arr[tic->t_res_num].r_len = len;
347 tic->t_res_arr[tic->t_res_num].r_type = type;
348 tic->t_res_arr_sum += len;
354 struct xfs_mount *mp)
357 * Do not write to the log on norecovery mounts, if the data or log
358 * devices are read-only, or if the filesystem is shutdown. Read-only
359 * mounts allow internal writes for log recovery and unmount purposes,
360 * so don't restrict that case.
362 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
364 if (xfs_readonly_buftarg(mp->m_ddev_targp))
366 if (xfs_readonly_buftarg(mp->m_log->l_targ))
368 if (XFS_FORCED_SHUTDOWN(mp))
374 * Replenish the byte reservation required by moving the grant write head.
378 struct xfs_mount *mp,
379 struct xlog_ticket *tic)
381 struct xlog *log = mp->m_log;
385 if (XLOG_FORCED_SHUTDOWN(log))
388 XFS_STATS_INC(mp, xs_try_logspace);
391 * This is a new transaction on the ticket, so we need to change the
392 * transaction ID so that the next transaction has a different TID in
393 * the log. Just add one to the existing tid so that we can see chains
394 * of rolling transactions in the log easily.
398 xlog_grant_push_ail(log, tic->t_unit_res);
400 tic->t_curr_res = tic->t_unit_res;
401 xlog_tic_reset_res(tic);
406 trace_xfs_log_regrant(log, tic);
408 error = xlog_grant_head_check(log, &log->l_write_head, tic,
413 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
414 trace_xfs_log_regrant_exit(log, tic);
415 xlog_verify_grant_tail(log);
420 * If we are failing, make sure the ticket doesn't have any current
421 * reservations. We don't want to add this back when the ticket/
422 * transaction gets cancelled.
425 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
430 * Reserve log space and return a ticket corresponding to the reservation.
432 * Each reservation is going to reserve extra space for a log record header.
433 * When writes happen to the on-disk log, we don't subtract the length of the
434 * log record header from any reservation. By wasting space in each
435 * reservation, we prevent over allocation problems.
439 struct xfs_mount *mp,
442 struct xlog_ticket **ticp,
446 struct xlog *log = mp->m_log;
447 struct xlog_ticket *tic;
451 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
453 if (XLOG_FORCED_SHUTDOWN(log))
456 XFS_STATS_INC(mp, xs_try_logspace);
458 ASSERT(*ticp == NULL);
459 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent);
462 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
465 trace_xfs_log_reserve(log, tic);
467 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
472 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
473 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
474 trace_xfs_log_reserve_exit(log, tic);
475 xlog_verify_grant_tail(log);
480 * If we are failing, make sure the ticket doesn't have any current
481 * reservations. We don't want to add this back when the ticket/
482 * transaction gets cancelled.
485 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
490 * Flush iclog to disk if this is the last reference to the given iclog and the
491 * it is in the WANT_SYNC state.
493 * If the caller passes in a non-zero @old_tail_lsn and the current log tail
494 * does not match, there may be metadata on disk that must be persisted before
495 * this iclog is written. To satisfy that requirement, set the
496 * XLOG_ICL_NEED_FLUSH flag as a condition for writing this iclog with the new
499 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
500 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
501 * written to stable storage, and implies that a commit record is contained
502 * within the iclog. We need to ensure that the log tail does not move beyond
503 * the tail that the first commit record in the iclog ordered against, otherwise
504 * correct recovery of that checkpoint becomes dependent on future operations
505 * performed on this iclog.
507 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
508 * current tail into iclog. Once the iclog tail is set, future operations must
509 * not modify it, otherwise they potentially violate ordering constraints for
510 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
511 * the iclog will get zeroed on activation of the iclog after sync, so we
512 * always capture the tail lsn on the iclog on the first NEED_FUA release
513 * regardless of the number of active reference counts on this iclog.
517 xlog_state_release_iclog(
519 struct xlog_in_core *iclog,
520 xfs_lsn_t old_tail_lsn)
523 lockdep_assert_held(&log->l_icloglock);
525 trace_xlog_iclog_release(iclog, _RET_IP_);
526 if (iclog->ic_state == XLOG_STATE_IOERROR)
530 * Grabbing the current log tail needs to be atomic w.r.t. the writing
531 * of the tail LSN into the iclog so we guarantee that the log tail does
532 * not move between deciding if a cache flush is required and writing
533 * the LSN into the iclog below.
535 if (old_tail_lsn || iclog->ic_state == XLOG_STATE_WANT_SYNC) {
536 tail_lsn = xlog_assign_tail_lsn(log->l_mp);
538 if (old_tail_lsn && tail_lsn != old_tail_lsn)
539 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
541 if ((iclog->ic_flags & XLOG_ICL_NEED_FUA) &&
542 !iclog->ic_header.h_tail_lsn)
543 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
546 if (!atomic_dec_and_test(&iclog->ic_refcnt))
549 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
550 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
554 iclog->ic_state = XLOG_STATE_SYNCING;
555 if (!iclog->ic_header.h_tail_lsn)
556 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
557 xlog_verify_tail_lsn(log, iclog);
558 trace_xlog_iclog_syncing(iclog, _RET_IP_);
560 spin_unlock(&log->l_icloglock);
561 xlog_sync(log, iclog);
562 spin_lock(&log->l_icloglock);
567 * Mount a log filesystem
569 * mp - ubiquitous xfs mount point structure
570 * log_target - buftarg of on-disk log device
571 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
572 * num_bblocks - Number of BBSIZE blocks in on-disk log
574 * Return error or zero.
579 xfs_buftarg_t *log_target,
580 xfs_daddr_t blk_offset,
583 bool fatal = xfs_sb_version_hascrc(&mp->m_sb);
587 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
588 xfs_notice(mp, "Mounting V%d Filesystem",
589 XFS_SB_VERSION_NUM(&mp->m_sb));
592 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
593 XFS_SB_VERSION_NUM(&mp->m_sb));
594 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
597 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
598 if (IS_ERR(mp->m_log)) {
599 error = PTR_ERR(mp->m_log);
604 * Validate the given log space and drop a critical message via syslog
605 * if the log size is too small that would lead to some unexpected
606 * situations in transaction log space reservation stage.
608 * Note: we can't just reject the mount if the validation fails. This
609 * would mean that people would have to downgrade their kernel just to
610 * remedy the situation as there is no way to grow the log (short of
611 * black magic surgery with xfs_db).
613 * We can, however, reject mounts for CRC format filesystems, as the
614 * mkfs binary being used to make the filesystem should never create a
615 * filesystem with a log that is too small.
617 min_logfsbs = xfs_log_calc_minimum_size(mp);
619 if (mp->m_sb.sb_logblocks < min_logfsbs) {
621 "Log size %d blocks too small, minimum size is %d blocks",
622 mp->m_sb.sb_logblocks, min_logfsbs);
624 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
626 "Log size %d blocks too large, maximum size is %lld blocks",
627 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
629 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
631 "log size %lld bytes too large, maximum size is %lld bytes",
632 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
635 } else if (mp->m_sb.sb_logsunit > 1 &&
636 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
638 "log stripe unit %u bytes must be a multiple of block size",
639 mp->m_sb.sb_logsunit);
645 * Log check errors are always fatal on v5; or whenever bad
646 * metadata leads to a crash.
649 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
653 xfs_crit(mp, "Log size out of supported range.");
655 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
659 * Initialize the AIL now we have a log.
661 error = xfs_trans_ail_init(mp);
663 xfs_warn(mp, "AIL initialisation failed: error %d", error);
666 mp->m_log->l_ailp = mp->m_ail;
669 * skip log recovery on a norecovery mount. pretend it all
672 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
673 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
676 mp->m_flags &= ~XFS_MOUNT_RDONLY;
678 error = xlog_recover(mp->m_log);
681 mp->m_flags |= XFS_MOUNT_RDONLY;
683 xfs_warn(mp, "log mount/recovery failed: error %d",
685 xlog_recover_cancel(mp->m_log);
686 goto out_destroy_ail;
690 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
693 goto out_destroy_ail;
695 /* Normal transactions can now occur */
696 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
699 * Now the log has been fully initialised and we know were our
700 * space grant counters are, we can initialise the permanent ticket
701 * needed for delayed logging to work.
703 xlog_cil_init_post_recovery(mp->m_log);
708 xfs_trans_ail_destroy(mp);
710 xlog_dealloc_log(mp->m_log);
716 * Finish the recovery of the file system. This is separate from the
717 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
718 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
721 * If we finish recovery successfully, start the background log work. If we are
722 * not doing recovery, then we have a RO filesystem and we don't need to start
726 xfs_log_mount_finish(
727 struct xfs_mount *mp)
730 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
731 bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
733 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
734 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
736 } else if (readonly) {
737 /* Allow unlinked processing to proceed */
738 mp->m_flags &= ~XFS_MOUNT_RDONLY;
742 * During the second phase of log recovery, we need iget and
743 * iput to behave like they do for an active filesystem.
744 * xfs_fs_drop_inode needs to be able to prevent the deletion
745 * of inodes before we're done replaying log items on those
746 * inodes. Turn it off immediately after recovery finishes
747 * so that we don't leak the quota inodes if subsequent mount
750 * We let all inodes involved in redo item processing end up on
751 * the LRU instead of being evicted immediately so that if we do
752 * something to an unlinked inode, the irele won't cause
753 * premature truncation and freeing of the inode, which results
754 * in log recovery failure. We have to evict the unreferenced
755 * lru inodes after clearing SB_ACTIVE because we don't
756 * otherwise clean up the lru if there's a subsequent failure in
757 * xfs_mountfs, which leads to us leaking the inodes if nothing
758 * else (e.g. quotacheck) references the inodes before the
759 * mount failure occurs.
761 mp->m_super->s_flags |= SB_ACTIVE;
762 error = xlog_recover_finish(mp->m_log);
764 xfs_log_work_queue(mp);
765 mp->m_super->s_flags &= ~SB_ACTIVE;
766 evict_inodes(mp->m_super);
769 * Drain the buffer LRU after log recovery. This is required for v4
770 * filesystems to avoid leaving around buffers with NULL verifier ops,
771 * but we do it unconditionally to make sure we're always in a clean
772 * cache state after mount.
774 * Don't push in the error case because the AIL may have pending intents
775 * that aren't removed until recovery is cancelled.
777 if (!error && recovered) {
778 xfs_log_force(mp, XFS_LOG_SYNC);
779 xfs_ail_push_all_sync(mp->m_ail);
781 xfs_buftarg_drain(mp->m_ddev_targp);
784 mp->m_flags |= XFS_MOUNT_RDONLY;
786 /* Make sure the log is dead if we're returning failure. */
787 ASSERT(!error || (mp->m_log->l_flags & XLOG_IO_ERROR));
793 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
797 xfs_log_mount_cancel(
798 struct xfs_mount *mp)
800 xlog_recover_cancel(mp->m_log);
805 * Flush out the iclog to disk ensuring that device caches are flushed and
806 * the iclog hits stable storage before any completion waiters are woken.
810 struct xlog_in_core *iclog)
812 atomic_inc(&iclog->ic_refcnt);
813 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
814 if (iclog->ic_state == XLOG_STATE_ACTIVE)
815 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
816 return xlog_state_release_iclog(iclog->ic_log, iclog, 0);
820 * Wait for the iclog and all prior iclogs to be written disk as required by the
821 * log force state machine. Waiting on ic_force_wait ensures iclog completions
822 * have been ordered and callbacks run before we are woken here, hence
823 * guaranteeing that all the iclogs up to this one are on stable storage.
827 struct xlog_in_core *iclog)
828 __releases(iclog->ic_log->l_icloglock)
830 struct xlog *log = iclog->ic_log;
832 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
833 if (!XLOG_FORCED_SHUTDOWN(log) &&
834 iclog->ic_state != XLOG_STATE_ACTIVE &&
835 iclog->ic_state != XLOG_STATE_DIRTY) {
836 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
837 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
839 spin_unlock(&log->l_icloglock);
842 if (XLOG_FORCED_SHUTDOWN(log))
848 * Write out an unmount record using the ticket provided. We have to account for
849 * the data space used in the unmount ticket as this write is not done from a
850 * transaction context that has already done the accounting for us.
853 xlog_write_unmount_record(
855 struct xlog_ticket *ticket)
857 struct xfs_unmount_log_format ulf = {
858 .magic = XLOG_UNMOUNT_TYPE,
860 struct xfs_log_iovec reg = {
862 .i_len = sizeof(ulf),
863 .i_type = XLOG_REG_TYPE_UNMOUNT,
865 struct xfs_log_vec vec = {
870 /* account for space used by record data */
871 ticket->t_curr_res -= sizeof(ulf);
873 return xlog_write(log, &vec, ticket, NULL, NULL, XLOG_UNMOUNT_TRANS);
877 * Mark the filesystem clean by writing an unmount record to the head of the
884 struct xfs_mount *mp = log->l_mp;
885 struct xlog_in_core *iclog;
886 struct xlog_ticket *tic = NULL;
889 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
893 error = xlog_write_unmount_record(log, tic);
895 * At this point, we're umounting anyway, so there's no point in
896 * transitioning log state to IOERROR. Just continue...
900 xfs_alert(mp, "%s: unmount record failed", __func__);
902 spin_lock(&log->l_icloglock);
903 iclog = log->l_iclog;
904 error = xlog_force_iclog(iclog);
905 xlog_wait_on_iclog(iclog);
908 trace_xfs_log_umount_write(log, tic);
909 xfs_log_ticket_ungrant(log, tic);
914 xfs_log_unmount_verify_iclog(
917 struct xlog_in_core *iclog = log->l_iclog;
920 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
921 ASSERT(iclog->ic_offset == 0);
922 } while ((iclog = iclog->ic_next) != log->l_iclog);
926 * Unmount record used to have a string "Unmount filesystem--" in the
927 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
928 * We just write the magic number now since that particular field isn't
929 * currently architecture converted and "Unmount" is a bit foo.
930 * As far as I know, there weren't any dependencies on the old behaviour.
933 xfs_log_unmount_write(
934 struct xfs_mount *mp)
936 struct xlog *log = mp->m_log;
938 if (!xfs_log_writable(mp))
941 xfs_log_force(mp, XFS_LOG_SYNC);
943 if (XLOG_FORCED_SHUTDOWN(log))
947 * If we think the summary counters are bad, avoid writing the unmount
948 * record to force log recovery at next mount, after which the summary
949 * counters will be recalculated. Refer to xlog_check_unmount_rec for
952 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
953 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
954 xfs_alert(mp, "%s: will fix summary counters at next mount",
959 xfs_log_unmount_verify_iclog(log);
960 xlog_unmount_write(log);
964 * Empty the log for unmount/freeze.
966 * To do this, we first need to shut down the background log work so it is not
967 * trying to cover the log as we clean up. We then need to unpin all objects in
968 * the log so we can then flush them out. Once they have completed their IO and
969 * run the callbacks removing themselves from the AIL, we can cover the log.
973 struct xfs_mount *mp)
975 cancel_delayed_work_sync(&mp->m_log->l_work);
976 xfs_log_force(mp, XFS_LOG_SYNC);
979 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
980 * will push it, xfs_buftarg_wait() will not wait for it. Further,
981 * xfs_buf_iowait() cannot be used because it was pushed with the
982 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
983 * the IO to complete.
985 xfs_ail_push_all_sync(mp->m_ail);
986 xfs_buftarg_wait(mp->m_ddev_targp);
987 xfs_buf_lock(mp->m_sb_bp);
988 xfs_buf_unlock(mp->m_sb_bp);
990 return xfs_log_cover(mp);
995 struct xfs_mount *mp)
998 xfs_log_unmount_write(mp);
1002 * Shut down and release the AIL and Log.
1004 * During unmount, we need to ensure we flush all the dirty metadata objects
1005 * from the AIL so that the log is empty before we write the unmount record to
1006 * the log. Once this is done, we can tear down the AIL and the log.
1010 struct xfs_mount *mp)
1014 xfs_buftarg_drain(mp->m_ddev_targp);
1016 xfs_trans_ail_destroy(mp);
1018 xfs_sysfs_del(&mp->m_log->l_kobj);
1020 xlog_dealloc_log(mp->m_log);
1025 struct xfs_mount *mp,
1026 struct xfs_log_item *item,
1028 const struct xfs_item_ops *ops)
1030 item->li_mountp = mp;
1031 item->li_ailp = mp->m_ail;
1032 item->li_type = type;
1036 INIT_LIST_HEAD(&item->li_ail);
1037 INIT_LIST_HEAD(&item->li_cil);
1038 INIT_LIST_HEAD(&item->li_bio_list);
1039 INIT_LIST_HEAD(&item->li_trans);
1043 * Wake up processes waiting for log space after we have moved the log tail.
1047 struct xfs_mount *mp)
1049 struct xlog *log = mp->m_log;
1052 if (XLOG_FORCED_SHUTDOWN(log))
1055 if (!list_empty_careful(&log->l_write_head.waiters)) {
1056 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1058 spin_lock(&log->l_write_head.lock);
1059 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1060 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1061 spin_unlock(&log->l_write_head.lock);
1064 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1065 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1067 spin_lock(&log->l_reserve_head.lock);
1068 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1069 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1070 spin_unlock(&log->l_reserve_head.lock);
1075 * Determine if we have a transaction that has gone to disk that needs to be
1076 * covered. To begin the transition to the idle state firstly the log needs to
1077 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1078 * we start attempting to cover the log.
1080 * Only if we are then in a state where covering is needed, the caller is
1081 * informed that dummy transactions are required to move the log into the idle
1084 * If there are any items in the AIl or CIL, then we do not want to attempt to
1085 * cover the log as we may be in a situation where there isn't log space
1086 * available to run a dummy transaction and this can lead to deadlocks when the
1087 * tail of the log is pinned by an item that is modified in the CIL. Hence
1088 * there's no point in running a dummy transaction at this point because we
1089 * can't start trying to idle the log until both the CIL and AIL are empty.
1092 xfs_log_need_covered(
1093 struct xfs_mount *mp)
1095 struct xlog *log = mp->m_log;
1096 bool needed = false;
1098 if (!xlog_cil_empty(log))
1101 spin_lock(&log->l_icloglock);
1102 switch (log->l_covered_state) {
1103 case XLOG_STATE_COVER_DONE:
1104 case XLOG_STATE_COVER_DONE2:
1105 case XLOG_STATE_COVER_IDLE:
1107 case XLOG_STATE_COVER_NEED:
1108 case XLOG_STATE_COVER_NEED2:
1109 if (xfs_ail_min_lsn(log->l_ailp))
1111 if (!xlog_iclogs_empty(log))
1115 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1116 log->l_covered_state = XLOG_STATE_COVER_DONE;
1118 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1124 spin_unlock(&log->l_icloglock);
1129 * Explicitly cover the log. This is similar to background log covering but
1130 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1131 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1132 * must all be empty.
1136 struct xfs_mount *mp)
1141 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1142 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1143 XFS_FORCED_SHUTDOWN(mp));
1145 if (!xfs_log_writable(mp))
1149 * xfs_log_need_covered() is not idempotent because it progresses the
1150 * state machine if the log requires covering. Therefore, we must call
1151 * this function once and use the result until we've issued an sb sync.
1152 * Do so first to make that abundantly clear.
1154 * Fall into the covering sequence if the log needs covering or the
1155 * mount has lazy superblock accounting to sync to disk. The sb sync
1156 * used for covering accumulates the in-core counters, so covering
1157 * handles this for us.
1159 need_covered = xfs_log_need_covered(mp);
1160 if (!need_covered && !xfs_sb_version_haslazysbcount(&mp->m_sb))
1164 * To cover the log, commit the superblock twice (at most) in
1165 * independent checkpoints. The first serves as a reference for the
1166 * tail pointer. The sync transaction and AIL push empties the AIL and
1167 * updates the in-core tail to the LSN of the first checkpoint. The
1168 * second commit updates the on-disk tail with the in-core LSN,
1169 * covering the log. Push the AIL one more time to leave it empty, as
1173 error = xfs_sync_sb(mp, true);
1176 xfs_ail_push_all_sync(mp->m_ail);
1177 } while (xfs_log_need_covered(mp));
1183 * We may be holding the log iclog lock upon entering this routine.
1186 xlog_assign_tail_lsn_locked(
1187 struct xfs_mount *mp)
1189 struct xlog *log = mp->m_log;
1190 struct xfs_log_item *lip;
1193 assert_spin_locked(&mp->m_ail->ail_lock);
1196 * To make sure we always have a valid LSN for the log tail we keep
1197 * track of the last LSN which was committed in log->l_last_sync_lsn,
1198 * and use that when the AIL was empty.
1200 lip = xfs_ail_min(mp->m_ail);
1202 tail_lsn = lip->li_lsn;
1204 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1205 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1206 atomic64_set(&log->l_tail_lsn, tail_lsn);
1211 xlog_assign_tail_lsn(
1212 struct xfs_mount *mp)
1216 spin_lock(&mp->m_ail->ail_lock);
1217 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1218 spin_unlock(&mp->m_ail->ail_lock);
1224 * Return the space in the log between the tail and the head. The head
1225 * is passed in the cycle/bytes formal parms. In the special case where
1226 * the reserve head has wrapped passed the tail, this calculation is no
1227 * longer valid. In this case, just return 0 which means there is no space
1228 * in the log. This works for all places where this function is called
1229 * with the reserve head. Of course, if the write head were to ever
1230 * wrap the tail, we should blow up. Rather than catch this case here,
1231 * we depend on other ASSERTions in other parts of the code. XXXmiken
1233 * This code also handles the case where the reservation head is behind
1234 * the tail. The details of this case are described below, but the end
1235 * result is that we return the size of the log as the amount of space left.
1248 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1249 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1250 tail_bytes = BBTOB(tail_bytes);
1251 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1252 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1253 else if (tail_cycle + 1 < head_cycle)
1255 else if (tail_cycle < head_cycle) {
1256 ASSERT(tail_cycle == (head_cycle - 1));
1257 free_bytes = tail_bytes - head_bytes;
1260 * The reservation head is behind the tail.
1261 * In this case we just want to return the size of the
1262 * log as the amount of space left.
1264 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1265 xfs_alert(log->l_mp,
1266 " tail_cycle = %d, tail_bytes = %d",
1267 tail_cycle, tail_bytes);
1268 xfs_alert(log->l_mp,
1269 " GH cycle = %d, GH bytes = %d",
1270 head_cycle, head_bytes);
1272 free_bytes = log->l_logsize;
1280 struct work_struct *work)
1282 struct xlog_in_core *iclog =
1283 container_of(work, struct xlog_in_core, ic_end_io_work);
1284 struct xlog *log = iclog->ic_log;
1287 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1289 /* treat writes with injected CRC errors as failed */
1290 if (iclog->ic_fail_crc)
1295 * Race to shutdown the filesystem if we see an error.
1297 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1298 xfs_alert(log->l_mp, "log I/O error %d", error);
1299 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1302 xlog_state_done_syncing(iclog);
1303 bio_uninit(&iclog->ic_bio);
1306 * Drop the lock to signal that we are done. Nothing references the
1307 * iclog after this, so an unmount waiting on this lock can now tear it
1308 * down safely. As such, it is unsafe to reference the iclog after the
1309 * unlock as we could race with it being freed.
1311 up(&iclog->ic_sema);
1315 * Return size of each in-core log record buffer.
1317 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1319 * If the filesystem blocksize is too large, we may need to choose a
1320 * larger size since the directory code currently logs entire blocks.
1323 xlog_get_iclog_buffer_size(
1324 struct xfs_mount *mp,
1327 if (mp->m_logbufs <= 0)
1328 mp->m_logbufs = XLOG_MAX_ICLOGS;
1329 if (mp->m_logbsize <= 0)
1330 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1332 log->l_iclog_bufs = mp->m_logbufs;
1333 log->l_iclog_size = mp->m_logbsize;
1336 * # headers = size / 32k - one header holds cycles from 32k of data.
1338 log->l_iclog_heads =
1339 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1340 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1345 struct xfs_mount *mp)
1347 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1348 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1352 * Every sync period we need to unpin all items in the AIL and push them to
1353 * disk. If there is nothing dirty, then we might need to cover the log to
1354 * indicate that the filesystem is idle.
1358 struct work_struct *work)
1360 struct xlog *log = container_of(to_delayed_work(work),
1361 struct xlog, l_work);
1362 struct xfs_mount *mp = log->l_mp;
1364 /* dgc: errors ignored - not fatal and nowhere to report them */
1365 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1367 * Dump a transaction into the log that contains no real change.
1368 * This is needed to stamp the current tail LSN into the log
1369 * during the covering operation.
1371 * We cannot use an inode here for this - that will push dirty
1372 * state back up into the VFS and then periodic inode flushing
1373 * will prevent log covering from making progress. Hence we
1374 * synchronously log the superblock instead to ensure the
1375 * superblock is immediately unpinned and can be written back.
1377 xfs_sync_sb(mp, true);
1379 xfs_log_force(mp, 0);
1381 /* start pushing all the metadata that is currently dirty */
1382 xfs_ail_push_all(mp->m_ail);
1384 /* queue us up again */
1385 xfs_log_work_queue(mp);
1389 * This routine initializes some of the log structure for a given mount point.
1390 * Its primary purpose is to fill in enough, so recovery can occur. However,
1391 * some other stuff may be filled in too.
1393 STATIC struct xlog *
1395 struct xfs_mount *mp,
1396 struct xfs_buftarg *log_target,
1397 xfs_daddr_t blk_offset,
1401 xlog_rec_header_t *head;
1402 xlog_in_core_t **iclogp;
1403 xlog_in_core_t *iclog, *prev_iclog=NULL;
1405 int error = -ENOMEM;
1408 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1410 xfs_warn(mp, "Log allocation failed: No memory!");
1415 log->l_targ = log_target;
1416 log->l_logsize = BBTOB(num_bblks);
1417 log->l_logBBstart = blk_offset;
1418 log->l_logBBsize = num_bblks;
1419 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1420 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1421 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1423 log->l_prev_block = -1;
1424 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1425 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1426 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1427 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1429 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1)
1430 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1432 log->l_iclog_roundoff = BBSIZE;
1434 xlog_grant_head_init(&log->l_reserve_head);
1435 xlog_grant_head_init(&log->l_write_head);
1437 error = -EFSCORRUPTED;
1438 if (xfs_sb_version_hassector(&mp->m_sb)) {
1439 log2_size = mp->m_sb.sb_logsectlog;
1440 if (log2_size < BBSHIFT) {
1441 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1442 log2_size, BBSHIFT);
1446 log2_size -= BBSHIFT;
1447 if (log2_size > mp->m_sectbb_log) {
1448 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1449 log2_size, mp->m_sectbb_log);
1453 /* for larger sector sizes, must have v2 or external log */
1454 if (log2_size && log->l_logBBstart > 0 &&
1455 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1457 "log sector size (0x%x) invalid for configuration.",
1462 log->l_sectBBsize = 1 << log2_size;
1464 xlog_get_iclog_buffer_size(mp, log);
1466 spin_lock_init(&log->l_icloglock);
1467 init_waitqueue_head(&log->l_flush_wait);
1469 iclogp = &log->l_iclog;
1471 * The amount of memory to allocate for the iclog structure is
1472 * rather funky due to the way the structure is defined. It is
1473 * done this way so that we can use different sizes for machines
1474 * with different amounts of memory. See the definition of
1475 * xlog_in_core_t in xfs_log_priv.h for details.
1477 ASSERT(log->l_iclog_size >= 4096);
1478 for (i = 0; i < log->l_iclog_bufs; i++) {
1479 int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1480 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1481 sizeof(struct bio_vec);
1483 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1485 goto out_free_iclog;
1488 iclog->ic_prev = prev_iclog;
1491 iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1492 KM_MAYFAIL | KM_ZERO);
1493 if (!iclog->ic_data)
1494 goto out_free_iclog;
1496 log->l_iclog_bak[i] = &iclog->ic_header;
1498 head = &iclog->ic_header;
1499 memset(head, 0, sizeof(xlog_rec_header_t));
1500 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1501 head->h_version = cpu_to_be32(
1502 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1503 head->h_size = cpu_to_be32(log->l_iclog_size);
1505 head->h_fmt = cpu_to_be32(XLOG_FMT);
1506 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1508 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1509 iclog->ic_state = XLOG_STATE_ACTIVE;
1510 iclog->ic_log = log;
1511 atomic_set(&iclog->ic_refcnt, 0);
1512 INIT_LIST_HEAD(&iclog->ic_callbacks);
1513 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1515 init_waitqueue_head(&iclog->ic_force_wait);
1516 init_waitqueue_head(&iclog->ic_write_wait);
1517 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1518 sema_init(&iclog->ic_sema, 1);
1520 iclogp = &iclog->ic_next;
1522 *iclogp = log->l_iclog; /* complete ring */
1523 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1525 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1526 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1528 0, mp->m_super->s_id);
1529 if (!log->l_ioend_workqueue)
1530 goto out_free_iclog;
1532 error = xlog_cil_init(log);
1534 goto out_destroy_workqueue;
1537 out_destroy_workqueue:
1538 destroy_workqueue(log->l_ioend_workqueue);
1540 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1541 prev_iclog = iclog->ic_next;
1542 kmem_free(iclog->ic_data);
1544 if (prev_iclog == log->l_iclog)
1550 return ERR_PTR(error);
1551 } /* xlog_alloc_log */
1554 * Write out the commit record of a transaction associated with the given
1555 * ticket to close off a running log write. Return the lsn of the commit record.
1560 struct xlog_ticket *ticket,
1561 struct xlog_in_core **iclog,
1564 struct xfs_log_iovec reg = {
1567 .i_type = XLOG_REG_TYPE_COMMIT,
1569 struct xfs_log_vec vec = {
1575 if (XLOG_FORCED_SHUTDOWN(log))
1578 error = xlog_write(log, &vec, ticket, lsn, iclog, XLOG_COMMIT_TRANS);
1580 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1585 * Compute the LSN that we'd need to push the log tail towards in order to have
1586 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1587 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1588 * log free space already meets all three thresholds, this function returns
1592 xlog_grant_push_threshold(
1596 xfs_lsn_t threshold_lsn = 0;
1597 xfs_lsn_t last_sync_lsn;
1600 int threshold_block;
1601 int threshold_cycle;
1604 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1606 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1607 free_blocks = BTOBBT(free_bytes);
1610 * Set the threshold for the minimum number of free blocks in the
1611 * log to the maximum of what the caller needs, one quarter of the
1612 * log, and 256 blocks.
1614 free_threshold = BTOBB(need_bytes);
1615 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1616 free_threshold = max(free_threshold, 256);
1617 if (free_blocks >= free_threshold)
1618 return NULLCOMMITLSN;
1620 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1622 threshold_block += free_threshold;
1623 if (threshold_block >= log->l_logBBsize) {
1624 threshold_block -= log->l_logBBsize;
1625 threshold_cycle += 1;
1627 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1630 * Don't pass in an lsn greater than the lsn of the last
1631 * log record known to be on disk. Use a snapshot of the last sync lsn
1632 * so that it doesn't change between the compare and the set.
1634 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1635 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1636 threshold_lsn = last_sync_lsn;
1638 return threshold_lsn;
1642 * Push the tail of the log if we need to do so to maintain the free log space
1643 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1644 * policy which pushes on an lsn which is further along in the log once we
1645 * reach the high water mark. In this manner, we would be creating a low water
1649 xlog_grant_push_ail(
1653 xfs_lsn_t threshold_lsn;
1655 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1656 if (threshold_lsn == NULLCOMMITLSN || XLOG_FORCED_SHUTDOWN(log))
1660 * Get the transaction layer to kick the dirty buffers out to
1661 * disk asynchronously. No point in trying to do this if
1662 * the filesystem is shutting down.
1664 xfs_ail_push(log->l_ailp, threshold_lsn);
1668 * Stamp cycle number in every block
1673 struct xlog_in_core *iclog,
1677 int size = iclog->ic_offset + roundoff;
1681 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1683 dp = iclog->ic_datap;
1684 for (i = 0; i < BTOBB(size); i++) {
1685 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1687 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1688 *(__be32 *)dp = cycle_lsn;
1692 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1693 xlog_in_core_2_t *xhdr = iclog->ic_data;
1695 for ( ; i < BTOBB(size); i++) {
1696 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1697 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1698 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1699 *(__be32 *)dp = cycle_lsn;
1703 for (i = 1; i < log->l_iclog_heads; i++)
1704 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1709 * Calculate the checksum for a log buffer.
1711 * This is a little more complicated than it should be because the various
1712 * headers and the actual data are non-contiguous.
1717 struct xlog_rec_header *rhead,
1723 /* first generate the crc for the record header ... */
1724 crc = xfs_start_cksum_update((char *)rhead,
1725 sizeof(struct xlog_rec_header),
1726 offsetof(struct xlog_rec_header, h_crc));
1728 /* ... then for additional cycle data for v2 logs ... */
1729 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1730 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1734 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1736 for (i = 1; i < xheads; i++) {
1737 crc = crc32c(crc, &xhdr[i].hic_xheader,
1738 sizeof(struct xlog_rec_ext_header));
1742 /* ... and finally for the payload */
1743 crc = crc32c(crc, dp, size);
1745 return xfs_end_cksum(crc);
1752 struct xlog_in_core *iclog = bio->bi_private;
1754 queue_work(iclog->ic_log->l_ioend_workqueue,
1755 &iclog->ic_end_io_work);
1759 xlog_map_iclog_data(
1765 struct page *page = kmem_to_page(data);
1766 unsigned int off = offset_in_page(data);
1767 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1769 if (bio_add_page(bio, page, len, off) != len)
1782 struct xlog_in_core *iclog,
1786 ASSERT(bno < log->l_logBBsize);
1787 trace_xlog_iclog_write(iclog, _RET_IP_);
1790 * We lock the iclogbufs here so that we can serialise against I/O
1791 * completion during unmount. We might be processing a shutdown
1792 * triggered during unmount, and that can occur asynchronously to the
1793 * unmount thread, and hence we need to ensure that completes before
1794 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1795 * across the log IO to archieve that.
1797 down(&iclog->ic_sema);
1798 if (unlikely(iclog->ic_state == XLOG_STATE_IOERROR)) {
1800 * It would seem logical to return EIO here, but we rely on
1801 * the log state machine to propagate I/O errors instead of
1802 * doing it here. We kick of the state machine and unlock
1803 * the buffer manually, the code needs to be kept in sync
1804 * with the I/O completion path.
1806 xlog_state_done_syncing(iclog);
1807 up(&iclog->ic_sema);
1811 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1812 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1813 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1814 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1815 iclog->ic_bio.bi_private = iclog;
1818 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1819 * IOs coming immediately after this one. This prevents the block layer
1820 * writeback throttle from throttling log writes behind background
1821 * metadata writeback and causing priority inversions.
1823 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE;
1824 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1825 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1827 * For external log devices, we also need to flush the data
1828 * device cache first to ensure all metadata writeback covered
1829 * by the LSN in this iclog is on stable storage. This is slow,
1830 * but it *must* complete before we issue the external log IO.
1832 if (log->l_targ != log->l_mp->m_ddev_targp)
1833 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev);
1835 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1836 iclog->ic_bio.bi_opf |= REQ_FUA;
1838 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1840 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1841 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1844 if (is_vmalloc_addr(iclog->ic_data))
1845 flush_kernel_vmap_range(iclog->ic_data, count);
1848 * If this log buffer would straddle the end of the log we will have
1849 * to split it up into two bios, so that we can continue at the start.
1851 if (bno + BTOBB(count) > log->l_logBBsize) {
1854 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1855 GFP_NOIO, &fs_bio_set);
1856 bio_chain(split, &iclog->ic_bio);
1859 /* restart at logical offset zero for the remainder */
1860 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1863 submit_bio(&iclog->ic_bio);
1867 * We need to bump cycle number for the part of the iclog that is
1868 * written to the start of the log. Watch out for the header magic
1869 * number case, though.
1878 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1881 for (i = split_offset; i < count; i += BBSIZE) {
1882 uint32_t cycle = get_unaligned_be32(data + i);
1884 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1886 put_unaligned_be32(cycle, data + i);
1891 xlog_calc_iclog_size(
1893 struct xlog_in_core *iclog,
1896 uint32_t count_init, count;
1898 /* Add for LR header */
1899 count_init = log->l_iclog_hsize + iclog->ic_offset;
1900 count = roundup(count_init, log->l_iclog_roundoff);
1902 *roundoff = count - count_init;
1904 ASSERT(count >= count_init);
1905 ASSERT(*roundoff < log->l_iclog_roundoff);
1910 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1911 * fashion. Previously, we should have moved the current iclog
1912 * ptr in the log to point to the next available iclog. This allows further
1913 * write to continue while this code syncs out an iclog ready to go.
1914 * Before an in-core log can be written out, the data section must be scanned
1915 * to save away the 1st word of each BBSIZE block into the header. We replace
1916 * it with the current cycle count. Each BBSIZE block is tagged with the
1917 * cycle count because there in an implicit assumption that drives will
1918 * guarantee that entire 512 byte blocks get written at once. In other words,
1919 * we can't have part of a 512 byte block written and part not written. By
1920 * tagging each block, we will know which blocks are valid when recovering
1921 * after an unclean shutdown.
1923 * This routine is single threaded on the iclog. No other thread can be in
1924 * this routine with the same iclog. Changing contents of iclog can there-
1925 * fore be done without grabbing the state machine lock. Updating the global
1926 * log will require grabbing the lock though.
1928 * The entire log manager uses a logical block numbering scheme. Only
1929 * xlog_write_iclog knows about the fact that the log may not start with
1930 * block zero on a given device.
1935 struct xlog_in_core *iclog)
1937 unsigned int count; /* byte count of bwrite */
1938 unsigned int roundoff; /* roundoff to BB or stripe */
1942 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1943 trace_xlog_iclog_sync(iclog, _RET_IP_);
1945 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1947 /* move grant heads by roundoff in sync */
1948 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1949 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1951 /* put cycle number in every block */
1952 xlog_pack_data(log, iclog, roundoff);
1954 /* real byte length */
1955 size = iclog->ic_offset;
1956 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1958 iclog->ic_header.h_len = cpu_to_be32(size);
1960 XFS_STATS_INC(log->l_mp, xs_log_writes);
1961 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1963 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1965 /* Do we need to split this write into 2 parts? */
1966 if (bno + BTOBB(count) > log->l_logBBsize)
1967 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1969 /* calculcate the checksum */
1970 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1971 iclog->ic_datap, size);
1973 * Intentionally corrupt the log record CRC based on the error injection
1974 * frequency, if defined. This facilitates testing log recovery in the
1975 * event of torn writes. Hence, set the IOABORT state to abort the log
1976 * write on I/O completion and shutdown the fs. The subsequent mount
1977 * detects the bad CRC and attempts to recover.
1980 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1981 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1982 iclog->ic_fail_crc = true;
1984 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1985 be64_to_cpu(iclog->ic_header.h_lsn));
1988 xlog_verify_iclog(log, iclog, count);
1989 xlog_write_iclog(log, iclog, bno, count);
1993 * Deallocate a log structure
1999 xlog_in_core_t *iclog, *next_iclog;
2002 xlog_cil_destroy(log);
2005 * Cycle all the iclogbuf locks to make sure all log IO completion
2006 * is done before we tear down these buffers.
2008 iclog = log->l_iclog;
2009 for (i = 0; i < log->l_iclog_bufs; i++) {
2010 down(&iclog->ic_sema);
2011 up(&iclog->ic_sema);
2012 iclog = iclog->ic_next;
2015 iclog = log->l_iclog;
2016 for (i = 0; i < log->l_iclog_bufs; i++) {
2017 next_iclog = iclog->ic_next;
2018 kmem_free(iclog->ic_data);
2023 log->l_mp->m_log = NULL;
2024 destroy_workqueue(log->l_ioend_workqueue);
2029 * Update counters atomically now that memcpy is done.
2032 xlog_state_finish_copy(
2034 struct xlog_in_core *iclog,
2038 lockdep_assert_held(&log->l_icloglock);
2040 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2041 iclog->ic_offset += copy_bytes;
2045 * print out info relating to regions written which consume
2050 struct xfs_mount *mp,
2051 struct xlog_ticket *ticket)
2054 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2056 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2057 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2058 static char *res_type_str[] = {
2059 REG_TYPE_STR(BFORMAT, "bformat"),
2060 REG_TYPE_STR(BCHUNK, "bchunk"),
2061 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2062 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2063 REG_TYPE_STR(IFORMAT, "iformat"),
2064 REG_TYPE_STR(ICORE, "icore"),
2065 REG_TYPE_STR(IEXT, "iext"),
2066 REG_TYPE_STR(IBROOT, "ibroot"),
2067 REG_TYPE_STR(ILOCAL, "ilocal"),
2068 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2069 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2070 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2071 REG_TYPE_STR(QFORMAT, "qformat"),
2072 REG_TYPE_STR(DQUOT, "dquot"),
2073 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2074 REG_TYPE_STR(LRHEADER, "LR header"),
2075 REG_TYPE_STR(UNMOUNT, "unmount"),
2076 REG_TYPE_STR(COMMIT, "commit"),
2077 REG_TYPE_STR(TRANSHDR, "trans header"),
2078 REG_TYPE_STR(ICREATE, "inode create"),
2079 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2080 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2081 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2082 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2083 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2084 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2086 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2089 xfs_warn(mp, "ticket reservation summary:");
2090 xfs_warn(mp, " unit res = %d bytes",
2091 ticket->t_unit_res);
2092 xfs_warn(mp, " current res = %d bytes",
2093 ticket->t_curr_res);
2094 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2095 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2096 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2097 ticket->t_res_num_ophdrs, ophdr_spc);
2098 xfs_warn(mp, " ophdr + reg = %u bytes",
2099 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2100 xfs_warn(mp, " num regions = %u",
2103 for (i = 0; i < ticket->t_res_num; i++) {
2104 uint r_type = ticket->t_res_arr[i].r_type;
2105 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2106 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2107 "bad-rtype" : res_type_str[r_type]),
2108 ticket->t_res_arr[i].r_len);
2113 * Print a summary of the transaction.
2117 struct xfs_trans *tp)
2119 struct xfs_mount *mp = tp->t_mountp;
2120 struct xfs_log_item *lip;
2122 /* dump core transaction and ticket info */
2123 xfs_warn(mp, "transaction summary:");
2124 xfs_warn(mp, " log res = %d", tp->t_log_res);
2125 xfs_warn(mp, " log count = %d", tp->t_log_count);
2126 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2128 xlog_print_tic_res(mp, tp->t_ticket);
2130 /* dump each log item */
2131 list_for_each_entry(lip, &tp->t_items, li_trans) {
2132 struct xfs_log_vec *lv = lip->li_lv;
2133 struct xfs_log_iovec *vec;
2136 xfs_warn(mp, "log item: ");
2137 xfs_warn(mp, " type = 0x%x", lip->li_type);
2138 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2141 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2142 xfs_warn(mp, " size = %d", lv->lv_size);
2143 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2144 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2146 /* dump each iovec for the log item */
2147 vec = lv->lv_iovecp;
2148 for (i = 0; i < lv->lv_niovecs; i++) {
2149 int dumplen = min(vec->i_len, 32);
2151 xfs_warn(mp, " iovec[%d]", i);
2152 xfs_warn(mp, " type = 0x%x", vec->i_type);
2153 xfs_warn(mp, " len = %d", vec->i_len);
2154 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2155 xfs_hex_dump(vec->i_addr, dumplen);
2163 * Calculate the potential space needed by the log vector. We may need a start
2164 * record, and each region gets its own struct xlog_op_header and may need to be
2165 * double word aligned.
2168 xlog_write_calc_vec_length(
2169 struct xlog_ticket *ticket,
2170 struct xfs_log_vec *log_vector,
2173 struct xfs_log_vec *lv;
2178 if (optype & XLOG_START_TRANS)
2181 for (lv = log_vector; lv; lv = lv->lv_next) {
2182 /* we don't write ordered log vectors */
2183 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2186 headers += lv->lv_niovecs;
2188 for (i = 0; i < lv->lv_niovecs; i++) {
2189 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2192 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2196 ticket->t_res_num_ophdrs += headers;
2197 len += headers * sizeof(struct xlog_op_header);
2203 xlog_write_start_rec(
2204 struct xlog_op_header *ophdr,
2205 struct xlog_ticket *ticket)
2207 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2208 ophdr->oh_clientid = ticket->t_clientid;
2210 ophdr->oh_flags = XLOG_START_TRANS;
2214 static xlog_op_header_t *
2215 xlog_write_setup_ophdr(
2217 struct xlog_op_header *ophdr,
2218 struct xlog_ticket *ticket,
2221 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2222 ophdr->oh_clientid = ticket->t_clientid;
2225 /* are we copying a commit or unmount record? */
2226 ophdr->oh_flags = flags;
2229 * We've seen logs corrupted with bad transaction client ids. This
2230 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2231 * and shut down the filesystem.
2233 switch (ophdr->oh_clientid) {
2234 case XFS_TRANSACTION:
2240 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2241 ophdr->oh_clientid, ticket);
2249 * Set up the parameters of the region copy into the log. This has
2250 * to handle region write split across multiple log buffers - this
2251 * state is kept external to this function so that this code can
2252 * be written in an obvious, self documenting manner.
2255 xlog_write_setup_copy(
2256 struct xlog_ticket *ticket,
2257 struct xlog_op_header *ophdr,
2258 int space_available,
2262 int *last_was_partial_copy,
2263 int *bytes_consumed)
2267 still_to_copy = space_required - *bytes_consumed;
2268 *copy_off = *bytes_consumed;
2270 if (still_to_copy <= space_available) {
2271 /* write of region completes here */
2272 *copy_len = still_to_copy;
2273 ophdr->oh_len = cpu_to_be32(*copy_len);
2274 if (*last_was_partial_copy)
2275 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2276 *last_was_partial_copy = 0;
2277 *bytes_consumed = 0;
2281 /* partial write of region, needs extra log op header reservation */
2282 *copy_len = space_available;
2283 ophdr->oh_len = cpu_to_be32(*copy_len);
2284 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2285 if (*last_was_partial_copy)
2286 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2287 *bytes_consumed += *copy_len;
2288 (*last_was_partial_copy)++;
2290 /* account for new log op header */
2291 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2292 ticket->t_res_num_ophdrs++;
2294 return sizeof(struct xlog_op_header);
2298 xlog_write_copy_finish(
2300 struct xlog_in_core *iclog,
2305 int *partial_copy_len,
2307 struct xlog_in_core **commit_iclog)
2311 if (*partial_copy) {
2313 * This iclog has already been marked WANT_SYNC by
2314 * xlog_state_get_iclog_space.
2316 spin_lock(&log->l_icloglock);
2317 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2324 *partial_copy_len = 0;
2326 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2327 /* no more space in this iclog - push it. */
2328 spin_lock(&log->l_icloglock);
2329 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2333 if (iclog->ic_state == XLOG_STATE_ACTIVE)
2334 xlog_state_switch_iclogs(log, iclog, 0);
2336 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2337 iclog->ic_state == XLOG_STATE_IOERROR);
2340 spin_unlock(&log->l_icloglock);
2341 ASSERT(flags & XLOG_COMMIT_TRANS);
2342 *commit_iclog = iclog;
2348 error = xlog_state_release_iclog(log, iclog, 0);
2349 spin_unlock(&log->l_icloglock);
2354 * Write some region out to in-core log
2356 * This will be called when writing externally provided regions or when
2357 * writing out a commit record for a given transaction.
2359 * General algorithm:
2360 * 1. Find total length of this write. This may include adding to the
2361 * lengths passed in.
2362 * 2. Check whether we violate the tickets reservation.
2363 * 3. While writing to this iclog
2364 * A. Reserve as much space in this iclog as can get
2365 * B. If this is first write, save away start lsn
2366 * C. While writing this region:
2367 * 1. If first write of transaction, write start record
2368 * 2. Write log operation header (header per region)
2369 * 3. Find out if we can fit entire region into this iclog
2370 * 4. Potentially, verify destination memcpy ptr
2371 * 5. Memcpy (partial) region
2372 * 6. If partial copy, release iclog; otherwise, continue
2373 * copying more regions into current iclog
2374 * 4. Mark want sync bit (in simulation mode)
2375 * 5. Release iclog for potential flush to on-disk log.
2378 * 1. Panic if reservation is overrun. This should never happen since
2379 * reservation amounts are generated internal to the filesystem.
2381 * 1. Tickets are single threaded data structures.
2382 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2383 * syncing routine. When a single log_write region needs to span
2384 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2385 * on all log operation writes which don't contain the end of the
2386 * region. The XLOG_END_TRANS bit is used for the in-core log
2387 * operation which contains the end of the continued log_write region.
2388 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2389 * we don't really know exactly how much space will be used. As a result,
2390 * we don't update ic_offset until the end when we know exactly how many
2391 * bytes have been written out.
2396 struct xfs_log_vec *log_vector,
2397 struct xlog_ticket *ticket,
2398 xfs_lsn_t *start_lsn,
2399 struct xlog_in_core **commit_iclog,
2402 struct xlog_in_core *iclog = NULL;
2403 struct xfs_log_vec *lv = log_vector;
2404 struct xfs_log_iovec *vecp = lv->lv_iovecp;
2407 int partial_copy = 0;
2408 int partial_copy_len = 0;
2415 * If this is a commit or unmount transaction, we don't need a start
2416 * record to be written. We do, however, have to account for the
2417 * commit or unmount header that gets written. Hence we always have
2418 * to account for an extra xlog_op_header here.
2420 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2421 if (ticket->t_curr_res < 0) {
2422 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2423 "ctx ticket reservation ran out. Need to up reservation");
2424 xlog_print_tic_res(log->l_mp, ticket);
2425 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2428 len = xlog_write_calc_vec_length(ticket, log_vector, optype);
2431 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2435 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2436 &contwr, &log_offset);
2440 ASSERT(log_offset <= iclog->ic_size - 1);
2441 ptr = iclog->ic_datap + log_offset;
2443 /* Start_lsn is the first lsn written to. */
2444 if (start_lsn && !*start_lsn)
2445 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2448 * This loop writes out as many regions as can fit in the amount
2449 * of space which was allocated by xlog_state_get_iclog_space().
2451 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2452 struct xfs_log_iovec *reg;
2453 struct xlog_op_header *ophdr;
2456 bool ordered = false;
2457 bool wrote_start_rec = false;
2459 /* ordered log vectors have no regions to write */
2460 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2461 ASSERT(lv->lv_niovecs == 0);
2467 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2468 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2471 * Before we start formatting log vectors, we need to
2472 * write a start record. Only do this for the first
2473 * iclog we write to.
2475 if (optype & XLOG_START_TRANS) {
2476 xlog_write_start_rec(ptr, ticket);
2477 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2478 sizeof(struct xlog_op_header));
2479 optype &= ~XLOG_START_TRANS;
2480 wrote_start_rec = true;
2483 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, optype);
2487 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2488 sizeof(struct xlog_op_header));
2490 len += xlog_write_setup_copy(ticket, ophdr,
2491 iclog->ic_size-log_offset,
2493 ©_off, ©_len,
2496 xlog_verify_dest_ptr(log, ptr);
2501 * Unmount records just log an opheader, so can have
2502 * empty payloads with no data region to copy. Hence we
2503 * only copy the payload if the vector says it has data
2506 ASSERT(copy_len >= 0);
2508 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2509 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2512 copy_len += sizeof(struct xlog_op_header);
2514 if (wrote_start_rec) {
2515 copy_len += sizeof(struct xlog_op_header);
2518 data_cnt += contwr ? copy_len : 0;
2520 error = xlog_write_copy_finish(log, iclog, optype,
2521 &record_cnt, &data_cnt,
2530 * if we had a partial copy, we need to get more iclog
2531 * space but we don't want to increment the region
2532 * index because there is still more is this region to
2535 * If we completed writing this region, and we flushed
2536 * the iclog (indicated by resetting of the record
2537 * count), then we also need to get more log space. If
2538 * this was the last record, though, we are done and
2544 if (++index == lv->lv_niovecs) {
2549 vecp = lv->lv_iovecp;
2551 if (record_cnt == 0 && !ordered) {
2561 spin_lock(&log->l_icloglock);
2562 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2564 ASSERT(optype & XLOG_COMMIT_TRANS);
2565 *commit_iclog = iclog;
2567 error = xlog_state_release_iclog(log, iclog, 0);
2569 spin_unlock(&log->l_icloglock);
2575 xlog_state_activate_iclog(
2576 struct xlog_in_core *iclog,
2577 int *iclogs_changed)
2579 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2580 trace_xlog_iclog_activate(iclog, _RET_IP_);
2583 * If the number of ops in this iclog indicate it just contains the
2584 * dummy transaction, we can change state into IDLE (the second time
2585 * around). Otherwise we should change the state into NEED a dummy.
2586 * We don't need to cover the dummy.
2588 if (*iclogs_changed == 0 &&
2589 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2590 *iclogs_changed = 1;
2593 * We have two dirty iclogs so start over. This could also be
2594 * num of ops indicating this is not the dummy going out.
2596 *iclogs_changed = 2;
2599 iclog->ic_state = XLOG_STATE_ACTIVE;
2600 iclog->ic_offset = 0;
2601 iclog->ic_header.h_num_logops = 0;
2602 memset(iclog->ic_header.h_cycle_data, 0,
2603 sizeof(iclog->ic_header.h_cycle_data));
2604 iclog->ic_header.h_lsn = 0;
2605 iclog->ic_header.h_tail_lsn = 0;
2609 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2610 * ACTIVE after iclog I/O has completed.
2613 xlog_state_activate_iclogs(
2615 int *iclogs_changed)
2617 struct xlog_in_core *iclog = log->l_iclog;
2620 if (iclog->ic_state == XLOG_STATE_DIRTY)
2621 xlog_state_activate_iclog(iclog, iclogs_changed);
2623 * The ordering of marking iclogs ACTIVE must be maintained, so
2624 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2626 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2628 } while ((iclog = iclog->ic_next) != log->l_iclog);
2637 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2638 * wrote the first covering record (DONE). We go to IDLE if we just
2639 * wrote the second covering record (DONE2) and remain in IDLE until a
2640 * non-covering write occurs.
2642 switch (prev_state) {
2643 case XLOG_STATE_COVER_IDLE:
2644 if (iclogs_changed == 1)
2645 return XLOG_STATE_COVER_IDLE;
2647 case XLOG_STATE_COVER_NEED:
2648 case XLOG_STATE_COVER_NEED2:
2650 case XLOG_STATE_COVER_DONE:
2651 if (iclogs_changed == 1)
2652 return XLOG_STATE_COVER_NEED2;
2654 case XLOG_STATE_COVER_DONE2:
2655 if (iclogs_changed == 1)
2656 return XLOG_STATE_COVER_IDLE;
2662 return XLOG_STATE_COVER_NEED;
2666 xlog_state_clean_iclog(
2668 struct xlog_in_core *dirty_iclog)
2670 int iclogs_changed = 0;
2672 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2674 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2676 xlog_state_activate_iclogs(log, &iclogs_changed);
2677 wake_up_all(&dirty_iclog->ic_force_wait);
2679 if (iclogs_changed) {
2680 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2686 xlog_get_lowest_lsn(
2689 struct xlog_in_core *iclog = log->l_iclog;
2690 xfs_lsn_t lowest_lsn = 0, lsn;
2693 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2694 iclog->ic_state == XLOG_STATE_DIRTY)
2697 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2698 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2700 } while ((iclog = iclog->ic_next) != log->l_iclog);
2706 * Completion of a iclog IO does not imply that a transaction has completed, as
2707 * transactions can be large enough to span many iclogs. We cannot change the
2708 * tail of the log half way through a transaction as this may be the only
2709 * transaction in the log and moving the tail to point to the middle of it
2710 * will prevent recovery from finding the start of the transaction. Hence we
2711 * should only update the last_sync_lsn if this iclog contains transaction
2712 * completion callbacks on it.
2714 * We have to do this before we drop the icloglock to ensure we are the only one
2715 * that can update it.
2717 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2718 * the reservation grant head pushing. This is due to the fact that the push
2719 * target is bound by the current last_sync_lsn value. Hence if we have a large
2720 * amount of log space bound up in this committing transaction then the
2721 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2722 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2723 * should push the AIL to ensure the push target (and hence the grant head) is
2724 * no longer bound by the old log head location and can move forwards and make
2728 xlog_state_set_callback(
2730 struct xlog_in_core *iclog,
2731 xfs_lsn_t header_lsn)
2733 trace_xlog_iclog_callback(iclog, _RET_IP_);
2734 iclog->ic_state = XLOG_STATE_CALLBACK;
2736 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2739 if (list_empty_careful(&iclog->ic_callbacks))
2742 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2743 xlog_grant_push_ail(log, 0);
2747 * Return true if we need to stop processing, false to continue to the next
2748 * iclog. The caller will need to run callbacks if the iclog is returned in the
2749 * XLOG_STATE_CALLBACK state.
2752 xlog_state_iodone_process_iclog(
2754 struct xlog_in_core *iclog,
2757 xfs_lsn_t lowest_lsn;
2758 xfs_lsn_t header_lsn;
2760 switch (iclog->ic_state) {
2761 case XLOG_STATE_ACTIVE:
2762 case XLOG_STATE_DIRTY:
2764 * Skip all iclogs in the ACTIVE & DIRTY states:
2767 case XLOG_STATE_IOERROR:
2769 * Between marking a filesystem SHUTDOWN and stopping the log,
2770 * we do flush all iclogs to disk (if there wasn't a log I/O
2771 * error). So, we do want things to go smoothly in case of just
2772 * a SHUTDOWN w/o a LOG_IO_ERROR.
2776 case XLOG_STATE_DONE_SYNC:
2778 * Now that we have an iclog that is in the DONE_SYNC state, do
2779 * one more check here to see if we have chased our tail around.
2780 * If this is not the lowest lsn iclog, then we will leave it
2781 * for another completion to process.
2783 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2784 lowest_lsn = xlog_get_lowest_lsn(log);
2785 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2787 xlog_state_set_callback(log, iclog, header_lsn);
2791 * Can only perform callbacks in order. Since this iclog is not
2792 * in the DONE_SYNC state, we skip the rest and just try to
2800 xlog_state_do_callback(
2803 struct xlog_in_core *iclog;
2804 struct xlog_in_core *first_iclog;
2805 bool cycled_icloglock;
2810 spin_lock(&log->l_icloglock);
2813 * Scan all iclogs starting with the one pointed to by the
2814 * log. Reset this starting point each time the log is
2815 * unlocked (during callbacks).
2817 * Keep looping through iclogs until one full pass is made
2818 * without running any callbacks.
2820 first_iclog = log->l_iclog;
2821 iclog = log->l_iclog;
2822 cycled_icloglock = false;
2829 if (xlog_state_iodone_process_iclog(log, iclog,
2833 if (iclog->ic_state != XLOG_STATE_CALLBACK &&
2834 iclog->ic_state != XLOG_STATE_IOERROR) {
2835 iclog = iclog->ic_next;
2838 list_splice_init(&iclog->ic_callbacks, &cb_list);
2839 spin_unlock(&log->l_icloglock);
2841 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2842 xlog_cil_process_committed(&cb_list);
2843 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2844 cycled_icloglock = true;
2846 spin_lock(&log->l_icloglock);
2847 if (XLOG_FORCED_SHUTDOWN(log))
2848 wake_up_all(&iclog->ic_force_wait);
2850 xlog_state_clean_iclog(log, iclog);
2851 iclog = iclog->ic_next;
2852 } while (first_iclog != iclog);
2854 if (repeats > 5000) {
2855 flushcnt += repeats;
2858 "%s: possible infinite loop (%d iterations)",
2859 __func__, flushcnt);
2861 } while (!ioerror && cycled_icloglock);
2863 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE ||
2864 log->l_iclog->ic_state == XLOG_STATE_IOERROR)
2865 wake_up_all(&log->l_flush_wait);
2867 spin_unlock(&log->l_icloglock);
2872 * Finish transitioning this iclog to the dirty state.
2874 * Make sure that we completely execute this routine only when this is
2875 * the last call to the iclog. There is a good chance that iclog flushes,
2876 * when we reach the end of the physical log, get turned into 2 separate
2877 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2878 * routine. By using the reference count bwritecnt, we guarantee that only
2879 * the second completion goes through.
2881 * Callbacks could take time, so they are done outside the scope of the
2882 * global state machine log lock.
2885 xlog_state_done_syncing(
2886 struct xlog_in_core *iclog)
2888 struct xlog *log = iclog->ic_log;
2890 spin_lock(&log->l_icloglock);
2891 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2892 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2895 * If we got an error, either on the first buffer, or in the case of
2896 * split log writes, on the second, we shut down the file system and
2897 * no iclogs should ever be attempted to be written to disk again.
2899 if (!XLOG_FORCED_SHUTDOWN(log)) {
2900 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2901 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2905 * Someone could be sleeping prior to writing out the next
2906 * iclog buffer, we wake them all, one will get to do the
2907 * I/O, the others get to wait for the result.
2909 wake_up_all(&iclog->ic_write_wait);
2910 spin_unlock(&log->l_icloglock);
2911 xlog_state_do_callback(log);
2915 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2916 * sleep. We wait on the flush queue on the head iclog as that should be
2917 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2918 * we will wait here and all new writes will sleep until a sync completes.
2920 * The in-core logs are used in a circular fashion. They are not used
2921 * out-of-order even when an iclog past the head is free.
2924 * * log_offset where xlog_write() can start writing into the in-core
2926 * * in-core log pointer to which xlog_write() should write.
2927 * * boolean indicating this is a continued write to an in-core log.
2928 * If this is the last write, then the in-core log's offset field
2929 * needs to be incremented, depending on the amount of data which
2933 xlog_state_get_iclog_space(
2936 struct xlog_in_core **iclogp,
2937 struct xlog_ticket *ticket,
2938 int *continued_write,
2942 xlog_rec_header_t *head;
2943 xlog_in_core_t *iclog;
2946 spin_lock(&log->l_icloglock);
2947 if (XLOG_FORCED_SHUTDOWN(log)) {
2948 spin_unlock(&log->l_icloglock);
2952 iclog = log->l_iclog;
2953 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2954 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2956 /* Wait for log writes to have flushed */
2957 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2961 head = &iclog->ic_header;
2963 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2964 log_offset = iclog->ic_offset;
2966 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2968 /* On the 1st write to an iclog, figure out lsn. This works
2969 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2970 * committing to. If the offset is set, that's how many blocks
2973 if (log_offset == 0) {
2974 ticket->t_curr_res -= log->l_iclog_hsize;
2975 xlog_tic_add_region(ticket,
2977 XLOG_REG_TYPE_LRHEADER);
2978 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2979 head->h_lsn = cpu_to_be64(
2980 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2981 ASSERT(log->l_curr_block >= 0);
2984 /* If there is enough room to write everything, then do it. Otherwise,
2985 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2986 * bit is on, so this will get flushed out. Don't update ic_offset
2987 * until you know exactly how many bytes get copied. Therefore, wait
2988 * until later to update ic_offset.
2990 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2991 * can fit into remaining data section.
2993 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2996 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2999 * If we are the only one writing to this iclog, sync it to
3000 * disk. We need to do an atomic compare and decrement here to
3001 * avoid racing with concurrent atomic_dec_and_lock() calls in
3002 * xlog_state_release_iclog() when there is more than one
3003 * reference to the iclog.
3005 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
3006 error = xlog_state_release_iclog(log, iclog, 0);
3007 spin_unlock(&log->l_icloglock);
3013 /* Do we have enough room to write the full amount in the remainder
3014 * of this iclog? Or must we continue a write on the next iclog and
3015 * mark this iclog as completely taken? In the case where we switch
3016 * iclogs (to mark it taken), this particular iclog will release/sync
3017 * to disk in xlog_write().
3019 if (len <= iclog->ic_size - iclog->ic_offset) {
3020 *continued_write = 0;
3021 iclog->ic_offset += len;
3023 *continued_write = 1;
3024 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3028 ASSERT(iclog->ic_offset <= iclog->ic_size);
3029 spin_unlock(&log->l_icloglock);
3031 *logoffsetp = log_offset;
3036 * The first cnt-1 times a ticket goes through here we don't need to move the
3037 * grant write head because the permanent reservation has reserved cnt times the
3038 * unit amount. Release part of current permanent unit reservation and reset
3039 * current reservation to be one units worth. Also move grant reservation head
3043 xfs_log_ticket_regrant(
3045 struct xlog_ticket *ticket)
3047 trace_xfs_log_ticket_regrant(log, ticket);
3049 if (ticket->t_cnt > 0)
3052 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3053 ticket->t_curr_res);
3054 xlog_grant_sub_space(log, &log->l_write_head.grant,
3055 ticket->t_curr_res);
3056 ticket->t_curr_res = ticket->t_unit_res;
3057 xlog_tic_reset_res(ticket);
3059 trace_xfs_log_ticket_regrant_sub(log, ticket);
3061 /* just return if we still have some of the pre-reserved space */
3062 if (!ticket->t_cnt) {
3063 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3064 ticket->t_unit_res);
3065 trace_xfs_log_ticket_regrant_exit(log, ticket);
3067 ticket->t_curr_res = ticket->t_unit_res;
3068 xlog_tic_reset_res(ticket);
3071 xfs_log_ticket_put(ticket);
3075 * Give back the space left from a reservation.
3077 * All the information we need to make a correct determination of space left
3078 * is present. For non-permanent reservations, things are quite easy. The
3079 * count should have been decremented to zero. We only need to deal with the
3080 * space remaining in the current reservation part of the ticket. If the
3081 * ticket contains a permanent reservation, there may be left over space which
3082 * needs to be released. A count of N means that N-1 refills of the current
3083 * reservation can be done before we need to ask for more space. The first
3084 * one goes to fill up the first current reservation. Once we run out of
3085 * space, the count will stay at zero and the only space remaining will be
3086 * in the current reservation field.
3089 xfs_log_ticket_ungrant(
3091 struct xlog_ticket *ticket)
3095 trace_xfs_log_ticket_ungrant(log, ticket);
3097 if (ticket->t_cnt > 0)
3100 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3103 * If this is a permanent reservation ticket, we may be able to free
3104 * up more space based on the remaining count.
3106 bytes = ticket->t_curr_res;
3107 if (ticket->t_cnt > 0) {
3108 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3109 bytes += ticket->t_unit_res*ticket->t_cnt;
3112 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3113 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3115 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3117 xfs_log_space_wake(log->l_mp);
3118 xfs_log_ticket_put(ticket);
3122 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3123 * the current iclog pointer to the next iclog in the ring.
3126 xlog_state_switch_iclogs(
3128 struct xlog_in_core *iclog,
3131 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3132 assert_spin_locked(&log->l_icloglock);
3133 trace_xlog_iclog_switch(iclog, _RET_IP_);
3136 eventual_size = iclog->ic_offset;
3137 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3138 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3139 log->l_prev_block = log->l_curr_block;
3140 log->l_prev_cycle = log->l_curr_cycle;
3142 /* roll log?: ic_offset changed later */
3143 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3145 /* Round up to next log-sunit */
3146 if (log->l_iclog_roundoff > BBSIZE) {
3147 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3148 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3151 if (log->l_curr_block >= log->l_logBBsize) {
3153 * Rewind the current block before the cycle is bumped to make
3154 * sure that the combined LSN never transiently moves forward
3155 * when the log wraps to the next cycle. This is to support the
3156 * unlocked sample of these fields from xlog_valid_lsn(). Most
3157 * other cases should acquire l_icloglock.
3159 log->l_curr_block -= log->l_logBBsize;
3160 ASSERT(log->l_curr_block >= 0);
3162 log->l_curr_cycle++;
3163 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3164 log->l_curr_cycle++;
3166 ASSERT(iclog == log->l_iclog);
3167 log->l_iclog = iclog->ic_next;
3171 * Force the iclog to disk and check if the iclog has been completed before
3172 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3173 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3174 * If completion has already occurred, tell the caller so that it can avoid an
3175 * unnecessary wait on the iclog.
3178 xlog_force_and_check_iclog(
3179 struct xlog_in_core *iclog,
3182 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3186 error = xlog_force_iclog(iclog);
3191 * If the iclog has already been completed and reused the header LSN
3192 * will have been rewritten by completion
3194 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3200 * Write out all data in the in-core log as of this exact moment in time.
3202 * Data may be written to the in-core log during this call. However,
3203 * we don't guarantee this data will be written out. A change from past
3204 * implementation means this routine will *not* write out zero length LRs.
3206 * Basically, we try and perform an intelligent scan of the in-core logs.
3207 * If we determine there is no flushable data, we just return. There is no
3208 * flushable data if:
3210 * 1. the current iclog is active and has no data; the previous iclog
3211 * is in the active or dirty state.
3212 * 2. the current iclog is drity, and the previous iclog is in the
3213 * active or dirty state.
3217 * 1. the current iclog is not in the active nor dirty state.
3218 * 2. the current iclog dirty, and the previous iclog is not in the
3219 * active nor dirty state.
3220 * 3. the current iclog is active, and there is another thread writing
3221 * to this particular iclog.
3222 * 4. a) the current iclog is active and has no other writers
3223 * b) when we return from flushing out this iclog, it is still
3224 * not in the active nor dirty state.
3228 struct xfs_mount *mp,
3231 struct xlog *log = mp->m_log;
3232 struct xlog_in_core *iclog;
3234 XFS_STATS_INC(mp, xs_log_force);
3235 trace_xfs_log_force(mp, 0, _RET_IP_);
3237 xlog_cil_force(log);
3239 spin_lock(&log->l_icloglock);
3240 iclog = log->l_iclog;
3241 if (iclog->ic_state == XLOG_STATE_IOERROR)
3244 trace_xlog_iclog_force(iclog, _RET_IP_);
3246 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3247 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3248 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3250 * If the head is dirty or (active and empty), then we need to
3251 * look at the previous iclog.
3253 * If the previous iclog is active or dirty we are done. There
3254 * is nothing to sync out. Otherwise, we attach ourselves to the
3255 * previous iclog and go to sleep.
3257 iclog = iclog->ic_prev;
3258 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3259 if (atomic_read(&iclog->ic_refcnt) == 0) {
3260 /* We have exclusive access to this iclog. */
3263 if (xlog_force_and_check_iclog(iclog, &completed))
3270 * Someone else is still writing to this iclog, so we
3271 * need to ensure that when they release the iclog it
3272 * gets synced immediately as we may be waiting on it.
3274 xlog_state_switch_iclogs(log, iclog, 0);
3279 * The iclog we are about to wait on may contain the checkpoint pushed
3280 * by the above xlog_cil_force() call, but it may not have been pushed
3281 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3282 * are flushed when this iclog is written.
3284 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3285 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3287 if (flags & XFS_LOG_SYNC)
3288 return xlog_wait_on_iclog(iclog);
3290 spin_unlock(&log->l_icloglock);
3293 spin_unlock(&log->l_icloglock);
3305 struct xlog_in_core *iclog;
3308 spin_lock(&log->l_icloglock);
3309 iclog = log->l_iclog;
3310 if (iclog->ic_state == XLOG_STATE_IOERROR)
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 in-core log to disk for a specific LSN.
3384 * Find in-core log with lsn.
3385 * If it is in the DIRTY state, just return.
3386 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3387 * state and go to sleep or return.
3388 * If it is in any other state, go to sleep or return.
3390 * Synchronous forces are implemented with a wait queue. All callers trying
3391 * to force a given lsn to disk must wait on the queue attached to the
3392 * specific in-core log. When given in-core log finally completes its write
3393 * to disk, that thread will wake up all threads waiting on the queue.
3397 struct xfs_mount *mp,
3402 struct xlog *log = mp->m_log;
3407 XFS_STATS_INC(mp, xs_log_force);
3408 trace_xfs_log_force(mp, seq, _RET_IP_);
3410 lsn = xlog_cil_force_seq(log, seq);
3411 if (lsn == NULLCOMMITLSN)
3414 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3415 if (ret == -EAGAIN) {
3416 XFS_STATS_INC(mp, xs_log_force_sleep);
3417 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3423 * Free a used ticket when its refcount falls to zero.
3427 xlog_ticket_t *ticket)
3429 ASSERT(atomic_read(&ticket->t_ref) > 0);
3430 if (atomic_dec_and_test(&ticket->t_ref))
3431 kmem_cache_free(xfs_log_ticket_zone, ticket);
3436 xlog_ticket_t *ticket)
3438 ASSERT(atomic_read(&ticket->t_ref) > 0);
3439 atomic_inc(&ticket->t_ref);
3444 * Figure out the total log space unit (in bytes) that would be
3445 * required for a log ticket.
3456 * Permanent reservations have up to 'cnt'-1 active log operations
3457 * in the log. A unit in this case is the amount of space for one
3458 * of these log operations. Normal reservations have a cnt of 1
3459 * and their unit amount is the total amount of space required.
3461 * The following lines of code account for non-transaction data
3462 * which occupy space in the on-disk log.
3464 * Normal form of a transaction is:
3465 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3466 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3468 * We need to account for all the leadup data and trailer data
3469 * around the transaction data.
3470 * And then we need to account for the worst case in terms of using
3472 * The worst case will happen if:
3473 * - the placement of the transaction happens to be such that the
3474 * roundoff is at its maximum
3475 * - the transaction data is synced before the commit record is synced
3476 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3477 * Therefore the commit record is in its own Log Record.
3478 * This can happen as the commit record is called with its
3479 * own region to xlog_write().
3480 * This then means that in the worst case, roundoff can happen for
3481 * the commit-rec as well.
3482 * The commit-rec is smaller than padding in this scenario and so it is
3483 * not added separately.
3486 /* for trans header */
3487 unit_bytes += sizeof(xlog_op_header_t);
3488 unit_bytes += sizeof(xfs_trans_header_t);
3491 unit_bytes += sizeof(xlog_op_header_t);
3494 * for LR headers - the space for data in an iclog is the size minus
3495 * the space used for the headers. If we use the iclog size, then we
3496 * undercalculate the number of headers required.
3498 * Furthermore - the addition of op headers for split-recs might
3499 * increase the space required enough to require more log and op
3500 * headers, so take that into account too.
3502 * IMPORTANT: This reservation makes the assumption that if this
3503 * transaction is the first in an iclog and hence has the LR headers
3504 * accounted to it, then the remaining space in the iclog is
3505 * exclusively for this transaction. i.e. if the transaction is larger
3506 * than the iclog, it will be the only thing in that iclog.
3507 * Fundamentally, this means we must pass the entire log vector to
3508 * xlog_write to guarantee this.
3510 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3511 num_headers = howmany(unit_bytes, iclog_space);
3513 /* for split-recs - ophdrs added when data split over LRs */
3514 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3516 /* add extra header reservations if we overrun */
3517 while (!num_headers ||
3518 howmany(unit_bytes, iclog_space) > num_headers) {
3519 unit_bytes += sizeof(xlog_op_header_t);
3522 unit_bytes += log->l_iclog_hsize * num_headers;
3524 /* for commit-rec LR header - note: padding will subsume the ophdr */
3525 unit_bytes += log->l_iclog_hsize;
3527 /* roundoff padding for transaction data and one for commit record */
3528 unit_bytes += 2 * log->l_iclog_roundoff;
3534 xfs_log_calc_unit_res(
3535 struct xfs_mount *mp,
3538 return xlog_calc_unit_res(mp->m_log, unit_bytes);
3542 * Allocate and initialise a new log ticket.
3544 struct xlog_ticket *
3552 struct xlog_ticket *tic;
3555 tic = kmem_cache_zalloc(xfs_log_ticket_zone, GFP_NOFS | __GFP_NOFAIL);
3557 unit_res = xlog_calc_unit_res(log, unit_bytes);
3559 atomic_set(&tic->t_ref, 1);
3560 tic->t_task = current;
3561 INIT_LIST_HEAD(&tic->t_queue);
3562 tic->t_unit_res = unit_res;
3563 tic->t_curr_res = unit_res;
3566 tic->t_tid = prandom_u32();
3567 tic->t_clientid = client;
3569 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3571 xlog_tic_reset_res(tic);
3578 * Make sure that the destination ptr is within the valid data region of
3579 * one of the iclogs. This uses backup pointers stored in a different
3580 * part of the log in case we trash the log structure.
3583 xlog_verify_dest_ptr(
3590 for (i = 0; i < log->l_iclog_bufs; i++) {
3591 if (ptr >= log->l_iclog_bak[i] &&
3592 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3597 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3601 * Check to make sure the grant write head didn't just over lap the tail. If
3602 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3603 * the cycles differ by exactly one and check the byte count.
3605 * This check is run unlocked, so can give false positives. Rather than assert
3606 * on failures, use a warn-once flag and a panic tag to allow the admin to
3607 * determine if they want to panic the machine when such an error occurs. For
3608 * debug kernels this will have the same effect as using an assert but, unlinke
3609 * an assert, it can be turned off at runtime.
3612 xlog_verify_grant_tail(
3615 int tail_cycle, tail_blocks;
3618 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3619 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3620 if (tail_cycle != cycle) {
3621 if (cycle - 1 != tail_cycle &&
3622 !(log->l_flags & XLOG_TAIL_WARN)) {
3623 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3624 "%s: cycle - 1 != tail_cycle", __func__);
3625 log->l_flags |= XLOG_TAIL_WARN;
3628 if (space > BBTOB(tail_blocks) &&
3629 !(log->l_flags & XLOG_TAIL_WARN)) {
3630 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3631 "%s: space > BBTOB(tail_blocks)", __func__);
3632 log->l_flags |= XLOG_TAIL_WARN;
3637 /* check if it will fit */
3639 xlog_verify_tail_lsn(
3641 struct xlog_in_core *iclog)
3643 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3646 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3648 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3649 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3650 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3652 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3654 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3655 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3657 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3658 if (blocks < BTOBB(iclog->ic_offset) + 1)
3659 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3664 * Perform a number of checks on the iclog before writing to disk.
3666 * 1. Make sure the iclogs are still circular
3667 * 2. Make sure we have a good magic number
3668 * 3. Make sure we don't have magic numbers in the data
3669 * 4. Check fields of each log operation header for:
3670 * A. Valid client identifier
3671 * B. tid ptr value falls in valid ptr space (user space code)
3672 * C. Length in log record header is correct according to the
3673 * individual operation headers within record.
3674 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3675 * log, check the preceding blocks of the physical log to make sure all
3676 * the cycle numbers agree with the current cycle number.
3681 struct xlog_in_core *iclog,
3684 xlog_op_header_t *ophead;
3685 xlog_in_core_t *icptr;
3686 xlog_in_core_2_t *xhdr;
3687 void *base_ptr, *ptr, *p;
3688 ptrdiff_t field_offset;
3690 int len, i, j, k, op_len;
3693 /* check validity of iclog pointers */
3694 spin_lock(&log->l_icloglock);
3695 icptr = log->l_iclog;
3696 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3699 if (icptr != log->l_iclog)
3700 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3701 spin_unlock(&log->l_icloglock);
3703 /* check log magic numbers */
3704 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3705 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3707 base_ptr = ptr = &iclog->ic_header;
3708 p = &iclog->ic_header;
3709 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3710 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3711 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3716 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3717 base_ptr = ptr = iclog->ic_datap;
3719 xhdr = iclog->ic_data;
3720 for (i = 0; i < len; i++) {
3723 /* clientid is only 1 byte */
3724 p = &ophead->oh_clientid;
3725 field_offset = p - base_ptr;
3726 if (field_offset & 0x1ff) {
3727 clientid = ophead->oh_clientid;
3729 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3730 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3731 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3732 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3733 clientid = xlog_get_client_id(
3734 xhdr[j].hic_xheader.xh_cycle_data[k]);
3736 clientid = xlog_get_client_id(
3737 iclog->ic_header.h_cycle_data[idx]);
3740 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3742 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3743 __func__, clientid, ophead,
3744 (unsigned long)field_offset);
3747 p = &ophead->oh_len;
3748 field_offset = p - base_ptr;
3749 if (field_offset & 0x1ff) {
3750 op_len = be32_to_cpu(ophead->oh_len);
3752 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3753 (uintptr_t)iclog->ic_datap);
3754 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3755 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3756 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3757 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3759 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3762 ptr += sizeof(xlog_op_header_t) + op_len;
3768 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3774 xlog_in_core_t *iclog, *ic;
3776 iclog = log->l_iclog;
3777 if (iclog->ic_state != XLOG_STATE_IOERROR) {
3779 * Mark all the incore logs IOERROR.
3780 * From now on, no log flushes will result.
3784 ic->ic_state = XLOG_STATE_IOERROR;
3786 } while (ic != iclog);
3790 * Return non-zero, if state transition has already happened.
3796 * This is called from xfs_force_shutdown, when we're forcibly
3797 * shutting down the filesystem, typically because of an IO error.
3798 * Our main objectives here are to make sure that:
3799 * a. if !logerror, flush the logs to disk. Anything modified
3800 * after this is ignored.
3801 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3802 * parties to find out, 'atomically'.
3803 * c. those who're sleeping on log reservations, pinned objects and
3804 * other resources get woken up, and be told the bad news.
3805 * d. nothing new gets queued up after (b) and (c) are done.
3807 * Note: for the !logerror case we need to flush the regions held in memory out
3808 * to disk first. This needs to be done before the log is marked as shutdown,
3809 * otherwise the iclog writes will fail.
3812 xfs_log_force_umount(
3813 struct xfs_mount *mp,
3822 * If this happens during log recovery, don't worry about
3823 * locking; the log isn't open for business yet.
3826 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3827 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3829 mp->m_sb_bp->b_flags |= XBF_DONE;
3834 * Somebody could've already done the hard work for us.
3835 * No need to get locks for this.
3837 if (logerror && log->l_iclog->ic_state == XLOG_STATE_IOERROR) {
3838 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3843 * Flush all the completed transactions to disk before marking the log
3844 * being shut down. We need to do it in this order to ensure that
3845 * completed operations are safely on disk before we shut down, and that
3846 * we don't have to issue any buffer IO after the shutdown flags are set
3847 * to guarantee this.
3850 xfs_log_force(mp, XFS_LOG_SYNC);
3853 * mark the filesystem and the as in a shutdown state and wake
3854 * everybody up to tell them the bad news.
3856 spin_lock(&log->l_icloglock);
3857 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3859 mp->m_sb_bp->b_flags |= XBF_DONE;
3862 * Mark the log and the iclogs with IO error flags to prevent any
3863 * further log IO from being issued or completed.
3865 log->l_flags |= XLOG_IO_ERROR;
3866 retval = xlog_state_ioerror(log);
3867 spin_unlock(&log->l_icloglock);
3870 * We don't want anybody waiting for log reservations after this. That
3871 * means we have to wake up everybody queued up on reserveq as well as
3872 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3873 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3874 * action is protected by the grant locks.
3876 xlog_grant_head_wake_all(&log->l_reserve_head);
3877 xlog_grant_head_wake_all(&log->l_write_head);
3880 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3881 * as if the log writes were completed. The abort handling in the log
3882 * item committed callback functions will do this again under lock to
3885 spin_lock(&log->l_cilp->xc_push_lock);
3886 wake_up_all(&log->l_cilp->xc_commit_wait);
3887 spin_unlock(&log->l_cilp->xc_push_lock);
3888 xlog_state_do_callback(log);
3890 /* return non-zero if log IOERROR transition had already happened */
3898 xlog_in_core_t *iclog;
3900 iclog = log->l_iclog;
3902 /* endianness does not matter here, zero is zero in
3905 if (iclog->ic_header.h_num_logops)
3907 iclog = iclog->ic_next;
3908 } while (iclog != log->l_iclog);
3913 * Verify that an LSN stamped into a piece of metadata is valid. This is
3914 * intended for use in read verifiers on v5 superblocks.
3918 struct xfs_mount *mp,
3921 struct xlog *log = mp->m_log;
3925 * norecovery mode skips mount-time log processing and unconditionally
3926 * resets the in-core LSN. We can't validate in this mode, but
3927 * modifications are not allowed anyways so just return true.
3929 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3933 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3934 * handled by recovery and thus safe to ignore here.
3936 if (lsn == NULLCOMMITLSN)
3939 valid = xlog_valid_lsn(mp->m_log, lsn);
3941 /* warn the user about what's gone wrong before verifier failure */
3943 spin_lock(&log->l_icloglock);
3945 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3946 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3947 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3948 log->l_curr_cycle, log->l_curr_block);
3949 spin_unlock(&log->l_icloglock);
3956 xfs_log_in_recovery(
3957 struct xfs_mount *mp)
3959 struct xlog *log = mp->m_log;
3961 return log->l_flags & XLOG_ACTIVE_RECOVERY;