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 xlog_ticket *ticket,
31 struct xlog_in_core **iclog,
32 xfs_lsn_t *commitlsnp);
37 struct xfs_buftarg *log_target,
38 xfs_daddr_t blk_offset,
48 /* local state machine functions */
49 STATIC void xlog_state_done_syncing(
50 struct xlog_in_core *iclog,
53 xlog_state_get_iclog_space(
56 struct xlog_in_core **iclog,
57 struct xlog_ticket *ticket,
61 xlog_state_release_iclog(
63 struct xlog_in_core *iclog);
65 xlog_state_switch_iclogs(
67 struct xlog_in_core *iclog,
72 struct xlog_in_core *iclog);
79 xlog_regrant_reserve_log_space(
81 struct xlog_ticket *ticket);
83 xlog_ungrant_log_space(
85 struct xlog_ticket *ticket);
93 xlog_verify_grant_tail(
98 struct xlog_in_core *iclog,
101 xlog_verify_tail_lsn(
103 struct xlog_in_core *iclog,
106 #define xlog_verify_dest_ptr(a,b)
107 #define xlog_verify_grant_tail(a)
108 #define xlog_verify_iclog(a,b,c)
109 #define xlog_verify_tail_lsn(a,b,c)
117 xlog_grant_sub_space(
122 int64_t head_val = atomic64_read(head);
128 xlog_crack_grant_head_val(head_val, &cycle, &space);
132 space += log->l_logsize;
137 new = xlog_assign_grant_head_val(cycle, space);
138 head_val = atomic64_cmpxchg(head, old, new);
139 } while (head_val != old);
143 xlog_grant_add_space(
148 int64_t head_val = atomic64_read(head);
155 xlog_crack_grant_head_val(head_val, &cycle, &space);
157 tmp = log->l_logsize - space;
166 new = xlog_assign_grant_head_val(cycle, space);
167 head_val = atomic64_cmpxchg(head, old, new);
168 } while (head_val != old);
172 xlog_grant_head_init(
173 struct xlog_grant_head *head)
175 xlog_assign_grant_head(&head->grant, 1, 0);
176 INIT_LIST_HEAD(&head->waiters);
177 spin_lock_init(&head->lock);
181 xlog_grant_head_wake_all(
182 struct xlog_grant_head *head)
184 struct xlog_ticket *tic;
186 spin_lock(&head->lock);
187 list_for_each_entry(tic, &head->waiters, t_queue)
188 wake_up_process(tic->t_task);
189 spin_unlock(&head->lock);
193 xlog_ticket_reservation(
195 struct xlog_grant_head *head,
196 struct xlog_ticket *tic)
198 if (head == &log->l_write_head) {
199 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
200 return tic->t_unit_res;
202 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
203 return tic->t_unit_res * tic->t_cnt;
205 return tic->t_unit_res;
210 xlog_grant_head_wake(
212 struct xlog_grant_head *head,
215 struct xlog_ticket *tic;
217 bool woken_task = false;
219 list_for_each_entry(tic, &head->waiters, t_queue) {
222 * There is a chance that the size of the CIL checkpoints in
223 * progress at the last AIL push target calculation resulted in
224 * limiting the target to the log head (l_last_sync_lsn) at the
225 * time. This may not reflect where the log head is now as the
226 * CIL checkpoints may have completed.
228 * Hence when we are woken here, it may be that the head of the
229 * log that has moved rather than the tail. As the tail didn't
230 * move, there still won't be space available for the
231 * reservation we require. However, if the AIL has already
232 * pushed to the target defined by the old log head location, we
233 * will hang here waiting for something else to update the AIL
236 * Therefore, if there isn't space to wake the first waiter on
237 * the grant head, we need to push the AIL again to ensure the
238 * target reflects both the current log tail and log head
239 * position before we wait for the tail to move again.
242 need_bytes = xlog_ticket_reservation(log, head, tic);
243 if (*free_bytes < need_bytes) {
245 xlog_grant_push_ail(log, need_bytes);
249 *free_bytes -= need_bytes;
250 trace_xfs_log_grant_wake_up(log, tic);
251 wake_up_process(tic->t_task);
259 xlog_grant_head_wait(
261 struct xlog_grant_head *head,
262 struct xlog_ticket *tic,
263 int need_bytes) __releases(&head->lock)
264 __acquires(&head->lock)
266 list_add_tail(&tic->t_queue, &head->waiters);
269 if (XLOG_FORCED_SHUTDOWN(log))
271 xlog_grant_push_ail(log, need_bytes);
273 __set_current_state(TASK_UNINTERRUPTIBLE);
274 spin_unlock(&head->lock);
276 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
278 trace_xfs_log_grant_sleep(log, tic);
280 trace_xfs_log_grant_wake(log, tic);
282 spin_lock(&head->lock);
283 if (XLOG_FORCED_SHUTDOWN(log))
285 } while (xlog_space_left(log, &head->grant) < need_bytes);
287 list_del_init(&tic->t_queue);
290 list_del_init(&tic->t_queue);
295 * Atomically get the log space required for a log ticket.
297 * Once a ticket gets put onto head->waiters, it will only return after the
298 * needed reservation is satisfied.
300 * This function is structured so that it has a lock free fast path. This is
301 * necessary because every new transaction reservation will come through this
302 * path. Hence any lock will be globally hot if we take it unconditionally on
305 * As tickets are only ever moved on and off head->waiters under head->lock, we
306 * only need to take that lock if we are going to add the ticket to the queue
307 * and sleep. We can avoid taking the lock if the ticket was never added to
308 * head->waiters because the t_queue list head will be empty and we hold the
309 * only reference to it so it can safely be checked unlocked.
312 xlog_grant_head_check(
314 struct xlog_grant_head *head,
315 struct xlog_ticket *tic,
321 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
324 * If there are other waiters on the queue then give them a chance at
325 * logspace before us. Wake up the first waiters, if we do not wake
326 * up all the waiters then go to sleep waiting for more free space,
327 * otherwise try to get some space for this transaction.
329 *need_bytes = xlog_ticket_reservation(log, head, tic);
330 free_bytes = xlog_space_left(log, &head->grant);
331 if (!list_empty_careful(&head->waiters)) {
332 spin_lock(&head->lock);
333 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
334 free_bytes < *need_bytes) {
335 error = xlog_grant_head_wait(log, head, tic,
338 spin_unlock(&head->lock);
339 } else if (free_bytes < *need_bytes) {
340 spin_lock(&head->lock);
341 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
342 spin_unlock(&head->lock);
349 xlog_tic_reset_res(xlog_ticket_t *tic)
352 tic->t_res_arr_sum = 0;
353 tic->t_res_num_ophdrs = 0;
357 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
359 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
360 /* add to overflow and start again */
361 tic->t_res_o_flow += tic->t_res_arr_sum;
363 tic->t_res_arr_sum = 0;
366 tic->t_res_arr[tic->t_res_num].r_len = len;
367 tic->t_res_arr[tic->t_res_num].r_type = type;
368 tic->t_res_arr_sum += len;
373 * Replenish the byte reservation required by moving the grant write head.
377 struct xfs_mount *mp,
378 struct xlog_ticket *tic)
380 struct xlog *log = mp->m_log;
384 if (XLOG_FORCED_SHUTDOWN(log))
387 XFS_STATS_INC(mp, xs_try_logspace);
390 * This is a new transaction on the ticket, so we need to change the
391 * transaction ID so that the next transaction has a different TID in
392 * the log. Just add one to the existing tid so that we can see chains
393 * of rolling transactions in the log easily.
397 xlog_grant_push_ail(log, tic->t_unit_res);
399 tic->t_curr_res = tic->t_unit_res;
400 xlog_tic_reset_res(tic);
405 trace_xfs_log_regrant(log, tic);
407 error = xlog_grant_head_check(log, &log->l_write_head, tic,
412 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
413 trace_xfs_log_regrant_exit(log, tic);
414 xlog_verify_grant_tail(log);
419 * If we are failing, make sure the ticket doesn't have any current
420 * reservations. We don't want to add this back when the ticket/
421 * transaction gets cancelled.
424 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
429 * Reserve log space and return a ticket corresponding to the reservation.
431 * Each reservation is going to reserve extra space for a log record header.
432 * When writes happen to the on-disk log, we don't subtract the length of the
433 * log record header from any reservation. By wasting space in each
434 * reservation, we prevent over allocation problems.
438 struct xfs_mount *mp,
441 struct xlog_ticket **ticp,
445 struct xlog *log = mp->m_log;
446 struct xlog_ticket *tic;
450 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
452 if (XLOG_FORCED_SHUTDOWN(log))
455 XFS_STATS_INC(mp, xs_try_logspace);
457 ASSERT(*ticp == NULL);
458 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent, 0);
461 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
464 trace_xfs_log_reserve(log, tic);
466 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
471 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
472 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
473 trace_xfs_log_reserve_exit(log, tic);
474 xlog_verify_grant_tail(log);
479 * If we are failing, make sure the ticket doesn't have any current
480 * reservations. We don't want to add this back when the ticket/
481 * transaction gets cancelled.
484 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
492 * 1. currblock field gets updated at startup and after in-core logs
493 * marked as with WANT_SYNC.
497 * This routine is called when a user of a log manager ticket is done with
498 * the reservation. If the ticket was ever used, then a commit record for
499 * the associated transaction is written out as a log operation header with
500 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
501 * a given ticket. If the ticket was one with a permanent reservation, then
502 * a few operations are done differently. Permanent reservation tickets by
503 * default don't release the reservation. They just commit the current
504 * transaction with the belief that the reservation is still needed. A flag
505 * must be passed in before permanent reservations are actually released.
506 * When these type of tickets are not released, they need to be set into
507 * the inited state again. By doing this, a start record will be written
508 * out when the next write occurs.
512 struct xfs_mount *mp,
513 struct xlog_ticket *ticket,
514 struct xlog_in_core **iclog,
517 struct xlog *log = mp->m_log;
520 if (XLOG_FORCED_SHUTDOWN(log) ||
522 * If nothing was ever written, don't write out commit record.
523 * If we get an error, just continue and give back the log ticket.
525 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
526 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
527 lsn = (xfs_lsn_t) -1;
533 trace_xfs_log_done_nonperm(log, ticket);
536 * Release ticket if not permanent reservation or a specific
537 * request has been made to release a permanent reservation.
539 xlog_ungrant_log_space(log, ticket);
541 trace_xfs_log_done_perm(log, ticket);
543 xlog_regrant_reserve_log_space(log, ticket);
544 /* If this ticket was a permanent reservation and we aren't
545 * trying to release it, reset the inited flags; so next time
546 * we write, a start record will be written out.
548 ticket->t_flags |= XLOG_TIC_INITED;
551 xfs_log_ticket_put(ticket);
556 xfs_log_release_iclog(
557 struct xfs_mount *mp,
558 struct xlog_in_core *iclog)
560 if (xlog_state_release_iclog(mp->m_log, iclog)) {
561 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
569 * Mount a log filesystem
571 * mp - ubiquitous xfs mount point structure
572 * log_target - buftarg of on-disk log device
573 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
574 * num_bblocks - Number of BBSIZE blocks in on-disk log
576 * Return error or zero.
581 xfs_buftarg_t *log_target,
582 xfs_daddr_t blk_offset,
585 bool fatal = xfs_sb_version_hascrc(&mp->m_sb);
589 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
590 xfs_notice(mp, "Mounting V%d Filesystem",
591 XFS_SB_VERSION_NUM(&mp->m_sb));
594 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
595 XFS_SB_VERSION_NUM(&mp->m_sb));
596 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
599 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
600 if (IS_ERR(mp->m_log)) {
601 error = PTR_ERR(mp->m_log);
606 * Validate the given log space and drop a critical message via syslog
607 * if the log size is too small that would lead to some unexpected
608 * situations in transaction log space reservation stage.
610 * Note: we can't just reject the mount if the validation fails. This
611 * would mean that people would have to downgrade their kernel just to
612 * remedy the situation as there is no way to grow the log (short of
613 * black magic surgery with xfs_db).
615 * We can, however, reject mounts for CRC format filesystems, as the
616 * mkfs binary being used to make the filesystem should never create a
617 * filesystem with a log that is too small.
619 min_logfsbs = xfs_log_calc_minimum_size(mp);
621 if (mp->m_sb.sb_logblocks < min_logfsbs) {
623 "Log size %d blocks too small, minimum size is %d blocks",
624 mp->m_sb.sb_logblocks, min_logfsbs);
626 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
628 "Log size %d blocks too large, maximum size is %lld blocks",
629 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
631 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
633 "log size %lld bytes too large, maximum size is %lld bytes",
634 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
637 } else if (mp->m_sb.sb_logsunit > 1 &&
638 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
640 "log stripe unit %u bytes must be a multiple of block size",
641 mp->m_sb.sb_logsunit);
647 * Log check errors are always fatal on v5; or whenever bad
648 * metadata leads to a crash.
651 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
655 xfs_crit(mp, "Log size out of supported range.");
657 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
661 * Initialize the AIL now we have a log.
663 error = xfs_trans_ail_init(mp);
665 xfs_warn(mp, "AIL initialisation failed: error %d", error);
668 mp->m_log->l_ailp = mp->m_ail;
671 * skip log recovery on a norecovery mount. pretend it all
674 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
675 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
678 mp->m_flags &= ~XFS_MOUNT_RDONLY;
680 error = xlog_recover(mp->m_log);
683 mp->m_flags |= XFS_MOUNT_RDONLY;
685 xfs_warn(mp, "log mount/recovery failed: error %d",
687 xlog_recover_cancel(mp->m_log);
688 goto out_destroy_ail;
692 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
695 goto out_destroy_ail;
697 /* Normal transactions can now occur */
698 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
701 * Now the log has been fully initialised and we know were our
702 * space grant counters are, we can initialise the permanent ticket
703 * needed for delayed logging to work.
705 xlog_cil_init_post_recovery(mp->m_log);
710 xfs_trans_ail_destroy(mp);
712 xlog_dealloc_log(mp->m_log);
718 * Finish the recovery of the file system. This is separate from the
719 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
720 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
723 * If we finish recovery successfully, start the background log work. If we are
724 * not doing recovery, then we have a RO filesystem and we don't need to start
728 xfs_log_mount_finish(
729 struct xfs_mount *mp)
732 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
733 bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
735 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
736 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
738 } else if (readonly) {
739 /* Allow unlinked processing to proceed */
740 mp->m_flags &= ~XFS_MOUNT_RDONLY;
744 * During the second phase of log recovery, we need iget and
745 * iput to behave like they do for an active filesystem.
746 * xfs_fs_drop_inode needs to be able to prevent the deletion
747 * of inodes before we're done replaying log items on those
748 * inodes. Turn it off immediately after recovery finishes
749 * so that we don't leak the quota inodes if subsequent mount
752 * We let all inodes involved in redo item processing end up on
753 * the LRU instead of being evicted immediately so that if we do
754 * something to an unlinked inode, the irele won't cause
755 * premature truncation and freeing of the inode, which results
756 * in log recovery failure. We have to evict the unreferenced
757 * lru inodes after clearing SB_ACTIVE because we don't
758 * otherwise clean up the lru if there's a subsequent failure in
759 * xfs_mountfs, which leads to us leaking the inodes if nothing
760 * else (e.g. quotacheck) references the inodes before the
761 * mount failure occurs.
763 mp->m_super->s_flags |= SB_ACTIVE;
764 error = xlog_recover_finish(mp->m_log);
766 xfs_log_work_queue(mp);
767 mp->m_super->s_flags &= ~SB_ACTIVE;
768 evict_inodes(mp->m_super);
771 * Drain the buffer LRU after log recovery. This is required for v4
772 * filesystems to avoid leaving around buffers with NULL verifier ops,
773 * but we do it unconditionally to make sure we're always in a clean
774 * cache state after mount.
776 * Don't push in the error case because the AIL may have pending intents
777 * that aren't removed until recovery is cancelled.
779 if (!error && recovered) {
780 xfs_log_force(mp, XFS_LOG_SYNC);
781 xfs_ail_push_all_sync(mp->m_ail);
783 xfs_wait_buftarg(mp->m_ddev_targp);
786 mp->m_flags |= XFS_MOUNT_RDONLY;
792 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
796 xfs_log_mount_cancel(
797 struct xfs_mount *mp)
799 xlog_recover_cancel(mp->m_log);
804 * Final log writes as part of unmount.
806 * Mark the filesystem clean as unmount happens. Note that during relocation
807 * this routine needs to be executed as part of source-bag while the
808 * deallocation must not be done until source-end.
811 /* Actually write the unmount record to disk. */
813 xfs_log_write_unmount_record(
814 struct xfs_mount *mp)
816 /* the data section must be 32 bit size aligned */
817 struct xfs_unmount_log_format magic = {
818 .magic = XLOG_UNMOUNT_TYPE,
820 struct xfs_log_iovec reg = {
822 .i_len = sizeof(magic),
823 .i_type = XLOG_REG_TYPE_UNMOUNT,
825 struct xfs_log_vec vec = {
829 struct xlog *log = mp->m_log;
830 struct xlog_in_core *iclog;
831 struct xlog_ticket *tic = NULL;
833 uint flags = XLOG_UNMOUNT_TRANS;
836 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
841 * If we think the summary counters are bad, clear the unmount header
842 * flag in the unmount record so that the summary counters will be
843 * recalculated during log recovery at next mount. Refer to
844 * xlog_check_unmount_rec for more details.
846 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
847 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
848 xfs_alert(mp, "%s: will fix summary counters at next mount",
850 flags &= ~XLOG_UNMOUNT_TRANS;
853 /* remove inited flag, and account for space used */
855 tic->t_curr_res -= sizeof(magic);
856 error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
858 * At this point, we're umounting anyway, so there's no point in
859 * transitioning log state to IOERROR. Just continue...
863 xfs_alert(mp, "%s: unmount record failed", __func__);
865 spin_lock(&log->l_icloglock);
866 iclog = log->l_iclog;
867 atomic_inc(&iclog->ic_refcnt);
868 xlog_state_want_sync(log, iclog);
869 spin_unlock(&log->l_icloglock);
870 error = xlog_state_release_iclog(log, iclog);
872 spin_lock(&log->l_icloglock);
873 switch (iclog->ic_state) {
875 if (!XLOG_FORCED_SHUTDOWN(log)) {
876 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
880 case XLOG_STATE_ACTIVE:
881 case XLOG_STATE_DIRTY:
882 spin_unlock(&log->l_icloglock);
887 trace_xfs_log_umount_write(log, tic);
888 xlog_ungrant_log_space(log, tic);
889 xfs_log_ticket_put(tic);
894 * Unmount record used to have a string "Unmount filesystem--" in the
895 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
896 * We just write the magic number now since that particular field isn't
897 * currently architecture converted and "Unmount" is a bit foo.
898 * As far as I know, there weren't any dependencies on the old behaviour.
902 xfs_log_unmount_write(xfs_mount_t *mp)
904 struct xlog *log = mp->m_log;
905 xlog_in_core_t *iclog;
907 xlog_in_core_t *first_iclog;
912 * Don't write out unmount record on norecovery mounts or ro devices.
913 * Or, if we are doing a forced umount (typically because of IO errors).
915 if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
916 xfs_readonly_buftarg(log->l_targ)) {
917 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
921 error = xfs_log_force(mp, XFS_LOG_SYNC);
922 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
925 first_iclog = iclog = log->l_iclog;
927 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
928 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
929 ASSERT(iclog->ic_offset == 0);
931 iclog = iclog->ic_next;
932 } while (iclog != first_iclog);
934 if (! (XLOG_FORCED_SHUTDOWN(log))) {
935 xfs_log_write_unmount_record(mp);
938 * We're already in forced_shutdown mode, couldn't
939 * even attempt to write out the unmount transaction.
941 * Go through the motions of sync'ing and releasing
942 * the iclog, even though no I/O will actually happen,
943 * we need to wait for other log I/Os that may already
944 * be in progress. Do this as a separate section of
945 * code so we'll know if we ever get stuck here that
946 * we're in this odd situation of trying to unmount
947 * a file system that went into forced_shutdown as
948 * the result of an unmount..
950 spin_lock(&log->l_icloglock);
951 iclog = log->l_iclog;
952 atomic_inc(&iclog->ic_refcnt);
954 xlog_state_want_sync(log, iclog);
955 spin_unlock(&log->l_icloglock);
956 error = xlog_state_release_iclog(log, iclog);
958 spin_lock(&log->l_icloglock);
960 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
961 || iclog->ic_state == XLOG_STATE_DIRTY
962 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
964 xlog_wait(&iclog->ic_force_wait,
967 spin_unlock(&log->l_icloglock);
972 } /* xfs_log_unmount_write */
975 * Empty the log for unmount/freeze.
977 * To do this, we first need to shut down the background log work so it is not
978 * trying to cover the log as we clean up. We then need to unpin all objects in
979 * the log so we can then flush them out. Once they have completed their IO and
980 * run the callbacks removing themselves from the AIL, we can write the unmount
985 struct xfs_mount *mp)
987 cancel_delayed_work_sync(&mp->m_log->l_work);
988 xfs_log_force(mp, XFS_LOG_SYNC);
991 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
992 * will push it, xfs_wait_buftarg() will not wait for it. Further,
993 * xfs_buf_iowait() cannot be used because it was pushed with the
994 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
995 * the IO to complete.
997 xfs_ail_push_all_sync(mp->m_ail);
998 xfs_wait_buftarg(mp->m_ddev_targp);
999 xfs_buf_lock(mp->m_sb_bp);
1000 xfs_buf_unlock(mp->m_sb_bp);
1002 xfs_log_unmount_write(mp);
1006 * Shut down and release the AIL and Log.
1008 * During unmount, we need to ensure we flush all the dirty metadata objects
1009 * from the AIL so that the log is empty before we write the unmount record to
1010 * the log. Once this is done, we can tear down the AIL and the log.
1014 struct xfs_mount *mp)
1016 xfs_log_quiesce(mp);
1018 xfs_trans_ail_destroy(mp);
1020 xfs_sysfs_del(&mp->m_log->l_kobj);
1022 xlog_dealloc_log(mp->m_log);
1027 struct xfs_mount *mp,
1028 struct xfs_log_item *item,
1030 const struct xfs_item_ops *ops)
1032 item->li_mountp = mp;
1033 item->li_ailp = mp->m_ail;
1034 item->li_type = type;
1038 INIT_LIST_HEAD(&item->li_ail);
1039 INIT_LIST_HEAD(&item->li_cil);
1040 INIT_LIST_HEAD(&item->li_bio_list);
1041 INIT_LIST_HEAD(&item->li_trans);
1045 * Wake up processes waiting for log space after we have moved the log tail.
1049 struct xfs_mount *mp)
1051 struct xlog *log = mp->m_log;
1054 if (XLOG_FORCED_SHUTDOWN(log))
1057 if (!list_empty_careful(&log->l_write_head.waiters)) {
1058 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1060 spin_lock(&log->l_write_head.lock);
1061 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1062 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1063 spin_unlock(&log->l_write_head.lock);
1066 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1067 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1069 spin_lock(&log->l_reserve_head.lock);
1070 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1071 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1072 spin_unlock(&log->l_reserve_head.lock);
1077 * Determine if we have a transaction that has gone to disk that needs to be
1078 * covered. To begin the transition to the idle state firstly the log needs to
1079 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1080 * we start attempting to cover the log.
1082 * Only if we are then in a state where covering is needed, the caller is
1083 * informed that dummy transactions are required to move the log into the idle
1086 * If there are any items in the AIl or CIL, then we do not want to attempt to
1087 * cover the log as we may be in a situation where there isn't log space
1088 * available to run a dummy transaction and this can lead to deadlocks when the
1089 * tail of the log is pinned by an item that is modified in the CIL. Hence
1090 * there's no point in running a dummy transaction at this point because we
1091 * can't start trying to idle the log until both the CIL and AIL are empty.
1094 xfs_log_need_covered(xfs_mount_t *mp)
1096 struct xlog *log = mp->m_log;
1099 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1102 if (!xlog_cil_empty(log))
1105 spin_lock(&log->l_icloglock);
1106 switch (log->l_covered_state) {
1107 case XLOG_STATE_COVER_DONE:
1108 case XLOG_STATE_COVER_DONE2:
1109 case XLOG_STATE_COVER_IDLE:
1111 case XLOG_STATE_COVER_NEED:
1112 case XLOG_STATE_COVER_NEED2:
1113 if (xfs_ail_min_lsn(log->l_ailp))
1115 if (!xlog_iclogs_empty(log))
1119 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1120 log->l_covered_state = XLOG_STATE_COVER_DONE;
1122 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1128 spin_unlock(&log->l_icloglock);
1133 * We may be holding the log iclog lock upon entering this routine.
1136 xlog_assign_tail_lsn_locked(
1137 struct xfs_mount *mp)
1139 struct xlog *log = mp->m_log;
1140 struct xfs_log_item *lip;
1143 assert_spin_locked(&mp->m_ail->ail_lock);
1146 * To make sure we always have a valid LSN for the log tail we keep
1147 * track of the last LSN which was committed in log->l_last_sync_lsn,
1148 * and use that when the AIL was empty.
1150 lip = xfs_ail_min(mp->m_ail);
1152 tail_lsn = lip->li_lsn;
1154 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1155 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1156 atomic64_set(&log->l_tail_lsn, tail_lsn);
1161 xlog_assign_tail_lsn(
1162 struct xfs_mount *mp)
1166 spin_lock(&mp->m_ail->ail_lock);
1167 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1168 spin_unlock(&mp->m_ail->ail_lock);
1174 * Return the space in the log between the tail and the head. The head
1175 * is passed in the cycle/bytes formal parms. In the special case where
1176 * the reserve head has wrapped passed the tail, this calculation is no
1177 * longer valid. In this case, just return 0 which means there is no space
1178 * in the log. This works for all places where this function is called
1179 * with the reserve head. Of course, if the write head were to ever
1180 * wrap the tail, we should blow up. Rather than catch this case here,
1181 * we depend on other ASSERTions in other parts of the code. XXXmiken
1183 * This code also handles the case where the reservation head is behind
1184 * the tail. The details of this case are described below, but the end
1185 * result is that we return the size of the log as the amount of space left.
1198 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1199 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1200 tail_bytes = BBTOB(tail_bytes);
1201 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1202 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1203 else if (tail_cycle + 1 < head_cycle)
1205 else if (tail_cycle < head_cycle) {
1206 ASSERT(tail_cycle == (head_cycle - 1));
1207 free_bytes = tail_bytes - head_bytes;
1210 * The reservation head is behind the tail.
1211 * In this case we just want to return the size of the
1212 * log as the amount of space left.
1214 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1215 xfs_alert(log->l_mp,
1216 " tail_cycle = %d, tail_bytes = %d",
1217 tail_cycle, tail_bytes);
1218 xfs_alert(log->l_mp,
1219 " GH cycle = %d, GH bytes = %d",
1220 head_cycle, head_bytes);
1222 free_bytes = log->l_logsize;
1230 struct work_struct *work)
1232 struct xlog_in_core *iclog =
1233 container_of(work, struct xlog_in_core, ic_end_io_work);
1234 struct xlog *log = iclog->ic_log;
1235 bool aborted = false;
1238 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1240 /* treat writes with injected CRC errors as failed */
1241 if (iclog->ic_fail_crc)
1246 * Race to shutdown the filesystem if we see an error.
1248 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1249 xfs_alert(log->l_mp, "log I/O error %d", error);
1250 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1252 * This flag will be propagated to the trans-committed
1253 * callback routines to let them know that the log-commit
1257 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1261 xlog_state_done_syncing(iclog, aborted);
1262 bio_uninit(&iclog->ic_bio);
1265 * Drop the lock to signal that we are done. Nothing references the
1266 * iclog after this, so an unmount waiting on this lock can now tear it
1267 * down safely. As such, it is unsafe to reference the iclog after the
1268 * unlock as we could race with it being freed.
1270 up(&iclog->ic_sema);
1274 * Return size of each in-core log record buffer.
1276 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1278 * If the filesystem blocksize is too large, we may need to choose a
1279 * larger size since the directory code currently logs entire blocks.
1282 xlog_get_iclog_buffer_size(
1283 struct xfs_mount *mp,
1286 if (mp->m_logbufs <= 0)
1287 mp->m_logbufs = XLOG_MAX_ICLOGS;
1288 if (mp->m_logbsize <= 0)
1289 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1291 log->l_iclog_bufs = mp->m_logbufs;
1292 log->l_iclog_size = mp->m_logbsize;
1295 * # headers = size / 32k - one header holds cycles from 32k of data.
1297 log->l_iclog_heads =
1298 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1299 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1304 struct xfs_mount *mp)
1306 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1307 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1311 * Every sync period we need to unpin all items in the AIL and push them to
1312 * disk. If there is nothing dirty, then we might need to cover the log to
1313 * indicate that the filesystem is idle.
1317 struct work_struct *work)
1319 struct xlog *log = container_of(to_delayed_work(work),
1320 struct xlog, l_work);
1321 struct xfs_mount *mp = log->l_mp;
1323 /* dgc: errors ignored - not fatal and nowhere to report them */
1324 if (xfs_log_need_covered(mp)) {
1326 * Dump a transaction into the log that contains no real change.
1327 * This is needed to stamp the current tail LSN into the log
1328 * during the covering operation.
1330 * We cannot use an inode here for this - that will push dirty
1331 * state back up into the VFS and then periodic inode flushing
1332 * will prevent log covering from making progress. Hence we
1333 * synchronously log the superblock instead to ensure the
1334 * superblock is immediately unpinned and can be written back.
1336 xfs_sync_sb(mp, true);
1338 xfs_log_force(mp, 0);
1340 /* start pushing all the metadata that is currently dirty */
1341 xfs_ail_push_all(mp->m_ail);
1343 /* queue us up again */
1344 xfs_log_work_queue(mp);
1348 * This routine initializes some of the log structure for a given mount point.
1349 * Its primary purpose is to fill in enough, so recovery can occur. However,
1350 * some other stuff may be filled in too.
1352 STATIC struct xlog *
1354 struct xfs_mount *mp,
1355 struct xfs_buftarg *log_target,
1356 xfs_daddr_t blk_offset,
1360 xlog_rec_header_t *head;
1361 xlog_in_core_t **iclogp;
1362 xlog_in_core_t *iclog, *prev_iclog=NULL;
1364 int error = -ENOMEM;
1367 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1369 xfs_warn(mp, "Log allocation failed: No memory!");
1374 log->l_targ = log_target;
1375 log->l_logsize = BBTOB(num_bblks);
1376 log->l_logBBstart = blk_offset;
1377 log->l_logBBsize = num_bblks;
1378 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1379 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1380 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1382 log->l_prev_block = -1;
1383 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1384 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1385 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1386 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1388 xlog_grant_head_init(&log->l_reserve_head);
1389 xlog_grant_head_init(&log->l_write_head);
1391 error = -EFSCORRUPTED;
1392 if (xfs_sb_version_hassector(&mp->m_sb)) {
1393 log2_size = mp->m_sb.sb_logsectlog;
1394 if (log2_size < BBSHIFT) {
1395 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1396 log2_size, BBSHIFT);
1400 log2_size -= BBSHIFT;
1401 if (log2_size > mp->m_sectbb_log) {
1402 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1403 log2_size, mp->m_sectbb_log);
1407 /* for larger sector sizes, must have v2 or external log */
1408 if (log2_size && log->l_logBBstart > 0 &&
1409 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1411 "log sector size (0x%x) invalid for configuration.",
1416 log->l_sectBBsize = 1 << log2_size;
1418 xlog_get_iclog_buffer_size(mp, log);
1420 spin_lock_init(&log->l_icloglock);
1421 init_waitqueue_head(&log->l_flush_wait);
1423 iclogp = &log->l_iclog;
1425 * The amount of memory to allocate for the iclog structure is
1426 * rather funky due to the way the structure is defined. It is
1427 * done this way so that we can use different sizes for machines
1428 * with different amounts of memory. See the definition of
1429 * xlog_in_core_t in xfs_log_priv.h for details.
1431 ASSERT(log->l_iclog_size >= 4096);
1432 for (i = 0; i < log->l_iclog_bufs; i++) {
1433 int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1434 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1435 sizeof(struct bio_vec);
1437 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1439 goto out_free_iclog;
1442 iclog->ic_prev = prev_iclog;
1445 iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1447 if (!iclog->ic_data)
1448 goto out_free_iclog;
1450 log->l_iclog_bak[i] = &iclog->ic_header;
1452 head = &iclog->ic_header;
1453 memset(head, 0, sizeof(xlog_rec_header_t));
1454 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1455 head->h_version = cpu_to_be32(
1456 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1457 head->h_size = cpu_to_be32(log->l_iclog_size);
1459 head->h_fmt = cpu_to_be32(XLOG_FMT);
1460 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1462 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1463 iclog->ic_state = XLOG_STATE_ACTIVE;
1464 iclog->ic_log = log;
1465 atomic_set(&iclog->ic_refcnt, 0);
1466 spin_lock_init(&iclog->ic_callback_lock);
1467 INIT_LIST_HEAD(&iclog->ic_callbacks);
1468 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1470 init_waitqueue_head(&iclog->ic_force_wait);
1471 init_waitqueue_head(&iclog->ic_write_wait);
1472 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1473 sema_init(&iclog->ic_sema, 1);
1475 iclogp = &iclog->ic_next;
1477 *iclogp = log->l_iclog; /* complete ring */
1478 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1480 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1481 WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0,
1483 if (!log->l_ioend_workqueue)
1484 goto out_free_iclog;
1486 error = xlog_cil_init(log);
1488 goto out_destroy_workqueue;
1491 out_destroy_workqueue:
1492 destroy_workqueue(log->l_ioend_workqueue);
1494 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1495 prev_iclog = iclog->ic_next;
1496 kmem_free(iclog->ic_data);
1502 return ERR_PTR(error);
1503 } /* xlog_alloc_log */
1507 * Write out the commit record of a transaction associated with the given
1508 * ticket. Return the lsn of the commit record.
1513 struct xlog_ticket *ticket,
1514 struct xlog_in_core **iclog,
1515 xfs_lsn_t *commitlsnp)
1517 struct xfs_mount *mp = log->l_mp;
1519 struct xfs_log_iovec reg = {
1522 .i_type = XLOG_REG_TYPE_COMMIT,
1524 struct xfs_log_vec vec = {
1529 ASSERT_ALWAYS(iclog);
1530 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1533 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1538 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1539 * log space. This code pushes on the lsn which would supposedly free up
1540 * the 25% which we want to leave free. We may need to adopt a policy which
1541 * pushes on an lsn which is further along in the log once we reach the high
1542 * water mark. In this manner, we would be creating a low water mark.
1545 xlog_grant_push_ail(
1549 xfs_lsn_t threshold_lsn = 0;
1550 xfs_lsn_t last_sync_lsn;
1553 int threshold_block;
1554 int threshold_cycle;
1557 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1559 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1560 free_blocks = BTOBBT(free_bytes);
1563 * Set the threshold for the minimum number of free blocks in the
1564 * log to the maximum of what the caller needs, one quarter of the
1565 * log, and 256 blocks.
1567 free_threshold = BTOBB(need_bytes);
1568 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1569 free_threshold = max(free_threshold, 256);
1570 if (free_blocks >= free_threshold)
1573 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1575 threshold_block += free_threshold;
1576 if (threshold_block >= log->l_logBBsize) {
1577 threshold_block -= log->l_logBBsize;
1578 threshold_cycle += 1;
1580 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1583 * Don't pass in an lsn greater than the lsn of the last
1584 * log record known to be on disk. Use a snapshot of the last sync lsn
1585 * so that it doesn't change between the compare and the set.
1587 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1588 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1589 threshold_lsn = last_sync_lsn;
1592 * Get the transaction layer to kick the dirty buffers out to
1593 * disk asynchronously. No point in trying to do this if
1594 * the filesystem is shutting down.
1596 if (!XLOG_FORCED_SHUTDOWN(log))
1597 xfs_ail_push(log->l_ailp, threshold_lsn);
1601 * Stamp cycle number in every block
1606 struct xlog_in_core *iclog,
1610 int size = iclog->ic_offset + roundoff;
1614 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1616 dp = iclog->ic_datap;
1617 for (i = 0; i < BTOBB(size); i++) {
1618 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1620 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1621 *(__be32 *)dp = cycle_lsn;
1625 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1626 xlog_in_core_2_t *xhdr = iclog->ic_data;
1628 for ( ; i < BTOBB(size); i++) {
1629 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1630 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1631 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1632 *(__be32 *)dp = cycle_lsn;
1636 for (i = 1; i < log->l_iclog_heads; i++)
1637 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1642 * Calculate the checksum for a log buffer.
1644 * This is a little more complicated than it should be because the various
1645 * headers and the actual data are non-contiguous.
1650 struct xlog_rec_header *rhead,
1656 /* first generate the crc for the record header ... */
1657 crc = xfs_start_cksum_update((char *)rhead,
1658 sizeof(struct xlog_rec_header),
1659 offsetof(struct xlog_rec_header, h_crc));
1661 /* ... then for additional cycle data for v2 logs ... */
1662 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1663 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1667 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1668 if (size % XLOG_HEADER_CYCLE_SIZE)
1671 for (i = 1; i < xheads; i++) {
1672 crc = crc32c(crc, &xhdr[i].hic_xheader,
1673 sizeof(struct xlog_rec_ext_header));
1677 /* ... and finally for the payload */
1678 crc = crc32c(crc, dp, size);
1680 return xfs_end_cksum(crc);
1687 struct xlog_in_core *iclog = bio->bi_private;
1689 queue_work(iclog->ic_log->l_ioend_workqueue,
1690 &iclog->ic_end_io_work);
1694 xlog_map_iclog_data(
1700 struct page *page = kmem_to_page(data);
1701 unsigned int off = offset_in_page(data);
1702 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1704 WARN_ON_ONCE(bio_add_page(bio, page, len, off) != len);
1714 struct xlog_in_core *iclog,
1719 ASSERT(bno < log->l_logBBsize);
1722 * We lock the iclogbufs here so that we can serialise against I/O
1723 * completion during unmount. We might be processing a shutdown
1724 * triggered during unmount, and that can occur asynchronously to the
1725 * unmount thread, and hence we need to ensure that completes before
1726 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1727 * across the log IO to archieve that.
1729 down(&iclog->ic_sema);
1730 if (unlikely(iclog->ic_state & XLOG_STATE_IOERROR)) {
1732 * It would seem logical to return EIO here, but we rely on
1733 * the log state machine to propagate I/O errors instead of
1734 * doing it here. We kick of the state machine and unlock
1735 * the buffer manually, the code needs to be kept in sync
1736 * with the I/O completion path.
1738 xlog_state_done_syncing(iclog, XFS_LI_ABORTED);
1739 up(&iclog->ic_sema);
1743 iclog->ic_io_size = count;
1745 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1746 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1747 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1748 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1749 iclog->ic_bio.bi_private = iclog;
1750 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_FUA;
1752 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1754 xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, iclog->ic_io_size);
1755 if (is_vmalloc_addr(iclog->ic_data))
1756 flush_kernel_vmap_range(iclog->ic_data, iclog->ic_io_size);
1759 * If this log buffer would straddle the end of the log we will have
1760 * to split it up into two bios, so that we can continue at the start.
1762 if (bno + BTOBB(count) > log->l_logBBsize) {
1765 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1766 GFP_NOIO, &fs_bio_set);
1767 bio_chain(split, &iclog->ic_bio);
1770 /* restart at logical offset zero for the remainder */
1771 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1774 submit_bio(&iclog->ic_bio);
1778 * We need to bump cycle number for the part of the iclog that is
1779 * written to the start of the log. Watch out for the header magic
1780 * number case, though.
1789 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1792 for (i = split_offset; i < count; i += BBSIZE) {
1793 uint32_t cycle = get_unaligned_be32(data + i);
1795 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1797 put_unaligned_be32(cycle, data + i);
1802 xlog_calc_iclog_size(
1804 struct xlog_in_core *iclog,
1807 uint32_t count_init, count;
1810 use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1811 log->l_mp->m_sb.sb_logsunit > 1;
1813 /* Add for LR header */
1814 count_init = log->l_iclog_hsize + iclog->ic_offset;
1816 /* Round out the log write size */
1818 /* we have a v2 stripe unit to use */
1819 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1821 count = BBTOB(BTOBB(count_init));
1824 ASSERT(count >= count_init);
1825 *roundoff = count - count_init;
1828 ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1830 ASSERT(*roundoff < BBTOB(1));
1835 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1836 * fashion. Previously, we should have moved the current iclog
1837 * ptr in the log to point to the next available iclog. This allows further
1838 * write to continue while this code syncs out an iclog ready to go.
1839 * Before an in-core log can be written out, the data section must be scanned
1840 * to save away the 1st word of each BBSIZE block into the header. We replace
1841 * it with the current cycle count. Each BBSIZE block is tagged with the
1842 * cycle count because there in an implicit assumption that drives will
1843 * guarantee that entire 512 byte blocks get written at once. In other words,
1844 * we can't have part of a 512 byte block written and part not written. By
1845 * tagging each block, we will know which blocks are valid when recovering
1846 * after an unclean shutdown.
1848 * This routine is single threaded on the iclog. No other thread can be in
1849 * this routine with the same iclog. Changing contents of iclog can there-
1850 * fore be done without grabbing the state machine lock. Updating the global
1851 * log will require grabbing the lock though.
1853 * The entire log manager uses a logical block numbering scheme. Only
1854 * xlog_write_iclog knows about the fact that the log may not start with
1855 * block zero on a given device.
1860 struct xlog_in_core *iclog)
1862 unsigned int count; /* byte count of bwrite */
1863 unsigned int roundoff; /* roundoff to BB or stripe */
1866 bool need_flush = true, split = false;
1868 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1870 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1872 /* move grant heads by roundoff in sync */
1873 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1874 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1876 /* put cycle number in every block */
1877 xlog_pack_data(log, iclog, roundoff);
1879 /* real byte length */
1880 size = iclog->ic_offset;
1881 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1883 iclog->ic_header.h_len = cpu_to_be32(size);
1885 XFS_STATS_INC(log->l_mp, xs_log_writes);
1886 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1888 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1890 /* Do we need to split this write into 2 parts? */
1891 if (bno + BTOBB(count) > log->l_logBBsize) {
1892 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1896 /* calculcate the checksum */
1897 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1898 iclog->ic_datap, size);
1900 * Intentionally corrupt the log record CRC based on the error injection
1901 * frequency, if defined. This facilitates testing log recovery in the
1902 * event of torn writes. Hence, set the IOABORT state to abort the log
1903 * write on I/O completion and shutdown the fs. The subsequent mount
1904 * detects the bad CRC and attempts to recover.
1907 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1908 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1909 iclog->ic_fail_crc = true;
1911 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1912 be64_to_cpu(iclog->ic_header.h_lsn));
1917 * Flush the data device before flushing the log to make sure all meta
1918 * data written back from the AIL actually made it to disk before
1919 * stamping the new log tail LSN into the log buffer. For an external
1920 * log we need to issue the flush explicitly, and unfortunately
1921 * synchronously here; for an internal log we can simply use the block
1922 * layer state machine for preflushes.
1924 if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1925 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1929 xlog_verify_iclog(log, iclog, count);
1930 xlog_write_iclog(log, iclog, bno, count, need_flush);
1934 * Deallocate a log structure
1940 xlog_in_core_t *iclog, *next_iclog;
1943 xlog_cil_destroy(log);
1946 * Cycle all the iclogbuf locks to make sure all log IO completion
1947 * is done before we tear down these buffers.
1949 iclog = log->l_iclog;
1950 for (i = 0; i < log->l_iclog_bufs; i++) {
1951 down(&iclog->ic_sema);
1952 up(&iclog->ic_sema);
1953 iclog = iclog->ic_next;
1956 iclog = log->l_iclog;
1957 for (i = 0; i < log->l_iclog_bufs; i++) {
1958 next_iclog = iclog->ic_next;
1959 kmem_free(iclog->ic_data);
1964 log->l_mp->m_log = NULL;
1965 destroy_workqueue(log->l_ioend_workqueue);
1967 } /* xlog_dealloc_log */
1970 * Update counters atomically now that memcpy is done.
1974 xlog_state_finish_copy(
1976 struct xlog_in_core *iclog,
1980 spin_lock(&log->l_icloglock);
1982 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1983 iclog->ic_offset += copy_bytes;
1985 spin_unlock(&log->l_icloglock);
1986 } /* xlog_state_finish_copy */
1992 * print out info relating to regions written which consume
1997 struct xfs_mount *mp,
1998 struct xlog_ticket *ticket)
2001 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2003 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2004 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2005 static char *res_type_str[] = {
2006 REG_TYPE_STR(BFORMAT, "bformat"),
2007 REG_TYPE_STR(BCHUNK, "bchunk"),
2008 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2009 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2010 REG_TYPE_STR(IFORMAT, "iformat"),
2011 REG_TYPE_STR(ICORE, "icore"),
2012 REG_TYPE_STR(IEXT, "iext"),
2013 REG_TYPE_STR(IBROOT, "ibroot"),
2014 REG_TYPE_STR(ILOCAL, "ilocal"),
2015 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2016 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2017 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2018 REG_TYPE_STR(QFORMAT, "qformat"),
2019 REG_TYPE_STR(DQUOT, "dquot"),
2020 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2021 REG_TYPE_STR(LRHEADER, "LR header"),
2022 REG_TYPE_STR(UNMOUNT, "unmount"),
2023 REG_TYPE_STR(COMMIT, "commit"),
2024 REG_TYPE_STR(TRANSHDR, "trans header"),
2025 REG_TYPE_STR(ICREATE, "inode create"),
2026 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2027 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2028 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2029 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2030 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2031 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2033 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2036 xfs_warn(mp, "ticket reservation summary:");
2037 xfs_warn(mp, " unit res = %d bytes",
2038 ticket->t_unit_res);
2039 xfs_warn(mp, " current res = %d bytes",
2040 ticket->t_curr_res);
2041 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2042 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2043 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2044 ticket->t_res_num_ophdrs, ophdr_spc);
2045 xfs_warn(mp, " ophdr + reg = %u bytes",
2046 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2047 xfs_warn(mp, " num regions = %u",
2050 for (i = 0; i < ticket->t_res_num; i++) {
2051 uint r_type = ticket->t_res_arr[i].r_type;
2052 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2053 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2054 "bad-rtype" : res_type_str[r_type]),
2055 ticket->t_res_arr[i].r_len);
2060 * Print a summary of the transaction.
2064 struct xfs_trans *tp)
2066 struct xfs_mount *mp = tp->t_mountp;
2067 struct xfs_log_item *lip;
2069 /* dump core transaction and ticket info */
2070 xfs_warn(mp, "transaction summary:");
2071 xfs_warn(mp, " log res = %d", tp->t_log_res);
2072 xfs_warn(mp, " log count = %d", tp->t_log_count);
2073 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2075 xlog_print_tic_res(mp, tp->t_ticket);
2077 /* dump each log item */
2078 list_for_each_entry(lip, &tp->t_items, li_trans) {
2079 struct xfs_log_vec *lv = lip->li_lv;
2080 struct xfs_log_iovec *vec;
2083 xfs_warn(mp, "log item: ");
2084 xfs_warn(mp, " type = 0x%x", lip->li_type);
2085 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2088 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2089 xfs_warn(mp, " size = %d", lv->lv_size);
2090 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2091 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2093 /* dump each iovec for the log item */
2094 vec = lv->lv_iovecp;
2095 for (i = 0; i < lv->lv_niovecs; i++) {
2096 int dumplen = min(vec->i_len, 32);
2098 xfs_warn(mp, " iovec[%d]", i);
2099 xfs_warn(mp, " type = 0x%x", vec->i_type);
2100 xfs_warn(mp, " len = %d", vec->i_len);
2101 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2102 xfs_hex_dump(vec->i_addr, dumplen);
2110 * Calculate the potential space needed by the log vector. Each region gets
2111 * its own xlog_op_header_t and may need to be double word aligned.
2114 xlog_write_calc_vec_length(
2115 struct xlog_ticket *ticket,
2116 struct xfs_log_vec *log_vector)
2118 struct xfs_log_vec *lv;
2123 /* acct for start rec of xact */
2124 if (ticket->t_flags & XLOG_TIC_INITED)
2127 for (lv = log_vector; lv; lv = lv->lv_next) {
2128 /* we don't write ordered log vectors */
2129 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2132 headers += lv->lv_niovecs;
2134 for (i = 0; i < lv->lv_niovecs; i++) {
2135 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2138 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2142 ticket->t_res_num_ophdrs += headers;
2143 len += headers * sizeof(struct xlog_op_header);
2149 * If first write for transaction, insert start record We can't be trying to
2150 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2153 xlog_write_start_rec(
2154 struct xlog_op_header *ophdr,
2155 struct xlog_ticket *ticket)
2157 if (!(ticket->t_flags & XLOG_TIC_INITED))
2160 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2161 ophdr->oh_clientid = ticket->t_clientid;
2163 ophdr->oh_flags = XLOG_START_TRANS;
2166 ticket->t_flags &= ~XLOG_TIC_INITED;
2168 return sizeof(struct xlog_op_header);
2171 static xlog_op_header_t *
2172 xlog_write_setup_ophdr(
2174 struct xlog_op_header *ophdr,
2175 struct xlog_ticket *ticket,
2178 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2179 ophdr->oh_clientid = ticket->t_clientid;
2182 /* are we copying a commit or unmount record? */
2183 ophdr->oh_flags = flags;
2186 * We've seen logs corrupted with bad transaction client ids. This
2187 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2188 * and shut down the filesystem.
2190 switch (ophdr->oh_clientid) {
2191 case XFS_TRANSACTION:
2197 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2198 ophdr->oh_clientid, ticket);
2206 * Set up the parameters of the region copy into the log. This has
2207 * to handle region write split across multiple log buffers - this
2208 * state is kept external to this function so that this code can
2209 * be written in an obvious, self documenting manner.
2212 xlog_write_setup_copy(
2213 struct xlog_ticket *ticket,
2214 struct xlog_op_header *ophdr,
2215 int space_available,
2219 int *last_was_partial_copy,
2220 int *bytes_consumed)
2224 still_to_copy = space_required - *bytes_consumed;
2225 *copy_off = *bytes_consumed;
2227 if (still_to_copy <= space_available) {
2228 /* write of region completes here */
2229 *copy_len = still_to_copy;
2230 ophdr->oh_len = cpu_to_be32(*copy_len);
2231 if (*last_was_partial_copy)
2232 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2233 *last_was_partial_copy = 0;
2234 *bytes_consumed = 0;
2238 /* partial write of region, needs extra log op header reservation */
2239 *copy_len = space_available;
2240 ophdr->oh_len = cpu_to_be32(*copy_len);
2241 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2242 if (*last_was_partial_copy)
2243 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2244 *bytes_consumed += *copy_len;
2245 (*last_was_partial_copy)++;
2247 /* account for new log op header */
2248 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2249 ticket->t_res_num_ophdrs++;
2251 return sizeof(struct xlog_op_header);
2255 xlog_write_copy_finish(
2257 struct xlog_in_core *iclog,
2262 int *partial_copy_len,
2264 struct xlog_in_core **commit_iclog)
2266 if (*partial_copy) {
2268 * This iclog has already been marked WANT_SYNC by
2269 * xlog_state_get_iclog_space.
2271 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2274 return xlog_state_release_iclog(log, iclog);
2278 *partial_copy_len = 0;
2280 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2281 /* no more space in this iclog - push it. */
2282 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2286 spin_lock(&log->l_icloglock);
2287 xlog_state_want_sync(log, iclog);
2288 spin_unlock(&log->l_icloglock);
2291 return xlog_state_release_iclog(log, iclog);
2292 ASSERT(flags & XLOG_COMMIT_TRANS);
2293 *commit_iclog = iclog;
2300 * Write some region out to in-core log
2302 * This will be called when writing externally provided regions or when
2303 * writing out a commit record for a given transaction.
2305 * General algorithm:
2306 * 1. Find total length of this write. This may include adding to the
2307 * lengths passed in.
2308 * 2. Check whether we violate the tickets reservation.
2309 * 3. While writing to this iclog
2310 * A. Reserve as much space in this iclog as can get
2311 * B. If this is first write, save away start lsn
2312 * C. While writing this region:
2313 * 1. If first write of transaction, write start record
2314 * 2. Write log operation header (header per region)
2315 * 3. Find out if we can fit entire region into this iclog
2316 * 4. Potentially, verify destination memcpy ptr
2317 * 5. Memcpy (partial) region
2318 * 6. If partial copy, release iclog; otherwise, continue
2319 * copying more regions into current iclog
2320 * 4. Mark want sync bit (in simulation mode)
2321 * 5. Release iclog for potential flush to on-disk log.
2324 * 1. Panic if reservation is overrun. This should never happen since
2325 * reservation amounts are generated internal to the filesystem.
2327 * 1. Tickets are single threaded data structures.
2328 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2329 * syncing routine. When a single log_write region needs to span
2330 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2331 * on all log operation writes which don't contain the end of the
2332 * region. The XLOG_END_TRANS bit is used for the in-core log
2333 * operation which contains the end of the continued log_write region.
2334 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2335 * we don't really know exactly how much space will be used. As a result,
2336 * we don't update ic_offset until the end when we know exactly how many
2337 * bytes have been written out.
2342 struct xfs_log_vec *log_vector,
2343 struct xlog_ticket *ticket,
2344 xfs_lsn_t *start_lsn,
2345 struct xlog_in_core **commit_iclog,
2348 struct xlog_in_core *iclog = NULL;
2349 struct xfs_log_iovec *vecp;
2350 struct xfs_log_vec *lv;
2353 int partial_copy = 0;
2354 int partial_copy_len = 0;
2362 len = xlog_write_calc_vec_length(ticket, log_vector);
2365 * Region headers and bytes are already accounted for.
2366 * We only need to take into account start records and
2367 * split regions in this function.
2369 if (ticket->t_flags & XLOG_TIC_INITED)
2370 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2373 * Commit record headers need to be accounted for. These
2374 * come in as separate writes so are easy to detect.
2376 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2377 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2379 if (ticket->t_curr_res < 0) {
2380 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2381 "ctx ticket reservation ran out. Need to up reservation");
2382 xlog_print_tic_res(log->l_mp, ticket);
2383 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2388 vecp = lv->lv_iovecp;
2389 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2393 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2394 &contwr, &log_offset);
2398 ASSERT(log_offset <= iclog->ic_size - 1);
2399 ptr = iclog->ic_datap + log_offset;
2401 /* start_lsn is the first lsn written to. That's all we need. */
2403 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2406 * This loop writes out as many regions as can fit in the amount
2407 * of space which was allocated by xlog_state_get_iclog_space().
2409 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2410 struct xfs_log_iovec *reg;
2411 struct xlog_op_header *ophdr;
2415 bool ordered = false;
2417 /* ordered log vectors have no regions to write */
2418 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2419 ASSERT(lv->lv_niovecs == 0);
2425 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2426 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2428 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2429 if (start_rec_copy) {
2431 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2435 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2439 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2440 sizeof(struct xlog_op_header));
2442 len += xlog_write_setup_copy(ticket, ophdr,
2443 iclog->ic_size-log_offset,
2445 ©_off, ©_len,
2448 xlog_verify_dest_ptr(log, ptr);
2453 * Unmount records just log an opheader, so can have
2454 * empty payloads with no data region to copy. Hence we
2455 * only copy the payload if the vector says it has data
2458 ASSERT(copy_len >= 0);
2460 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2461 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2464 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2466 data_cnt += contwr ? copy_len : 0;
2468 error = xlog_write_copy_finish(log, iclog, flags,
2469 &record_cnt, &data_cnt,
2478 * if we had a partial copy, we need to get more iclog
2479 * space but we don't want to increment the region
2480 * index because there is still more is this region to
2483 * If we completed writing this region, and we flushed
2484 * the iclog (indicated by resetting of the record
2485 * count), then we also need to get more log space. If
2486 * this was the last record, though, we are done and
2492 if (++index == lv->lv_niovecs) {
2497 vecp = lv->lv_iovecp;
2499 if (record_cnt == 0 && !ordered) {
2509 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2511 return xlog_state_release_iclog(log, iclog);
2513 ASSERT(flags & XLOG_COMMIT_TRANS);
2514 *commit_iclog = iclog;
2519 /*****************************************************************************
2521 * State Machine functions
2523 *****************************************************************************
2526 /* Clean iclogs starting from the head. This ordering must be
2527 * maintained, so an iclog doesn't become ACTIVE beyond one that
2528 * is SYNCING. This is also required to maintain the notion that we use
2529 * a ordered wait queue to hold off would be writers to the log when every
2530 * iclog is trying to sync to disk.
2532 * State Change: DIRTY -> ACTIVE
2535 xlog_state_clean_log(
2538 xlog_in_core_t *iclog;
2541 iclog = log->l_iclog;
2543 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2544 iclog->ic_state = XLOG_STATE_ACTIVE;
2545 iclog->ic_offset = 0;
2546 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2548 * If the number of ops in this iclog indicate it just
2549 * contains the dummy transaction, we can
2550 * change state into IDLE (the second time around).
2551 * Otherwise we should change the state into
2553 * We don't need to cover the dummy.
2556 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2561 * We have two dirty iclogs so start over
2562 * This could also be num of ops indicates
2563 * this is not the dummy going out.
2567 iclog->ic_header.h_num_logops = 0;
2568 memset(iclog->ic_header.h_cycle_data, 0,
2569 sizeof(iclog->ic_header.h_cycle_data));
2570 iclog->ic_header.h_lsn = 0;
2571 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2574 break; /* stop cleaning */
2575 iclog = iclog->ic_next;
2576 } while (iclog != log->l_iclog);
2578 /* log is locked when we are called */
2580 * Change state for the dummy log recording.
2581 * We usually go to NEED. But we go to NEED2 if the changed indicates
2582 * we are done writing the dummy record.
2583 * If we are done with the second dummy recored (DONE2), then
2587 switch (log->l_covered_state) {
2588 case XLOG_STATE_COVER_IDLE:
2589 case XLOG_STATE_COVER_NEED:
2590 case XLOG_STATE_COVER_NEED2:
2591 log->l_covered_state = XLOG_STATE_COVER_NEED;
2594 case XLOG_STATE_COVER_DONE:
2596 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2598 log->l_covered_state = XLOG_STATE_COVER_NEED;
2601 case XLOG_STATE_COVER_DONE2:
2603 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2605 log->l_covered_state = XLOG_STATE_COVER_NEED;
2612 } /* xlog_state_clean_log */
2615 xlog_get_lowest_lsn(
2618 struct xlog_in_core *iclog = log->l_iclog;
2619 xfs_lsn_t lowest_lsn = 0, lsn;
2622 if (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))
2625 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2626 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2628 } while ((iclog = iclog->ic_next) != log->l_iclog);
2634 xlog_state_do_callback(
2637 struct xlog_in_core *ciclog)
2639 xlog_in_core_t *iclog;
2640 xlog_in_core_t *first_iclog; /* used to know when we've
2641 * processed all iclogs once */
2643 xfs_lsn_t lowest_lsn;
2644 int ioerrors; /* counter: iclogs with errors */
2645 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2646 int funcdidcallbacks; /* flag: function did callbacks */
2647 int repeats; /* for issuing console warnings if
2648 * looping too many times */
2651 spin_lock(&log->l_icloglock);
2652 first_iclog = iclog = log->l_iclog;
2654 funcdidcallbacks = 0;
2659 * Scan all iclogs starting with the one pointed to by the
2660 * log. Reset this starting point each time the log is
2661 * unlocked (during callbacks).
2663 * Keep looping through iclogs until one full pass is made
2664 * without running any callbacks.
2666 first_iclog = log->l_iclog;
2667 iclog = log->l_iclog;
2668 loopdidcallbacks = 0;
2673 /* skip all iclogs in the ACTIVE & DIRTY states */
2674 if (iclog->ic_state &
2675 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2676 iclog = iclog->ic_next;
2681 * Between marking a filesystem SHUTDOWN and stopping
2682 * the log, we do flush all iclogs to disk (if there
2683 * wasn't a log I/O error). So, we do want things to
2684 * go smoothly in case of just a SHUTDOWN w/o a
2687 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2689 * Can only perform callbacks in order. Since
2690 * this iclog is not in the DONE_SYNC/
2691 * DO_CALLBACK state, we skip the rest and
2692 * just try to clean up. If we set our iclog
2693 * to DO_CALLBACK, we will not process it when
2694 * we retry since a previous iclog is in the
2695 * CALLBACK and the state cannot change since
2696 * we are holding the l_icloglock.
2698 if (!(iclog->ic_state &
2699 (XLOG_STATE_DONE_SYNC |
2700 XLOG_STATE_DO_CALLBACK))) {
2701 if (ciclog && (ciclog->ic_state ==
2702 XLOG_STATE_DONE_SYNC)) {
2703 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2708 * We now have an iclog that is in either the
2709 * DO_CALLBACK or DONE_SYNC states. The other
2710 * states (WANT_SYNC, SYNCING, or CALLBACK were
2711 * caught by the above if and are going to
2712 * clean (i.e. we aren't doing their callbacks)
2717 * We will do one more check here to see if we
2718 * have chased our tail around.
2721 lowest_lsn = xlog_get_lowest_lsn(log);
2723 XFS_LSN_CMP(lowest_lsn,
2724 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2725 iclog = iclog->ic_next;
2726 continue; /* Leave this iclog for
2730 iclog->ic_state = XLOG_STATE_CALLBACK;
2734 * Completion of a iclog IO does not imply that
2735 * a transaction has completed, as transactions
2736 * can be large enough to span many iclogs. We
2737 * cannot change the tail of the log half way
2738 * through a transaction as this may be the only
2739 * transaction in the log and moving th etail to
2740 * point to the middle of it will prevent
2741 * recovery from finding the start of the
2742 * transaction. Hence we should only update the
2743 * last_sync_lsn if this iclog contains
2744 * transaction completion callbacks on it.
2746 * We have to do this before we drop the
2747 * icloglock to ensure we are the only one that
2750 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2751 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2752 if (!list_empty_careful(&iclog->ic_callbacks))
2753 atomic64_set(&log->l_last_sync_lsn,
2754 be64_to_cpu(iclog->ic_header.h_lsn));
2759 spin_unlock(&log->l_icloglock);
2762 * Keep processing entries in the callback list until
2763 * we come around and it is empty. We need to
2764 * atomically see that the list is empty and change the
2765 * state to DIRTY so that we don't miss any more
2766 * callbacks being added.
2768 spin_lock(&iclog->ic_callback_lock);
2769 while (!list_empty(&iclog->ic_callbacks)) {
2772 list_splice_init(&iclog->ic_callbacks, &tmp);
2774 spin_unlock(&iclog->ic_callback_lock);
2775 xlog_cil_process_committed(&tmp, aborted);
2776 spin_lock(&iclog->ic_callback_lock);
2782 spin_lock(&log->l_icloglock);
2783 spin_unlock(&iclog->ic_callback_lock);
2784 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2785 iclog->ic_state = XLOG_STATE_DIRTY;
2788 * Transition from DIRTY to ACTIVE if applicable.
2789 * NOP if STATE_IOERROR.
2791 xlog_state_clean_log(log);
2793 /* wake up threads waiting in xfs_log_force() */
2794 wake_up_all(&iclog->ic_force_wait);
2796 iclog = iclog->ic_next;
2797 } while (first_iclog != iclog);
2799 if (repeats > 5000) {
2800 flushcnt += repeats;
2803 "%s: possible infinite loop (%d iterations)",
2804 __func__, flushcnt);
2806 } while (!ioerrors && loopdidcallbacks);
2810 * Make one last gasp attempt to see if iclogs are being left in limbo.
2811 * If the above loop finds an iclog earlier than the current iclog and
2812 * in one of the syncing states, the current iclog is put into
2813 * DO_CALLBACK and the callbacks are deferred to the completion of the
2814 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2815 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2818 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2819 * for ic_state == SYNCING.
2821 if (funcdidcallbacks) {
2822 first_iclog = iclog = log->l_iclog;
2824 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2826 * Terminate the loop if iclogs are found in states
2827 * which will cause other threads to clean up iclogs.
2829 * SYNCING - i/o completion will go through logs
2830 * DONE_SYNC - interrupt thread should be waiting for
2832 * IOERROR - give up hope all ye who enter here
2834 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2835 iclog->ic_state & XLOG_STATE_SYNCING ||
2836 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2837 iclog->ic_state == XLOG_STATE_IOERROR )
2839 iclog = iclog->ic_next;
2840 } while (first_iclog != iclog);
2844 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2846 spin_unlock(&log->l_icloglock);
2849 wake_up_all(&log->l_flush_wait);
2854 * Finish transitioning this iclog to the dirty state.
2856 * Make sure that we completely execute this routine only when this is
2857 * the last call to the iclog. There is a good chance that iclog flushes,
2858 * when we reach the end of the physical log, get turned into 2 separate
2859 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2860 * routine. By using the reference count bwritecnt, we guarantee that only
2861 * the second completion goes through.
2863 * Callbacks could take time, so they are done outside the scope of the
2864 * global state machine log lock.
2867 xlog_state_done_syncing(
2868 struct xlog_in_core *iclog,
2871 struct xlog *log = iclog->ic_log;
2873 spin_lock(&log->l_icloglock);
2875 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2876 iclog->ic_state == XLOG_STATE_IOERROR);
2877 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2880 * If we got an error, either on the first buffer, or in the case of
2881 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2882 * and none should ever be attempted to be written to disk
2885 if (iclog->ic_state != XLOG_STATE_IOERROR)
2886 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2889 * Someone could be sleeping prior to writing out the next
2890 * iclog buffer, we wake them all, one will get to do the
2891 * I/O, the others get to wait for the result.
2893 wake_up_all(&iclog->ic_write_wait);
2894 spin_unlock(&log->l_icloglock);
2895 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2896 } /* xlog_state_done_syncing */
2900 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2901 * sleep. We wait on the flush queue on the head iclog as that should be
2902 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2903 * we will wait here and all new writes will sleep until a sync completes.
2905 * The in-core logs are used in a circular fashion. They are not used
2906 * out-of-order even when an iclog past the head is free.
2909 * * log_offset where xlog_write() can start writing into the in-core
2911 * * in-core log pointer to which xlog_write() should write.
2912 * * boolean indicating this is a continued write to an in-core log.
2913 * If this is the last write, then the in-core log's offset field
2914 * needs to be incremented, depending on the amount of data which
2918 xlog_state_get_iclog_space(
2921 struct xlog_in_core **iclogp,
2922 struct xlog_ticket *ticket,
2923 int *continued_write,
2927 xlog_rec_header_t *head;
2928 xlog_in_core_t *iclog;
2932 spin_lock(&log->l_icloglock);
2933 if (XLOG_FORCED_SHUTDOWN(log)) {
2934 spin_unlock(&log->l_icloglock);
2938 iclog = log->l_iclog;
2939 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2940 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2942 /* Wait for log writes to have flushed */
2943 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2947 head = &iclog->ic_header;
2949 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2950 log_offset = iclog->ic_offset;
2952 /* On the 1st write to an iclog, figure out lsn. This works
2953 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2954 * committing to. If the offset is set, that's how many blocks
2957 if (log_offset == 0) {
2958 ticket->t_curr_res -= log->l_iclog_hsize;
2959 xlog_tic_add_region(ticket,
2961 XLOG_REG_TYPE_LRHEADER);
2962 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2963 head->h_lsn = cpu_to_be64(
2964 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2965 ASSERT(log->l_curr_block >= 0);
2968 /* If there is enough room to write everything, then do it. Otherwise,
2969 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2970 * bit is on, so this will get flushed out. Don't update ic_offset
2971 * until you know exactly how many bytes get copied. Therefore, wait
2972 * until later to update ic_offset.
2974 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2975 * can fit into remaining data section.
2977 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2978 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2981 * If I'm the only one writing to this iclog, sync it to disk.
2982 * We need to do an atomic compare and decrement here to avoid
2983 * racing with concurrent atomic_dec_and_lock() calls in
2984 * xlog_state_release_iclog() when there is more than one
2985 * reference to the iclog.
2987 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2988 /* we are the only one */
2989 spin_unlock(&log->l_icloglock);
2990 error = xlog_state_release_iclog(log, iclog);
2994 spin_unlock(&log->l_icloglock);
2999 /* Do we have enough room to write the full amount in the remainder
3000 * of this iclog? Or must we continue a write on the next iclog and
3001 * mark this iclog as completely taken? In the case where we switch
3002 * iclogs (to mark it taken), this particular iclog will release/sync
3003 * to disk in xlog_write().
3005 if (len <= iclog->ic_size - iclog->ic_offset) {
3006 *continued_write = 0;
3007 iclog->ic_offset += len;
3009 *continued_write = 1;
3010 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3014 ASSERT(iclog->ic_offset <= iclog->ic_size);
3015 spin_unlock(&log->l_icloglock);
3017 *logoffsetp = log_offset;
3019 } /* xlog_state_get_iclog_space */
3021 /* The first cnt-1 times through here we don't need to
3022 * move the grant write head because the permanent
3023 * reservation has reserved cnt times the unit amount.
3024 * Release part of current permanent unit reservation and
3025 * reset current reservation to be one units worth. Also
3026 * move grant reservation head forward.
3029 xlog_regrant_reserve_log_space(
3031 struct xlog_ticket *ticket)
3033 trace_xfs_log_regrant_reserve_enter(log, ticket);
3035 if (ticket->t_cnt > 0)
3038 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3039 ticket->t_curr_res);
3040 xlog_grant_sub_space(log, &log->l_write_head.grant,
3041 ticket->t_curr_res);
3042 ticket->t_curr_res = ticket->t_unit_res;
3043 xlog_tic_reset_res(ticket);
3045 trace_xfs_log_regrant_reserve_sub(log, ticket);
3047 /* just return if we still have some of the pre-reserved space */
3048 if (ticket->t_cnt > 0)
3051 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3052 ticket->t_unit_res);
3054 trace_xfs_log_regrant_reserve_exit(log, ticket);
3056 ticket->t_curr_res = ticket->t_unit_res;
3057 xlog_tic_reset_res(ticket);
3058 } /* xlog_regrant_reserve_log_space */
3062 * Give back the space left from a reservation.
3064 * All the information we need to make a correct determination of space left
3065 * is present. For non-permanent reservations, things are quite easy. The
3066 * count should have been decremented to zero. We only need to deal with the
3067 * space remaining in the current reservation part of the ticket. If the
3068 * ticket contains a permanent reservation, there may be left over space which
3069 * needs to be released. A count of N means that N-1 refills of the current
3070 * reservation can be done before we need to ask for more space. The first
3071 * one goes to fill up the first current reservation. Once we run out of
3072 * space, the count will stay at zero and the only space remaining will be
3073 * in the current reservation field.
3076 xlog_ungrant_log_space(
3078 struct xlog_ticket *ticket)
3082 if (ticket->t_cnt > 0)
3085 trace_xfs_log_ungrant_enter(log, ticket);
3086 trace_xfs_log_ungrant_sub(log, ticket);
3089 * If this is a permanent reservation ticket, we may be able to free
3090 * up more space based on the remaining count.
3092 bytes = ticket->t_curr_res;
3093 if (ticket->t_cnt > 0) {
3094 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3095 bytes += ticket->t_unit_res*ticket->t_cnt;
3098 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3099 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3101 trace_xfs_log_ungrant_exit(log, ticket);
3103 xfs_log_space_wake(log->l_mp);
3107 * Flush iclog to disk if this is the last reference to the given iclog and
3108 * the WANT_SYNC bit is set.
3110 * When this function is entered, the iclog is not necessarily in the
3111 * WANT_SYNC state. It may be sitting around waiting to get filled.
3116 xlog_state_release_iclog(
3118 struct xlog_in_core *iclog)
3120 int sync = 0; /* do we sync? */
3122 if (iclog->ic_state & XLOG_STATE_IOERROR)
3125 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3126 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3129 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3130 spin_unlock(&log->l_icloglock);
3133 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3134 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3136 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3137 /* update tail before writing to iclog */
3138 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3140 iclog->ic_state = XLOG_STATE_SYNCING;
3141 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3142 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3143 /* cycle incremented when incrementing curr_block */
3145 spin_unlock(&log->l_icloglock);
3148 * We let the log lock go, so it's possible that we hit a log I/O
3149 * error or some other SHUTDOWN condition that marks the iclog
3150 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3151 * this iclog has consistent data, so we ignore IOERROR
3152 * flags after this point.
3155 xlog_sync(log, iclog);
3157 } /* xlog_state_release_iclog */
3161 * This routine will mark the current iclog in the ring as WANT_SYNC
3162 * and move the current iclog pointer to the next iclog in the ring.
3163 * When this routine is called from xlog_state_get_iclog_space(), the
3164 * exact size of the iclog has not yet been determined. All we know is
3165 * that every data block. We have run out of space in this log record.
3168 xlog_state_switch_iclogs(
3170 struct xlog_in_core *iclog,
3173 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3175 eventual_size = iclog->ic_offset;
3176 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3177 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3178 log->l_prev_block = log->l_curr_block;
3179 log->l_prev_cycle = log->l_curr_cycle;
3181 /* roll log?: ic_offset changed later */
3182 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3184 /* Round up to next log-sunit */
3185 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3186 log->l_mp->m_sb.sb_logsunit > 1) {
3187 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3188 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3191 if (log->l_curr_block >= log->l_logBBsize) {
3193 * Rewind the current block before the cycle is bumped to make
3194 * sure that the combined LSN never transiently moves forward
3195 * when the log wraps to the next cycle. This is to support the
3196 * unlocked sample of these fields from xlog_valid_lsn(). Most
3197 * other cases should acquire l_icloglock.
3199 log->l_curr_block -= log->l_logBBsize;
3200 ASSERT(log->l_curr_block >= 0);
3202 log->l_curr_cycle++;
3203 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3204 log->l_curr_cycle++;
3206 ASSERT(iclog == log->l_iclog);
3207 log->l_iclog = iclog->ic_next;
3208 } /* xlog_state_switch_iclogs */
3211 * Write out all data in the in-core log as of this exact moment in time.
3213 * Data may be written to the in-core log during this call. However,
3214 * we don't guarantee this data will be written out. A change from past
3215 * implementation means this routine will *not* write out zero length LRs.
3217 * Basically, we try and perform an intelligent scan of the in-core logs.
3218 * If we determine there is no flushable data, we just return. There is no
3219 * flushable data if:
3221 * 1. the current iclog is active and has no data; the previous iclog
3222 * is in the active or dirty state.
3223 * 2. the current iclog is drity, and the previous iclog is in the
3224 * active or dirty state.
3228 * 1. the current iclog is not in the active nor dirty state.
3229 * 2. the current iclog dirty, and the previous iclog is not in the
3230 * active nor dirty state.
3231 * 3. the current iclog is active, and there is another thread writing
3232 * to this particular iclog.
3233 * 4. a) the current iclog is active and has no other writers
3234 * b) when we return from flushing out this iclog, it is still
3235 * not in the active nor dirty state.
3239 struct xfs_mount *mp,
3242 struct xlog *log = mp->m_log;
3243 struct xlog_in_core *iclog;
3246 XFS_STATS_INC(mp, xs_log_force);
3247 trace_xfs_log_force(mp, 0, _RET_IP_);
3249 xlog_cil_force(log);
3251 spin_lock(&log->l_icloglock);
3252 iclog = log->l_iclog;
3253 if (iclog->ic_state & XLOG_STATE_IOERROR)
3256 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3257 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3258 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3260 * If the head is dirty or (active and empty), then we need to
3261 * look at the previous iclog.
3263 * If the previous iclog is active or dirty we are done. There
3264 * is nothing to sync out. Otherwise, we attach ourselves to the
3265 * previous iclog and go to sleep.
3267 iclog = iclog->ic_prev;
3268 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3269 iclog->ic_state == XLOG_STATE_DIRTY)
3271 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3272 if (atomic_read(&iclog->ic_refcnt) == 0) {
3274 * We are the only one with access to this iclog.
3276 * Flush it out now. There should be a roundoff of zero
3277 * to show that someone has already taken care of the
3278 * roundoff from the previous sync.
3280 atomic_inc(&iclog->ic_refcnt);
3281 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3282 xlog_state_switch_iclogs(log, iclog, 0);
3283 spin_unlock(&log->l_icloglock);
3285 if (xlog_state_release_iclog(log, iclog))
3288 spin_lock(&log->l_icloglock);
3289 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3290 iclog->ic_state == XLOG_STATE_DIRTY)
3294 * Someone else is writing to this iclog.
3296 * Use its call to flush out the data. However, the
3297 * other thread may not force out this LR, so we mark
3300 xlog_state_switch_iclogs(log, iclog, 0);
3304 * If the head iclog is not active nor dirty, we just attach
3305 * ourselves to the head and go to sleep if necessary.
3310 if (!(flags & XFS_LOG_SYNC))
3313 if (iclog->ic_state & XLOG_STATE_IOERROR)
3315 XFS_STATS_INC(mp, xs_log_force_sleep);
3316 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3317 if (iclog->ic_state & XLOG_STATE_IOERROR)
3322 spin_unlock(&log->l_icloglock);
3325 spin_unlock(&log->l_icloglock);
3330 __xfs_log_force_lsn(
3331 struct xfs_mount *mp,
3337 struct xlog *log = mp->m_log;
3338 struct xlog_in_core *iclog;
3340 spin_lock(&log->l_icloglock);
3341 iclog = log->l_iclog;
3342 if (iclog->ic_state & XLOG_STATE_IOERROR)
3345 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3346 iclog = iclog->ic_next;
3347 if (iclog == log->l_iclog)
3351 if (iclog->ic_state == XLOG_STATE_DIRTY)
3354 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3356 * We sleep here if we haven't already slept (e.g. this is the
3357 * first time we've looked at the correct iclog buf) and the
3358 * buffer before us is going to be sync'ed. The reason for this
3359 * is that if we are doing sync transactions here, by waiting
3360 * for the previous I/O to complete, we can allow a few more
3361 * transactions into this iclog before we close it down.
3363 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3364 * refcnt so we can release the log (which drops the ref count).
3365 * The state switch keeps new transaction commits from using
3366 * this buffer. When the current commits finish writing into
3367 * the buffer, the refcount will drop to zero and the buffer
3370 if (!already_slept &&
3371 (iclog->ic_prev->ic_state &
3372 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3373 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3375 XFS_STATS_INC(mp, xs_log_force_sleep);
3377 xlog_wait(&iclog->ic_prev->ic_write_wait,
3381 atomic_inc(&iclog->ic_refcnt);
3382 xlog_state_switch_iclogs(log, iclog, 0);
3383 spin_unlock(&log->l_icloglock);
3384 if (xlog_state_release_iclog(log, iclog))
3388 spin_lock(&log->l_icloglock);
3391 if (!(flags & XFS_LOG_SYNC) ||
3392 (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3395 if (iclog->ic_state & XLOG_STATE_IOERROR)
3398 XFS_STATS_INC(mp, xs_log_force_sleep);
3399 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3400 if (iclog->ic_state & XLOG_STATE_IOERROR)
3405 spin_unlock(&log->l_icloglock);
3408 spin_unlock(&log->l_icloglock);
3413 * Force the in-core log to disk for a specific LSN.
3415 * Find in-core log with lsn.
3416 * If it is in the DIRTY state, just return.
3417 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3418 * state and go to sleep or return.
3419 * If it is in any other state, go to sleep or return.
3421 * Synchronous forces are implemented with a wait queue. All callers trying
3422 * to force a given lsn to disk must wait on the queue attached to the
3423 * specific in-core log. When given in-core log finally completes its write
3424 * to disk, that thread will wake up all threads waiting on the queue.
3428 struct xfs_mount *mp,
3436 XFS_STATS_INC(mp, xs_log_force);
3437 trace_xfs_log_force(mp, lsn, _RET_IP_);
3439 lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3440 if (lsn == NULLCOMMITLSN)
3443 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3445 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3450 * Called when we want to mark the current iclog as being ready to sync to
3454 xlog_state_want_sync(
3456 struct xlog_in_core *iclog)
3458 assert_spin_locked(&log->l_icloglock);
3460 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3461 xlog_state_switch_iclogs(log, iclog, 0);
3463 ASSERT(iclog->ic_state &
3464 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3469 /*****************************************************************************
3473 *****************************************************************************
3477 * Free a used ticket when its refcount falls to zero.
3481 xlog_ticket_t *ticket)
3483 ASSERT(atomic_read(&ticket->t_ref) > 0);
3484 if (atomic_dec_and_test(&ticket->t_ref))
3485 kmem_zone_free(xfs_log_ticket_zone, ticket);
3490 xlog_ticket_t *ticket)
3492 ASSERT(atomic_read(&ticket->t_ref) > 0);
3493 atomic_inc(&ticket->t_ref);
3498 * Figure out the total log space unit (in bytes) that would be
3499 * required for a log ticket.
3502 xfs_log_calc_unit_res(
3503 struct xfs_mount *mp,
3506 struct xlog *log = mp->m_log;
3511 * Permanent reservations have up to 'cnt'-1 active log operations
3512 * in the log. A unit in this case is the amount of space for one
3513 * of these log operations. Normal reservations have a cnt of 1
3514 * and their unit amount is the total amount of space required.
3516 * The following lines of code account for non-transaction data
3517 * which occupy space in the on-disk log.
3519 * Normal form of a transaction is:
3520 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3521 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3523 * We need to account for all the leadup data and trailer data
3524 * around the transaction data.
3525 * And then we need to account for the worst case in terms of using
3527 * The worst case will happen if:
3528 * - the placement of the transaction happens to be such that the
3529 * roundoff is at its maximum
3530 * - the transaction data is synced before the commit record is synced
3531 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3532 * Therefore the commit record is in its own Log Record.
3533 * This can happen as the commit record is called with its
3534 * own region to xlog_write().
3535 * This then means that in the worst case, roundoff can happen for
3536 * the commit-rec as well.
3537 * The commit-rec is smaller than padding in this scenario and so it is
3538 * not added separately.
3541 /* for trans header */
3542 unit_bytes += sizeof(xlog_op_header_t);
3543 unit_bytes += sizeof(xfs_trans_header_t);
3546 unit_bytes += sizeof(xlog_op_header_t);
3549 * for LR headers - the space for data in an iclog is the size minus
3550 * the space used for the headers. If we use the iclog size, then we
3551 * undercalculate the number of headers required.
3553 * Furthermore - the addition of op headers for split-recs might
3554 * increase the space required enough to require more log and op
3555 * headers, so take that into account too.
3557 * IMPORTANT: This reservation makes the assumption that if this
3558 * transaction is the first in an iclog and hence has the LR headers
3559 * accounted to it, then the remaining space in the iclog is
3560 * exclusively for this transaction. i.e. if the transaction is larger
3561 * than the iclog, it will be the only thing in that iclog.
3562 * Fundamentally, this means we must pass the entire log vector to
3563 * xlog_write to guarantee this.
3565 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3566 num_headers = howmany(unit_bytes, iclog_space);
3568 /* for split-recs - ophdrs added when data split over LRs */
3569 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3571 /* add extra header reservations if we overrun */
3572 while (!num_headers ||
3573 howmany(unit_bytes, iclog_space) > num_headers) {
3574 unit_bytes += sizeof(xlog_op_header_t);
3577 unit_bytes += log->l_iclog_hsize * num_headers;
3579 /* for commit-rec LR header - note: padding will subsume the ophdr */
3580 unit_bytes += log->l_iclog_hsize;
3582 /* for roundoff padding for transaction data and one for commit record */
3583 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3584 /* log su roundoff */
3585 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3588 unit_bytes += 2 * BBSIZE;
3595 * Allocate and initialise a new log ticket.
3597 struct xlog_ticket *
3604 xfs_km_flags_t alloc_flags)
3606 struct xlog_ticket *tic;
3609 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3613 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3615 atomic_set(&tic->t_ref, 1);
3616 tic->t_task = current;
3617 INIT_LIST_HEAD(&tic->t_queue);
3618 tic->t_unit_res = unit_res;
3619 tic->t_curr_res = unit_res;
3622 tic->t_tid = prandom_u32();
3623 tic->t_clientid = client;
3624 tic->t_flags = XLOG_TIC_INITED;
3626 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3628 xlog_tic_reset_res(tic);
3634 /******************************************************************************
3636 * Log debug routines
3638 ******************************************************************************
3642 * Make sure that the destination ptr is within the valid data region of
3643 * one of the iclogs. This uses backup pointers stored in a different
3644 * part of the log in case we trash the log structure.
3647 xlog_verify_dest_ptr(
3654 for (i = 0; i < log->l_iclog_bufs; i++) {
3655 if (ptr >= log->l_iclog_bak[i] &&
3656 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3661 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3665 * Check to make sure the grant write head didn't just over lap the tail. If
3666 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3667 * the cycles differ by exactly one and check the byte count.
3669 * This check is run unlocked, so can give false positives. Rather than assert
3670 * on failures, use a warn-once flag and a panic tag to allow the admin to
3671 * determine if they want to panic the machine when such an error occurs. For
3672 * debug kernels this will have the same effect as using an assert but, unlinke
3673 * an assert, it can be turned off at runtime.
3676 xlog_verify_grant_tail(
3679 int tail_cycle, tail_blocks;
3682 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3683 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3684 if (tail_cycle != cycle) {
3685 if (cycle - 1 != tail_cycle &&
3686 !(log->l_flags & XLOG_TAIL_WARN)) {
3687 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3688 "%s: cycle - 1 != tail_cycle", __func__);
3689 log->l_flags |= XLOG_TAIL_WARN;
3692 if (space > BBTOB(tail_blocks) &&
3693 !(log->l_flags & XLOG_TAIL_WARN)) {
3694 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3695 "%s: space > BBTOB(tail_blocks)", __func__);
3696 log->l_flags |= XLOG_TAIL_WARN;
3701 /* check if it will fit */
3703 xlog_verify_tail_lsn(
3705 struct xlog_in_core *iclog,
3710 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3712 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3713 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3714 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3716 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3718 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3719 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3721 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3722 if (blocks < BTOBB(iclog->ic_offset) + 1)
3723 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3725 } /* xlog_verify_tail_lsn */
3728 * Perform a number of checks on the iclog before writing to disk.
3730 * 1. Make sure the iclogs are still circular
3731 * 2. Make sure we have a good magic number
3732 * 3. Make sure we don't have magic numbers in the data
3733 * 4. Check fields of each log operation header for:
3734 * A. Valid client identifier
3735 * B. tid ptr value falls in valid ptr space (user space code)
3736 * C. Length in log record header is correct according to the
3737 * individual operation headers within record.
3738 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3739 * log, check the preceding blocks of the physical log to make sure all
3740 * the cycle numbers agree with the current cycle number.
3745 struct xlog_in_core *iclog,
3748 xlog_op_header_t *ophead;
3749 xlog_in_core_t *icptr;
3750 xlog_in_core_2_t *xhdr;
3751 void *base_ptr, *ptr, *p;
3752 ptrdiff_t field_offset;
3754 int len, i, j, k, op_len;
3757 /* check validity of iclog pointers */
3758 spin_lock(&log->l_icloglock);
3759 icptr = log->l_iclog;
3760 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3763 if (icptr != log->l_iclog)
3764 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3765 spin_unlock(&log->l_icloglock);
3767 /* check log magic numbers */
3768 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3769 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3771 base_ptr = ptr = &iclog->ic_header;
3772 p = &iclog->ic_header;
3773 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3774 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3775 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3780 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3781 base_ptr = ptr = iclog->ic_datap;
3783 xhdr = iclog->ic_data;
3784 for (i = 0; i < len; i++) {
3787 /* clientid is only 1 byte */
3788 p = &ophead->oh_clientid;
3789 field_offset = p - base_ptr;
3790 if (field_offset & 0x1ff) {
3791 clientid = ophead->oh_clientid;
3793 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3794 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3795 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3796 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3797 clientid = xlog_get_client_id(
3798 xhdr[j].hic_xheader.xh_cycle_data[k]);
3800 clientid = xlog_get_client_id(
3801 iclog->ic_header.h_cycle_data[idx]);
3804 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3806 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3807 __func__, clientid, ophead,
3808 (unsigned long)field_offset);
3811 p = &ophead->oh_len;
3812 field_offset = p - base_ptr;
3813 if (field_offset & 0x1ff) {
3814 op_len = be32_to_cpu(ophead->oh_len);
3816 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3817 (uintptr_t)iclog->ic_datap);
3818 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3819 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3820 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3821 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3823 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3826 ptr += sizeof(xlog_op_header_t) + op_len;
3828 } /* xlog_verify_iclog */
3832 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3838 xlog_in_core_t *iclog, *ic;
3840 iclog = log->l_iclog;
3841 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3843 * Mark all the incore logs IOERROR.
3844 * From now on, no log flushes will result.
3848 ic->ic_state = XLOG_STATE_IOERROR;
3850 } while (ic != iclog);
3854 * Return non-zero, if state transition has already happened.
3860 * This is called from xfs_force_shutdown, when we're forcibly
3861 * shutting down the filesystem, typically because of an IO error.
3862 * Our main objectives here are to make sure that:
3863 * a. if !logerror, flush the logs to disk. Anything modified
3864 * after this is ignored.
3865 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3866 * parties to find out, 'atomically'.
3867 * c. those who're sleeping on log reservations, pinned objects and
3868 * other resources get woken up, and be told the bad news.
3869 * d. nothing new gets queued up after (b) and (c) are done.
3871 * Note: for the !logerror case we need to flush the regions held in memory out
3872 * to disk first. This needs to be done before the log is marked as shutdown,
3873 * otherwise the iclog writes will fail.
3876 xfs_log_force_umount(
3877 struct xfs_mount *mp,
3886 * If this happens during log recovery, don't worry about
3887 * locking; the log isn't open for business yet.
3890 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3891 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3893 mp->m_sb_bp->b_flags |= XBF_DONE;
3898 * Somebody could've already done the hard work for us.
3899 * No need to get locks for this.
3901 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3902 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3907 * Flush all the completed transactions to disk before marking the log
3908 * being shut down. We need to do it in this order to ensure that
3909 * completed operations are safely on disk before we shut down, and that
3910 * we don't have to issue any buffer IO after the shutdown flags are set
3911 * to guarantee this.
3914 xfs_log_force(mp, XFS_LOG_SYNC);
3917 * mark the filesystem and the as in a shutdown state and wake
3918 * everybody up to tell them the bad news.
3920 spin_lock(&log->l_icloglock);
3921 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3923 mp->m_sb_bp->b_flags |= XBF_DONE;
3926 * Mark the log and the iclogs with IO error flags to prevent any
3927 * further log IO from being issued or completed.
3929 log->l_flags |= XLOG_IO_ERROR;
3930 retval = xlog_state_ioerror(log);
3931 spin_unlock(&log->l_icloglock);
3934 * We don't want anybody waiting for log reservations after this. That
3935 * means we have to wake up everybody queued up on reserveq as well as
3936 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3937 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3938 * action is protected by the grant locks.
3940 xlog_grant_head_wake_all(&log->l_reserve_head);
3941 xlog_grant_head_wake_all(&log->l_write_head);
3944 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3945 * as if the log writes were completed. The abort handling in the log
3946 * item committed callback functions will do this again under lock to
3949 wake_up_all(&log->l_cilp->xc_commit_wait);
3950 xlog_state_do_callback(log, true, NULL);
3952 #ifdef XFSERRORDEBUG
3954 xlog_in_core_t *iclog;
3956 spin_lock(&log->l_icloglock);
3957 iclog = log->l_iclog;
3959 ASSERT(iclog->ic_callback == 0);
3960 iclog = iclog->ic_next;
3961 } while (iclog != log->l_iclog);
3962 spin_unlock(&log->l_icloglock);
3965 /* return non-zero if log IOERROR transition had already happened */
3973 xlog_in_core_t *iclog;
3975 iclog = log->l_iclog;
3977 /* endianness does not matter here, zero is zero in
3980 if (iclog->ic_header.h_num_logops)
3982 iclog = iclog->ic_next;
3983 } while (iclog != log->l_iclog);
3988 * Verify that an LSN stamped into a piece of metadata is valid. This is
3989 * intended for use in read verifiers on v5 superblocks.
3993 struct xfs_mount *mp,
3996 struct xlog *log = mp->m_log;
4000 * norecovery mode skips mount-time log processing and unconditionally
4001 * resets the in-core LSN. We can't validate in this mode, but
4002 * modifications are not allowed anyways so just return true.
4004 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4008 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4009 * handled by recovery and thus safe to ignore here.
4011 if (lsn == NULLCOMMITLSN)
4014 valid = xlog_valid_lsn(mp->m_log, lsn);
4016 /* warn the user about what's gone wrong before verifier failure */
4018 spin_lock(&log->l_icloglock);
4020 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4021 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4022 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4023 log->l_curr_cycle, log->l_curr_block);
4024 spin_unlock(&log->l_icloglock);
4031 xfs_log_in_recovery(
4032 struct xfs_mount *mp)
4034 struct xlog *log = mp->m_log;
4036 return log->l_flags & XLOG_ACTIVE_RECOVERY;