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)
776 xlog_recover_cancel(mp->m_log);
781 * Final log writes as part of unmount.
783 * Mark the filesystem clean as unmount happens. Note that during relocation
784 * this routine needs to be executed as part of source-bag while the
785 * deallocation must not be done until source-end.
788 /* Actually write the unmount record to disk. */
790 xfs_log_write_unmount_record(
791 struct xfs_mount *mp)
793 /* the data section must be 32 bit size aligned */
794 struct xfs_unmount_log_format magic = {
795 .magic = XLOG_UNMOUNT_TYPE,
797 struct xfs_log_iovec reg = {
799 .i_len = sizeof(magic),
800 .i_type = XLOG_REG_TYPE_UNMOUNT,
802 struct xfs_log_vec vec = {
806 struct xlog *log = mp->m_log;
807 struct xlog_in_core *iclog;
808 struct xlog_ticket *tic = NULL;
810 uint flags = XLOG_UNMOUNT_TRANS;
813 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
818 * If we think the summary counters are bad, clear the unmount header
819 * flag in the unmount record so that the summary counters will be
820 * recalculated during log recovery at next mount. Refer to
821 * xlog_check_unmount_rec for more details.
823 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
824 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
825 xfs_alert(mp, "%s: will fix summary counters at next mount",
827 flags &= ~XLOG_UNMOUNT_TRANS;
830 /* remove inited flag, and account for space used */
832 tic->t_curr_res -= sizeof(magic);
833 error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
835 * At this point, we're umounting anyway, so there's no point in
836 * transitioning log state to IOERROR. Just continue...
840 xfs_alert(mp, "%s: unmount record failed", __func__);
842 spin_lock(&log->l_icloglock);
843 iclog = log->l_iclog;
844 atomic_inc(&iclog->ic_refcnt);
845 xlog_state_want_sync(log, iclog);
846 spin_unlock(&log->l_icloglock);
847 error = xlog_state_release_iclog(log, iclog);
849 spin_lock(&log->l_icloglock);
850 switch (iclog->ic_state) {
852 if (!XLOG_FORCED_SHUTDOWN(log)) {
853 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
857 case XLOG_STATE_ACTIVE:
858 case XLOG_STATE_DIRTY:
859 spin_unlock(&log->l_icloglock);
864 trace_xfs_log_umount_write(log, tic);
865 xlog_ungrant_log_space(log, tic);
866 xfs_log_ticket_put(tic);
871 * Unmount record used to have a string "Unmount filesystem--" in the
872 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
873 * We just write the magic number now since that particular field isn't
874 * currently architecture converted and "Unmount" is a bit foo.
875 * As far as I know, there weren't any dependencies on the old behaviour.
879 xfs_log_unmount_write(xfs_mount_t *mp)
881 struct xlog *log = mp->m_log;
882 xlog_in_core_t *iclog;
884 xlog_in_core_t *first_iclog;
889 * Don't write out unmount record on norecovery mounts or ro devices.
890 * Or, if we are doing a forced umount (typically because of IO errors).
892 if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
893 xfs_readonly_buftarg(log->l_targ)) {
894 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
898 error = xfs_log_force(mp, XFS_LOG_SYNC);
899 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
902 first_iclog = iclog = log->l_iclog;
904 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
905 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
906 ASSERT(iclog->ic_offset == 0);
908 iclog = iclog->ic_next;
909 } while (iclog != first_iclog);
911 if (! (XLOG_FORCED_SHUTDOWN(log))) {
912 xfs_log_write_unmount_record(mp);
915 * We're already in forced_shutdown mode, couldn't
916 * even attempt to write out the unmount transaction.
918 * Go through the motions of sync'ing and releasing
919 * the iclog, even though no I/O will actually happen,
920 * we need to wait for other log I/Os that may already
921 * be in progress. Do this as a separate section of
922 * code so we'll know if we ever get stuck here that
923 * we're in this odd situation of trying to unmount
924 * a file system that went into forced_shutdown as
925 * the result of an unmount..
927 spin_lock(&log->l_icloglock);
928 iclog = log->l_iclog;
929 atomic_inc(&iclog->ic_refcnt);
931 xlog_state_want_sync(log, iclog);
932 spin_unlock(&log->l_icloglock);
933 error = xlog_state_release_iclog(log, iclog);
935 spin_lock(&log->l_icloglock);
937 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
938 || iclog->ic_state == XLOG_STATE_DIRTY
939 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
941 xlog_wait(&iclog->ic_force_wait,
944 spin_unlock(&log->l_icloglock);
949 } /* xfs_log_unmount_write */
952 * Empty the log for unmount/freeze.
954 * To do this, we first need to shut down the background log work so it is not
955 * trying to cover the log as we clean up. We then need to unpin all objects in
956 * the log so we can then flush them out. Once they have completed their IO and
957 * run the callbacks removing themselves from the AIL, we can write the unmount
962 struct xfs_mount *mp)
964 cancel_delayed_work_sync(&mp->m_log->l_work);
965 xfs_log_force(mp, XFS_LOG_SYNC);
968 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
969 * will push it, xfs_wait_buftarg() will not wait for it. Further,
970 * xfs_buf_iowait() cannot be used because it was pushed with the
971 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
972 * the IO to complete.
974 xfs_ail_push_all_sync(mp->m_ail);
975 xfs_wait_buftarg(mp->m_ddev_targp);
976 xfs_buf_lock(mp->m_sb_bp);
977 xfs_buf_unlock(mp->m_sb_bp);
979 xfs_log_unmount_write(mp);
983 * Shut down and release the AIL and Log.
985 * During unmount, we need to ensure we flush all the dirty metadata objects
986 * from the AIL so that the log is empty before we write the unmount record to
987 * the log. Once this is done, we can tear down the AIL and the log.
991 struct xfs_mount *mp)
995 xfs_trans_ail_destroy(mp);
997 xfs_sysfs_del(&mp->m_log->l_kobj);
999 xlog_dealloc_log(mp->m_log);
1004 struct xfs_mount *mp,
1005 struct xfs_log_item *item,
1007 const struct xfs_item_ops *ops)
1009 item->li_mountp = mp;
1010 item->li_ailp = mp->m_ail;
1011 item->li_type = type;
1015 INIT_LIST_HEAD(&item->li_ail);
1016 INIT_LIST_HEAD(&item->li_cil);
1017 INIT_LIST_HEAD(&item->li_bio_list);
1018 INIT_LIST_HEAD(&item->li_trans);
1022 * Wake up processes waiting for log space after we have moved the log tail.
1026 struct xfs_mount *mp)
1028 struct xlog *log = mp->m_log;
1031 if (XLOG_FORCED_SHUTDOWN(log))
1034 if (!list_empty_careful(&log->l_write_head.waiters)) {
1035 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1037 spin_lock(&log->l_write_head.lock);
1038 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1039 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1040 spin_unlock(&log->l_write_head.lock);
1043 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1044 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1046 spin_lock(&log->l_reserve_head.lock);
1047 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1048 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1049 spin_unlock(&log->l_reserve_head.lock);
1054 * Determine if we have a transaction that has gone to disk that needs to be
1055 * covered. To begin the transition to the idle state firstly the log needs to
1056 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1057 * we start attempting to cover the log.
1059 * Only if we are then in a state where covering is needed, the caller is
1060 * informed that dummy transactions are required to move the log into the idle
1063 * If there are any items in the AIl or CIL, then we do not want to attempt to
1064 * cover the log as we may be in a situation where there isn't log space
1065 * available to run a dummy transaction and this can lead to deadlocks when the
1066 * tail of the log is pinned by an item that is modified in the CIL. Hence
1067 * there's no point in running a dummy transaction at this point because we
1068 * can't start trying to idle the log until both the CIL and AIL are empty.
1071 xfs_log_need_covered(xfs_mount_t *mp)
1073 struct xlog *log = mp->m_log;
1076 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1079 if (!xlog_cil_empty(log))
1082 spin_lock(&log->l_icloglock);
1083 switch (log->l_covered_state) {
1084 case XLOG_STATE_COVER_DONE:
1085 case XLOG_STATE_COVER_DONE2:
1086 case XLOG_STATE_COVER_IDLE:
1088 case XLOG_STATE_COVER_NEED:
1089 case XLOG_STATE_COVER_NEED2:
1090 if (xfs_ail_min_lsn(log->l_ailp))
1092 if (!xlog_iclogs_empty(log))
1096 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1097 log->l_covered_state = XLOG_STATE_COVER_DONE;
1099 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1105 spin_unlock(&log->l_icloglock);
1110 * We may be holding the log iclog lock upon entering this routine.
1113 xlog_assign_tail_lsn_locked(
1114 struct xfs_mount *mp)
1116 struct xlog *log = mp->m_log;
1117 struct xfs_log_item *lip;
1120 assert_spin_locked(&mp->m_ail->ail_lock);
1123 * To make sure we always have a valid LSN for the log tail we keep
1124 * track of the last LSN which was committed in log->l_last_sync_lsn,
1125 * and use that when the AIL was empty.
1127 lip = xfs_ail_min(mp->m_ail);
1129 tail_lsn = lip->li_lsn;
1131 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1132 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1133 atomic64_set(&log->l_tail_lsn, tail_lsn);
1138 xlog_assign_tail_lsn(
1139 struct xfs_mount *mp)
1143 spin_lock(&mp->m_ail->ail_lock);
1144 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1145 spin_unlock(&mp->m_ail->ail_lock);
1151 * Return the space in the log between the tail and the head. The head
1152 * is passed in the cycle/bytes formal parms. In the special case where
1153 * the reserve head has wrapped passed the tail, this calculation is no
1154 * longer valid. In this case, just return 0 which means there is no space
1155 * in the log. This works for all places where this function is called
1156 * with the reserve head. Of course, if the write head were to ever
1157 * wrap the tail, we should blow up. Rather than catch this case here,
1158 * we depend on other ASSERTions in other parts of the code. XXXmiken
1160 * This code also handles the case where the reservation head is behind
1161 * the tail. The details of this case are described below, but the end
1162 * result is that we return the size of the log as the amount of space left.
1175 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1176 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1177 tail_bytes = BBTOB(tail_bytes);
1178 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1179 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1180 else if (tail_cycle + 1 < head_cycle)
1182 else if (tail_cycle < head_cycle) {
1183 ASSERT(tail_cycle == (head_cycle - 1));
1184 free_bytes = tail_bytes - head_bytes;
1187 * The reservation head is behind the tail.
1188 * In this case we just want to return the size of the
1189 * log as the amount of space left.
1191 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1192 xfs_alert(log->l_mp,
1193 " tail_cycle = %d, tail_bytes = %d",
1194 tail_cycle, tail_bytes);
1195 xfs_alert(log->l_mp,
1196 " GH cycle = %d, GH bytes = %d",
1197 head_cycle, head_bytes);
1199 free_bytes = log->l_logsize;
1207 struct work_struct *work)
1209 struct xlog_in_core *iclog =
1210 container_of(work, struct xlog_in_core, ic_end_io_work);
1211 struct xlog *log = iclog->ic_log;
1212 bool aborted = false;
1215 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1217 /* treat writes with injected CRC errors as failed */
1218 if (iclog->ic_fail_crc)
1223 * Race to shutdown the filesystem if we see an error.
1225 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1226 xfs_alert(log->l_mp, "log I/O error %d", error);
1227 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1229 * This flag will be propagated to the trans-committed
1230 * callback routines to let them know that the log-commit
1234 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1238 xlog_state_done_syncing(iclog, aborted);
1239 bio_uninit(&iclog->ic_bio);
1242 * Drop the lock to signal that we are done. Nothing references the
1243 * iclog after this, so an unmount waiting on this lock can now tear it
1244 * down safely. As such, it is unsafe to reference the iclog after the
1245 * unlock as we could race with it being freed.
1247 up(&iclog->ic_sema);
1251 * Return size of each in-core log record buffer.
1253 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1255 * If the filesystem blocksize is too large, we may need to choose a
1256 * larger size since the directory code currently logs entire blocks.
1259 xlog_get_iclog_buffer_size(
1260 struct xfs_mount *mp,
1263 if (mp->m_logbufs <= 0)
1264 mp->m_logbufs = XLOG_MAX_ICLOGS;
1265 if (mp->m_logbsize <= 0)
1266 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1268 log->l_iclog_bufs = mp->m_logbufs;
1269 log->l_iclog_size = mp->m_logbsize;
1272 * # headers = size / 32k - one header holds cycles from 32k of data.
1274 log->l_iclog_heads =
1275 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1276 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1281 struct xfs_mount *mp)
1283 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1284 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1288 * Every sync period we need to unpin all items in the AIL and push them to
1289 * disk. If there is nothing dirty, then we might need to cover the log to
1290 * indicate that the filesystem is idle.
1294 struct work_struct *work)
1296 struct xlog *log = container_of(to_delayed_work(work),
1297 struct xlog, l_work);
1298 struct xfs_mount *mp = log->l_mp;
1300 /* dgc: errors ignored - not fatal and nowhere to report them */
1301 if (xfs_log_need_covered(mp)) {
1303 * Dump a transaction into the log that contains no real change.
1304 * This is needed to stamp the current tail LSN into the log
1305 * during the covering operation.
1307 * We cannot use an inode here for this - that will push dirty
1308 * state back up into the VFS and then periodic inode flushing
1309 * will prevent log covering from making progress. Hence we
1310 * synchronously log the superblock instead to ensure the
1311 * superblock is immediately unpinned and can be written back.
1313 xfs_sync_sb(mp, true);
1315 xfs_log_force(mp, 0);
1317 /* start pushing all the metadata that is currently dirty */
1318 xfs_ail_push_all(mp->m_ail);
1320 /* queue us up again */
1321 xfs_log_work_queue(mp);
1325 * This routine initializes some of the log structure for a given mount point.
1326 * Its primary purpose is to fill in enough, so recovery can occur. However,
1327 * some other stuff may be filled in too.
1329 STATIC struct xlog *
1331 struct xfs_mount *mp,
1332 struct xfs_buftarg *log_target,
1333 xfs_daddr_t blk_offset,
1337 xlog_rec_header_t *head;
1338 xlog_in_core_t **iclogp;
1339 xlog_in_core_t *iclog, *prev_iclog=NULL;
1341 int error = -ENOMEM;
1344 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1346 xfs_warn(mp, "Log allocation failed: No memory!");
1351 log->l_targ = log_target;
1352 log->l_logsize = BBTOB(num_bblks);
1353 log->l_logBBstart = blk_offset;
1354 log->l_logBBsize = num_bblks;
1355 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1356 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1357 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1359 log->l_prev_block = -1;
1360 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1361 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1362 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1363 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1365 xlog_grant_head_init(&log->l_reserve_head);
1366 xlog_grant_head_init(&log->l_write_head);
1368 error = -EFSCORRUPTED;
1369 if (xfs_sb_version_hassector(&mp->m_sb)) {
1370 log2_size = mp->m_sb.sb_logsectlog;
1371 if (log2_size < BBSHIFT) {
1372 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1373 log2_size, BBSHIFT);
1377 log2_size -= BBSHIFT;
1378 if (log2_size > mp->m_sectbb_log) {
1379 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1380 log2_size, mp->m_sectbb_log);
1384 /* for larger sector sizes, must have v2 or external log */
1385 if (log2_size && log->l_logBBstart > 0 &&
1386 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1388 "log sector size (0x%x) invalid for configuration.",
1393 log->l_sectBBsize = 1 << log2_size;
1395 xlog_get_iclog_buffer_size(mp, log);
1397 spin_lock_init(&log->l_icloglock);
1398 init_waitqueue_head(&log->l_flush_wait);
1400 iclogp = &log->l_iclog;
1402 * The amount of memory to allocate for the iclog structure is
1403 * rather funky due to the way the structure is defined. It is
1404 * done this way so that we can use different sizes for machines
1405 * with different amounts of memory. See the definition of
1406 * xlog_in_core_t in xfs_log_priv.h for details.
1408 ASSERT(log->l_iclog_size >= 4096);
1409 for (i = 0; i < log->l_iclog_bufs; i++) {
1410 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1411 sizeof(struct bio_vec);
1413 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1415 goto out_free_iclog;
1418 iclog->ic_prev = prev_iclog;
1421 iclog->ic_data = kmem_alloc_large(log->l_iclog_size,
1423 if (!iclog->ic_data)
1424 goto out_free_iclog;
1426 log->l_iclog_bak[i] = &iclog->ic_header;
1428 head = &iclog->ic_header;
1429 memset(head, 0, sizeof(xlog_rec_header_t));
1430 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1431 head->h_version = cpu_to_be32(
1432 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1433 head->h_size = cpu_to_be32(log->l_iclog_size);
1435 head->h_fmt = cpu_to_be32(XLOG_FMT);
1436 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1438 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1439 iclog->ic_state = XLOG_STATE_ACTIVE;
1440 iclog->ic_log = log;
1441 atomic_set(&iclog->ic_refcnt, 0);
1442 spin_lock_init(&iclog->ic_callback_lock);
1443 INIT_LIST_HEAD(&iclog->ic_callbacks);
1444 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1446 init_waitqueue_head(&iclog->ic_force_wait);
1447 init_waitqueue_head(&iclog->ic_write_wait);
1448 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1449 sema_init(&iclog->ic_sema, 1);
1451 iclogp = &iclog->ic_next;
1453 *iclogp = log->l_iclog; /* complete ring */
1454 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1456 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1457 WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0,
1459 if (!log->l_ioend_workqueue)
1460 goto out_free_iclog;
1462 error = xlog_cil_init(log);
1464 goto out_destroy_workqueue;
1467 out_destroy_workqueue:
1468 destroy_workqueue(log->l_ioend_workqueue);
1470 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1471 prev_iclog = iclog->ic_next;
1472 kmem_free(iclog->ic_data);
1478 return ERR_PTR(error);
1479 } /* xlog_alloc_log */
1483 * Write out the commit record of a transaction associated with the given
1484 * ticket. Return the lsn of the commit record.
1489 struct xlog_ticket *ticket,
1490 struct xlog_in_core **iclog,
1491 xfs_lsn_t *commitlsnp)
1493 struct xfs_mount *mp = log->l_mp;
1495 struct xfs_log_iovec reg = {
1498 .i_type = XLOG_REG_TYPE_COMMIT,
1500 struct xfs_log_vec vec = {
1505 ASSERT_ALWAYS(iclog);
1506 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1509 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1514 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1515 * log space. This code pushes on the lsn which would supposedly free up
1516 * the 25% which we want to leave free. We may need to adopt a policy which
1517 * pushes on an lsn which is further along in the log once we reach the high
1518 * water mark. In this manner, we would be creating a low water mark.
1521 xlog_grant_push_ail(
1525 xfs_lsn_t threshold_lsn = 0;
1526 xfs_lsn_t last_sync_lsn;
1529 int threshold_block;
1530 int threshold_cycle;
1533 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1535 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1536 free_blocks = BTOBBT(free_bytes);
1539 * Set the threshold for the minimum number of free blocks in the
1540 * log to the maximum of what the caller needs, one quarter of the
1541 * log, and 256 blocks.
1543 free_threshold = BTOBB(need_bytes);
1544 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1545 free_threshold = max(free_threshold, 256);
1546 if (free_blocks >= free_threshold)
1549 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1551 threshold_block += free_threshold;
1552 if (threshold_block >= log->l_logBBsize) {
1553 threshold_block -= log->l_logBBsize;
1554 threshold_cycle += 1;
1556 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1559 * Don't pass in an lsn greater than the lsn of the last
1560 * log record known to be on disk. Use a snapshot of the last sync lsn
1561 * so that it doesn't change between the compare and the set.
1563 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1564 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1565 threshold_lsn = last_sync_lsn;
1568 * Get the transaction layer to kick the dirty buffers out to
1569 * disk asynchronously. No point in trying to do this if
1570 * the filesystem is shutting down.
1572 if (!XLOG_FORCED_SHUTDOWN(log))
1573 xfs_ail_push(log->l_ailp, threshold_lsn);
1577 * Stamp cycle number in every block
1582 struct xlog_in_core *iclog,
1586 int size = iclog->ic_offset + roundoff;
1590 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1592 dp = iclog->ic_datap;
1593 for (i = 0; i < BTOBB(size); i++) {
1594 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1596 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1597 *(__be32 *)dp = cycle_lsn;
1601 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1602 xlog_in_core_2_t *xhdr = iclog->ic_data;
1604 for ( ; i < BTOBB(size); i++) {
1605 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1606 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1607 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1608 *(__be32 *)dp = cycle_lsn;
1612 for (i = 1; i < log->l_iclog_heads; i++)
1613 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1618 * Calculate the checksum for a log buffer.
1620 * This is a little more complicated than it should be because the various
1621 * headers and the actual data are non-contiguous.
1626 struct xlog_rec_header *rhead,
1632 /* first generate the crc for the record header ... */
1633 crc = xfs_start_cksum_update((char *)rhead,
1634 sizeof(struct xlog_rec_header),
1635 offsetof(struct xlog_rec_header, h_crc));
1637 /* ... then for additional cycle data for v2 logs ... */
1638 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1639 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1643 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1644 if (size % XLOG_HEADER_CYCLE_SIZE)
1647 for (i = 1; i < xheads; i++) {
1648 crc = crc32c(crc, &xhdr[i].hic_xheader,
1649 sizeof(struct xlog_rec_ext_header));
1653 /* ... and finally for the payload */
1654 crc = crc32c(crc, dp, size);
1656 return xfs_end_cksum(crc);
1663 struct xlog_in_core *iclog = bio->bi_private;
1665 queue_work(iclog->ic_log->l_ioend_workqueue,
1666 &iclog->ic_end_io_work);
1670 xlog_map_iclog_data(
1676 struct page *page = kmem_to_page(data);
1677 unsigned int off = offset_in_page(data);
1678 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1680 WARN_ON_ONCE(bio_add_page(bio, page, len, off) != len);
1690 struct xlog_in_core *iclog,
1695 ASSERT(bno < log->l_logBBsize);
1698 * We lock the iclogbufs here so that we can serialise against I/O
1699 * completion during unmount. We might be processing a shutdown
1700 * triggered during unmount, and that can occur asynchronously to the
1701 * unmount thread, and hence we need to ensure that completes before
1702 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1703 * across the log IO to archieve that.
1705 down(&iclog->ic_sema);
1706 if (unlikely(iclog->ic_state & XLOG_STATE_IOERROR)) {
1708 * It would seem logical to return EIO here, but we rely on
1709 * the log state machine to propagate I/O errors instead of
1710 * doing it here. We kick of the state machine and unlock
1711 * the buffer manually, the code needs to be kept in sync
1712 * with the I/O completion path.
1714 xlog_state_done_syncing(iclog, XFS_LI_ABORTED);
1715 up(&iclog->ic_sema);
1719 iclog->ic_io_size = count;
1721 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1722 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1723 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1724 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1725 iclog->ic_bio.bi_private = iclog;
1726 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_FUA;
1728 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1730 xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, iclog->ic_io_size);
1731 if (is_vmalloc_addr(iclog->ic_data))
1732 flush_kernel_vmap_range(iclog->ic_data, iclog->ic_io_size);
1735 * If this log buffer would straddle the end of the log we will have
1736 * to split it up into two bios, so that we can continue at the start.
1738 if (bno + BTOBB(count) > log->l_logBBsize) {
1741 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1742 GFP_NOIO, &fs_bio_set);
1743 bio_chain(split, &iclog->ic_bio);
1746 /* restart at logical offset zero for the remainder */
1747 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1750 submit_bio(&iclog->ic_bio);
1754 * We need to bump cycle number for the part of the iclog that is
1755 * written to the start of the log. Watch out for the header magic
1756 * number case, though.
1765 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1768 for (i = split_offset; i < count; i += BBSIZE) {
1769 uint32_t cycle = get_unaligned_be32(data + i);
1771 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1773 put_unaligned_be32(cycle, data + i);
1778 xlog_calc_iclog_size(
1780 struct xlog_in_core *iclog,
1783 uint32_t count_init, count;
1786 use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1787 log->l_mp->m_sb.sb_logsunit > 1;
1789 /* Add for LR header */
1790 count_init = log->l_iclog_hsize + iclog->ic_offset;
1792 /* Round out the log write size */
1794 /* we have a v2 stripe unit to use */
1795 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1797 count = BBTOB(BTOBB(count_init));
1800 ASSERT(count >= count_init);
1801 *roundoff = count - count_init;
1804 ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1806 ASSERT(*roundoff < BBTOB(1));
1811 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1812 * fashion. Previously, we should have moved the current iclog
1813 * ptr in the log to point to the next available iclog. This allows further
1814 * write to continue while this code syncs out an iclog ready to go.
1815 * Before an in-core log can be written out, the data section must be scanned
1816 * to save away the 1st word of each BBSIZE block into the header. We replace
1817 * it with the current cycle count. Each BBSIZE block is tagged with the
1818 * cycle count because there in an implicit assumption that drives will
1819 * guarantee that entire 512 byte blocks get written at once. In other words,
1820 * we can't have part of a 512 byte block written and part not written. By
1821 * tagging each block, we will know which blocks are valid when recovering
1822 * after an unclean shutdown.
1824 * This routine is single threaded on the iclog. No other thread can be in
1825 * this routine with the same iclog. Changing contents of iclog can there-
1826 * fore be done without grabbing the state machine lock. Updating the global
1827 * log will require grabbing the lock though.
1829 * The entire log manager uses a logical block numbering scheme. Only
1830 * xlog_write_iclog knows about the fact that the log may not start with
1831 * block zero on a given device.
1836 struct xlog_in_core *iclog)
1838 unsigned int count; /* byte count of bwrite */
1839 unsigned int roundoff; /* roundoff to BB or stripe */
1842 bool need_flush = true, split = false;
1844 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1846 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1848 /* move grant heads by roundoff in sync */
1849 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1850 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1852 /* put cycle number in every block */
1853 xlog_pack_data(log, iclog, roundoff);
1855 /* real byte length */
1856 size = iclog->ic_offset;
1857 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1859 iclog->ic_header.h_len = cpu_to_be32(size);
1861 XFS_STATS_INC(log->l_mp, xs_log_writes);
1862 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1864 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1866 /* Do we need to split this write into 2 parts? */
1867 if (bno + BTOBB(count) > log->l_logBBsize) {
1868 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1872 /* calculcate the checksum */
1873 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1874 iclog->ic_datap, size);
1876 * Intentionally corrupt the log record CRC based on the error injection
1877 * frequency, if defined. This facilitates testing log recovery in the
1878 * event of torn writes. Hence, set the IOABORT state to abort the log
1879 * write on I/O completion and shutdown the fs. The subsequent mount
1880 * detects the bad CRC and attempts to recover.
1883 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1884 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1885 iclog->ic_fail_crc = true;
1887 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1888 be64_to_cpu(iclog->ic_header.h_lsn));
1893 * Flush the data device before flushing the log to make sure all meta
1894 * data written back from the AIL actually made it to disk before
1895 * stamping the new log tail LSN into the log buffer. For an external
1896 * log we need to issue the flush explicitly, and unfortunately
1897 * synchronously here; for an internal log we can simply use the block
1898 * layer state machine for preflushes.
1900 if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1901 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1905 xlog_verify_iclog(log, iclog, count);
1906 xlog_write_iclog(log, iclog, bno, count, need_flush);
1910 * Deallocate a log structure
1916 xlog_in_core_t *iclog, *next_iclog;
1919 xlog_cil_destroy(log);
1922 * Cycle all the iclogbuf locks to make sure all log IO completion
1923 * is done before we tear down these buffers.
1925 iclog = log->l_iclog;
1926 for (i = 0; i < log->l_iclog_bufs; i++) {
1927 down(&iclog->ic_sema);
1928 up(&iclog->ic_sema);
1929 iclog = iclog->ic_next;
1932 iclog = log->l_iclog;
1933 for (i = 0; i < log->l_iclog_bufs; i++) {
1934 next_iclog = iclog->ic_next;
1935 kmem_free(iclog->ic_data);
1940 log->l_mp->m_log = NULL;
1941 destroy_workqueue(log->l_ioend_workqueue);
1943 } /* xlog_dealloc_log */
1946 * Update counters atomically now that memcpy is done.
1950 xlog_state_finish_copy(
1952 struct xlog_in_core *iclog,
1956 spin_lock(&log->l_icloglock);
1958 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1959 iclog->ic_offset += copy_bytes;
1961 spin_unlock(&log->l_icloglock);
1962 } /* xlog_state_finish_copy */
1968 * print out info relating to regions written which consume
1973 struct xfs_mount *mp,
1974 struct xlog_ticket *ticket)
1977 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1979 /* match with XLOG_REG_TYPE_* in xfs_log.h */
1980 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
1981 static char *res_type_str[] = {
1982 REG_TYPE_STR(BFORMAT, "bformat"),
1983 REG_TYPE_STR(BCHUNK, "bchunk"),
1984 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
1985 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
1986 REG_TYPE_STR(IFORMAT, "iformat"),
1987 REG_TYPE_STR(ICORE, "icore"),
1988 REG_TYPE_STR(IEXT, "iext"),
1989 REG_TYPE_STR(IBROOT, "ibroot"),
1990 REG_TYPE_STR(ILOCAL, "ilocal"),
1991 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
1992 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
1993 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
1994 REG_TYPE_STR(QFORMAT, "qformat"),
1995 REG_TYPE_STR(DQUOT, "dquot"),
1996 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
1997 REG_TYPE_STR(LRHEADER, "LR header"),
1998 REG_TYPE_STR(UNMOUNT, "unmount"),
1999 REG_TYPE_STR(COMMIT, "commit"),
2000 REG_TYPE_STR(TRANSHDR, "trans header"),
2001 REG_TYPE_STR(ICREATE, "inode create"),
2002 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2003 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2004 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2005 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2006 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2007 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2009 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2012 xfs_warn(mp, "ticket reservation summary:");
2013 xfs_warn(mp, " unit res = %d bytes",
2014 ticket->t_unit_res);
2015 xfs_warn(mp, " current res = %d bytes",
2016 ticket->t_curr_res);
2017 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2018 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2019 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2020 ticket->t_res_num_ophdrs, ophdr_spc);
2021 xfs_warn(mp, " ophdr + reg = %u bytes",
2022 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2023 xfs_warn(mp, " num regions = %u",
2026 for (i = 0; i < ticket->t_res_num; i++) {
2027 uint r_type = ticket->t_res_arr[i].r_type;
2028 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2029 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2030 "bad-rtype" : res_type_str[r_type]),
2031 ticket->t_res_arr[i].r_len);
2036 * Print a summary of the transaction.
2040 struct xfs_trans *tp)
2042 struct xfs_mount *mp = tp->t_mountp;
2043 struct xfs_log_item *lip;
2045 /* dump core transaction and ticket info */
2046 xfs_warn(mp, "transaction summary:");
2047 xfs_warn(mp, " log res = %d", tp->t_log_res);
2048 xfs_warn(mp, " log count = %d", tp->t_log_count);
2049 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2051 xlog_print_tic_res(mp, tp->t_ticket);
2053 /* dump each log item */
2054 list_for_each_entry(lip, &tp->t_items, li_trans) {
2055 struct xfs_log_vec *lv = lip->li_lv;
2056 struct xfs_log_iovec *vec;
2059 xfs_warn(mp, "log item: ");
2060 xfs_warn(mp, " type = 0x%x", lip->li_type);
2061 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2064 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2065 xfs_warn(mp, " size = %d", lv->lv_size);
2066 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2067 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2069 /* dump each iovec for the log item */
2070 vec = lv->lv_iovecp;
2071 for (i = 0; i < lv->lv_niovecs; i++) {
2072 int dumplen = min(vec->i_len, 32);
2074 xfs_warn(mp, " iovec[%d]", i);
2075 xfs_warn(mp, " type = 0x%x", vec->i_type);
2076 xfs_warn(mp, " len = %d", vec->i_len);
2077 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2078 xfs_hex_dump(vec->i_addr, dumplen);
2086 * Calculate the potential space needed by the log vector. Each region gets
2087 * its own xlog_op_header_t and may need to be double word aligned.
2090 xlog_write_calc_vec_length(
2091 struct xlog_ticket *ticket,
2092 struct xfs_log_vec *log_vector)
2094 struct xfs_log_vec *lv;
2099 /* acct for start rec of xact */
2100 if (ticket->t_flags & XLOG_TIC_INITED)
2103 for (lv = log_vector; lv; lv = lv->lv_next) {
2104 /* we don't write ordered log vectors */
2105 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2108 headers += lv->lv_niovecs;
2110 for (i = 0; i < lv->lv_niovecs; i++) {
2111 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2114 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2118 ticket->t_res_num_ophdrs += headers;
2119 len += headers * sizeof(struct xlog_op_header);
2125 * If first write for transaction, insert start record We can't be trying to
2126 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2129 xlog_write_start_rec(
2130 struct xlog_op_header *ophdr,
2131 struct xlog_ticket *ticket)
2133 if (!(ticket->t_flags & XLOG_TIC_INITED))
2136 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2137 ophdr->oh_clientid = ticket->t_clientid;
2139 ophdr->oh_flags = XLOG_START_TRANS;
2142 ticket->t_flags &= ~XLOG_TIC_INITED;
2144 return sizeof(struct xlog_op_header);
2147 static xlog_op_header_t *
2148 xlog_write_setup_ophdr(
2150 struct xlog_op_header *ophdr,
2151 struct xlog_ticket *ticket,
2154 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2155 ophdr->oh_clientid = ticket->t_clientid;
2158 /* are we copying a commit or unmount record? */
2159 ophdr->oh_flags = flags;
2162 * We've seen logs corrupted with bad transaction client ids. This
2163 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2164 * and shut down the filesystem.
2166 switch (ophdr->oh_clientid) {
2167 case XFS_TRANSACTION:
2173 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2174 ophdr->oh_clientid, ticket);
2182 * Set up the parameters of the region copy into the log. This has
2183 * to handle region write split across multiple log buffers - this
2184 * state is kept external to this function so that this code can
2185 * be written in an obvious, self documenting manner.
2188 xlog_write_setup_copy(
2189 struct xlog_ticket *ticket,
2190 struct xlog_op_header *ophdr,
2191 int space_available,
2195 int *last_was_partial_copy,
2196 int *bytes_consumed)
2200 still_to_copy = space_required - *bytes_consumed;
2201 *copy_off = *bytes_consumed;
2203 if (still_to_copy <= space_available) {
2204 /* write of region completes here */
2205 *copy_len = still_to_copy;
2206 ophdr->oh_len = cpu_to_be32(*copy_len);
2207 if (*last_was_partial_copy)
2208 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2209 *last_was_partial_copy = 0;
2210 *bytes_consumed = 0;
2214 /* partial write of region, needs extra log op header reservation */
2215 *copy_len = space_available;
2216 ophdr->oh_len = cpu_to_be32(*copy_len);
2217 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2218 if (*last_was_partial_copy)
2219 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2220 *bytes_consumed += *copy_len;
2221 (*last_was_partial_copy)++;
2223 /* account for new log op header */
2224 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2225 ticket->t_res_num_ophdrs++;
2227 return sizeof(struct xlog_op_header);
2231 xlog_write_copy_finish(
2233 struct xlog_in_core *iclog,
2238 int *partial_copy_len,
2240 struct xlog_in_core **commit_iclog)
2242 if (*partial_copy) {
2244 * This iclog has already been marked WANT_SYNC by
2245 * xlog_state_get_iclog_space.
2247 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2250 return xlog_state_release_iclog(log, iclog);
2254 *partial_copy_len = 0;
2256 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2257 /* no more space in this iclog - push it. */
2258 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2262 spin_lock(&log->l_icloglock);
2263 xlog_state_want_sync(log, iclog);
2264 spin_unlock(&log->l_icloglock);
2267 return xlog_state_release_iclog(log, iclog);
2268 ASSERT(flags & XLOG_COMMIT_TRANS);
2269 *commit_iclog = iclog;
2276 * Write some region out to in-core log
2278 * This will be called when writing externally provided regions or when
2279 * writing out a commit record for a given transaction.
2281 * General algorithm:
2282 * 1. Find total length of this write. This may include adding to the
2283 * lengths passed in.
2284 * 2. Check whether we violate the tickets reservation.
2285 * 3. While writing to this iclog
2286 * A. Reserve as much space in this iclog as can get
2287 * B. If this is first write, save away start lsn
2288 * C. While writing this region:
2289 * 1. If first write of transaction, write start record
2290 * 2. Write log operation header (header per region)
2291 * 3. Find out if we can fit entire region into this iclog
2292 * 4. Potentially, verify destination memcpy ptr
2293 * 5. Memcpy (partial) region
2294 * 6. If partial copy, release iclog; otherwise, continue
2295 * copying more regions into current iclog
2296 * 4. Mark want sync bit (in simulation mode)
2297 * 5. Release iclog for potential flush to on-disk log.
2300 * 1. Panic if reservation is overrun. This should never happen since
2301 * reservation amounts are generated internal to the filesystem.
2303 * 1. Tickets are single threaded data structures.
2304 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2305 * syncing routine. When a single log_write region needs to span
2306 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2307 * on all log operation writes which don't contain the end of the
2308 * region. The XLOG_END_TRANS bit is used for the in-core log
2309 * operation which contains the end of the continued log_write region.
2310 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2311 * we don't really know exactly how much space will be used. As a result,
2312 * we don't update ic_offset until the end when we know exactly how many
2313 * bytes have been written out.
2318 struct xfs_log_vec *log_vector,
2319 struct xlog_ticket *ticket,
2320 xfs_lsn_t *start_lsn,
2321 struct xlog_in_core **commit_iclog,
2324 struct xlog_in_core *iclog = NULL;
2325 struct xfs_log_iovec *vecp;
2326 struct xfs_log_vec *lv;
2329 int partial_copy = 0;
2330 int partial_copy_len = 0;
2338 len = xlog_write_calc_vec_length(ticket, log_vector);
2341 * Region headers and bytes are already accounted for.
2342 * We only need to take into account start records and
2343 * split regions in this function.
2345 if (ticket->t_flags & XLOG_TIC_INITED)
2346 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2349 * Commit record headers need to be accounted for. These
2350 * come in as separate writes so are easy to detect.
2352 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2353 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2355 if (ticket->t_curr_res < 0) {
2356 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2357 "ctx ticket reservation ran out. Need to up reservation");
2358 xlog_print_tic_res(log->l_mp, ticket);
2359 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2364 vecp = lv->lv_iovecp;
2365 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2369 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2370 &contwr, &log_offset);
2374 ASSERT(log_offset <= iclog->ic_size - 1);
2375 ptr = iclog->ic_datap + log_offset;
2377 /* start_lsn is the first lsn written to. That's all we need. */
2379 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2382 * This loop writes out as many regions as can fit in the amount
2383 * of space which was allocated by xlog_state_get_iclog_space().
2385 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2386 struct xfs_log_iovec *reg;
2387 struct xlog_op_header *ophdr;
2391 bool ordered = false;
2393 /* ordered log vectors have no regions to write */
2394 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2395 ASSERT(lv->lv_niovecs == 0);
2401 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2402 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2404 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2405 if (start_rec_copy) {
2407 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2411 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2415 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2416 sizeof(struct xlog_op_header));
2418 len += xlog_write_setup_copy(ticket, ophdr,
2419 iclog->ic_size-log_offset,
2421 ©_off, ©_len,
2424 xlog_verify_dest_ptr(log, ptr);
2429 * Unmount records just log an opheader, so can have
2430 * empty payloads with no data region to copy. Hence we
2431 * only copy the payload if the vector says it has data
2434 ASSERT(copy_len >= 0);
2436 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2437 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2440 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2442 data_cnt += contwr ? copy_len : 0;
2444 error = xlog_write_copy_finish(log, iclog, flags,
2445 &record_cnt, &data_cnt,
2454 * if we had a partial copy, we need to get more iclog
2455 * space but we don't want to increment the region
2456 * index because there is still more is this region to
2459 * If we completed writing this region, and we flushed
2460 * the iclog (indicated by resetting of the record
2461 * count), then we also need to get more log space. If
2462 * this was the last record, though, we are done and
2468 if (++index == lv->lv_niovecs) {
2473 vecp = lv->lv_iovecp;
2475 if (record_cnt == 0 && !ordered) {
2485 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2487 return xlog_state_release_iclog(log, iclog);
2489 ASSERT(flags & XLOG_COMMIT_TRANS);
2490 *commit_iclog = iclog;
2495 /*****************************************************************************
2497 * State Machine functions
2499 *****************************************************************************
2502 /* Clean iclogs starting from the head. This ordering must be
2503 * maintained, so an iclog doesn't become ACTIVE beyond one that
2504 * is SYNCING. This is also required to maintain the notion that we use
2505 * a ordered wait queue to hold off would be writers to the log when every
2506 * iclog is trying to sync to disk.
2508 * State Change: DIRTY -> ACTIVE
2511 xlog_state_clean_log(
2514 xlog_in_core_t *iclog;
2517 iclog = log->l_iclog;
2519 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2520 iclog->ic_state = XLOG_STATE_ACTIVE;
2521 iclog->ic_offset = 0;
2522 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2524 * If the number of ops in this iclog indicate it just
2525 * contains the dummy transaction, we can
2526 * change state into IDLE (the second time around).
2527 * Otherwise we should change the state into
2529 * We don't need to cover the dummy.
2532 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2537 * We have two dirty iclogs so start over
2538 * This could also be num of ops indicates
2539 * this is not the dummy going out.
2543 iclog->ic_header.h_num_logops = 0;
2544 memset(iclog->ic_header.h_cycle_data, 0,
2545 sizeof(iclog->ic_header.h_cycle_data));
2546 iclog->ic_header.h_lsn = 0;
2547 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2550 break; /* stop cleaning */
2551 iclog = iclog->ic_next;
2552 } while (iclog != log->l_iclog);
2554 /* log is locked when we are called */
2556 * Change state for the dummy log recording.
2557 * We usually go to NEED. But we go to NEED2 if the changed indicates
2558 * we are done writing the dummy record.
2559 * If we are done with the second dummy recored (DONE2), then
2563 switch (log->l_covered_state) {
2564 case XLOG_STATE_COVER_IDLE:
2565 case XLOG_STATE_COVER_NEED:
2566 case XLOG_STATE_COVER_NEED2:
2567 log->l_covered_state = XLOG_STATE_COVER_NEED;
2570 case XLOG_STATE_COVER_DONE:
2572 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2574 log->l_covered_state = XLOG_STATE_COVER_NEED;
2577 case XLOG_STATE_COVER_DONE2:
2579 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2581 log->l_covered_state = XLOG_STATE_COVER_NEED;
2588 } /* xlog_state_clean_log */
2591 xlog_get_lowest_lsn(
2594 struct xlog_in_core *iclog = log->l_iclog;
2595 xfs_lsn_t lowest_lsn = 0, lsn;
2598 if (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))
2601 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2602 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2604 } while ((iclog = iclog->ic_next) != log->l_iclog);
2610 xlog_state_do_callback(
2613 struct xlog_in_core *ciclog)
2615 xlog_in_core_t *iclog;
2616 xlog_in_core_t *first_iclog; /* used to know when we've
2617 * processed all iclogs once */
2619 xfs_lsn_t lowest_lsn;
2620 int ioerrors; /* counter: iclogs with errors */
2621 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2622 int funcdidcallbacks; /* flag: function did callbacks */
2623 int repeats; /* for issuing console warnings if
2624 * looping too many times */
2627 spin_lock(&log->l_icloglock);
2628 first_iclog = iclog = log->l_iclog;
2630 funcdidcallbacks = 0;
2635 * Scan all iclogs starting with the one pointed to by the
2636 * log. Reset this starting point each time the log is
2637 * unlocked (during callbacks).
2639 * Keep looping through iclogs until one full pass is made
2640 * without running any callbacks.
2642 first_iclog = log->l_iclog;
2643 iclog = log->l_iclog;
2644 loopdidcallbacks = 0;
2649 /* skip all iclogs in the ACTIVE & DIRTY states */
2650 if (iclog->ic_state &
2651 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2652 iclog = iclog->ic_next;
2657 * Between marking a filesystem SHUTDOWN and stopping
2658 * the log, we do flush all iclogs to disk (if there
2659 * wasn't a log I/O error). So, we do want things to
2660 * go smoothly in case of just a SHUTDOWN w/o a
2663 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2665 * Can only perform callbacks in order. Since
2666 * this iclog is not in the DONE_SYNC/
2667 * DO_CALLBACK state, we skip the rest and
2668 * just try to clean up. If we set our iclog
2669 * to DO_CALLBACK, we will not process it when
2670 * we retry since a previous iclog is in the
2671 * CALLBACK and the state cannot change since
2672 * we are holding the l_icloglock.
2674 if (!(iclog->ic_state &
2675 (XLOG_STATE_DONE_SYNC |
2676 XLOG_STATE_DO_CALLBACK))) {
2677 if (ciclog && (ciclog->ic_state ==
2678 XLOG_STATE_DONE_SYNC)) {
2679 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2684 * We now have an iclog that is in either the
2685 * DO_CALLBACK or DONE_SYNC states. The other
2686 * states (WANT_SYNC, SYNCING, or CALLBACK were
2687 * caught by the above if and are going to
2688 * clean (i.e. we aren't doing their callbacks)
2693 * We will do one more check here to see if we
2694 * have chased our tail around.
2697 lowest_lsn = xlog_get_lowest_lsn(log);
2699 XFS_LSN_CMP(lowest_lsn,
2700 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2701 iclog = iclog->ic_next;
2702 continue; /* Leave this iclog for
2706 iclog->ic_state = XLOG_STATE_CALLBACK;
2710 * Completion of a iclog IO does not imply that
2711 * a transaction has completed, as transactions
2712 * can be large enough to span many iclogs. We
2713 * cannot change the tail of the log half way
2714 * through a transaction as this may be the only
2715 * transaction in the log and moving th etail to
2716 * point to the middle of it will prevent
2717 * recovery from finding the start of the
2718 * transaction. Hence we should only update the
2719 * last_sync_lsn if this iclog contains
2720 * transaction completion callbacks on it.
2722 * We have to do this before we drop the
2723 * icloglock to ensure we are the only one that
2726 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2727 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2728 if (!list_empty_careful(&iclog->ic_callbacks))
2729 atomic64_set(&log->l_last_sync_lsn,
2730 be64_to_cpu(iclog->ic_header.h_lsn));
2735 spin_unlock(&log->l_icloglock);
2738 * Keep processing entries in the callback list until
2739 * we come around and it is empty. We need to
2740 * atomically see that the list is empty and change the
2741 * state to DIRTY so that we don't miss any more
2742 * callbacks being added.
2744 spin_lock(&iclog->ic_callback_lock);
2745 while (!list_empty(&iclog->ic_callbacks)) {
2748 list_splice_init(&iclog->ic_callbacks, &tmp);
2750 spin_unlock(&iclog->ic_callback_lock);
2751 xlog_cil_process_committed(&tmp, aborted);
2752 spin_lock(&iclog->ic_callback_lock);
2758 spin_lock(&log->l_icloglock);
2759 spin_unlock(&iclog->ic_callback_lock);
2760 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2761 iclog->ic_state = XLOG_STATE_DIRTY;
2764 * Transition from DIRTY to ACTIVE if applicable.
2765 * NOP if STATE_IOERROR.
2767 xlog_state_clean_log(log);
2769 /* wake up threads waiting in xfs_log_force() */
2770 wake_up_all(&iclog->ic_force_wait);
2772 iclog = iclog->ic_next;
2773 } while (first_iclog != iclog);
2775 if (repeats > 5000) {
2776 flushcnt += repeats;
2779 "%s: possible infinite loop (%d iterations)",
2780 __func__, flushcnt);
2782 } while (!ioerrors && loopdidcallbacks);
2786 * Make one last gasp attempt to see if iclogs are being left in limbo.
2787 * If the above loop finds an iclog earlier than the current iclog and
2788 * in one of the syncing states, the current iclog is put into
2789 * DO_CALLBACK and the callbacks are deferred to the completion of the
2790 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2791 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2794 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2795 * for ic_state == SYNCING.
2797 if (funcdidcallbacks) {
2798 first_iclog = iclog = log->l_iclog;
2800 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2802 * Terminate the loop if iclogs are found in states
2803 * which will cause other threads to clean up iclogs.
2805 * SYNCING - i/o completion will go through logs
2806 * DONE_SYNC - interrupt thread should be waiting for
2808 * IOERROR - give up hope all ye who enter here
2810 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2811 iclog->ic_state & XLOG_STATE_SYNCING ||
2812 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2813 iclog->ic_state == XLOG_STATE_IOERROR )
2815 iclog = iclog->ic_next;
2816 } while (first_iclog != iclog);
2820 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2822 spin_unlock(&log->l_icloglock);
2825 wake_up_all(&log->l_flush_wait);
2830 * Finish transitioning this iclog to the dirty state.
2832 * Make sure that we completely execute this routine only when this is
2833 * the last call to the iclog. There is a good chance that iclog flushes,
2834 * when we reach the end of the physical log, get turned into 2 separate
2835 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2836 * routine. By using the reference count bwritecnt, we guarantee that only
2837 * the second completion goes through.
2839 * Callbacks could take time, so they are done outside the scope of the
2840 * global state machine log lock.
2843 xlog_state_done_syncing(
2844 struct xlog_in_core *iclog,
2847 struct xlog *log = iclog->ic_log;
2849 spin_lock(&log->l_icloglock);
2851 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2852 iclog->ic_state == XLOG_STATE_IOERROR);
2853 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2856 * If we got an error, either on the first buffer, or in the case of
2857 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2858 * and none should ever be attempted to be written to disk
2861 if (iclog->ic_state != XLOG_STATE_IOERROR)
2862 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2865 * Someone could be sleeping prior to writing out the next
2866 * iclog buffer, we wake them all, one will get to do the
2867 * I/O, the others get to wait for the result.
2869 wake_up_all(&iclog->ic_write_wait);
2870 spin_unlock(&log->l_icloglock);
2871 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2872 } /* xlog_state_done_syncing */
2876 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2877 * sleep. We wait on the flush queue on the head iclog as that should be
2878 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2879 * we will wait here and all new writes will sleep until a sync completes.
2881 * The in-core logs are used in a circular fashion. They are not used
2882 * out-of-order even when an iclog past the head is free.
2885 * * log_offset where xlog_write() can start writing into the in-core
2887 * * in-core log pointer to which xlog_write() should write.
2888 * * boolean indicating this is a continued write to an in-core log.
2889 * If this is the last write, then the in-core log's offset field
2890 * needs to be incremented, depending on the amount of data which
2894 xlog_state_get_iclog_space(
2897 struct xlog_in_core **iclogp,
2898 struct xlog_ticket *ticket,
2899 int *continued_write,
2903 xlog_rec_header_t *head;
2904 xlog_in_core_t *iclog;
2908 spin_lock(&log->l_icloglock);
2909 if (XLOG_FORCED_SHUTDOWN(log)) {
2910 spin_unlock(&log->l_icloglock);
2914 iclog = log->l_iclog;
2915 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2916 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2918 /* Wait for log writes to have flushed */
2919 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2923 head = &iclog->ic_header;
2925 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2926 log_offset = iclog->ic_offset;
2928 /* On the 1st write to an iclog, figure out lsn. This works
2929 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2930 * committing to. If the offset is set, that's how many blocks
2933 if (log_offset == 0) {
2934 ticket->t_curr_res -= log->l_iclog_hsize;
2935 xlog_tic_add_region(ticket,
2937 XLOG_REG_TYPE_LRHEADER);
2938 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2939 head->h_lsn = cpu_to_be64(
2940 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2941 ASSERT(log->l_curr_block >= 0);
2944 /* If there is enough room to write everything, then do it. Otherwise,
2945 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2946 * bit is on, so this will get flushed out. Don't update ic_offset
2947 * until you know exactly how many bytes get copied. Therefore, wait
2948 * until later to update ic_offset.
2950 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2951 * can fit into remaining data section.
2953 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2954 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2957 * If I'm the only one writing to this iclog, sync it to disk.
2958 * We need to do an atomic compare and decrement here to avoid
2959 * racing with concurrent atomic_dec_and_lock() calls in
2960 * xlog_state_release_iclog() when there is more than one
2961 * reference to the iclog.
2963 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2964 /* we are the only one */
2965 spin_unlock(&log->l_icloglock);
2966 error = xlog_state_release_iclog(log, iclog);
2970 spin_unlock(&log->l_icloglock);
2975 /* Do we have enough room to write the full amount in the remainder
2976 * of this iclog? Or must we continue a write on the next iclog and
2977 * mark this iclog as completely taken? In the case where we switch
2978 * iclogs (to mark it taken), this particular iclog will release/sync
2979 * to disk in xlog_write().
2981 if (len <= iclog->ic_size - iclog->ic_offset) {
2982 *continued_write = 0;
2983 iclog->ic_offset += len;
2985 *continued_write = 1;
2986 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2990 ASSERT(iclog->ic_offset <= iclog->ic_size);
2991 spin_unlock(&log->l_icloglock);
2993 *logoffsetp = log_offset;
2995 } /* xlog_state_get_iclog_space */
2997 /* The first cnt-1 times through here we don't need to
2998 * move the grant write head because the permanent
2999 * reservation has reserved cnt times the unit amount.
3000 * Release part of current permanent unit reservation and
3001 * reset current reservation to be one units worth. Also
3002 * move grant reservation head forward.
3005 xlog_regrant_reserve_log_space(
3007 struct xlog_ticket *ticket)
3009 trace_xfs_log_regrant_reserve_enter(log, ticket);
3011 if (ticket->t_cnt > 0)
3014 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3015 ticket->t_curr_res);
3016 xlog_grant_sub_space(log, &log->l_write_head.grant,
3017 ticket->t_curr_res);
3018 ticket->t_curr_res = ticket->t_unit_res;
3019 xlog_tic_reset_res(ticket);
3021 trace_xfs_log_regrant_reserve_sub(log, ticket);
3023 /* just return if we still have some of the pre-reserved space */
3024 if (ticket->t_cnt > 0)
3027 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3028 ticket->t_unit_res);
3030 trace_xfs_log_regrant_reserve_exit(log, ticket);
3032 ticket->t_curr_res = ticket->t_unit_res;
3033 xlog_tic_reset_res(ticket);
3034 } /* xlog_regrant_reserve_log_space */
3038 * Give back the space left from a reservation.
3040 * All the information we need to make a correct determination of space left
3041 * is present. For non-permanent reservations, things are quite easy. The
3042 * count should have been decremented to zero. We only need to deal with the
3043 * space remaining in the current reservation part of the ticket. If the
3044 * ticket contains a permanent reservation, there may be left over space which
3045 * needs to be released. A count of N means that N-1 refills of the current
3046 * reservation can be done before we need to ask for more space. The first
3047 * one goes to fill up the first current reservation. Once we run out of
3048 * space, the count will stay at zero and the only space remaining will be
3049 * in the current reservation field.
3052 xlog_ungrant_log_space(
3054 struct xlog_ticket *ticket)
3058 if (ticket->t_cnt > 0)
3061 trace_xfs_log_ungrant_enter(log, ticket);
3062 trace_xfs_log_ungrant_sub(log, ticket);
3065 * If this is a permanent reservation ticket, we may be able to free
3066 * up more space based on the remaining count.
3068 bytes = ticket->t_curr_res;
3069 if (ticket->t_cnt > 0) {
3070 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3071 bytes += ticket->t_unit_res*ticket->t_cnt;
3074 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3075 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3077 trace_xfs_log_ungrant_exit(log, ticket);
3079 xfs_log_space_wake(log->l_mp);
3083 * Flush iclog to disk if this is the last reference to the given iclog and
3084 * the WANT_SYNC bit is set.
3086 * When this function is entered, the iclog is not necessarily in the
3087 * WANT_SYNC state. It may be sitting around waiting to get filled.
3092 xlog_state_release_iclog(
3094 struct xlog_in_core *iclog)
3096 int sync = 0; /* do we sync? */
3098 if (iclog->ic_state & XLOG_STATE_IOERROR)
3101 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3102 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3105 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3106 spin_unlock(&log->l_icloglock);
3109 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3110 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3112 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3113 /* update tail before writing to iclog */
3114 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3116 iclog->ic_state = XLOG_STATE_SYNCING;
3117 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3118 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3119 /* cycle incremented when incrementing curr_block */
3121 spin_unlock(&log->l_icloglock);
3124 * We let the log lock go, so it's possible that we hit a log I/O
3125 * error or some other SHUTDOWN condition that marks the iclog
3126 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3127 * this iclog has consistent data, so we ignore IOERROR
3128 * flags after this point.
3131 xlog_sync(log, iclog);
3133 } /* xlog_state_release_iclog */
3137 * This routine will mark the current iclog in the ring as WANT_SYNC
3138 * and move the current iclog pointer to the next iclog in the ring.
3139 * When this routine is called from xlog_state_get_iclog_space(), the
3140 * exact size of the iclog has not yet been determined. All we know is
3141 * that every data block. We have run out of space in this log record.
3144 xlog_state_switch_iclogs(
3146 struct xlog_in_core *iclog,
3149 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3151 eventual_size = iclog->ic_offset;
3152 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3153 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3154 log->l_prev_block = log->l_curr_block;
3155 log->l_prev_cycle = log->l_curr_cycle;
3157 /* roll log?: ic_offset changed later */
3158 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3160 /* Round up to next log-sunit */
3161 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3162 log->l_mp->m_sb.sb_logsunit > 1) {
3163 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3164 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3167 if (log->l_curr_block >= log->l_logBBsize) {
3169 * Rewind the current block before the cycle is bumped to make
3170 * sure that the combined LSN never transiently moves forward
3171 * when the log wraps to the next cycle. This is to support the
3172 * unlocked sample of these fields from xlog_valid_lsn(). Most
3173 * other cases should acquire l_icloglock.
3175 log->l_curr_block -= log->l_logBBsize;
3176 ASSERT(log->l_curr_block >= 0);
3178 log->l_curr_cycle++;
3179 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3180 log->l_curr_cycle++;
3182 ASSERT(iclog == log->l_iclog);
3183 log->l_iclog = iclog->ic_next;
3184 } /* xlog_state_switch_iclogs */
3187 * Write out all data in the in-core log as of this exact moment in time.
3189 * Data may be written to the in-core log during this call. However,
3190 * we don't guarantee this data will be written out. A change from past
3191 * implementation means this routine will *not* write out zero length LRs.
3193 * Basically, we try and perform an intelligent scan of the in-core logs.
3194 * If we determine there is no flushable data, we just return. There is no
3195 * flushable data if:
3197 * 1. the current iclog is active and has no data; the previous iclog
3198 * is in the active or dirty state.
3199 * 2. the current iclog is drity, and the previous iclog is in the
3200 * active or dirty state.
3204 * 1. the current iclog is not in the active nor dirty state.
3205 * 2. the current iclog dirty, and the previous iclog is not in the
3206 * active nor dirty state.
3207 * 3. the current iclog is active, and there is another thread writing
3208 * to this particular iclog.
3209 * 4. a) the current iclog is active and has no other writers
3210 * b) when we return from flushing out this iclog, it is still
3211 * not in the active nor dirty state.
3215 struct xfs_mount *mp,
3218 struct xlog *log = mp->m_log;
3219 struct xlog_in_core *iclog;
3222 XFS_STATS_INC(mp, xs_log_force);
3223 trace_xfs_log_force(mp, 0, _RET_IP_);
3225 xlog_cil_force(log);
3227 spin_lock(&log->l_icloglock);
3228 iclog = log->l_iclog;
3229 if (iclog->ic_state & XLOG_STATE_IOERROR)
3232 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3233 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3234 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3236 * If the head is dirty or (active and empty), then we need to
3237 * look at the previous iclog.
3239 * If the previous iclog is active or dirty we are done. There
3240 * is nothing to sync out. Otherwise, we attach ourselves to the
3241 * previous iclog and go to sleep.
3243 iclog = iclog->ic_prev;
3244 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3245 iclog->ic_state == XLOG_STATE_DIRTY)
3247 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3248 if (atomic_read(&iclog->ic_refcnt) == 0) {
3250 * We are the only one with access to this iclog.
3252 * Flush it out now. There should be a roundoff of zero
3253 * to show that someone has already taken care of the
3254 * roundoff from the previous sync.
3256 atomic_inc(&iclog->ic_refcnt);
3257 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3258 xlog_state_switch_iclogs(log, iclog, 0);
3259 spin_unlock(&log->l_icloglock);
3261 if (xlog_state_release_iclog(log, iclog))
3264 spin_lock(&log->l_icloglock);
3265 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3266 iclog->ic_state == XLOG_STATE_DIRTY)
3270 * Someone else is writing to this iclog.
3272 * Use its call to flush out the data. However, the
3273 * other thread may not force out this LR, so we mark
3276 xlog_state_switch_iclogs(log, iclog, 0);
3280 * If the head iclog is not active nor dirty, we just attach
3281 * ourselves to the head and go to sleep if necessary.
3286 if (!(flags & XFS_LOG_SYNC))
3289 if (iclog->ic_state & XLOG_STATE_IOERROR)
3291 XFS_STATS_INC(mp, xs_log_force_sleep);
3292 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3293 if (iclog->ic_state & XLOG_STATE_IOERROR)
3298 spin_unlock(&log->l_icloglock);
3301 spin_unlock(&log->l_icloglock);
3306 __xfs_log_force_lsn(
3307 struct xfs_mount *mp,
3313 struct xlog *log = mp->m_log;
3314 struct xlog_in_core *iclog;
3316 spin_lock(&log->l_icloglock);
3317 iclog = log->l_iclog;
3318 if (iclog->ic_state & XLOG_STATE_IOERROR)
3321 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3322 iclog = iclog->ic_next;
3323 if (iclog == log->l_iclog)
3327 if (iclog->ic_state == XLOG_STATE_DIRTY)
3330 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3332 * We sleep here if we haven't already slept (e.g. this is the
3333 * first time we've looked at the correct iclog buf) and the
3334 * buffer before us is going to be sync'ed. The reason for this
3335 * is that if we are doing sync transactions here, by waiting
3336 * for the previous I/O to complete, we can allow a few more
3337 * transactions into this iclog before we close it down.
3339 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3340 * refcnt so we can release the log (which drops the ref count).
3341 * The state switch keeps new transaction commits from using
3342 * this buffer. When the current commits finish writing into
3343 * the buffer, the refcount will drop to zero and the buffer
3346 if (!already_slept &&
3347 (iclog->ic_prev->ic_state &
3348 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3349 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3351 XFS_STATS_INC(mp, xs_log_force_sleep);
3353 xlog_wait(&iclog->ic_prev->ic_write_wait,
3357 atomic_inc(&iclog->ic_refcnt);
3358 xlog_state_switch_iclogs(log, iclog, 0);
3359 spin_unlock(&log->l_icloglock);
3360 if (xlog_state_release_iclog(log, iclog))
3364 spin_lock(&log->l_icloglock);
3367 if (!(flags & XFS_LOG_SYNC) ||
3368 (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3371 if (iclog->ic_state & XLOG_STATE_IOERROR)
3374 XFS_STATS_INC(mp, xs_log_force_sleep);
3375 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3376 if (iclog->ic_state & XLOG_STATE_IOERROR)
3381 spin_unlock(&log->l_icloglock);
3384 spin_unlock(&log->l_icloglock);
3389 * Force the in-core log to disk for a specific LSN.
3391 * Find in-core log with lsn.
3392 * If it is in the DIRTY state, just return.
3393 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3394 * state and go to sleep or return.
3395 * If it is in any other state, go to sleep or return.
3397 * Synchronous forces are implemented with a wait queue. All callers trying
3398 * to force a given lsn to disk must wait on the queue attached to the
3399 * specific in-core log. When given in-core log finally completes its write
3400 * to disk, that thread will wake up all threads waiting on the queue.
3404 struct xfs_mount *mp,
3412 XFS_STATS_INC(mp, xs_log_force);
3413 trace_xfs_log_force(mp, lsn, _RET_IP_);
3415 lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3416 if (lsn == NULLCOMMITLSN)
3419 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3421 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3426 * Called when we want to mark the current iclog as being ready to sync to
3430 xlog_state_want_sync(
3432 struct xlog_in_core *iclog)
3434 assert_spin_locked(&log->l_icloglock);
3436 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3437 xlog_state_switch_iclogs(log, iclog, 0);
3439 ASSERT(iclog->ic_state &
3440 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3445 /*****************************************************************************
3449 *****************************************************************************
3453 * Free a used ticket when its refcount falls to zero.
3457 xlog_ticket_t *ticket)
3459 ASSERT(atomic_read(&ticket->t_ref) > 0);
3460 if (atomic_dec_and_test(&ticket->t_ref))
3461 kmem_zone_free(xfs_log_ticket_zone, ticket);
3466 xlog_ticket_t *ticket)
3468 ASSERT(atomic_read(&ticket->t_ref) > 0);
3469 atomic_inc(&ticket->t_ref);
3474 * Figure out the total log space unit (in bytes) that would be
3475 * required for a log ticket.
3478 xfs_log_calc_unit_res(
3479 struct xfs_mount *mp,
3482 struct xlog *log = mp->m_log;
3487 * Permanent reservations have up to 'cnt'-1 active log operations
3488 * in the log. A unit in this case is the amount of space for one
3489 * of these log operations. Normal reservations have a cnt of 1
3490 * and their unit amount is the total amount of space required.
3492 * The following lines of code account for non-transaction data
3493 * which occupy space in the on-disk log.
3495 * Normal form of a transaction is:
3496 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3497 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3499 * We need to account for all the leadup data and trailer data
3500 * around the transaction data.
3501 * And then we need to account for the worst case in terms of using
3503 * The worst case will happen if:
3504 * - the placement of the transaction happens to be such that the
3505 * roundoff is at its maximum
3506 * - the transaction data is synced before the commit record is synced
3507 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3508 * Therefore the commit record is in its own Log Record.
3509 * This can happen as the commit record is called with its
3510 * own region to xlog_write().
3511 * This then means that in the worst case, roundoff can happen for
3512 * the commit-rec as well.
3513 * The commit-rec is smaller than padding in this scenario and so it is
3514 * not added separately.
3517 /* for trans header */
3518 unit_bytes += sizeof(xlog_op_header_t);
3519 unit_bytes += sizeof(xfs_trans_header_t);
3522 unit_bytes += sizeof(xlog_op_header_t);
3525 * for LR headers - the space for data in an iclog is the size minus
3526 * the space used for the headers. If we use the iclog size, then we
3527 * undercalculate the number of headers required.
3529 * Furthermore - the addition of op headers for split-recs might
3530 * increase the space required enough to require more log and op
3531 * headers, so take that into account too.
3533 * IMPORTANT: This reservation makes the assumption that if this
3534 * transaction is the first in an iclog and hence has the LR headers
3535 * accounted to it, then the remaining space in the iclog is
3536 * exclusively for this transaction. i.e. if the transaction is larger
3537 * than the iclog, it will be the only thing in that iclog.
3538 * Fundamentally, this means we must pass the entire log vector to
3539 * xlog_write to guarantee this.
3541 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3542 num_headers = howmany(unit_bytes, iclog_space);
3544 /* for split-recs - ophdrs added when data split over LRs */
3545 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3547 /* add extra header reservations if we overrun */
3548 while (!num_headers ||
3549 howmany(unit_bytes, iclog_space) > num_headers) {
3550 unit_bytes += sizeof(xlog_op_header_t);
3553 unit_bytes += log->l_iclog_hsize * num_headers;
3555 /* for commit-rec LR header - note: padding will subsume the ophdr */
3556 unit_bytes += log->l_iclog_hsize;
3558 /* for roundoff padding for transaction data and one for commit record */
3559 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3560 /* log su roundoff */
3561 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3564 unit_bytes += 2 * BBSIZE;
3571 * Allocate and initialise a new log ticket.
3573 struct xlog_ticket *
3580 xfs_km_flags_t alloc_flags)
3582 struct xlog_ticket *tic;
3585 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3589 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3591 atomic_set(&tic->t_ref, 1);
3592 tic->t_task = current;
3593 INIT_LIST_HEAD(&tic->t_queue);
3594 tic->t_unit_res = unit_res;
3595 tic->t_curr_res = unit_res;
3598 tic->t_tid = prandom_u32();
3599 tic->t_clientid = client;
3600 tic->t_flags = XLOG_TIC_INITED;
3602 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3604 xlog_tic_reset_res(tic);
3610 /******************************************************************************
3612 * Log debug routines
3614 ******************************************************************************
3618 * Make sure that the destination ptr is within the valid data region of
3619 * one of the iclogs. This uses backup pointers stored in a different
3620 * part of the log in case we trash the log structure.
3623 xlog_verify_dest_ptr(
3630 for (i = 0; i < log->l_iclog_bufs; i++) {
3631 if (ptr >= log->l_iclog_bak[i] &&
3632 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3637 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3641 * Check to make sure the grant write head didn't just over lap the tail. If
3642 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3643 * the cycles differ by exactly one and check the byte count.
3645 * This check is run unlocked, so can give false positives. Rather than assert
3646 * on failures, use a warn-once flag and a panic tag to allow the admin to
3647 * determine if they want to panic the machine when such an error occurs. For
3648 * debug kernels this will have the same effect as using an assert but, unlinke
3649 * an assert, it can be turned off at runtime.
3652 xlog_verify_grant_tail(
3655 int tail_cycle, tail_blocks;
3658 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3659 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3660 if (tail_cycle != cycle) {
3661 if (cycle - 1 != tail_cycle &&
3662 !(log->l_flags & XLOG_TAIL_WARN)) {
3663 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3664 "%s: cycle - 1 != tail_cycle", __func__);
3665 log->l_flags |= XLOG_TAIL_WARN;
3668 if (space > BBTOB(tail_blocks) &&
3669 !(log->l_flags & XLOG_TAIL_WARN)) {
3670 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3671 "%s: space > BBTOB(tail_blocks)", __func__);
3672 log->l_flags |= XLOG_TAIL_WARN;
3677 /* check if it will fit */
3679 xlog_verify_tail_lsn(
3681 struct xlog_in_core *iclog,
3686 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3688 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3689 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3690 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3692 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3694 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3695 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3697 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3698 if (blocks < BTOBB(iclog->ic_offset) + 1)
3699 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3701 } /* xlog_verify_tail_lsn */
3704 * Perform a number of checks on the iclog before writing to disk.
3706 * 1. Make sure the iclogs are still circular
3707 * 2. Make sure we have a good magic number
3708 * 3. Make sure we don't have magic numbers in the data
3709 * 4. Check fields of each log operation header for:
3710 * A. Valid client identifier
3711 * B. tid ptr value falls in valid ptr space (user space code)
3712 * C. Length in log record header is correct according to the
3713 * individual operation headers within record.
3714 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3715 * log, check the preceding blocks of the physical log to make sure all
3716 * the cycle numbers agree with the current cycle number.
3721 struct xlog_in_core *iclog,
3724 xlog_op_header_t *ophead;
3725 xlog_in_core_t *icptr;
3726 xlog_in_core_2_t *xhdr;
3727 void *base_ptr, *ptr, *p;
3728 ptrdiff_t field_offset;
3730 int len, i, j, k, op_len;
3733 /* check validity of iclog pointers */
3734 spin_lock(&log->l_icloglock);
3735 icptr = log->l_iclog;
3736 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3739 if (icptr != log->l_iclog)
3740 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3741 spin_unlock(&log->l_icloglock);
3743 /* check log magic numbers */
3744 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3745 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3747 base_ptr = ptr = &iclog->ic_header;
3748 p = &iclog->ic_header;
3749 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3750 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3751 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3756 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3757 base_ptr = ptr = iclog->ic_datap;
3759 xhdr = iclog->ic_data;
3760 for (i = 0; i < len; i++) {
3763 /* clientid is only 1 byte */
3764 p = &ophead->oh_clientid;
3765 field_offset = p - base_ptr;
3766 if (field_offset & 0x1ff) {
3767 clientid = ophead->oh_clientid;
3769 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3770 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3771 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3772 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3773 clientid = xlog_get_client_id(
3774 xhdr[j].hic_xheader.xh_cycle_data[k]);
3776 clientid = xlog_get_client_id(
3777 iclog->ic_header.h_cycle_data[idx]);
3780 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3782 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3783 __func__, clientid, ophead,
3784 (unsigned long)field_offset);
3787 p = &ophead->oh_len;
3788 field_offset = p - base_ptr;
3789 if (field_offset & 0x1ff) {
3790 op_len = be32_to_cpu(ophead->oh_len);
3792 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3793 (uintptr_t)iclog->ic_datap);
3794 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3795 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3796 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3797 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3799 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3802 ptr += sizeof(xlog_op_header_t) + op_len;
3804 } /* xlog_verify_iclog */
3808 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3814 xlog_in_core_t *iclog, *ic;
3816 iclog = log->l_iclog;
3817 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3819 * Mark all the incore logs IOERROR.
3820 * From now on, no log flushes will result.
3824 ic->ic_state = XLOG_STATE_IOERROR;
3826 } while (ic != iclog);
3830 * Return non-zero, if state transition has already happened.
3836 * This is called from xfs_force_shutdown, when we're forcibly
3837 * shutting down the filesystem, typically because of an IO error.
3838 * Our main objectives here are to make sure that:
3839 * a. if !logerror, flush the logs to disk. Anything modified
3840 * after this is ignored.
3841 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3842 * parties to find out, 'atomically'.
3843 * c. those who're sleeping on log reservations, pinned objects and
3844 * other resources get woken up, and be told the bad news.
3845 * d. nothing new gets queued up after (b) and (c) are done.
3847 * Note: for the !logerror case we need to flush the regions held in memory out
3848 * to disk first. This needs to be done before the log is marked as shutdown,
3849 * otherwise the iclog writes will fail.
3852 xfs_log_force_umount(
3853 struct xfs_mount *mp,
3862 * If this happens during log recovery, don't worry about
3863 * locking; the log isn't open for business yet.
3866 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3867 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3869 mp->m_sb_bp->b_flags |= XBF_DONE;
3874 * Somebody could've already done the hard work for us.
3875 * No need to get locks for this.
3877 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3878 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3883 * Flush all the completed transactions to disk before marking the log
3884 * being shut down. We need to do it in this order to ensure that
3885 * completed operations are safely on disk before we shut down, and that
3886 * we don't have to issue any buffer IO after the shutdown flags are set
3887 * to guarantee this.
3890 xfs_log_force(mp, XFS_LOG_SYNC);
3893 * mark the filesystem and the as in a shutdown state and wake
3894 * everybody up to tell them the bad news.
3896 spin_lock(&log->l_icloglock);
3897 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3899 mp->m_sb_bp->b_flags |= XBF_DONE;
3902 * Mark the log and the iclogs with IO error flags to prevent any
3903 * further log IO from being issued or completed.
3905 log->l_flags |= XLOG_IO_ERROR;
3906 retval = xlog_state_ioerror(log);
3907 spin_unlock(&log->l_icloglock);
3910 * We don't want anybody waiting for log reservations after this. That
3911 * means we have to wake up everybody queued up on reserveq as well as
3912 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3913 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3914 * action is protected by the grant locks.
3916 xlog_grant_head_wake_all(&log->l_reserve_head);
3917 xlog_grant_head_wake_all(&log->l_write_head);
3920 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3921 * as if the log writes were completed. The abort handling in the log
3922 * item committed callback functions will do this again under lock to
3925 wake_up_all(&log->l_cilp->xc_commit_wait);
3926 xlog_state_do_callback(log, true, NULL);
3928 #ifdef XFSERRORDEBUG
3930 xlog_in_core_t *iclog;
3932 spin_lock(&log->l_icloglock);
3933 iclog = log->l_iclog;
3935 ASSERT(iclog->ic_callback == 0);
3936 iclog = iclog->ic_next;
3937 } while (iclog != log->l_iclog);
3938 spin_unlock(&log->l_icloglock);
3941 /* return non-zero if log IOERROR transition had already happened */
3949 xlog_in_core_t *iclog;
3951 iclog = log->l_iclog;
3953 /* endianness does not matter here, zero is zero in
3956 if (iclog->ic_header.h_num_logops)
3958 iclog = iclog->ic_next;
3959 } while (iclog != log->l_iclog);
3964 * Verify that an LSN stamped into a piece of metadata is valid. This is
3965 * intended for use in read verifiers on v5 superblocks.
3969 struct xfs_mount *mp,
3972 struct xlog *log = mp->m_log;
3976 * norecovery mode skips mount-time log processing and unconditionally
3977 * resets the in-core LSN. We can't validate in this mode, but
3978 * modifications are not allowed anyways so just return true.
3980 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3984 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3985 * handled by recovery and thus safe to ignore here.
3987 if (lsn == NULLCOMMITLSN)
3990 valid = xlog_valid_lsn(mp->m_log, lsn);
3992 /* warn the user about what's gone wrong before verifier failure */
3994 spin_lock(&log->l_icloglock);
3996 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3997 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3998 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3999 log->l_curr_cycle, log->l_curr_block);
4000 spin_unlock(&log->l_icloglock);
4007 xfs_log_in_recovery(
4008 struct xfs_mount *mp)
4010 struct xlog *log = mp->m_log;
4012 return log->l_flags & XLOG_ACTIVE_RECOVERY;