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;
218 list_for_each_entry(tic, &head->waiters, t_queue) {
219 need_bytes = xlog_ticket_reservation(log, head, tic);
220 if (*free_bytes < need_bytes)
223 *free_bytes -= need_bytes;
224 trace_xfs_log_grant_wake_up(log, tic);
225 wake_up_process(tic->t_task);
232 xlog_grant_head_wait(
234 struct xlog_grant_head *head,
235 struct xlog_ticket *tic,
236 int need_bytes) __releases(&head->lock)
237 __acquires(&head->lock)
239 list_add_tail(&tic->t_queue, &head->waiters);
242 if (XLOG_FORCED_SHUTDOWN(log))
244 xlog_grant_push_ail(log, need_bytes);
246 __set_current_state(TASK_UNINTERRUPTIBLE);
247 spin_unlock(&head->lock);
249 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
251 trace_xfs_log_grant_sleep(log, tic);
253 trace_xfs_log_grant_wake(log, tic);
255 spin_lock(&head->lock);
256 if (XLOG_FORCED_SHUTDOWN(log))
258 } while (xlog_space_left(log, &head->grant) < need_bytes);
260 list_del_init(&tic->t_queue);
263 list_del_init(&tic->t_queue);
268 * Atomically get the log space required for a log ticket.
270 * Once a ticket gets put onto head->waiters, it will only return after the
271 * needed reservation is satisfied.
273 * This function is structured so that it has a lock free fast path. This is
274 * necessary because every new transaction reservation will come through this
275 * path. Hence any lock will be globally hot if we take it unconditionally on
278 * As tickets are only ever moved on and off head->waiters under head->lock, we
279 * only need to take that lock if we are going to add the ticket to the queue
280 * and sleep. We can avoid taking the lock if the ticket was never added to
281 * head->waiters because the t_queue list head will be empty and we hold the
282 * only reference to it so it can safely be checked unlocked.
285 xlog_grant_head_check(
287 struct xlog_grant_head *head,
288 struct xlog_ticket *tic,
294 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
297 * If there are other waiters on the queue then give them a chance at
298 * logspace before us. Wake up the first waiters, if we do not wake
299 * up all the waiters then go to sleep waiting for more free space,
300 * otherwise try to get some space for this transaction.
302 *need_bytes = xlog_ticket_reservation(log, head, tic);
303 free_bytes = xlog_space_left(log, &head->grant);
304 if (!list_empty_careful(&head->waiters)) {
305 spin_lock(&head->lock);
306 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
307 free_bytes < *need_bytes) {
308 error = xlog_grant_head_wait(log, head, tic,
311 spin_unlock(&head->lock);
312 } else if (free_bytes < *need_bytes) {
313 spin_lock(&head->lock);
314 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
315 spin_unlock(&head->lock);
322 xlog_tic_reset_res(xlog_ticket_t *tic)
325 tic->t_res_arr_sum = 0;
326 tic->t_res_num_ophdrs = 0;
330 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
332 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
333 /* add to overflow and start again */
334 tic->t_res_o_flow += tic->t_res_arr_sum;
336 tic->t_res_arr_sum = 0;
339 tic->t_res_arr[tic->t_res_num].r_len = len;
340 tic->t_res_arr[tic->t_res_num].r_type = type;
341 tic->t_res_arr_sum += len;
346 * Replenish the byte reservation required by moving the grant write head.
350 struct xfs_mount *mp,
351 struct xlog_ticket *tic)
353 struct xlog *log = mp->m_log;
357 if (XLOG_FORCED_SHUTDOWN(log))
360 XFS_STATS_INC(mp, xs_try_logspace);
363 * This is a new transaction on the ticket, so we need to change the
364 * transaction ID so that the next transaction has a different TID in
365 * the log. Just add one to the existing tid so that we can see chains
366 * of rolling transactions in the log easily.
370 xlog_grant_push_ail(log, tic->t_unit_res);
372 tic->t_curr_res = tic->t_unit_res;
373 xlog_tic_reset_res(tic);
378 trace_xfs_log_regrant(log, tic);
380 error = xlog_grant_head_check(log, &log->l_write_head, tic,
385 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
386 trace_xfs_log_regrant_exit(log, tic);
387 xlog_verify_grant_tail(log);
392 * If we are failing, make sure the ticket doesn't have any current
393 * reservations. We don't want to add this back when the ticket/
394 * transaction gets cancelled.
397 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
402 * Reserve log space and return a ticket corresponding to the reservation.
404 * Each reservation is going to reserve extra space for a log record header.
405 * When writes happen to the on-disk log, we don't subtract the length of the
406 * log record header from any reservation. By wasting space in each
407 * reservation, we prevent over allocation problems.
411 struct xfs_mount *mp,
414 struct xlog_ticket **ticp,
418 struct xlog *log = mp->m_log;
419 struct xlog_ticket *tic;
423 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
425 if (XLOG_FORCED_SHUTDOWN(log))
428 XFS_STATS_INC(mp, xs_try_logspace);
430 ASSERT(*ticp == NULL);
431 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
432 KM_SLEEP | KM_MAYFAIL);
438 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
441 trace_xfs_log_reserve(log, tic);
443 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
448 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
449 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
450 trace_xfs_log_reserve_exit(log, tic);
451 xlog_verify_grant_tail(log);
456 * If we are failing, make sure the ticket doesn't have any current
457 * reservations. We don't want to add this back when the ticket/
458 * transaction gets cancelled.
461 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
469 * 1. currblock field gets updated at startup and after in-core logs
470 * marked as with WANT_SYNC.
474 * This routine is called when a user of a log manager ticket is done with
475 * the reservation. If the ticket was ever used, then a commit record for
476 * the associated transaction is written out as a log operation header with
477 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
478 * a given ticket. If the ticket was one with a permanent reservation, then
479 * a few operations are done differently. Permanent reservation tickets by
480 * default don't release the reservation. They just commit the current
481 * transaction with the belief that the reservation is still needed. A flag
482 * must be passed in before permanent reservations are actually released.
483 * When these type of tickets are not released, they need to be set into
484 * the inited state again. By doing this, a start record will be written
485 * out when the next write occurs.
489 struct xfs_mount *mp,
490 struct xlog_ticket *ticket,
491 struct xlog_in_core **iclog,
494 struct xlog *log = mp->m_log;
497 if (XLOG_FORCED_SHUTDOWN(log) ||
499 * If nothing was ever written, don't write out commit record.
500 * If we get an error, just continue and give back the log ticket.
502 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
503 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
504 lsn = (xfs_lsn_t) -1;
510 trace_xfs_log_done_nonperm(log, ticket);
513 * Release ticket if not permanent reservation or a specific
514 * request has been made to release a permanent reservation.
516 xlog_ungrant_log_space(log, ticket);
518 trace_xfs_log_done_perm(log, ticket);
520 xlog_regrant_reserve_log_space(log, ticket);
521 /* If this ticket was a permanent reservation and we aren't
522 * trying to release it, reset the inited flags; so next time
523 * we write, a start record will be written out.
525 ticket->t_flags |= XLOG_TIC_INITED;
528 xfs_log_ticket_put(ticket);
533 xfs_log_release_iclog(
534 struct xfs_mount *mp,
535 struct xlog_in_core *iclog)
537 if (xlog_state_release_iclog(mp->m_log, iclog)) {
538 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
546 * Mount a log filesystem
548 * mp - ubiquitous xfs mount point structure
549 * log_target - buftarg of on-disk log device
550 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
551 * num_bblocks - Number of BBSIZE blocks in on-disk log
553 * Return error or zero.
558 xfs_buftarg_t *log_target,
559 xfs_daddr_t blk_offset,
562 bool fatal = xfs_sb_version_hascrc(&mp->m_sb);
566 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
567 xfs_notice(mp, "Mounting V%d Filesystem",
568 XFS_SB_VERSION_NUM(&mp->m_sb));
571 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
572 XFS_SB_VERSION_NUM(&mp->m_sb));
573 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
576 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
577 if (IS_ERR(mp->m_log)) {
578 error = PTR_ERR(mp->m_log);
583 * Validate the given log space and drop a critical message via syslog
584 * if the log size is too small that would lead to some unexpected
585 * situations in transaction log space reservation stage.
587 * Note: we can't just reject the mount if the validation fails. This
588 * would mean that people would have to downgrade their kernel just to
589 * remedy the situation as there is no way to grow the log (short of
590 * black magic surgery with xfs_db).
592 * We can, however, reject mounts for CRC format filesystems, as the
593 * mkfs binary being used to make the filesystem should never create a
594 * filesystem with a log that is too small.
596 min_logfsbs = xfs_log_calc_minimum_size(mp);
598 if (mp->m_sb.sb_logblocks < min_logfsbs) {
600 "Log size %d blocks too small, minimum size is %d blocks",
601 mp->m_sb.sb_logblocks, min_logfsbs);
603 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
605 "Log size %d blocks too large, maximum size is %lld blocks",
606 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
608 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
610 "log size %lld bytes too large, maximum size is %lld bytes",
611 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
614 } else if (mp->m_sb.sb_logsunit > 1 &&
615 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
617 "log stripe unit %u bytes must be a multiple of block size",
618 mp->m_sb.sb_logsunit);
624 * Log check errors are always fatal on v5; or whenever bad
625 * metadata leads to a crash.
628 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
632 xfs_crit(mp, "Log size out of supported range.");
634 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
638 * Initialize the AIL now we have a log.
640 error = xfs_trans_ail_init(mp);
642 xfs_warn(mp, "AIL initialisation failed: error %d", error);
645 mp->m_log->l_ailp = mp->m_ail;
648 * skip log recovery on a norecovery mount. pretend it all
651 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
652 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
655 mp->m_flags &= ~XFS_MOUNT_RDONLY;
657 error = xlog_recover(mp->m_log);
660 mp->m_flags |= XFS_MOUNT_RDONLY;
662 xfs_warn(mp, "log mount/recovery failed: error %d",
664 xlog_recover_cancel(mp->m_log);
665 goto out_destroy_ail;
669 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
672 goto out_destroy_ail;
674 /* Normal transactions can now occur */
675 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
678 * Now the log has been fully initialised and we know were our
679 * space grant counters are, we can initialise the permanent ticket
680 * needed for delayed logging to work.
682 xlog_cil_init_post_recovery(mp->m_log);
687 xfs_trans_ail_destroy(mp);
689 xlog_dealloc_log(mp->m_log);
695 * Finish the recovery of the file system. This is separate from the
696 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
697 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
700 * If we finish recovery successfully, start the background log work. If we are
701 * not doing recovery, then we have a RO filesystem and we don't need to start
705 xfs_log_mount_finish(
706 struct xfs_mount *mp)
709 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
710 bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
712 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
713 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
715 } else if (readonly) {
716 /* Allow unlinked processing to proceed */
717 mp->m_flags &= ~XFS_MOUNT_RDONLY;
721 * During the second phase of log recovery, we need iget and
722 * iput to behave like they do for an active filesystem.
723 * xfs_fs_drop_inode needs to be able to prevent the deletion
724 * of inodes before we're done replaying log items on those
725 * inodes. Turn it off immediately after recovery finishes
726 * so that we don't leak the quota inodes if subsequent mount
729 * We let all inodes involved in redo item processing end up on
730 * the LRU instead of being evicted immediately so that if we do
731 * something to an unlinked inode, the irele won't cause
732 * premature truncation and freeing of the inode, which results
733 * in log recovery failure. We have to evict the unreferenced
734 * lru inodes after clearing SB_ACTIVE because we don't
735 * otherwise clean up the lru if there's a subsequent failure in
736 * xfs_mountfs, which leads to us leaking the inodes if nothing
737 * else (e.g. quotacheck) references the inodes before the
738 * mount failure occurs.
740 mp->m_super->s_flags |= SB_ACTIVE;
741 error = xlog_recover_finish(mp->m_log);
743 xfs_log_work_queue(mp);
744 mp->m_super->s_flags &= ~SB_ACTIVE;
745 evict_inodes(mp->m_super);
748 * Drain the buffer LRU after log recovery. This is required for v4
749 * filesystems to avoid leaving around buffers with NULL verifier ops,
750 * but we do it unconditionally to make sure we're always in a clean
751 * cache state after mount.
753 * Don't push in the error case because the AIL may have pending intents
754 * that aren't removed until recovery is cancelled.
756 if (!error && recovered) {
757 xfs_log_force(mp, XFS_LOG_SYNC);
758 xfs_ail_push_all_sync(mp->m_ail);
760 xfs_wait_buftarg(mp->m_ddev_targp);
763 mp->m_flags |= XFS_MOUNT_RDONLY;
769 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
773 xfs_log_mount_cancel(
774 struct xfs_mount *mp)
778 error = xlog_recover_cancel(mp->m_log);
785 * Final log writes as part of unmount.
787 * Mark the filesystem clean as unmount happens. Note that during relocation
788 * this routine needs to be executed as part of source-bag while the
789 * deallocation must not be done until source-end.
792 /* Actually write the unmount record to disk. */
794 xfs_log_write_unmount_record(
795 struct xfs_mount *mp)
797 /* the data section must be 32 bit size aligned */
798 struct xfs_unmount_log_format magic = {
799 .magic = XLOG_UNMOUNT_TYPE,
801 struct xfs_log_iovec reg = {
803 .i_len = sizeof(magic),
804 .i_type = XLOG_REG_TYPE_UNMOUNT,
806 struct xfs_log_vec vec = {
810 struct xlog *log = mp->m_log;
811 struct xlog_in_core *iclog;
812 struct xlog_ticket *tic = NULL;
814 uint flags = XLOG_UNMOUNT_TRANS;
817 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
822 * If we think the summary counters are bad, clear the unmount header
823 * flag in the unmount record so that the summary counters will be
824 * recalculated during log recovery at next mount. Refer to
825 * xlog_check_unmount_rec for more details.
827 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
828 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
829 xfs_alert(mp, "%s: will fix summary counters at next mount",
831 flags &= ~XLOG_UNMOUNT_TRANS;
834 /* remove inited flag, and account for space used */
836 tic->t_curr_res -= sizeof(magic);
837 error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
839 * At this point, we're umounting anyway, so there's no point in
840 * transitioning log state to IOERROR. Just continue...
844 xfs_alert(mp, "%s: unmount record failed", __func__);
846 spin_lock(&log->l_icloglock);
847 iclog = log->l_iclog;
848 atomic_inc(&iclog->ic_refcnt);
849 xlog_state_want_sync(log, iclog);
850 spin_unlock(&log->l_icloglock);
851 error = xlog_state_release_iclog(log, iclog);
853 spin_lock(&log->l_icloglock);
854 switch (iclog->ic_state) {
856 if (!XLOG_FORCED_SHUTDOWN(log)) {
857 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
861 case XLOG_STATE_ACTIVE:
862 case XLOG_STATE_DIRTY:
863 spin_unlock(&log->l_icloglock);
868 trace_xfs_log_umount_write(log, tic);
869 xlog_ungrant_log_space(log, tic);
870 xfs_log_ticket_put(tic);
875 * Unmount record used to have a string "Unmount filesystem--" in the
876 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
877 * We just write the magic number now since that particular field isn't
878 * currently architecture converted and "Unmount" is a bit foo.
879 * As far as I know, there weren't any dependencies on the old behaviour.
883 xfs_log_unmount_write(xfs_mount_t *mp)
885 struct xlog *log = mp->m_log;
886 xlog_in_core_t *iclog;
888 xlog_in_core_t *first_iclog;
893 * Don't write out unmount record on norecovery mounts or ro devices.
894 * Or, if we are doing a forced umount (typically because of IO errors).
896 if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
897 xfs_readonly_buftarg(log->l_targ)) {
898 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
902 error = xfs_log_force(mp, XFS_LOG_SYNC);
903 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
906 first_iclog = iclog = log->l_iclog;
908 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
909 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
910 ASSERT(iclog->ic_offset == 0);
912 iclog = iclog->ic_next;
913 } while (iclog != first_iclog);
915 if (! (XLOG_FORCED_SHUTDOWN(log))) {
916 xfs_log_write_unmount_record(mp);
919 * We're already in forced_shutdown mode, couldn't
920 * even attempt to write out the unmount transaction.
922 * Go through the motions of sync'ing and releasing
923 * the iclog, even though no I/O will actually happen,
924 * we need to wait for other log I/Os that may already
925 * be in progress. Do this as a separate section of
926 * code so we'll know if we ever get stuck here that
927 * we're in this odd situation of trying to unmount
928 * a file system that went into forced_shutdown as
929 * the result of an unmount..
931 spin_lock(&log->l_icloglock);
932 iclog = log->l_iclog;
933 atomic_inc(&iclog->ic_refcnt);
935 xlog_state_want_sync(log, iclog);
936 spin_unlock(&log->l_icloglock);
937 error = xlog_state_release_iclog(log, iclog);
939 spin_lock(&log->l_icloglock);
941 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
942 || iclog->ic_state == XLOG_STATE_DIRTY
943 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
945 xlog_wait(&iclog->ic_force_wait,
948 spin_unlock(&log->l_icloglock);
953 } /* xfs_log_unmount_write */
956 * Empty the log for unmount/freeze.
958 * To do this, we first need to shut down the background log work so it is not
959 * trying to cover the log as we clean up. We then need to unpin all objects in
960 * the log so we can then flush them out. Once they have completed their IO and
961 * run the callbacks removing themselves from the AIL, we can write the unmount
966 struct xfs_mount *mp)
968 cancel_delayed_work_sync(&mp->m_log->l_work);
969 xfs_log_force(mp, XFS_LOG_SYNC);
972 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
973 * will push it, xfs_wait_buftarg() will not wait for it. Further,
974 * xfs_buf_iowait() cannot be used because it was pushed with the
975 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
976 * the IO to complete.
978 xfs_ail_push_all_sync(mp->m_ail);
979 xfs_wait_buftarg(mp->m_ddev_targp);
980 xfs_buf_lock(mp->m_sb_bp);
981 xfs_buf_unlock(mp->m_sb_bp);
983 xfs_log_unmount_write(mp);
987 * Shut down and release the AIL and Log.
989 * During unmount, we need to ensure we flush all the dirty metadata objects
990 * from the AIL so that the log is empty before we write the unmount record to
991 * the log. Once this is done, we can tear down the AIL and the log.
995 struct xfs_mount *mp)
999 xfs_trans_ail_destroy(mp);
1001 xfs_sysfs_del(&mp->m_log->l_kobj);
1003 xlog_dealloc_log(mp->m_log);
1008 struct xfs_mount *mp,
1009 struct xfs_log_item *item,
1011 const struct xfs_item_ops *ops)
1013 item->li_mountp = mp;
1014 item->li_ailp = mp->m_ail;
1015 item->li_type = type;
1019 INIT_LIST_HEAD(&item->li_ail);
1020 INIT_LIST_HEAD(&item->li_cil);
1021 INIT_LIST_HEAD(&item->li_bio_list);
1022 INIT_LIST_HEAD(&item->li_trans);
1026 * Wake up processes waiting for log space after we have moved the log tail.
1030 struct xfs_mount *mp)
1032 struct xlog *log = mp->m_log;
1035 if (XLOG_FORCED_SHUTDOWN(log))
1038 if (!list_empty_careful(&log->l_write_head.waiters)) {
1039 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1041 spin_lock(&log->l_write_head.lock);
1042 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1043 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1044 spin_unlock(&log->l_write_head.lock);
1047 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1048 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1050 spin_lock(&log->l_reserve_head.lock);
1051 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1052 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1053 spin_unlock(&log->l_reserve_head.lock);
1058 * Determine if we have a transaction that has gone to disk that needs to be
1059 * covered. To begin the transition to the idle state firstly the log needs to
1060 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1061 * we start attempting to cover the log.
1063 * Only if we are then in a state where covering is needed, the caller is
1064 * informed that dummy transactions are required to move the log into the idle
1067 * If there are any items in the AIl or CIL, then we do not want to attempt to
1068 * cover the log as we may be in a situation where there isn't log space
1069 * available to run a dummy transaction and this can lead to deadlocks when the
1070 * tail of the log is pinned by an item that is modified in the CIL. Hence
1071 * there's no point in running a dummy transaction at this point because we
1072 * can't start trying to idle the log until both the CIL and AIL are empty.
1075 xfs_log_need_covered(xfs_mount_t *mp)
1077 struct xlog *log = mp->m_log;
1080 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1083 if (!xlog_cil_empty(log))
1086 spin_lock(&log->l_icloglock);
1087 switch (log->l_covered_state) {
1088 case XLOG_STATE_COVER_DONE:
1089 case XLOG_STATE_COVER_DONE2:
1090 case XLOG_STATE_COVER_IDLE:
1092 case XLOG_STATE_COVER_NEED:
1093 case XLOG_STATE_COVER_NEED2:
1094 if (xfs_ail_min_lsn(log->l_ailp))
1096 if (!xlog_iclogs_empty(log))
1100 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1101 log->l_covered_state = XLOG_STATE_COVER_DONE;
1103 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1109 spin_unlock(&log->l_icloglock);
1114 * We may be holding the log iclog lock upon entering this routine.
1117 xlog_assign_tail_lsn_locked(
1118 struct xfs_mount *mp)
1120 struct xlog *log = mp->m_log;
1121 struct xfs_log_item *lip;
1124 assert_spin_locked(&mp->m_ail->ail_lock);
1127 * To make sure we always have a valid LSN for the log tail we keep
1128 * track of the last LSN which was committed in log->l_last_sync_lsn,
1129 * and use that when the AIL was empty.
1131 lip = xfs_ail_min(mp->m_ail);
1133 tail_lsn = lip->li_lsn;
1135 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1136 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1137 atomic64_set(&log->l_tail_lsn, tail_lsn);
1142 xlog_assign_tail_lsn(
1143 struct xfs_mount *mp)
1147 spin_lock(&mp->m_ail->ail_lock);
1148 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1149 spin_unlock(&mp->m_ail->ail_lock);
1155 * Return the space in the log between the tail and the head. The head
1156 * is passed in the cycle/bytes formal parms. In the special case where
1157 * the reserve head has wrapped passed the tail, this calculation is no
1158 * longer valid. In this case, just return 0 which means there is no space
1159 * in the log. This works for all places where this function is called
1160 * with the reserve head. Of course, if the write head were to ever
1161 * wrap the tail, we should blow up. Rather than catch this case here,
1162 * we depend on other ASSERTions in other parts of the code. XXXmiken
1164 * This code also handles the case where the reservation head is behind
1165 * the tail. The details of this case are described below, but the end
1166 * result is that we return the size of the log as the amount of space left.
1179 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1180 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1181 tail_bytes = BBTOB(tail_bytes);
1182 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1183 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1184 else if (tail_cycle + 1 < head_cycle)
1186 else if (tail_cycle < head_cycle) {
1187 ASSERT(tail_cycle == (head_cycle - 1));
1188 free_bytes = tail_bytes - head_bytes;
1191 * The reservation head is behind the tail.
1192 * In this case we just want to return the size of the
1193 * log as the amount of space left.
1195 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1196 xfs_alert(log->l_mp,
1197 " tail_cycle = %d, tail_bytes = %d",
1198 tail_cycle, tail_bytes);
1199 xfs_alert(log->l_mp,
1200 " GH cycle = %d, GH bytes = %d",
1201 head_cycle, head_bytes);
1203 free_bytes = log->l_logsize;
1211 struct work_struct *work)
1213 struct xlog_in_core *iclog =
1214 container_of(work, struct xlog_in_core, ic_end_io_work);
1215 struct xlog *log = iclog->ic_log;
1216 bool aborted = false;
1219 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1221 /* treat writes with injected CRC errors as failed */
1222 if (iclog->ic_fail_crc)
1227 * Race to shutdown the filesystem if we see an error.
1229 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1230 xfs_alert(log->l_mp, "log I/O error %d", error);
1231 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1233 * This flag will be propagated to the trans-committed
1234 * callback routines to let them know that the log-commit
1238 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1242 xlog_state_done_syncing(iclog, aborted);
1243 bio_uninit(&iclog->ic_bio);
1246 * Drop the lock to signal that we are done. Nothing references the
1247 * iclog after this, so an unmount waiting on this lock can now tear it
1248 * down safely. As such, it is unsafe to reference the iclog after the
1249 * unlock as we could race with it being freed.
1251 up(&iclog->ic_sema);
1255 * Return size of each in-core log record buffer.
1257 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1259 * If the filesystem blocksize is too large, we may need to choose a
1260 * larger size since the directory code currently logs entire blocks.
1263 xlog_get_iclog_buffer_size(
1264 struct xfs_mount *mp,
1267 if (mp->m_logbufs <= 0)
1268 mp->m_logbufs = XLOG_MAX_ICLOGS;
1269 if (mp->m_logbsize <= 0)
1270 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1272 log->l_iclog_bufs = mp->m_logbufs;
1273 log->l_iclog_size = mp->m_logbsize;
1276 * # headers = size / 32k - one header holds cycles from 32k of data.
1278 log->l_iclog_heads =
1279 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1280 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1285 struct xfs_mount *mp)
1287 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1288 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1292 * Every sync period we need to unpin all items in the AIL and push them to
1293 * disk. If there is nothing dirty, then we might need to cover the log to
1294 * indicate that the filesystem is idle.
1298 struct work_struct *work)
1300 struct xlog *log = container_of(to_delayed_work(work),
1301 struct xlog, l_work);
1302 struct xfs_mount *mp = log->l_mp;
1304 /* dgc: errors ignored - not fatal and nowhere to report them */
1305 if (xfs_log_need_covered(mp)) {
1307 * Dump a transaction into the log that contains no real change.
1308 * This is needed to stamp the current tail LSN into the log
1309 * during the covering operation.
1311 * We cannot use an inode here for this - that will push dirty
1312 * state back up into the VFS and then periodic inode flushing
1313 * will prevent log covering from making progress. Hence we
1314 * synchronously log the superblock instead to ensure the
1315 * superblock is immediately unpinned and can be written back.
1317 xfs_sync_sb(mp, true);
1319 xfs_log_force(mp, 0);
1321 /* start pushing all the metadata that is currently dirty */
1322 xfs_ail_push_all(mp->m_ail);
1324 /* queue us up again */
1325 xfs_log_work_queue(mp);
1329 * This routine initializes some of the log structure for a given mount point.
1330 * Its primary purpose is to fill in enough, so recovery can occur. However,
1331 * some other stuff may be filled in too.
1333 STATIC struct xlog *
1335 struct xfs_mount *mp,
1336 struct xfs_buftarg *log_target,
1337 xfs_daddr_t blk_offset,
1341 xlog_rec_header_t *head;
1342 xlog_in_core_t **iclogp;
1343 xlog_in_core_t *iclog, *prev_iclog=NULL;
1345 int error = -ENOMEM;
1348 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1350 xfs_warn(mp, "Log allocation failed: No memory!");
1355 log->l_targ = log_target;
1356 log->l_logsize = BBTOB(num_bblks);
1357 log->l_logBBstart = blk_offset;
1358 log->l_logBBsize = num_bblks;
1359 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1360 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1361 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1363 log->l_prev_block = -1;
1364 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1365 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1366 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1367 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1369 xlog_grant_head_init(&log->l_reserve_head);
1370 xlog_grant_head_init(&log->l_write_head);
1372 error = -EFSCORRUPTED;
1373 if (xfs_sb_version_hassector(&mp->m_sb)) {
1374 log2_size = mp->m_sb.sb_logsectlog;
1375 if (log2_size < BBSHIFT) {
1376 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1377 log2_size, BBSHIFT);
1381 log2_size -= BBSHIFT;
1382 if (log2_size > mp->m_sectbb_log) {
1383 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1384 log2_size, mp->m_sectbb_log);
1388 /* for larger sector sizes, must have v2 or external log */
1389 if (log2_size && log->l_logBBstart > 0 &&
1390 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1392 "log sector size (0x%x) invalid for configuration.",
1397 log->l_sectBBsize = 1 << log2_size;
1399 xlog_get_iclog_buffer_size(mp, log);
1401 spin_lock_init(&log->l_icloglock);
1402 init_waitqueue_head(&log->l_flush_wait);
1404 iclogp = &log->l_iclog;
1406 * The amount of memory to allocate for the iclog structure is
1407 * rather funky due to the way the structure is defined. It is
1408 * done this way so that we can use different sizes for machines
1409 * with different amounts of memory. See the definition of
1410 * xlog_in_core_t in xfs_log_priv.h for details.
1412 ASSERT(log->l_iclog_size >= 4096);
1413 for (i = 0; i < log->l_iclog_bufs; i++) {
1414 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE);
1416 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1418 goto out_free_iclog;
1421 iclog->ic_prev = prev_iclog;
1424 iclog->ic_data = kmem_alloc_large(log->l_iclog_size,
1426 if (!iclog->ic_data)
1427 goto out_free_iclog;
1429 log->l_iclog_bak[i] = &iclog->ic_header;
1431 head = &iclog->ic_header;
1432 memset(head, 0, sizeof(xlog_rec_header_t));
1433 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1434 head->h_version = cpu_to_be32(
1435 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1436 head->h_size = cpu_to_be32(log->l_iclog_size);
1438 head->h_fmt = cpu_to_be32(XLOG_FMT);
1439 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1441 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1442 iclog->ic_state = XLOG_STATE_ACTIVE;
1443 iclog->ic_log = log;
1444 atomic_set(&iclog->ic_refcnt, 0);
1445 spin_lock_init(&iclog->ic_callback_lock);
1446 INIT_LIST_HEAD(&iclog->ic_callbacks);
1447 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1449 init_waitqueue_head(&iclog->ic_force_wait);
1450 init_waitqueue_head(&iclog->ic_write_wait);
1451 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1452 sema_init(&iclog->ic_sema, 1);
1454 iclogp = &iclog->ic_next;
1456 *iclogp = log->l_iclog; /* complete ring */
1457 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1459 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1460 WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0,
1462 if (!log->l_ioend_workqueue)
1463 goto out_free_iclog;
1465 error = xlog_cil_init(log);
1467 goto out_destroy_workqueue;
1470 out_destroy_workqueue:
1471 destroy_workqueue(log->l_ioend_workqueue);
1473 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1474 prev_iclog = iclog->ic_next;
1475 kmem_free(iclog->ic_data);
1481 return ERR_PTR(error);
1482 } /* xlog_alloc_log */
1486 * Write out the commit record of a transaction associated with the given
1487 * ticket. Return the lsn of the commit record.
1492 struct xlog_ticket *ticket,
1493 struct xlog_in_core **iclog,
1494 xfs_lsn_t *commitlsnp)
1496 struct xfs_mount *mp = log->l_mp;
1498 struct xfs_log_iovec reg = {
1501 .i_type = XLOG_REG_TYPE_COMMIT,
1503 struct xfs_log_vec vec = {
1508 ASSERT_ALWAYS(iclog);
1509 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1512 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1517 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1518 * log space. This code pushes on the lsn which would supposedly free up
1519 * the 25% which we want to leave free. We may need to adopt a policy which
1520 * pushes on an lsn which is further along in the log once we reach the high
1521 * water mark. In this manner, we would be creating a low water mark.
1524 xlog_grant_push_ail(
1528 xfs_lsn_t threshold_lsn = 0;
1529 xfs_lsn_t last_sync_lsn;
1532 int threshold_block;
1533 int threshold_cycle;
1536 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1538 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1539 free_blocks = BTOBBT(free_bytes);
1542 * Set the threshold for the minimum number of free blocks in the
1543 * log to the maximum of what the caller needs, one quarter of the
1544 * log, and 256 blocks.
1546 free_threshold = BTOBB(need_bytes);
1547 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1548 free_threshold = max(free_threshold, 256);
1549 if (free_blocks >= free_threshold)
1552 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1554 threshold_block += free_threshold;
1555 if (threshold_block >= log->l_logBBsize) {
1556 threshold_block -= log->l_logBBsize;
1557 threshold_cycle += 1;
1559 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1562 * Don't pass in an lsn greater than the lsn of the last
1563 * log record known to be on disk. Use a snapshot of the last sync lsn
1564 * so that it doesn't change between the compare and the set.
1566 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1567 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1568 threshold_lsn = last_sync_lsn;
1571 * Get the transaction layer to kick the dirty buffers out to
1572 * disk asynchronously. No point in trying to do this if
1573 * the filesystem is shutting down.
1575 if (!XLOG_FORCED_SHUTDOWN(log))
1576 xfs_ail_push(log->l_ailp, threshold_lsn);
1580 * Stamp cycle number in every block
1585 struct xlog_in_core *iclog,
1589 int size = iclog->ic_offset + roundoff;
1593 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1595 dp = iclog->ic_datap;
1596 for (i = 0; i < BTOBB(size); i++) {
1597 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1599 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1600 *(__be32 *)dp = cycle_lsn;
1604 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1605 xlog_in_core_2_t *xhdr = iclog->ic_data;
1607 for ( ; i < BTOBB(size); i++) {
1608 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1609 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1610 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1611 *(__be32 *)dp = cycle_lsn;
1615 for (i = 1; i < log->l_iclog_heads; i++)
1616 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1621 * Calculate the checksum for a log buffer.
1623 * This is a little more complicated than it should be because the various
1624 * headers and the actual data are non-contiguous.
1629 struct xlog_rec_header *rhead,
1635 /* first generate the crc for the record header ... */
1636 crc = xfs_start_cksum_update((char *)rhead,
1637 sizeof(struct xlog_rec_header),
1638 offsetof(struct xlog_rec_header, h_crc));
1640 /* ... then for additional cycle data for v2 logs ... */
1641 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1642 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1646 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1647 if (size % XLOG_HEADER_CYCLE_SIZE)
1650 for (i = 1; i < xheads; i++) {
1651 crc = crc32c(crc, &xhdr[i].hic_xheader,
1652 sizeof(struct xlog_rec_ext_header));
1656 /* ... and finally for the payload */
1657 crc = crc32c(crc, dp, size);
1659 return xfs_end_cksum(crc);
1666 struct xlog_in_core *iclog = bio->bi_private;
1668 queue_work(iclog->ic_log->l_ioend_workqueue,
1669 &iclog->ic_end_io_work);
1673 xlog_map_iclog_data(
1679 struct page *page = kmem_to_page(data);
1680 unsigned int off = offset_in_page(data);
1681 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1683 WARN_ON_ONCE(bio_add_page(bio, page, len, off) != len);
1693 struct xlog_in_core *iclog,
1698 ASSERT(bno < log->l_logBBsize);
1701 * We lock the iclogbufs here so that we can serialise against I/O
1702 * completion during unmount. We might be processing a shutdown
1703 * triggered during unmount, and that can occur asynchronously to the
1704 * unmount thread, and hence we need to ensure that completes before
1705 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1706 * across the log IO to archieve that.
1708 down(&iclog->ic_sema);
1709 if (unlikely(iclog->ic_state & XLOG_STATE_IOERROR)) {
1711 * It would seem logical to return EIO here, but we rely on
1712 * the log state machine to propagate I/O errors instead of
1713 * doing it here. We kick of the state machine and unlock
1714 * the buffer manually, the code needs to be kept in sync
1715 * with the I/O completion path.
1717 xlog_state_done_syncing(iclog, XFS_LI_ABORTED);
1718 up(&iclog->ic_sema);
1722 iclog->ic_io_size = count;
1724 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1725 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1726 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1727 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1728 iclog->ic_bio.bi_private = iclog;
1729 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_FUA;
1731 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1733 xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, iclog->ic_io_size);
1734 if (is_vmalloc_addr(iclog->ic_data))
1735 flush_kernel_vmap_range(iclog->ic_data, iclog->ic_io_size);
1738 * If this log buffer would straddle the end of the log we will have
1739 * to split it up into two bios, so that we can continue at the start.
1741 if (bno + BTOBB(count) > log->l_logBBsize) {
1744 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1745 GFP_NOIO, &fs_bio_set);
1746 bio_chain(split, &iclog->ic_bio);
1749 /* restart at logical offset zero for the remainder */
1750 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1753 submit_bio(&iclog->ic_bio);
1757 * We need to bump cycle number for the part of the iclog that is
1758 * written to the start of the log. Watch out for the header magic
1759 * number case, though.
1768 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1771 for (i = split_offset; i < count; i += BBSIZE) {
1772 uint32_t cycle = get_unaligned_be32(data + i);
1774 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1776 put_unaligned_be32(cycle, data + i);
1781 xlog_calc_iclog_size(
1783 struct xlog_in_core *iclog,
1786 uint32_t count_init, count;
1789 use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1790 log->l_mp->m_sb.sb_logsunit > 1;
1792 /* Add for LR header */
1793 count_init = log->l_iclog_hsize + iclog->ic_offset;
1795 /* Round out the log write size */
1797 /* we have a v2 stripe unit to use */
1798 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1800 count = BBTOB(BTOBB(count_init));
1803 ASSERT(count >= count_init);
1804 *roundoff = count - count_init;
1807 ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1809 ASSERT(*roundoff < BBTOB(1));
1814 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1815 * fashion. Previously, we should have moved the current iclog
1816 * ptr in the log to point to the next available iclog. This allows further
1817 * write to continue while this code syncs out an iclog ready to go.
1818 * Before an in-core log can be written out, the data section must be scanned
1819 * to save away the 1st word of each BBSIZE block into the header. We replace
1820 * it with the current cycle count. Each BBSIZE block is tagged with the
1821 * cycle count because there in an implicit assumption that drives will
1822 * guarantee that entire 512 byte blocks get written at once. In other words,
1823 * we can't have part of a 512 byte block written and part not written. By
1824 * tagging each block, we will know which blocks are valid when recovering
1825 * after an unclean shutdown.
1827 * This routine is single threaded on the iclog. No other thread can be in
1828 * this routine with the same iclog. Changing contents of iclog can there-
1829 * fore be done without grabbing the state machine lock. Updating the global
1830 * log will require grabbing the lock though.
1832 * The entire log manager uses a logical block numbering scheme. Only
1833 * xlog_write_iclog knows about the fact that the log may not start with
1834 * block zero on a given device.
1839 struct xlog_in_core *iclog)
1841 unsigned int count; /* byte count of bwrite */
1842 unsigned int roundoff; /* roundoff to BB or stripe */
1845 bool need_flush = true, split = false;
1847 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1849 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1851 /* move grant heads by roundoff in sync */
1852 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1853 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1855 /* put cycle number in every block */
1856 xlog_pack_data(log, iclog, roundoff);
1858 /* real byte length */
1859 size = iclog->ic_offset;
1860 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1862 iclog->ic_header.h_len = cpu_to_be32(size);
1864 XFS_STATS_INC(log->l_mp, xs_log_writes);
1865 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1867 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1869 /* Do we need to split this write into 2 parts? */
1870 if (bno + BTOBB(count) > log->l_logBBsize) {
1871 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1875 /* calculcate the checksum */
1876 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1877 iclog->ic_datap, size);
1879 * Intentionally corrupt the log record CRC based on the error injection
1880 * frequency, if defined. This facilitates testing log recovery in the
1881 * event of torn writes. Hence, set the IOABORT state to abort the log
1882 * write on I/O completion and shutdown the fs. The subsequent mount
1883 * detects the bad CRC and attempts to recover.
1886 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1887 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1888 iclog->ic_fail_crc = true;
1890 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1891 be64_to_cpu(iclog->ic_header.h_lsn));
1896 * Flush the data device before flushing the log to make sure all meta
1897 * data written back from the AIL actually made it to disk before
1898 * stamping the new log tail LSN into the log buffer. For an external
1899 * log we need to issue the flush explicitly, and unfortunately
1900 * synchronously here; for an internal log we can simply use the block
1901 * layer state machine for preflushes.
1903 if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1904 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1908 xlog_verify_iclog(log, iclog, count);
1909 xlog_write_iclog(log, iclog, bno, count, need_flush);
1913 * Deallocate a log structure
1919 xlog_in_core_t *iclog, *next_iclog;
1922 xlog_cil_destroy(log);
1925 * Cycle all the iclogbuf locks to make sure all log IO completion
1926 * is done before we tear down these buffers.
1928 iclog = log->l_iclog;
1929 for (i = 0; i < log->l_iclog_bufs; i++) {
1930 down(&iclog->ic_sema);
1931 up(&iclog->ic_sema);
1932 iclog = iclog->ic_next;
1935 iclog = log->l_iclog;
1936 for (i = 0; i < log->l_iclog_bufs; i++) {
1937 next_iclog = iclog->ic_next;
1938 kmem_free(iclog->ic_data);
1943 log->l_mp->m_log = NULL;
1944 destroy_workqueue(log->l_ioend_workqueue);
1946 } /* xlog_dealloc_log */
1949 * Update counters atomically now that memcpy is done.
1953 xlog_state_finish_copy(
1955 struct xlog_in_core *iclog,
1959 spin_lock(&log->l_icloglock);
1961 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1962 iclog->ic_offset += copy_bytes;
1964 spin_unlock(&log->l_icloglock);
1965 } /* xlog_state_finish_copy */
1971 * print out info relating to regions written which consume
1976 struct xfs_mount *mp,
1977 struct xlog_ticket *ticket)
1980 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1982 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1983 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
1984 static char *res_type_str[] = {
1985 REG_TYPE_STR(BFORMAT, "bformat"),
1986 REG_TYPE_STR(BCHUNK, "bchunk"),
1987 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
1988 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
1989 REG_TYPE_STR(IFORMAT, "iformat"),
1990 REG_TYPE_STR(ICORE, "icore"),
1991 REG_TYPE_STR(IEXT, "iext"),
1992 REG_TYPE_STR(IBROOT, "ibroot"),
1993 REG_TYPE_STR(ILOCAL, "ilocal"),
1994 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
1995 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
1996 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
1997 REG_TYPE_STR(QFORMAT, "qformat"),
1998 REG_TYPE_STR(DQUOT, "dquot"),
1999 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2000 REG_TYPE_STR(LRHEADER, "LR header"),
2001 REG_TYPE_STR(UNMOUNT, "unmount"),
2002 REG_TYPE_STR(COMMIT, "commit"),
2003 REG_TYPE_STR(TRANSHDR, "trans header"),
2004 REG_TYPE_STR(ICREATE, "inode create"),
2005 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2006 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2007 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2008 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2009 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2010 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2012 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2015 xfs_warn(mp, "ticket reservation summary:");
2016 xfs_warn(mp, " unit res = %d bytes",
2017 ticket->t_unit_res);
2018 xfs_warn(mp, " current res = %d bytes",
2019 ticket->t_curr_res);
2020 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2021 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2022 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2023 ticket->t_res_num_ophdrs, ophdr_spc);
2024 xfs_warn(mp, " ophdr + reg = %u bytes",
2025 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2026 xfs_warn(mp, " num regions = %u",
2029 for (i = 0; i < ticket->t_res_num; i++) {
2030 uint r_type = ticket->t_res_arr[i].r_type;
2031 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2032 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2033 "bad-rtype" : res_type_str[r_type]),
2034 ticket->t_res_arr[i].r_len);
2039 * Print a summary of the transaction.
2043 struct xfs_trans *tp)
2045 struct xfs_mount *mp = tp->t_mountp;
2046 struct xfs_log_item *lip;
2048 /* dump core transaction and ticket info */
2049 xfs_warn(mp, "transaction summary:");
2050 xfs_warn(mp, " log res = %d", tp->t_log_res);
2051 xfs_warn(mp, " log count = %d", tp->t_log_count);
2052 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2054 xlog_print_tic_res(mp, tp->t_ticket);
2056 /* dump each log item */
2057 list_for_each_entry(lip, &tp->t_items, li_trans) {
2058 struct xfs_log_vec *lv = lip->li_lv;
2059 struct xfs_log_iovec *vec;
2062 xfs_warn(mp, "log item: ");
2063 xfs_warn(mp, " type = 0x%x", lip->li_type);
2064 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2067 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2068 xfs_warn(mp, " size = %d", lv->lv_size);
2069 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2070 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2072 /* dump each iovec for the log item */
2073 vec = lv->lv_iovecp;
2074 for (i = 0; i < lv->lv_niovecs; i++) {
2075 int dumplen = min(vec->i_len, 32);
2077 xfs_warn(mp, " iovec[%d]", i);
2078 xfs_warn(mp, " type = 0x%x", vec->i_type);
2079 xfs_warn(mp, " len = %d", vec->i_len);
2080 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2081 xfs_hex_dump(vec->i_addr, dumplen);
2089 * Calculate the potential space needed by the log vector. Each region gets
2090 * its own xlog_op_header_t and may need to be double word aligned.
2093 xlog_write_calc_vec_length(
2094 struct xlog_ticket *ticket,
2095 struct xfs_log_vec *log_vector)
2097 struct xfs_log_vec *lv;
2102 /* acct for start rec of xact */
2103 if (ticket->t_flags & XLOG_TIC_INITED)
2106 for (lv = log_vector; lv; lv = lv->lv_next) {
2107 /* we don't write ordered log vectors */
2108 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2111 headers += lv->lv_niovecs;
2113 for (i = 0; i < lv->lv_niovecs; i++) {
2114 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2117 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2121 ticket->t_res_num_ophdrs += headers;
2122 len += headers * sizeof(struct xlog_op_header);
2128 * If first write for transaction, insert start record We can't be trying to
2129 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2132 xlog_write_start_rec(
2133 struct xlog_op_header *ophdr,
2134 struct xlog_ticket *ticket)
2136 if (!(ticket->t_flags & XLOG_TIC_INITED))
2139 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2140 ophdr->oh_clientid = ticket->t_clientid;
2142 ophdr->oh_flags = XLOG_START_TRANS;
2145 ticket->t_flags &= ~XLOG_TIC_INITED;
2147 return sizeof(struct xlog_op_header);
2150 static xlog_op_header_t *
2151 xlog_write_setup_ophdr(
2153 struct xlog_op_header *ophdr,
2154 struct xlog_ticket *ticket,
2157 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2158 ophdr->oh_clientid = ticket->t_clientid;
2161 /* are we copying a commit or unmount record? */
2162 ophdr->oh_flags = flags;
2165 * We've seen logs corrupted with bad transaction client ids. This
2166 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2167 * and shut down the filesystem.
2169 switch (ophdr->oh_clientid) {
2170 case XFS_TRANSACTION:
2176 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2177 ophdr->oh_clientid, ticket);
2185 * Set up the parameters of the region copy into the log. This has
2186 * to handle region write split across multiple log buffers - this
2187 * state is kept external to this function so that this code can
2188 * be written in an obvious, self documenting manner.
2191 xlog_write_setup_copy(
2192 struct xlog_ticket *ticket,
2193 struct xlog_op_header *ophdr,
2194 int space_available,
2198 int *last_was_partial_copy,
2199 int *bytes_consumed)
2203 still_to_copy = space_required - *bytes_consumed;
2204 *copy_off = *bytes_consumed;
2206 if (still_to_copy <= space_available) {
2207 /* write of region completes here */
2208 *copy_len = still_to_copy;
2209 ophdr->oh_len = cpu_to_be32(*copy_len);
2210 if (*last_was_partial_copy)
2211 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2212 *last_was_partial_copy = 0;
2213 *bytes_consumed = 0;
2217 /* partial write of region, needs extra log op header reservation */
2218 *copy_len = space_available;
2219 ophdr->oh_len = cpu_to_be32(*copy_len);
2220 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2221 if (*last_was_partial_copy)
2222 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2223 *bytes_consumed += *copy_len;
2224 (*last_was_partial_copy)++;
2226 /* account for new log op header */
2227 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2228 ticket->t_res_num_ophdrs++;
2230 return sizeof(struct xlog_op_header);
2234 xlog_write_copy_finish(
2236 struct xlog_in_core *iclog,
2241 int *partial_copy_len,
2243 struct xlog_in_core **commit_iclog)
2245 if (*partial_copy) {
2247 * This iclog has already been marked WANT_SYNC by
2248 * xlog_state_get_iclog_space.
2250 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2253 return xlog_state_release_iclog(log, iclog);
2257 *partial_copy_len = 0;
2259 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2260 /* no more space in this iclog - push it. */
2261 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2265 spin_lock(&log->l_icloglock);
2266 xlog_state_want_sync(log, iclog);
2267 spin_unlock(&log->l_icloglock);
2270 return xlog_state_release_iclog(log, iclog);
2271 ASSERT(flags & XLOG_COMMIT_TRANS);
2272 *commit_iclog = iclog;
2279 * Write some region out to in-core log
2281 * This will be called when writing externally provided regions or when
2282 * writing out a commit record for a given transaction.
2284 * General algorithm:
2285 * 1. Find total length of this write. This may include adding to the
2286 * lengths passed in.
2287 * 2. Check whether we violate the tickets reservation.
2288 * 3. While writing to this iclog
2289 * A. Reserve as much space in this iclog as can get
2290 * B. If this is first write, save away start lsn
2291 * C. While writing this region:
2292 * 1. If first write of transaction, write start record
2293 * 2. Write log operation header (header per region)
2294 * 3. Find out if we can fit entire region into this iclog
2295 * 4. Potentially, verify destination memcpy ptr
2296 * 5. Memcpy (partial) region
2297 * 6. If partial copy, release iclog; otherwise, continue
2298 * copying more regions into current iclog
2299 * 4. Mark want sync bit (in simulation mode)
2300 * 5. Release iclog for potential flush to on-disk log.
2303 * 1. Panic if reservation is overrun. This should never happen since
2304 * reservation amounts are generated internal to the filesystem.
2306 * 1. Tickets are single threaded data structures.
2307 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2308 * syncing routine. When a single log_write region needs to span
2309 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2310 * on all log operation writes which don't contain the end of the
2311 * region. The XLOG_END_TRANS bit is used for the in-core log
2312 * operation which contains the end of the continued log_write region.
2313 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2314 * we don't really know exactly how much space will be used. As a result,
2315 * we don't update ic_offset until the end when we know exactly how many
2316 * bytes have been written out.
2321 struct xfs_log_vec *log_vector,
2322 struct xlog_ticket *ticket,
2323 xfs_lsn_t *start_lsn,
2324 struct xlog_in_core **commit_iclog,
2327 struct xlog_in_core *iclog = NULL;
2328 struct xfs_log_iovec *vecp;
2329 struct xfs_log_vec *lv;
2332 int partial_copy = 0;
2333 int partial_copy_len = 0;
2341 len = xlog_write_calc_vec_length(ticket, log_vector);
2344 * Region headers and bytes are already accounted for.
2345 * We only need to take into account start records and
2346 * split regions in this function.
2348 if (ticket->t_flags & XLOG_TIC_INITED)
2349 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2352 * Commit record headers need to be accounted for. These
2353 * come in as separate writes so are easy to detect.
2355 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2356 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2358 if (ticket->t_curr_res < 0) {
2359 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2360 "ctx ticket reservation ran out. Need to up reservation");
2361 xlog_print_tic_res(log->l_mp, ticket);
2362 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2367 vecp = lv->lv_iovecp;
2368 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2372 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2373 &contwr, &log_offset);
2377 ASSERT(log_offset <= iclog->ic_size - 1);
2378 ptr = iclog->ic_datap + log_offset;
2380 /* start_lsn is the first lsn written to. That's all we need. */
2382 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2385 * This loop writes out as many regions as can fit in the amount
2386 * of space which was allocated by xlog_state_get_iclog_space().
2388 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2389 struct xfs_log_iovec *reg;
2390 struct xlog_op_header *ophdr;
2394 bool ordered = false;
2396 /* ordered log vectors have no regions to write */
2397 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2398 ASSERT(lv->lv_niovecs == 0);
2404 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2405 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2407 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2408 if (start_rec_copy) {
2410 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2414 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2418 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2419 sizeof(struct xlog_op_header));
2421 len += xlog_write_setup_copy(ticket, ophdr,
2422 iclog->ic_size-log_offset,
2424 ©_off, ©_len,
2427 xlog_verify_dest_ptr(log, ptr);
2432 * Unmount records just log an opheader, so can have
2433 * empty payloads with no data region to copy. Hence we
2434 * only copy the payload if the vector says it has data
2437 ASSERT(copy_len >= 0);
2439 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2440 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2443 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2445 data_cnt += contwr ? copy_len : 0;
2447 error = xlog_write_copy_finish(log, iclog, flags,
2448 &record_cnt, &data_cnt,
2457 * if we had a partial copy, we need to get more iclog
2458 * space but we don't want to increment the region
2459 * index because there is still more is this region to
2462 * If we completed writing this region, and we flushed
2463 * the iclog (indicated by resetting of the record
2464 * count), then we also need to get more log space. If
2465 * this was the last record, though, we are done and
2471 if (++index == lv->lv_niovecs) {
2476 vecp = lv->lv_iovecp;
2478 if (record_cnt == 0 && !ordered) {
2488 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2490 return xlog_state_release_iclog(log, iclog);
2492 ASSERT(flags & XLOG_COMMIT_TRANS);
2493 *commit_iclog = iclog;
2498 /*****************************************************************************
2500 * State Machine functions
2502 *****************************************************************************
2505 /* Clean iclogs starting from the head. This ordering must be
2506 * maintained, so an iclog doesn't become ACTIVE beyond one that
2507 * is SYNCING. This is also required to maintain the notion that we use
2508 * a ordered wait queue to hold off would be writers to the log when every
2509 * iclog is trying to sync to disk.
2511 * State Change: DIRTY -> ACTIVE
2514 xlog_state_clean_log(
2517 xlog_in_core_t *iclog;
2520 iclog = log->l_iclog;
2522 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2523 iclog->ic_state = XLOG_STATE_ACTIVE;
2524 iclog->ic_offset = 0;
2525 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2527 * If the number of ops in this iclog indicate it just
2528 * contains the dummy transaction, we can
2529 * change state into IDLE (the second time around).
2530 * Otherwise we should change the state into
2532 * We don't need to cover the dummy.
2535 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2540 * We have two dirty iclogs so start over
2541 * This could also be num of ops indicates
2542 * this is not the dummy going out.
2546 iclog->ic_header.h_num_logops = 0;
2547 memset(iclog->ic_header.h_cycle_data, 0,
2548 sizeof(iclog->ic_header.h_cycle_data));
2549 iclog->ic_header.h_lsn = 0;
2550 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2553 break; /* stop cleaning */
2554 iclog = iclog->ic_next;
2555 } while (iclog != log->l_iclog);
2557 /* log is locked when we are called */
2559 * Change state for the dummy log recording.
2560 * We usually go to NEED. But we go to NEED2 if the changed indicates
2561 * we are done writing the dummy record.
2562 * If we are done with the second dummy recored (DONE2), then
2566 switch (log->l_covered_state) {
2567 case XLOG_STATE_COVER_IDLE:
2568 case XLOG_STATE_COVER_NEED:
2569 case XLOG_STATE_COVER_NEED2:
2570 log->l_covered_state = XLOG_STATE_COVER_NEED;
2573 case XLOG_STATE_COVER_DONE:
2575 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2577 log->l_covered_state = XLOG_STATE_COVER_NEED;
2580 case XLOG_STATE_COVER_DONE2:
2582 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2584 log->l_covered_state = XLOG_STATE_COVER_NEED;
2591 } /* xlog_state_clean_log */
2594 xlog_get_lowest_lsn(
2597 struct xlog_in_core *iclog = log->l_iclog;
2598 xfs_lsn_t lowest_lsn = 0, lsn;
2601 if (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))
2604 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2605 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2607 } while ((iclog = iclog->ic_next) != log->l_iclog);
2613 xlog_state_do_callback(
2616 struct xlog_in_core *ciclog)
2618 xlog_in_core_t *iclog;
2619 xlog_in_core_t *first_iclog; /* used to know when we've
2620 * processed all iclogs once */
2622 xfs_lsn_t lowest_lsn;
2623 int ioerrors; /* counter: iclogs with errors */
2624 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2625 int funcdidcallbacks; /* flag: function did callbacks */
2626 int repeats; /* for issuing console warnings if
2627 * looping too many times */
2630 spin_lock(&log->l_icloglock);
2631 first_iclog = iclog = log->l_iclog;
2633 funcdidcallbacks = 0;
2638 * Scan all iclogs starting with the one pointed to by the
2639 * log. Reset this starting point each time the log is
2640 * unlocked (during callbacks).
2642 * Keep looping through iclogs until one full pass is made
2643 * without running any callbacks.
2645 first_iclog = log->l_iclog;
2646 iclog = log->l_iclog;
2647 loopdidcallbacks = 0;
2652 /* skip all iclogs in the ACTIVE & DIRTY states */
2653 if (iclog->ic_state &
2654 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2655 iclog = iclog->ic_next;
2660 * Between marking a filesystem SHUTDOWN and stopping
2661 * the log, we do flush all iclogs to disk (if there
2662 * wasn't a log I/O error). So, we do want things to
2663 * go smoothly in case of just a SHUTDOWN w/o a
2666 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2668 * Can only perform callbacks in order. Since
2669 * this iclog is not in the DONE_SYNC/
2670 * DO_CALLBACK state, we skip the rest and
2671 * just try to clean up. If we set our iclog
2672 * to DO_CALLBACK, we will not process it when
2673 * we retry since a previous iclog is in the
2674 * CALLBACK and the state cannot change since
2675 * we are holding the l_icloglock.
2677 if (!(iclog->ic_state &
2678 (XLOG_STATE_DONE_SYNC |
2679 XLOG_STATE_DO_CALLBACK))) {
2680 if (ciclog && (ciclog->ic_state ==
2681 XLOG_STATE_DONE_SYNC)) {
2682 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2687 * We now have an iclog that is in either the
2688 * DO_CALLBACK or DONE_SYNC states. The other
2689 * states (WANT_SYNC, SYNCING, or CALLBACK were
2690 * caught by the above if and are going to
2691 * clean (i.e. we aren't doing their callbacks)
2696 * We will do one more check here to see if we
2697 * have chased our tail around.
2700 lowest_lsn = xlog_get_lowest_lsn(log);
2702 XFS_LSN_CMP(lowest_lsn,
2703 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2704 iclog = iclog->ic_next;
2705 continue; /* Leave this iclog for
2709 iclog->ic_state = XLOG_STATE_CALLBACK;
2713 * Completion of a iclog IO does not imply that
2714 * a transaction has completed, as transactions
2715 * can be large enough to span many iclogs. We
2716 * cannot change the tail of the log half way
2717 * through a transaction as this may be the only
2718 * transaction in the log and moving th etail to
2719 * point to the middle of it will prevent
2720 * recovery from finding the start of the
2721 * transaction. Hence we should only update the
2722 * last_sync_lsn if this iclog contains
2723 * transaction completion callbacks on it.
2725 * We have to do this before we drop the
2726 * icloglock to ensure we are the only one that
2729 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2730 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2731 if (!list_empty_careful(&iclog->ic_callbacks))
2732 atomic64_set(&log->l_last_sync_lsn,
2733 be64_to_cpu(iclog->ic_header.h_lsn));
2738 spin_unlock(&log->l_icloglock);
2741 * Keep processing entries in the callback list until
2742 * we come around and it is empty. We need to
2743 * atomically see that the list is empty and change the
2744 * state to DIRTY so that we don't miss any more
2745 * callbacks being added.
2747 spin_lock(&iclog->ic_callback_lock);
2748 while (!list_empty(&iclog->ic_callbacks)) {
2751 list_splice_init(&iclog->ic_callbacks, &tmp);
2753 spin_unlock(&iclog->ic_callback_lock);
2754 xlog_cil_process_committed(&tmp, aborted);
2755 spin_lock(&iclog->ic_callback_lock);
2761 spin_lock(&log->l_icloglock);
2762 spin_unlock(&iclog->ic_callback_lock);
2763 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2764 iclog->ic_state = XLOG_STATE_DIRTY;
2767 * Transition from DIRTY to ACTIVE if applicable.
2768 * NOP if STATE_IOERROR.
2770 xlog_state_clean_log(log);
2772 /* wake up threads waiting in xfs_log_force() */
2773 wake_up_all(&iclog->ic_force_wait);
2775 iclog = iclog->ic_next;
2776 } while (first_iclog != iclog);
2778 if (repeats > 5000) {
2779 flushcnt += repeats;
2782 "%s: possible infinite loop (%d iterations)",
2783 __func__, flushcnt);
2785 } while (!ioerrors && loopdidcallbacks);
2789 * Make one last gasp attempt to see if iclogs are being left in limbo.
2790 * If the above loop finds an iclog earlier than the current iclog and
2791 * in one of the syncing states, the current iclog is put into
2792 * DO_CALLBACK and the callbacks are deferred to the completion of the
2793 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2794 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2797 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2798 * for ic_state == SYNCING.
2800 if (funcdidcallbacks) {
2801 first_iclog = iclog = log->l_iclog;
2803 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2805 * Terminate the loop if iclogs are found in states
2806 * which will cause other threads to clean up iclogs.
2808 * SYNCING - i/o completion will go through logs
2809 * DONE_SYNC - interrupt thread should be waiting for
2811 * IOERROR - give up hope all ye who enter here
2813 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2814 iclog->ic_state & XLOG_STATE_SYNCING ||
2815 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2816 iclog->ic_state == XLOG_STATE_IOERROR )
2818 iclog = iclog->ic_next;
2819 } while (first_iclog != iclog);
2823 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2825 spin_unlock(&log->l_icloglock);
2828 wake_up_all(&log->l_flush_wait);
2833 * Finish transitioning this iclog to the dirty state.
2835 * Make sure that we completely execute this routine only when this is
2836 * the last call to the iclog. There is a good chance that iclog flushes,
2837 * when we reach the end of the physical log, get turned into 2 separate
2838 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2839 * routine. By using the reference count bwritecnt, we guarantee that only
2840 * the second completion goes through.
2842 * Callbacks could take time, so they are done outside the scope of the
2843 * global state machine log lock.
2846 xlog_state_done_syncing(
2847 struct xlog_in_core *iclog,
2850 struct xlog *log = iclog->ic_log;
2852 spin_lock(&log->l_icloglock);
2854 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2855 iclog->ic_state == XLOG_STATE_IOERROR);
2856 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2859 * If we got an error, either on the first buffer, or in the case of
2860 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2861 * and none should ever be attempted to be written to disk
2864 if (iclog->ic_state != XLOG_STATE_IOERROR)
2865 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2868 * Someone could be sleeping prior to writing out the next
2869 * iclog buffer, we wake them all, one will get to do the
2870 * I/O, the others get to wait for the result.
2872 wake_up_all(&iclog->ic_write_wait);
2873 spin_unlock(&log->l_icloglock);
2874 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2875 } /* xlog_state_done_syncing */
2879 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2880 * sleep. We wait on the flush queue on the head iclog as that should be
2881 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2882 * we will wait here and all new writes will sleep until a sync completes.
2884 * The in-core logs are used in a circular fashion. They are not used
2885 * out-of-order even when an iclog past the head is free.
2888 * * log_offset where xlog_write() can start writing into the in-core
2890 * * in-core log pointer to which xlog_write() should write.
2891 * * boolean indicating this is a continued write to an in-core log.
2892 * If this is the last write, then the in-core log's offset field
2893 * needs to be incremented, depending on the amount of data which
2897 xlog_state_get_iclog_space(
2900 struct xlog_in_core **iclogp,
2901 struct xlog_ticket *ticket,
2902 int *continued_write,
2906 xlog_rec_header_t *head;
2907 xlog_in_core_t *iclog;
2911 spin_lock(&log->l_icloglock);
2912 if (XLOG_FORCED_SHUTDOWN(log)) {
2913 spin_unlock(&log->l_icloglock);
2917 iclog = log->l_iclog;
2918 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2919 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2921 /* Wait for log writes to have flushed */
2922 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2926 head = &iclog->ic_header;
2928 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2929 log_offset = iclog->ic_offset;
2931 /* On the 1st write to an iclog, figure out lsn. This works
2932 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2933 * committing to. If the offset is set, that's how many blocks
2936 if (log_offset == 0) {
2937 ticket->t_curr_res -= log->l_iclog_hsize;
2938 xlog_tic_add_region(ticket,
2940 XLOG_REG_TYPE_LRHEADER);
2941 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2942 head->h_lsn = cpu_to_be64(
2943 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2944 ASSERT(log->l_curr_block >= 0);
2947 /* If there is enough room to write everything, then do it. Otherwise,
2948 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2949 * bit is on, so this will get flushed out. Don't update ic_offset
2950 * until you know exactly how many bytes get copied. Therefore, wait
2951 * until later to update ic_offset.
2953 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2954 * can fit into remaining data section.
2956 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2957 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2960 * If I'm the only one writing to this iclog, sync it to disk.
2961 * We need to do an atomic compare and decrement here to avoid
2962 * racing with concurrent atomic_dec_and_lock() calls in
2963 * xlog_state_release_iclog() when there is more than one
2964 * reference to the iclog.
2966 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2967 /* we are the only one */
2968 spin_unlock(&log->l_icloglock);
2969 error = xlog_state_release_iclog(log, iclog);
2973 spin_unlock(&log->l_icloglock);
2978 /* Do we have enough room to write the full amount in the remainder
2979 * of this iclog? Or must we continue a write on the next iclog and
2980 * mark this iclog as completely taken? In the case where we switch
2981 * iclogs (to mark it taken), this particular iclog will release/sync
2982 * to disk in xlog_write().
2984 if (len <= iclog->ic_size - iclog->ic_offset) {
2985 *continued_write = 0;
2986 iclog->ic_offset += len;
2988 *continued_write = 1;
2989 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2993 ASSERT(iclog->ic_offset <= iclog->ic_size);
2994 spin_unlock(&log->l_icloglock);
2996 *logoffsetp = log_offset;
2998 } /* xlog_state_get_iclog_space */
3000 /* The first cnt-1 times through here we don't need to
3001 * move the grant write head because the permanent
3002 * reservation has reserved cnt times the unit amount.
3003 * Release part of current permanent unit reservation and
3004 * reset current reservation to be one units worth. Also
3005 * move grant reservation head forward.
3008 xlog_regrant_reserve_log_space(
3010 struct xlog_ticket *ticket)
3012 trace_xfs_log_regrant_reserve_enter(log, ticket);
3014 if (ticket->t_cnt > 0)
3017 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3018 ticket->t_curr_res);
3019 xlog_grant_sub_space(log, &log->l_write_head.grant,
3020 ticket->t_curr_res);
3021 ticket->t_curr_res = ticket->t_unit_res;
3022 xlog_tic_reset_res(ticket);
3024 trace_xfs_log_regrant_reserve_sub(log, ticket);
3026 /* just return if we still have some of the pre-reserved space */
3027 if (ticket->t_cnt > 0)
3030 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3031 ticket->t_unit_res);
3033 trace_xfs_log_regrant_reserve_exit(log, ticket);
3035 ticket->t_curr_res = ticket->t_unit_res;
3036 xlog_tic_reset_res(ticket);
3037 } /* xlog_regrant_reserve_log_space */
3041 * Give back the space left from a reservation.
3043 * All the information we need to make a correct determination of space left
3044 * is present. For non-permanent reservations, things are quite easy. The
3045 * count should have been decremented to zero. We only need to deal with the
3046 * space remaining in the current reservation part of the ticket. If the
3047 * ticket contains a permanent reservation, there may be left over space which
3048 * needs to be released. A count of N means that N-1 refills of the current
3049 * reservation can be done before we need to ask for more space. The first
3050 * one goes to fill up the first current reservation. Once we run out of
3051 * space, the count will stay at zero and the only space remaining will be
3052 * in the current reservation field.
3055 xlog_ungrant_log_space(
3057 struct xlog_ticket *ticket)
3061 if (ticket->t_cnt > 0)
3064 trace_xfs_log_ungrant_enter(log, ticket);
3065 trace_xfs_log_ungrant_sub(log, ticket);
3068 * If this is a permanent reservation ticket, we may be able to free
3069 * up more space based on the remaining count.
3071 bytes = ticket->t_curr_res;
3072 if (ticket->t_cnt > 0) {
3073 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3074 bytes += ticket->t_unit_res*ticket->t_cnt;
3077 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3078 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3080 trace_xfs_log_ungrant_exit(log, ticket);
3082 xfs_log_space_wake(log->l_mp);
3086 * Flush iclog to disk if this is the last reference to the given iclog and
3087 * the WANT_SYNC bit is set.
3089 * When this function is entered, the iclog is not necessarily in the
3090 * WANT_SYNC state. It may be sitting around waiting to get filled.
3095 xlog_state_release_iclog(
3097 struct xlog_in_core *iclog)
3099 int sync = 0; /* do we sync? */
3101 if (iclog->ic_state & XLOG_STATE_IOERROR)
3104 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3105 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3108 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3109 spin_unlock(&log->l_icloglock);
3112 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3113 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3115 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3116 /* update tail before writing to iclog */
3117 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3119 iclog->ic_state = XLOG_STATE_SYNCING;
3120 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3121 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3122 /* cycle incremented when incrementing curr_block */
3124 spin_unlock(&log->l_icloglock);
3127 * We let the log lock go, so it's possible that we hit a log I/O
3128 * error or some other SHUTDOWN condition that marks the iclog
3129 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3130 * this iclog has consistent data, so we ignore IOERROR
3131 * flags after this point.
3134 xlog_sync(log, iclog);
3136 } /* xlog_state_release_iclog */
3140 * This routine will mark the current iclog in the ring as WANT_SYNC
3141 * and move the current iclog pointer to the next iclog in the ring.
3142 * When this routine is called from xlog_state_get_iclog_space(), the
3143 * exact size of the iclog has not yet been determined. All we know is
3144 * that every data block. We have run out of space in this log record.
3147 xlog_state_switch_iclogs(
3149 struct xlog_in_core *iclog,
3152 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3154 eventual_size = iclog->ic_offset;
3155 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3156 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3157 log->l_prev_block = log->l_curr_block;
3158 log->l_prev_cycle = log->l_curr_cycle;
3160 /* roll log?: ic_offset changed later */
3161 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3163 /* Round up to next log-sunit */
3164 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3165 log->l_mp->m_sb.sb_logsunit > 1) {
3166 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3167 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3170 if (log->l_curr_block >= log->l_logBBsize) {
3172 * Rewind the current block before the cycle is bumped to make
3173 * sure that the combined LSN never transiently moves forward
3174 * when the log wraps to the next cycle. This is to support the
3175 * unlocked sample of these fields from xlog_valid_lsn(). Most
3176 * other cases should acquire l_icloglock.
3178 log->l_curr_block -= log->l_logBBsize;
3179 ASSERT(log->l_curr_block >= 0);
3181 log->l_curr_cycle++;
3182 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3183 log->l_curr_cycle++;
3185 ASSERT(iclog == log->l_iclog);
3186 log->l_iclog = iclog->ic_next;
3187 } /* xlog_state_switch_iclogs */
3190 * Write out all data in the in-core log as of this exact moment in time.
3192 * Data may be written to the in-core log during this call. However,
3193 * we don't guarantee this data will be written out. A change from past
3194 * implementation means this routine will *not* write out zero length LRs.
3196 * Basically, we try and perform an intelligent scan of the in-core logs.
3197 * If we determine there is no flushable data, we just return. There is no
3198 * flushable data if:
3200 * 1. the current iclog is active and has no data; the previous iclog
3201 * is in the active or dirty state.
3202 * 2. the current iclog is drity, and the previous iclog is in the
3203 * active or dirty state.
3207 * 1. the current iclog is not in the active nor dirty state.
3208 * 2. the current iclog dirty, and the previous iclog is not in the
3209 * active nor dirty state.
3210 * 3. the current iclog is active, and there is another thread writing
3211 * to this particular iclog.
3212 * 4. a) the current iclog is active and has no other writers
3213 * b) when we return from flushing out this iclog, it is still
3214 * not in the active nor dirty state.
3218 struct xfs_mount *mp,
3221 struct xlog *log = mp->m_log;
3222 struct xlog_in_core *iclog;
3225 XFS_STATS_INC(mp, xs_log_force);
3226 trace_xfs_log_force(mp, 0, _RET_IP_);
3228 xlog_cil_force(log);
3230 spin_lock(&log->l_icloglock);
3231 iclog = log->l_iclog;
3232 if (iclog->ic_state & XLOG_STATE_IOERROR)
3235 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3236 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3237 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3239 * If the head is dirty or (active and empty), then we need to
3240 * look at the previous iclog.
3242 * If the previous iclog is active or dirty we are done. There
3243 * is nothing to sync out. Otherwise, we attach ourselves to the
3244 * previous iclog and go to sleep.
3246 iclog = iclog->ic_prev;
3247 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3248 iclog->ic_state == XLOG_STATE_DIRTY)
3250 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3251 if (atomic_read(&iclog->ic_refcnt) == 0) {
3253 * We are the only one with access to this iclog.
3255 * Flush it out now. There should be a roundoff of zero
3256 * to show that someone has already taken care of the
3257 * roundoff from the previous sync.
3259 atomic_inc(&iclog->ic_refcnt);
3260 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3261 xlog_state_switch_iclogs(log, iclog, 0);
3262 spin_unlock(&log->l_icloglock);
3264 if (xlog_state_release_iclog(log, iclog))
3267 spin_lock(&log->l_icloglock);
3268 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3269 iclog->ic_state == XLOG_STATE_DIRTY)
3273 * Someone else is writing to this iclog.
3275 * Use its call to flush out the data. However, the
3276 * other thread may not force out this LR, so we mark
3279 xlog_state_switch_iclogs(log, iclog, 0);
3283 * If the head iclog is not active nor dirty, we just attach
3284 * ourselves to the head and go to sleep if necessary.
3289 if (!(flags & XFS_LOG_SYNC))
3292 if (iclog->ic_state & XLOG_STATE_IOERROR)
3294 XFS_STATS_INC(mp, xs_log_force_sleep);
3295 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3296 if (iclog->ic_state & XLOG_STATE_IOERROR)
3301 spin_unlock(&log->l_icloglock);
3304 spin_unlock(&log->l_icloglock);
3309 __xfs_log_force_lsn(
3310 struct xfs_mount *mp,
3316 struct xlog *log = mp->m_log;
3317 struct xlog_in_core *iclog;
3319 spin_lock(&log->l_icloglock);
3320 iclog = log->l_iclog;
3321 if (iclog->ic_state & XLOG_STATE_IOERROR)
3324 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3325 iclog = iclog->ic_next;
3326 if (iclog == log->l_iclog)
3330 if (iclog->ic_state == XLOG_STATE_DIRTY)
3333 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3335 * We sleep here if we haven't already slept (e.g. this is the
3336 * first time we've looked at the correct iclog buf) and the
3337 * buffer before us is going to be sync'ed. The reason for this
3338 * is that if we are doing sync transactions here, by waiting
3339 * for the previous I/O to complete, we can allow a few more
3340 * transactions into this iclog before we close it down.
3342 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3343 * refcnt so we can release the log (which drops the ref count).
3344 * The state switch keeps new transaction commits from using
3345 * this buffer. When the current commits finish writing into
3346 * the buffer, the refcount will drop to zero and the buffer
3349 if (!already_slept &&
3350 (iclog->ic_prev->ic_state &
3351 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3352 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3354 XFS_STATS_INC(mp, xs_log_force_sleep);
3356 xlog_wait(&iclog->ic_prev->ic_write_wait,
3360 atomic_inc(&iclog->ic_refcnt);
3361 xlog_state_switch_iclogs(log, iclog, 0);
3362 spin_unlock(&log->l_icloglock);
3363 if (xlog_state_release_iclog(log, iclog))
3367 spin_lock(&log->l_icloglock);
3370 if (!(flags & XFS_LOG_SYNC) ||
3371 (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3374 if (iclog->ic_state & XLOG_STATE_IOERROR)
3377 XFS_STATS_INC(mp, xs_log_force_sleep);
3378 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3379 if (iclog->ic_state & XLOG_STATE_IOERROR)
3384 spin_unlock(&log->l_icloglock);
3387 spin_unlock(&log->l_icloglock);
3392 * Force the in-core log to disk for a specific LSN.
3394 * Find in-core log with lsn.
3395 * If it is in the DIRTY state, just return.
3396 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3397 * state and go to sleep or return.
3398 * If it is in any other state, go to sleep or return.
3400 * Synchronous forces are implemented with a wait queue. All callers trying
3401 * to force a given lsn to disk must wait on the queue attached to the
3402 * specific in-core log. When given in-core log finally completes its write
3403 * to disk, that thread will wake up all threads waiting on the queue.
3407 struct xfs_mount *mp,
3415 XFS_STATS_INC(mp, xs_log_force);
3416 trace_xfs_log_force(mp, lsn, _RET_IP_);
3418 lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3419 if (lsn == NULLCOMMITLSN)
3422 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3424 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3429 * Called when we want to mark the current iclog as being ready to sync to
3433 xlog_state_want_sync(
3435 struct xlog_in_core *iclog)
3437 assert_spin_locked(&log->l_icloglock);
3439 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3440 xlog_state_switch_iclogs(log, iclog, 0);
3442 ASSERT(iclog->ic_state &
3443 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3448 /*****************************************************************************
3452 *****************************************************************************
3456 * Free a used ticket when its refcount falls to zero.
3460 xlog_ticket_t *ticket)
3462 ASSERT(atomic_read(&ticket->t_ref) > 0);
3463 if (atomic_dec_and_test(&ticket->t_ref))
3464 kmem_zone_free(xfs_log_ticket_zone, ticket);
3469 xlog_ticket_t *ticket)
3471 ASSERT(atomic_read(&ticket->t_ref) > 0);
3472 atomic_inc(&ticket->t_ref);
3477 * Figure out the total log space unit (in bytes) that would be
3478 * required for a log ticket.
3481 xfs_log_calc_unit_res(
3482 struct xfs_mount *mp,
3485 struct xlog *log = mp->m_log;
3490 * Permanent reservations have up to 'cnt'-1 active log operations
3491 * in the log. A unit in this case is the amount of space for one
3492 * of these log operations. Normal reservations have a cnt of 1
3493 * and their unit amount is the total amount of space required.
3495 * The following lines of code account for non-transaction data
3496 * which occupy space in the on-disk log.
3498 * Normal form of a transaction is:
3499 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3500 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3502 * We need to account for all the leadup data and trailer data
3503 * around the transaction data.
3504 * And then we need to account for the worst case in terms of using
3506 * The worst case will happen if:
3507 * - the placement of the transaction happens to be such that the
3508 * roundoff is at its maximum
3509 * - the transaction data is synced before the commit record is synced
3510 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3511 * Therefore the commit record is in its own Log Record.
3512 * This can happen as the commit record is called with its
3513 * own region to xlog_write().
3514 * This then means that in the worst case, roundoff can happen for
3515 * the commit-rec as well.
3516 * The commit-rec is smaller than padding in this scenario and so it is
3517 * not added separately.
3520 /* for trans header */
3521 unit_bytes += sizeof(xlog_op_header_t);
3522 unit_bytes += sizeof(xfs_trans_header_t);
3525 unit_bytes += sizeof(xlog_op_header_t);
3528 * for LR headers - the space for data in an iclog is the size minus
3529 * the space used for the headers. If we use the iclog size, then we
3530 * undercalculate the number of headers required.
3532 * Furthermore - the addition of op headers for split-recs might
3533 * increase the space required enough to require more log and op
3534 * headers, so take that into account too.
3536 * IMPORTANT: This reservation makes the assumption that if this
3537 * transaction is the first in an iclog and hence has the LR headers
3538 * accounted to it, then the remaining space in the iclog is
3539 * exclusively for this transaction. i.e. if the transaction is larger
3540 * than the iclog, it will be the only thing in that iclog.
3541 * Fundamentally, this means we must pass the entire log vector to
3542 * xlog_write to guarantee this.
3544 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3545 num_headers = howmany(unit_bytes, iclog_space);
3547 /* for split-recs - ophdrs added when data split over LRs */
3548 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3550 /* add extra header reservations if we overrun */
3551 while (!num_headers ||
3552 howmany(unit_bytes, iclog_space) > num_headers) {
3553 unit_bytes += sizeof(xlog_op_header_t);
3556 unit_bytes += log->l_iclog_hsize * num_headers;
3558 /* for commit-rec LR header - note: padding will subsume the ophdr */
3559 unit_bytes += log->l_iclog_hsize;
3561 /* for roundoff padding for transaction data and one for commit record */
3562 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3563 /* log su roundoff */
3564 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3567 unit_bytes += 2 * BBSIZE;
3574 * Allocate and initialise a new log ticket.
3576 struct xlog_ticket *
3583 xfs_km_flags_t alloc_flags)
3585 struct xlog_ticket *tic;
3588 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3592 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3594 atomic_set(&tic->t_ref, 1);
3595 tic->t_task = current;
3596 INIT_LIST_HEAD(&tic->t_queue);
3597 tic->t_unit_res = unit_res;
3598 tic->t_curr_res = unit_res;
3601 tic->t_tid = prandom_u32();
3602 tic->t_clientid = client;
3603 tic->t_flags = XLOG_TIC_INITED;
3605 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3607 xlog_tic_reset_res(tic);
3613 /******************************************************************************
3615 * Log debug routines
3617 ******************************************************************************
3621 * Make sure that the destination ptr is within the valid data region of
3622 * one of the iclogs. This uses backup pointers stored in a different
3623 * part of the log in case we trash the log structure.
3626 xlog_verify_dest_ptr(
3633 for (i = 0; i < log->l_iclog_bufs; i++) {
3634 if (ptr >= log->l_iclog_bak[i] &&
3635 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3640 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3644 * Check to make sure the grant write head didn't just over lap the tail. If
3645 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3646 * the cycles differ by exactly one and check the byte count.
3648 * This check is run unlocked, so can give false positives. Rather than assert
3649 * on failures, use a warn-once flag and a panic tag to allow the admin to
3650 * determine if they want to panic the machine when such an error occurs. For
3651 * debug kernels this will have the same effect as using an assert but, unlinke
3652 * an assert, it can be turned off at runtime.
3655 xlog_verify_grant_tail(
3658 int tail_cycle, tail_blocks;
3661 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3662 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3663 if (tail_cycle != cycle) {
3664 if (cycle - 1 != tail_cycle &&
3665 !(log->l_flags & XLOG_TAIL_WARN)) {
3666 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3667 "%s: cycle - 1 != tail_cycle", __func__);
3668 log->l_flags |= XLOG_TAIL_WARN;
3671 if (space > BBTOB(tail_blocks) &&
3672 !(log->l_flags & XLOG_TAIL_WARN)) {
3673 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3674 "%s: space > BBTOB(tail_blocks)", __func__);
3675 log->l_flags |= XLOG_TAIL_WARN;
3680 /* check if it will fit */
3682 xlog_verify_tail_lsn(
3684 struct xlog_in_core *iclog,
3689 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3691 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3692 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3693 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3695 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3697 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3698 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3700 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3701 if (blocks < BTOBB(iclog->ic_offset) + 1)
3702 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3704 } /* xlog_verify_tail_lsn */
3707 * Perform a number of checks on the iclog before writing to disk.
3709 * 1. Make sure the iclogs are still circular
3710 * 2. Make sure we have a good magic number
3711 * 3. Make sure we don't have magic numbers in the data
3712 * 4. Check fields of each log operation header for:
3713 * A. Valid client identifier
3714 * B. tid ptr value falls in valid ptr space (user space code)
3715 * C. Length in log record header is correct according to the
3716 * individual operation headers within record.
3717 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3718 * log, check the preceding blocks of the physical log to make sure all
3719 * the cycle numbers agree with the current cycle number.
3724 struct xlog_in_core *iclog,
3727 xlog_op_header_t *ophead;
3728 xlog_in_core_t *icptr;
3729 xlog_in_core_2_t *xhdr;
3730 void *base_ptr, *ptr, *p;
3731 ptrdiff_t field_offset;
3733 int len, i, j, k, op_len;
3736 /* check validity of iclog pointers */
3737 spin_lock(&log->l_icloglock);
3738 icptr = log->l_iclog;
3739 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3742 if (icptr != log->l_iclog)
3743 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3744 spin_unlock(&log->l_icloglock);
3746 /* check log magic numbers */
3747 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3748 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3750 base_ptr = ptr = &iclog->ic_header;
3751 p = &iclog->ic_header;
3752 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3753 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3754 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3759 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3760 base_ptr = ptr = iclog->ic_datap;
3762 xhdr = iclog->ic_data;
3763 for (i = 0; i < len; i++) {
3766 /* clientid is only 1 byte */
3767 p = &ophead->oh_clientid;
3768 field_offset = p - base_ptr;
3769 if (field_offset & 0x1ff) {
3770 clientid = ophead->oh_clientid;
3772 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3773 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3774 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3775 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3776 clientid = xlog_get_client_id(
3777 xhdr[j].hic_xheader.xh_cycle_data[k]);
3779 clientid = xlog_get_client_id(
3780 iclog->ic_header.h_cycle_data[idx]);
3783 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3785 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3786 __func__, clientid, ophead,
3787 (unsigned long)field_offset);
3790 p = &ophead->oh_len;
3791 field_offset = p - base_ptr;
3792 if (field_offset & 0x1ff) {
3793 op_len = be32_to_cpu(ophead->oh_len);
3795 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3796 (uintptr_t)iclog->ic_datap);
3797 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3798 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3799 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3800 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3802 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3805 ptr += sizeof(xlog_op_header_t) + op_len;
3807 } /* xlog_verify_iclog */
3811 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3817 xlog_in_core_t *iclog, *ic;
3819 iclog = log->l_iclog;
3820 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3822 * Mark all the incore logs IOERROR.
3823 * From now on, no log flushes will result.
3827 ic->ic_state = XLOG_STATE_IOERROR;
3829 } while (ic != iclog);
3833 * Return non-zero, if state transition has already happened.
3839 * This is called from xfs_force_shutdown, when we're forcibly
3840 * shutting down the filesystem, typically because of an IO error.
3841 * Our main objectives here are to make sure that:
3842 * a. if !logerror, flush the logs to disk. Anything modified
3843 * after this is ignored.
3844 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3845 * parties to find out, 'atomically'.
3846 * c. those who're sleeping on log reservations, pinned objects and
3847 * other resources get woken up, and be told the bad news.
3848 * d. nothing new gets queued up after (b) and (c) are done.
3850 * Note: for the !logerror case we need to flush the regions held in memory out
3851 * to disk first. This needs to be done before the log is marked as shutdown,
3852 * otherwise the iclog writes will fail.
3855 xfs_log_force_umount(
3856 struct xfs_mount *mp,
3865 * If this happens during log recovery, don't worry about
3866 * locking; the log isn't open for business yet.
3869 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3870 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3872 mp->m_sb_bp->b_flags |= XBF_DONE;
3877 * Somebody could've already done the hard work for us.
3878 * No need to get locks for this.
3880 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3881 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3886 * Flush all the completed transactions to disk before marking the log
3887 * being shut down. We need to do it in this order to ensure that
3888 * completed operations are safely on disk before we shut down, and that
3889 * we don't have to issue any buffer IO after the shutdown flags are set
3890 * to guarantee this.
3893 xfs_log_force(mp, XFS_LOG_SYNC);
3896 * mark the filesystem and the as in a shutdown state and wake
3897 * everybody up to tell them the bad news.
3899 spin_lock(&log->l_icloglock);
3900 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3902 mp->m_sb_bp->b_flags |= XBF_DONE;
3905 * Mark the log and the iclogs with IO error flags to prevent any
3906 * further log IO from being issued or completed.
3908 log->l_flags |= XLOG_IO_ERROR;
3909 retval = xlog_state_ioerror(log);
3910 spin_unlock(&log->l_icloglock);
3913 * We don't want anybody waiting for log reservations after this. That
3914 * means we have to wake up everybody queued up on reserveq as well as
3915 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3916 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3917 * action is protected by the grant locks.
3919 xlog_grant_head_wake_all(&log->l_reserve_head);
3920 xlog_grant_head_wake_all(&log->l_write_head);
3923 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3924 * as if the log writes were completed. The abort handling in the log
3925 * item committed callback functions will do this again under lock to
3928 wake_up_all(&log->l_cilp->xc_commit_wait);
3929 xlog_state_do_callback(log, true, NULL);
3931 #ifdef XFSERRORDEBUG
3933 xlog_in_core_t *iclog;
3935 spin_lock(&log->l_icloglock);
3936 iclog = log->l_iclog;
3938 ASSERT(iclog->ic_callback == 0);
3939 iclog = iclog->ic_next;
3940 } while (iclog != log->l_iclog);
3941 spin_unlock(&log->l_icloglock);
3944 /* return non-zero if log IOERROR transition had already happened */
3952 xlog_in_core_t *iclog;
3954 iclog = log->l_iclog;
3956 /* endianness does not matter here, zero is zero in
3959 if (iclog->ic_header.h_num_logops)
3961 iclog = iclog->ic_next;
3962 } while (iclog != log->l_iclog);
3967 * Verify that an LSN stamped into a piece of metadata is valid. This is
3968 * intended for use in read verifiers on v5 superblocks.
3972 struct xfs_mount *mp,
3975 struct xlog *log = mp->m_log;
3979 * norecovery mode skips mount-time log processing and unconditionally
3980 * resets the in-core LSN. We can't validate in this mode, but
3981 * modifications are not allowed anyways so just return true.
3983 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3987 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3988 * handled by recovery and thus safe to ignore here.
3990 if (lsn == NULLCOMMITLSN)
3993 valid = xlog_valid_lsn(mp->m_log, lsn);
3995 /* warn the user about what's gone wrong before verifier failure */
3997 spin_lock(&log->l_icloglock);
3999 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4000 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4001 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4002 log->l_curr_cycle, log->l_curr_block);
4003 spin_unlock(&log->l_icloglock);
4010 xfs_log_in_recovery(
4011 struct xfs_mount *mp)
4013 struct xlog *log = mp->m_log;
4015 return log->l_flags & XLOG_ACTIVE_RECOVERY;