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_log_recover.h"
20 #include "xfs_inode.h"
21 #include "xfs_trace.h"
22 #include "xfs_fsops.h"
23 #include "xfs_cksum.h"
24 #include "xfs_sysfs.h"
26 #include "xfs_health.h"
28 kmem_zone_t *xfs_log_ticket_zone;
30 /* Local miscellaneous function prototypes */
34 struct xlog_ticket *ticket,
35 struct xlog_in_core **iclog,
36 xfs_lsn_t *commitlsnp);
41 struct xfs_buftarg *log_target,
42 xfs_daddr_t blk_offset,
52 /* local state machine functions */
53 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
55 xlog_state_do_callback(
58 struct xlog_in_core *iclog);
60 xlog_state_get_iclog_space(
63 struct xlog_in_core **iclog,
64 struct xlog_ticket *ticket,
68 xlog_state_release_iclog(
70 struct xlog_in_core *iclog);
72 xlog_state_switch_iclogs(
74 struct xlog_in_core *iclog,
79 struct xlog_in_core *iclog);
86 xlog_regrant_reserve_log_space(
88 struct xlog_ticket *ticket);
90 xlog_ungrant_log_space(
92 struct xlog_ticket *ticket);
100 xlog_verify_grant_tail(
105 struct xlog_in_core *iclog,
109 xlog_verify_tail_lsn(
111 struct xlog_in_core *iclog,
114 #define xlog_verify_dest_ptr(a,b)
115 #define xlog_verify_grant_tail(a)
116 #define xlog_verify_iclog(a,b,c,d)
117 #define xlog_verify_tail_lsn(a,b,c)
125 xlog_grant_sub_space(
130 int64_t head_val = atomic64_read(head);
136 xlog_crack_grant_head_val(head_val, &cycle, &space);
140 space += log->l_logsize;
145 new = xlog_assign_grant_head_val(cycle, space);
146 head_val = atomic64_cmpxchg(head, old, new);
147 } while (head_val != old);
151 xlog_grant_add_space(
156 int64_t head_val = atomic64_read(head);
163 xlog_crack_grant_head_val(head_val, &cycle, &space);
165 tmp = log->l_logsize - space;
174 new = xlog_assign_grant_head_val(cycle, space);
175 head_val = atomic64_cmpxchg(head, old, new);
176 } while (head_val != old);
180 xlog_grant_head_init(
181 struct xlog_grant_head *head)
183 xlog_assign_grant_head(&head->grant, 1, 0);
184 INIT_LIST_HEAD(&head->waiters);
185 spin_lock_init(&head->lock);
189 xlog_grant_head_wake_all(
190 struct xlog_grant_head *head)
192 struct xlog_ticket *tic;
194 spin_lock(&head->lock);
195 list_for_each_entry(tic, &head->waiters, t_queue)
196 wake_up_process(tic->t_task);
197 spin_unlock(&head->lock);
201 xlog_ticket_reservation(
203 struct xlog_grant_head *head,
204 struct xlog_ticket *tic)
206 if (head == &log->l_write_head) {
207 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
208 return tic->t_unit_res;
210 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
211 return tic->t_unit_res * tic->t_cnt;
213 return tic->t_unit_res;
218 xlog_grant_head_wake(
220 struct xlog_grant_head *head,
223 struct xlog_ticket *tic;
226 list_for_each_entry(tic, &head->waiters, t_queue) {
227 need_bytes = xlog_ticket_reservation(log, head, tic);
228 if (*free_bytes < need_bytes)
231 *free_bytes -= need_bytes;
232 trace_xfs_log_grant_wake_up(log, tic);
233 wake_up_process(tic->t_task);
240 xlog_grant_head_wait(
242 struct xlog_grant_head *head,
243 struct xlog_ticket *tic,
244 int need_bytes) __releases(&head->lock)
245 __acquires(&head->lock)
247 list_add_tail(&tic->t_queue, &head->waiters);
250 if (XLOG_FORCED_SHUTDOWN(log))
252 xlog_grant_push_ail(log, need_bytes);
254 __set_current_state(TASK_UNINTERRUPTIBLE);
255 spin_unlock(&head->lock);
257 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
259 trace_xfs_log_grant_sleep(log, tic);
261 trace_xfs_log_grant_wake(log, tic);
263 spin_lock(&head->lock);
264 if (XLOG_FORCED_SHUTDOWN(log))
266 } while (xlog_space_left(log, &head->grant) < need_bytes);
268 list_del_init(&tic->t_queue);
271 list_del_init(&tic->t_queue);
276 * Atomically get the log space required for a log ticket.
278 * Once a ticket gets put onto head->waiters, it will only return after the
279 * needed reservation is satisfied.
281 * This function is structured so that it has a lock free fast path. This is
282 * necessary because every new transaction reservation will come through this
283 * path. Hence any lock will be globally hot if we take it unconditionally on
286 * As tickets are only ever moved on and off head->waiters under head->lock, we
287 * only need to take that lock if we are going to add the ticket to the queue
288 * and sleep. We can avoid taking the lock if the ticket was never added to
289 * head->waiters because the t_queue list head will be empty and we hold the
290 * only reference to it so it can safely be checked unlocked.
293 xlog_grant_head_check(
295 struct xlog_grant_head *head,
296 struct xlog_ticket *tic,
302 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
305 * If there are other waiters on the queue then give them a chance at
306 * logspace before us. Wake up the first waiters, if we do not wake
307 * up all the waiters then go to sleep waiting for more free space,
308 * otherwise try to get some space for this transaction.
310 *need_bytes = xlog_ticket_reservation(log, head, tic);
311 free_bytes = xlog_space_left(log, &head->grant);
312 if (!list_empty_careful(&head->waiters)) {
313 spin_lock(&head->lock);
314 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
315 free_bytes < *need_bytes) {
316 error = xlog_grant_head_wait(log, head, tic,
319 spin_unlock(&head->lock);
320 } else if (free_bytes < *need_bytes) {
321 spin_lock(&head->lock);
322 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
323 spin_unlock(&head->lock);
330 xlog_tic_reset_res(xlog_ticket_t *tic)
333 tic->t_res_arr_sum = 0;
334 tic->t_res_num_ophdrs = 0;
338 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
340 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
341 /* add to overflow and start again */
342 tic->t_res_o_flow += tic->t_res_arr_sum;
344 tic->t_res_arr_sum = 0;
347 tic->t_res_arr[tic->t_res_num].r_len = len;
348 tic->t_res_arr[tic->t_res_num].r_type = type;
349 tic->t_res_arr_sum += len;
354 * Replenish the byte reservation required by moving the grant write head.
358 struct xfs_mount *mp,
359 struct xlog_ticket *tic)
361 struct xlog *log = mp->m_log;
365 if (XLOG_FORCED_SHUTDOWN(log))
368 XFS_STATS_INC(mp, xs_try_logspace);
371 * This is a new transaction on the ticket, so we need to change the
372 * transaction ID so that the next transaction has a different TID in
373 * the log. Just add one to the existing tid so that we can see chains
374 * of rolling transactions in the log easily.
378 xlog_grant_push_ail(log, tic->t_unit_res);
380 tic->t_curr_res = tic->t_unit_res;
381 xlog_tic_reset_res(tic);
386 trace_xfs_log_regrant(log, tic);
388 error = xlog_grant_head_check(log, &log->l_write_head, tic,
393 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
394 trace_xfs_log_regrant_exit(log, tic);
395 xlog_verify_grant_tail(log);
400 * If we are failing, make sure the ticket doesn't have any current
401 * reservations. We don't want to add this back when the ticket/
402 * transaction gets cancelled.
405 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
410 * Reserve log space and return a ticket corresponding to the reservation.
412 * Each reservation is going to reserve extra space for a log record header.
413 * When writes happen to the on-disk log, we don't subtract the length of the
414 * log record header from any reservation. By wasting space in each
415 * reservation, we prevent over allocation problems.
419 struct xfs_mount *mp,
422 struct xlog_ticket **ticp,
426 struct xlog *log = mp->m_log;
427 struct xlog_ticket *tic;
431 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
433 if (XLOG_FORCED_SHUTDOWN(log))
436 XFS_STATS_INC(mp, xs_try_logspace);
438 ASSERT(*ticp == NULL);
439 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
440 KM_SLEEP | KM_MAYFAIL);
446 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
449 trace_xfs_log_reserve(log, tic);
451 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
456 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
457 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
458 trace_xfs_log_reserve_exit(log, tic);
459 xlog_verify_grant_tail(log);
464 * If we are failing, make sure the ticket doesn't have any current
465 * reservations. We don't want to add this back when the ticket/
466 * transaction gets cancelled.
469 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
477 * 1. currblock field gets updated at startup and after in-core logs
478 * marked as with WANT_SYNC.
482 * This routine is called when a user of a log manager ticket is done with
483 * the reservation. If the ticket was ever used, then a commit record for
484 * the associated transaction is written out as a log operation header with
485 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with
486 * a given ticket. If the ticket was one with a permanent reservation, then
487 * a few operations are done differently. Permanent reservation tickets by
488 * default don't release the reservation. They just commit the current
489 * transaction with the belief that the reservation is still needed. A flag
490 * must be passed in before permanent reservations are actually released.
491 * When these type of tickets are not released, they need to be set into
492 * the inited state again. By doing this, a start record will be written
493 * out when the next write occurs.
497 struct xfs_mount *mp,
498 struct xlog_ticket *ticket,
499 struct xlog_in_core **iclog,
502 struct xlog *log = mp->m_log;
505 if (XLOG_FORCED_SHUTDOWN(log) ||
507 * If nothing was ever written, don't write out commit record.
508 * If we get an error, just continue and give back the log ticket.
510 (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
511 (xlog_commit_record(log, ticket, iclog, &lsn)))) {
512 lsn = (xfs_lsn_t) -1;
518 trace_xfs_log_done_nonperm(log, ticket);
521 * Release ticket if not permanent reservation or a specific
522 * request has been made to release a permanent reservation.
524 xlog_ungrant_log_space(log, ticket);
526 trace_xfs_log_done_perm(log, ticket);
528 xlog_regrant_reserve_log_space(log, ticket);
529 /* If this ticket was a permanent reservation and we aren't
530 * trying to release it, reset the inited flags; so next time
531 * we write, a start record will be written out.
533 ticket->t_flags |= XLOG_TIC_INITED;
536 xfs_log_ticket_put(ticket);
541 * Attaches a new iclog I/O completion callback routine during
542 * transaction commit. If the log is in error state, a non-zero
543 * return code is handed back and the caller is responsible for
544 * executing the callback at an appropriate time.
548 struct xlog_in_core *iclog,
549 xfs_log_callback_t *cb)
553 spin_lock(&iclog->ic_callback_lock);
554 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
556 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
557 (iclog->ic_state == XLOG_STATE_WANT_SYNC));
559 *(iclog->ic_callback_tail) = cb;
560 iclog->ic_callback_tail = &(cb->cb_next);
562 spin_unlock(&iclog->ic_callback_lock);
567 xfs_log_release_iclog(
568 struct xfs_mount *mp,
569 struct xlog_in_core *iclog)
571 if (xlog_state_release_iclog(mp->m_log, iclog)) {
572 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
580 * Mount a log filesystem
582 * mp - ubiquitous xfs mount point structure
583 * log_target - buftarg of on-disk log device
584 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
585 * num_bblocks - Number of BBSIZE blocks in on-disk log
587 * Return error or zero.
592 xfs_buftarg_t *log_target,
593 xfs_daddr_t blk_offset,
596 bool fatal = xfs_sb_version_hascrc(&mp->m_sb);
600 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
601 xfs_notice(mp, "Mounting V%d Filesystem",
602 XFS_SB_VERSION_NUM(&mp->m_sb));
605 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
606 XFS_SB_VERSION_NUM(&mp->m_sb));
607 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
610 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
611 if (IS_ERR(mp->m_log)) {
612 error = PTR_ERR(mp->m_log);
617 * Validate the given log space and drop a critical message via syslog
618 * if the log size is too small that would lead to some unexpected
619 * situations in transaction log space reservation stage.
621 * Note: we can't just reject the mount if the validation fails. This
622 * would mean that people would have to downgrade their kernel just to
623 * remedy the situation as there is no way to grow the log (short of
624 * black magic surgery with xfs_db).
626 * We can, however, reject mounts for CRC format filesystems, as the
627 * mkfs binary being used to make the filesystem should never create a
628 * filesystem with a log that is too small.
630 min_logfsbs = xfs_log_calc_minimum_size(mp);
632 if (mp->m_sb.sb_logblocks < min_logfsbs) {
634 "Log size %d blocks too small, minimum size is %d blocks",
635 mp->m_sb.sb_logblocks, min_logfsbs);
637 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
639 "Log size %d blocks too large, maximum size is %lld blocks",
640 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
642 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
644 "log size %lld bytes too large, maximum size is %lld bytes",
645 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
648 } else if (mp->m_sb.sb_logsunit > 1 &&
649 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
651 "log stripe unit %u bytes must be a multiple of block size",
652 mp->m_sb.sb_logsunit);
658 * Log check errors are always fatal on v5; or whenever bad
659 * metadata leads to a crash.
662 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
666 xfs_crit(mp, "Log size out of supported range.");
668 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
672 * Initialize the AIL now we have a log.
674 error = xfs_trans_ail_init(mp);
676 xfs_warn(mp, "AIL initialisation failed: error %d", error);
679 mp->m_log->l_ailp = mp->m_ail;
682 * skip log recovery on a norecovery mount. pretend it all
685 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
686 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
689 mp->m_flags &= ~XFS_MOUNT_RDONLY;
691 error = xlog_recover(mp->m_log);
694 mp->m_flags |= XFS_MOUNT_RDONLY;
696 xfs_warn(mp, "log mount/recovery failed: error %d",
698 xlog_recover_cancel(mp->m_log);
699 goto out_destroy_ail;
703 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
706 goto out_destroy_ail;
708 /* Normal transactions can now occur */
709 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
712 * Now the log has been fully initialised and we know were our
713 * space grant counters are, we can initialise the permanent ticket
714 * needed for delayed logging to work.
716 xlog_cil_init_post_recovery(mp->m_log);
721 xfs_trans_ail_destroy(mp);
723 xlog_dealloc_log(mp->m_log);
729 * Finish the recovery of the file system. This is separate from the
730 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
731 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
734 * If we finish recovery successfully, start the background log work. If we are
735 * not doing recovery, then we have a RO filesystem and we don't need to start
739 xfs_log_mount_finish(
740 struct xfs_mount *mp)
743 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
744 bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
746 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
747 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
749 } else if (readonly) {
750 /* Allow unlinked processing to proceed */
751 mp->m_flags &= ~XFS_MOUNT_RDONLY;
755 * During the second phase of log recovery, we need iget and
756 * iput to behave like they do for an active filesystem.
757 * xfs_fs_drop_inode needs to be able to prevent the deletion
758 * of inodes before we're done replaying log items on those
759 * inodes. Turn it off immediately after recovery finishes
760 * so that we don't leak the quota inodes if subsequent mount
763 * We let all inodes involved in redo item processing end up on
764 * the LRU instead of being evicted immediately so that if we do
765 * something to an unlinked inode, the irele won't cause
766 * premature truncation and freeing of the inode, which results
767 * in log recovery failure. We have to evict the unreferenced
768 * lru inodes after clearing SB_ACTIVE because we don't
769 * otherwise clean up the lru if there's a subsequent failure in
770 * xfs_mountfs, which leads to us leaking the inodes if nothing
771 * else (e.g. quotacheck) references the inodes before the
772 * mount failure occurs.
774 mp->m_super->s_flags |= SB_ACTIVE;
775 error = xlog_recover_finish(mp->m_log);
777 xfs_log_work_queue(mp);
778 mp->m_super->s_flags &= ~SB_ACTIVE;
779 evict_inodes(mp->m_super);
782 * Drain the buffer LRU after log recovery. This is required for v4
783 * filesystems to avoid leaving around buffers with NULL verifier ops,
784 * but we do it unconditionally to make sure we're always in a clean
785 * cache state after mount.
787 * Don't push in the error case because the AIL may have pending intents
788 * that aren't removed until recovery is cancelled.
790 if (!error && recovered) {
791 xfs_log_force(mp, XFS_LOG_SYNC);
792 xfs_ail_push_all_sync(mp->m_ail);
794 xfs_wait_buftarg(mp->m_ddev_targp);
797 mp->m_flags |= XFS_MOUNT_RDONLY;
803 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
807 xfs_log_mount_cancel(
808 struct xfs_mount *mp)
812 error = xlog_recover_cancel(mp->m_log);
819 * Final log writes as part of unmount.
821 * Mark the filesystem clean as unmount happens. Note that during relocation
822 * this routine needs to be executed as part of source-bag while the
823 * deallocation must not be done until source-end.
826 /* Actually write the unmount record to disk. */
828 xfs_log_write_unmount_record(
829 struct xfs_mount *mp)
831 /* the data section must be 32 bit size aligned */
832 struct xfs_unmount_log_format magic = {
833 .magic = XLOG_UNMOUNT_TYPE,
835 struct xfs_log_iovec reg = {
837 .i_len = sizeof(magic),
838 .i_type = XLOG_REG_TYPE_UNMOUNT,
840 struct xfs_log_vec vec = {
844 struct xlog *log = mp->m_log;
845 struct xlog_in_core *iclog;
846 struct xlog_ticket *tic = NULL;
848 uint flags = XLOG_UNMOUNT_TRANS;
851 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
856 * If we think the summary counters are bad, clear the unmount header
857 * flag in the unmount record so that the summary counters will be
858 * recalculated during log recovery at next mount. Refer to
859 * xlog_check_unmount_rec for more details.
861 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
862 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
863 xfs_alert(mp, "%s: will fix summary counters at next mount",
865 flags &= ~XLOG_UNMOUNT_TRANS;
868 /* remove inited flag, and account for space used */
870 tic->t_curr_res -= sizeof(magic);
871 error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
873 * At this point, we're umounting anyway, so there's no point in
874 * transitioning log state to IOERROR. Just continue...
878 xfs_alert(mp, "%s: unmount record failed", __func__);
880 spin_lock(&log->l_icloglock);
881 iclog = log->l_iclog;
882 atomic_inc(&iclog->ic_refcnt);
883 xlog_state_want_sync(log, iclog);
884 spin_unlock(&log->l_icloglock);
885 error = xlog_state_release_iclog(log, iclog);
887 spin_lock(&log->l_icloglock);
888 switch (iclog->ic_state) {
890 if (!XLOG_FORCED_SHUTDOWN(log)) {
891 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
895 case XLOG_STATE_ACTIVE:
896 case XLOG_STATE_DIRTY:
897 spin_unlock(&log->l_icloglock);
902 trace_xfs_log_umount_write(log, tic);
903 xlog_ungrant_log_space(log, tic);
904 xfs_log_ticket_put(tic);
909 * Unmount record used to have a string "Unmount filesystem--" in the
910 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
911 * We just write the magic number now since that particular field isn't
912 * currently architecture converted and "Unmount" is a bit foo.
913 * As far as I know, there weren't any dependencies on the old behaviour.
917 xfs_log_unmount_write(xfs_mount_t *mp)
919 struct xlog *log = mp->m_log;
920 xlog_in_core_t *iclog;
922 xlog_in_core_t *first_iclog;
927 * Don't write out unmount record on norecovery mounts or ro devices.
928 * Or, if we are doing a forced umount (typically because of IO errors).
930 if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
931 xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
932 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
936 error = xfs_log_force(mp, XFS_LOG_SYNC);
937 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
940 first_iclog = iclog = log->l_iclog;
942 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
943 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
944 ASSERT(iclog->ic_offset == 0);
946 iclog = iclog->ic_next;
947 } while (iclog != first_iclog);
949 if (! (XLOG_FORCED_SHUTDOWN(log))) {
950 xfs_log_write_unmount_record(mp);
953 * We're already in forced_shutdown mode, couldn't
954 * even attempt to write out the unmount transaction.
956 * Go through the motions of sync'ing and releasing
957 * the iclog, even though no I/O will actually happen,
958 * we need to wait for other log I/Os that may already
959 * be in progress. Do this as a separate section of
960 * code so we'll know if we ever get stuck here that
961 * we're in this odd situation of trying to unmount
962 * a file system that went into forced_shutdown as
963 * the result of an unmount..
965 spin_lock(&log->l_icloglock);
966 iclog = log->l_iclog;
967 atomic_inc(&iclog->ic_refcnt);
969 xlog_state_want_sync(log, iclog);
970 spin_unlock(&log->l_icloglock);
971 error = xlog_state_release_iclog(log, iclog);
973 spin_lock(&log->l_icloglock);
975 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE
976 || iclog->ic_state == XLOG_STATE_DIRTY
977 || iclog->ic_state == XLOG_STATE_IOERROR) ) {
979 xlog_wait(&iclog->ic_force_wait,
982 spin_unlock(&log->l_icloglock);
987 } /* xfs_log_unmount_write */
990 * Empty the log for unmount/freeze.
992 * To do this, we first need to shut down the background log work so it is not
993 * trying to cover the log as we clean up. We then need to unpin all objects in
994 * the log so we can then flush them out. Once they have completed their IO and
995 * run the callbacks removing themselves from the AIL, we can write the unmount
1000 struct xfs_mount *mp)
1002 cancel_delayed_work_sync(&mp->m_log->l_work);
1003 xfs_log_force(mp, XFS_LOG_SYNC);
1006 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1007 * will push it, xfs_wait_buftarg() will not wait for it. Further,
1008 * xfs_buf_iowait() cannot be used because it was pushed with the
1009 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1010 * the IO to complete.
1012 xfs_ail_push_all_sync(mp->m_ail);
1013 xfs_wait_buftarg(mp->m_ddev_targp);
1014 xfs_buf_lock(mp->m_sb_bp);
1015 xfs_buf_unlock(mp->m_sb_bp);
1017 xfs_log_unmount_write(mp);
1021 * Shut down and release the AIL and Log.
1023 * During unmount, we need to ensure we flush all the dirty metadata objects
1024 * from the AIL so that the log is empty before we write the unmount record to
1025 * the log. Once this is done, we can tear down the AIL and the log.
1029 struct xfs_mount *mp)
1031 xfs_log_quiesce(mp);
1033 xfs_trans_ail_destroy(mp);
1035 xfs_sysfs_del(&mp->m_log->l_kobj);
1037 xlog_dealloc_log(mp->m_log);
1042 struct xfs_mount *mp,
1043 struct xfs_log_item *item,
1045 const struct xfs_item_ops *ops)
1047 item->li_mountp = mp;
1048 item->li_ailp = mp->m_ail;
1049 item->li_type = type;
1053 INIT_LIST_HEAD(&item->li_ail);
1054 INIT_LIST_HEAD(&item->li_cil);
1055 INIT_LIST_HEAD(&item->li_bio_list);
1056 INIT_LIST_HEAD(&item->li_trans);
1060 * Wake up processes waiting for log space after we have moved the log tail.
1064 struct xfs_mount *mp)
1066 struct xlog *log = mp->m_log;
1069 if (XLOG_FORCED_SHUTDOWN(log))
1072 if (!list_empty_careful(&log->l_write_head.waiters)) {
1073 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1075 spin_lock(&log->l_write_head.lock);
1076 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1077 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1078 spin_unlock(&log->l_write_head.lock);
1081 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1082 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1084 spin_lock(&log->l_reserve_head.lock);
1085 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1086 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1087 spin_unlock(&log->l_reserve_head.lock);
1092 * Determine if we have a transaction that has gone to disk that needs to be
1093 * covered. To begin the transition to the idle state firstly the log needs to
1094 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1095 * we start attempting to cover the log.
1097 * Only if we are then in a state where covering is needed, the caller is
1098 * informed that dummy transactions are required to move the log into the idle
1101 * If there are any items in the AIl or CIL, then we do not want to attempt to
1102 * cover the log as we may be in a situation where there isn't log space
1103 * available to run a dummy transaction and this can lead to deadlocks when the
1104 * tail of the log is pinned by an item that is modified in the CIL. Hence
1105 * there's no point in running a dummy transaction at this point because we
1106 * can't start trying to idle the log until both the CIL and AIL are empty.
1109 xfs_log_need_covered(xfs_mount_t *mp)
1111 struct xlog *log = mp->m_log;
1114 if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1117 if (!xlog_cil_empty(log))
1120 spin_lock(&log->l_icloglock);
1121 switch (log->l_covered_state) {
1122 case XLOG_STATE_COVER_DONE:
1123 case XLOG_STATE_COVER_DONE2:
1124 case XLOG_STATE_COVER_IDLE:
1126 case XLOG_STATE_COVER_NEED:
1127 case XLOG_STATE_COVER_NEED2:
1128 if (xfs_ail_min_lsn(log->l_ailp))
1130 if (!xlog_iclogs_empty(log))
1134 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1135 log->l_covered_state = XLOG_STATE_COVER_DONE;
1137 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1143 spin_unlock(&log->l_icloglock);
1148 * We may be holding the log iclog lock upon entering this routine.
1151 xlog_assign_tail_lsn_locked(
1152 struct xfs_mount *mp)
1154 struct xlog *log = mp->m_log;
1155 struct xfs_log_item *lip;
1158 assert_spin_locked(&mp->m_ail->ail_lock);
1161 * To make sure we always have a valid LSN for the log tail we keep
1162 * track of the last LSN which was committed in log->l_last_sync_lsn,
1163 * and use that when the AIL was empty.
1165 lip = xfs_ail_min(mp->m_ail);
1167 tail_lsn = lip->li_lsn;
1169 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1170 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1171 atomic64_set(&log->l_tail_lsn, tail_lsn);
1176 xlog_assign_tail_lsn(
1177 struct xfs_mount *mp)
1181 spin_lock(&mp->m_ail->ail_lock);
1182 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1183 spin_unlock(&mp->m_ail->ail_lock);
1189 * Return the space in the log between the tail and the head. The head
1190 * is passed in the cycle/bytes formal parms. In the special case where
1191 * the reserve head has wrapped passed the tail, this calculation is no
1192 * longer valid. In this case, just return 0 which means there is no space
1193 * in the log. This works for all places where this function is called
1194 * with the reserve head. Of course, if the write head were to ever
1195 * wrap the tail, we should blow up. Rather than catch this case here,
1196 * we depend on other ASSERTions in other parts of the code. XXXmiken
1198 * This code also handles the case where the reservation head is behind
1199 * the tail. The details of this case are described below, but the end
1200 * result is that we return the size of the log as the amount of space left.
1213 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1214 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1215 tail_bytes = BBTOB(tail_bytes);
1216 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1217 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1218 else if (tail_cycle + 1 < head_cycle)
1220 else if (tail_cycle < head_cycle) {
1221 ASSERT(tail_cycle == (head_cycle - 1));
1222 free_bytes = tail_bytes - head_bytes;
1225 * The reservation head is behind the tail.
1226 * In this case we just want to return the size of the
1227 * log as the amount of space left.
1229 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1230 xfs_alert(log->l_mp,
1231 " tail_cycle = %d, tail_bytes = %d",
1232 tail_cycle, tail_bytes);
1233 xfs_alert(log->l_mp,
1234 " GH cycle = %d, GH bytes = %d",
1235 head_cycle, head_bytes);
1237 free_bytes = log->l_logsize;
1244 * Log function which is called when an io completes.
1246 * The log manager needs its own routine, in order to control what
1247 * happens with the buffer after the write completes.
1250 xlog_iodone(xfs_buf_t *bp)
1252 struct xlog_in_core *iclog = bp->b_log_item;
1253 struct xlog *l = iclog->ic_log;
1257 /* treat writes with injected CRC errors as failed */
1258 if (iclog->ic_fail_crc)
1263 * Race to shutdown the filesystem if we see an error.
1265 if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1266 xfs_buf_ioerror_alert(bp, __func__);
1268 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1270 * This flag will be propagated to the trans-committed
1271 * callback routines to let them know that the log-commit
1274 aborted = XFS_LI_ABORTED;
1275 } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1276 aborted = XFS_LI_ABORTED;
1279 /* log I/O is always issued ASYNC */
1280 ASSERT(bp->b_flags & XBF_ASYNC);
1281 xlog_state_done_syncing(iclog, aborted);
1284 * drop the buffer lock now that we are done. Nothing references
1285 * the buffer after this, so an unmount waiting on this lock can now
1286 * tear it down safely. As such, it is unsafe to reference the buffer
1287 * (bp) after the unlock as we could race with it being freed.
1293 * Return size of each in-core log record buffer.
1295 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1297 * If the filesystem blocksize is too large, we may need to choose a
1298 * larger size since the directory code currently logs entire blocks.
1301 xlog_get_iclog_buffer_size(
1302 struct xfs_mount *mp,
1305 if (mp->m_logbufs <= 0)
1306 mp->m_logbufs = XLOG_MAX_ICLOGS;
1307 if (mp->m_logbsize <= 0)
1308 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1310 log->l_iclog_bufs = mp->m_logbufs;
1311 log->l_iclog_size = mp->m_logbsize;
1314 * # headers = size / 32k - one header holds cycles from 32k of data.
1316 log->l_iclog_heads =
1317 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1318 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1323 struct xfs_mount *mp)
1325 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1326 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1330 * Every sync period we need to unpin all items in the AIL and push them to
1331 * disk. If there is nothing dirty, then we might need to cover the log to
1332 * indicate that the filesystem is idle.
1336 struct work_struct *work)
1338 struct xlog *log = container_of(to_delayed_work(work),
1339 struct xlog, l_work);
1340 struct xfs_mount *mp = log->l_mp;
1342 /* dgc: errors ignored - not fatal and nowhere to report them */
1343 if (xfs_log_need_covered(mp)) {
1345 * Dump a transaction into the log that contains no real change.
1346 * This is needed to stamp the current tail LSN into the log
1347 * during the covering operation.
1349 * We cannot use an inode here for this - that will push dirty
1350 * state back up into the VFS and then periodic inode flushing
1351 * will prevent log covering from making progress. Hence we
1352 * synchronously log the superblock instead to ensure the
1353 * superblock is immediately unpinned and can be written back.
1355 xfs_sync_sb(mp, true);
1357 xfs_log_force(mp, 0);
1359 /* start pushing all the metadata that is currently dirty */
1360 xfs_ail_push_all(mp->m_ail);
1362 /* queue us up again */
1363 xfs_log_work_queue(mp);
1367 * This routine initializes some of the log structure for a given mount point.
1368 * Its primary purpose is to fill in enough, so recovery can occur. However,
1369 * some other stuff may be filled in too.
1371 STATIC struct xlog *
1373 struct xfs_mount *mp,
1374 struct xfs_buftarg *log_target,
1375 xfs_daddr_t blk_offset,
1379 xlog_rec_header_t *head;
1380 xlog_in_core_t **iclogp;
1381 xlog_in_core_t *iclog, *prev_iclog=NULL;
1384 int error = -ENOMEM;
1387 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1389 xfs_warn(mp, "Log allocation failed: No memory!");
1394 log->l_targ = log_target;
1395 log->l_logsize = BBTOB(num_bblks);
1396 log->l_logBBstart = blk_offset;
1397 log->l_logBBsize = num_bblks;
1398 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1399 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1400 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1402 log->l_prev_block = -1;
1403 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1404 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1405 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1406 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1408 xlog_grant_head_init(&log->l_reserve_head);
1409 xlog_grant_head_init(&log->l_write_head);
1411 error = -EFSCORRUPTED;
1412 if (xfs_sb_version_hassector(&mp->m_sb)) {
1413 log2_size = mp->m_sb.sb_logsectlog;
1414 if (log2_size < BBSHIFT) {
1415 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1416 log2_size, BBSHIFT);
1420 log2_size -= BBSHIFT;
1421 if (log2_size > mp->m_sectbb_log) {
1422 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1423 log2_size, mp->m_sectbb_log);
1427 /* for larger sector sizes, must have v2 or external log */
1428 if (log2_size && log->l_logBBstart > 0 &&
1429 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1431 "log sector size (0x%x) invalid for configuration.",
1436 log->l_sectBBsize = 1 << log2_size;
1438 xlog_get_iclog_buffer_size(mp, log);
1441 * Use a NULL block for the extra log buffer used during splits so that
1442 * it will trigger errors if we ever try to do IO on it without first
1443 * having set it up properly.
1446 bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1447 BTOBB(log->l_iclog_size), XBF_NO_IOACCT);
1452 * The iclogbuf buffer locks are held over IO but we are not going to do
1453 * IO yet. Hence unlock the buffer so that the log IO path can grab it
1454 * when appropriately.
1456 ASSERT(xfs_buf_islocked(bp));
1459 /* use high priority wq for log I/O completion */
1460 bp->b_ioend_wq = mp->m_log_workqueue;
1461 bp->b_iodone = xlog_iodone;
1464 spin_lock_init(&log->l_icloglock);
1465 init_waitqueue_head(&log->l_flush_wait);
1467 iclogp = &log->l_iclog;
1469 * The amount of memory to allocate for the iclog structure is
1470 * rather funky due to the way the structure is defined. It is
1471 * done this way so that we can use different sizes for machines
1472 * with different amounts of memory. See the definition of
1473 * xlog_in_core_t in xfs_log_priv.h for details.
1475 ASSERT(log->l_iclog_size >= 4096);
1476 for (i=0; i < log->l_iclog_bufs; i++) {
1477 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1479 goto out_free_iclog;
1482 iclog->ic_prev = prev_iclog;
1485 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1486 BTOBB(log->l_iclog_size),
1489 goto out_free_iclog;
1491 ASSERT(xfs_buf_islocked(bp));
1494 /* use high priority wq for log I/O completion */
1495 bp->b_ioend_wq = mp->m_log_workqueue;
1496 bp->b_iodone = xlog_iodone;
1498 iclog->ic_data = bp->b_addr;
1500 log->l_iclog_bak[i] = &iclog->ic_header;
1502 head = &iclog->ic_header;
1503 memset(head, 0, sizeof(xlog_rec_header_t));
1504 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1505 head->h_version = cpu_to_be32(
1506 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1507 head->h_size = cpu_to_be32(log->l_iclog_size);
1509 head->h_fmt = cpu_to_be32(XLOG_FMT);
1510 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1512 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1513 iclog->ic_state = XLOG_STATE_ACTIVE;
1514 iclog->ic_log = log;
1515 atomic_set(&iclog->ic_refcnt, 0);
1516 spin_lock_init(&iclog->ic_callback_lock);
1517 iclog->ic_callback_tail = &(iclog->ic_callback);
1518 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1520 init_waitqueue_head(&iclog->ic_force_wait);
1521 init_waitqueue_head(&iclog->ic_write_wait);
1523 iclogp = &iclog->ic_next;
1525 *iclogp = log->l_iclog; /* complete ring */
1526 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1528 error = xlog_cil_init(log);
1530 goto out_free_iclog;
1534 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1535 prev_iclog = iclog->ic_next;
1537 xfs_buf_free(iclog->ic_bp);
1540 xfs_buf_free(log->l_xbuf);
1544 return ERR_PTR(error);
1545 } /* xlog_alloc_log */
1549 * Write out the commit record of a transaction associated with the given
1550 * ticket. Return the lsn of the commit record.
1555 struct xlog_ticket *ticket,
1556 struct xlog_in_core **iclog,
1557 xfs_lsn_t *commitlsnp)
1559 struct xfs_mount *mp = log->l_mp;
1561 struct xfs_log_iovec reg = {
1564 .i_type = XLOG_REG_TYPE_COMMIT,
1566 struct xfs_log_vec vec = {
1571 ASSERT_ALWAYS(iclog);
1572 error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1575 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1580 * Push on the buffer cache code if we ever use more than 75% of the on-disk
1581 * log space. This code pushes on the lsn which would supposedly free up
1582 * the 25% which we want to leave free. We may need to adopt a policy which
1583 * pushes on an lsn which is further along in the log once we reach the high
1584 * water mark. In this manner, we would be creating a low water mark.
1587 xlog_grant_push_ail(
1591 xfs_lsn_t threshold_lsn = 0;
1592 xfs_lsn_t last_sync_lsn;
1595 int threshold_block;
1596 int threshold_cycle;
1599 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1601 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1602 free_blocks = BTOBBT(free_bytes);
1605 * Set the threshold for the minimum number of free blocks in the
1606 * log to the maximum of what the caller needs, one quarter of the
1607 * log, and 256 blocks.
1609 free_threshold = BTOBB(need_bytes);
1610 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1611 free_threshold = max(free_threshold, 256);
1612 if (free_blocks >= free_threshold)
1615 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1617 threshold_block += free_threshold;
1618 if (threshold_block >= log->l_logBBsize) {
1619 threshold_block -= log->l_logBBsize;
1620 threshold_cycle += 1;
1622 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1625 * Don't pass in an lsn greater than the lsn of the last
1626 * log record known to be on disk. Use a snapshot of the last sync lsn
1627 * so that it doesn't change between the compare and the set.
1629 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1630 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1631 threshold_lsn = last_sync_lsn;
1634 * Get the transaction layer to kick the dirty buffers out to
1635 * disk asynchronously. No point in trying to do this if
1636 * the filesystem is shutting down.
1638 if (!XLOG_FORCED_SHUTDOWN(log))
1639 xfs_ail_push(log->l_ailp, threshold_lsn);
1643 * Stamp cycle number in every block
1648 struct xlog_in_core *iclog,
1652 int size = iclog->ic_offset + roundoff;
1656 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1658 dp = iclog->ic_datap;
1659 for (i = 0; i < BTOBB(size); i++) {
1660 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1662 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1663 *(__be32 *)dp = cycle_lsn;
1667 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1668 xlog_in_core_2_t *xhdr = iclog->ic_data;
1670 for ( ; i < BTOBB(size); i++) {
1671 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1672 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1673 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1674 *(__be32 *)dp = cycle_lsn;
1678 for (i = 1; i < log->l_iclog_heads; i++)
1679 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1684 * Calculate the checksum for a log buffer.
1686 * This is a little more complicated than it should be because the various
1687 * headers and the actual data are non-contiguous.
1692 struct xlog_rec_header *rhead,
1698 /* first generate the crc for the record header ... */
1699 crc = xfs_start_cksum_update((char *)rhead,
1700 sizeof(struct xlog_rec_header),
1701 offsetof(struct xlog_rec_header, h_crc));
1703 /* ... then for additional cycle data for v2 logs ... */
1704 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1705 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1709 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1710 if (size % XLOG_HEADER_CYCLE_SIZE)
1713 for (i = 1; i < xheads; i++) {
1714 crc = crc32c(crc, &xhdr[i].hic_xheader,
1715 sizeof(struct xlog_rec_ext_header));
1719 /* ... and finally for the payload */
1720 crc = crc32c(crc, dp, size);
1722 return xfs_end_cksum(crc);
1728 struct xlog_in_core *iclog,
1733 ASSERT(bno < log->l_logBBsize);
1734 ASSERT(bno + bp->b_io_length <= log->l_logBBsize);
1736 bp->b_maps[0].bm_bn = log->l_logBBstart + bno;
1737 bp->b_log_item = iclog;
1738 bp->b_flags &= ~XBF_FLUSH;
1739 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1741 bp->b_flags |= XBF_FLUSH;
1744 * We lock the iclogbufs here so that we can serialise against I/O
1745 * completion during unmount. We might be processing a shutdown
1746 * triggered during unmount, and that can occur asynchronously to the
1747 * unmount thread, and hence we need to ensure that completes before
1748 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1749 * across the log IO to archieve that.
1752 if (unlikely(iclog->ic_state & XLOG_STATE_IOERROR)) {
1753 xfs_buf_ioerror(bp, -EIO);
1757 * It would seem logical to return EIO here, but we rely on
1758 * the log state machine to propagate I/O errors instead of
1759 * doing it here. Similarly, IO completion will unlock the
1760 * buffer, so we don't do it here.
1769 * We need to bump cycle number for the part of the iclog that is
1770 * written to the start of the log. Watch out for the header magic
1771 * number case, though.
1780 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1783 for (i = split_offset; i < count; i += BBSIZE) {
1784 uint32_t cycle = get_unaligned_be32(data + i);
1786 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1788 put_unaligned_be32(cycle, data + i);
1791 return split_offset;
1795 xlog_calc_iclog_size(
1797 struct xlog_in_core *iclog,
1800 uint32_t count_init, count;
1803 use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1804 log->l_mp->m_sb.sb_logsunit > 1;
1806 /* Add for LR header */
1807 count_init = log->l_iclog_hsize + iclog->ic_offset;
1809 /* Round out the log write size */
1811 /* we have a v2 stripe unit to use */
1812 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1814 count = BBTOB(BTOBB(count_init));
1817 ASSERT(count >= count_init);
1818 *roundoff = count - count_init;
1821 ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1823 ASSERT(*roundoff < BBTOB(1));
1828 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1829 * fashion. Previously, we should have moved the current iclog
1830 * ptr in the log to point to the next available iclog. This allows further
1831 * write to continue while this code syncs out an iclog ready to go.
1832 * Before an in-core log can be written out, the data section must be scanned
1833 * to save away the 1st word of each BBSIZE block into the header. We replace
1834 * it with the current cycle count. Each BBSIZE block is tagged with the
1835 * cycle count because there in an implicit assumption that drives will
1836 * guarantee that entire 512 byte blocks get written at once. In other words,
1837 * we can't have part of a 512 byte block written and part not written. By
1838 * tagging each block, we will know which blocks are valid when recovering
1839 * after an unclean shutdown.
1841 * This routine is single threaded on the iclog. No other thread can be in
1842 * this routine with the same iclog. Changing contents of iclog can there-
1843 * fore be done without grabbing the state machine lock. Updating the global
1844 * log will require grabbing the lock though.
1846 * The entire log manager uses a logical block numbering scheme. Only
1847 * xlog_write_iclog knows about the fact that the log may not start with
1848 * block zero on a given device.
1853 struct xlog_in_core *iclog)
1855 unsigned int count; /* byte count of bwrite */
1856 unsigned int roundoff; /* roundoff to BB or stripe */
1858 unsigned int split = 0;
1860 bool need_flush = true;
1862 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1864 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1866 /* move grant heads by roundoff in sync */
1867 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1868 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1870 /* put cycle number in every block */
1871 xlog_pack_data(log, iclog, roundoff);
1873 /* real byte length */
1874 size = iclog->ic_offset;
1875 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1877 iclog->ic_header.h_len = cpu_to_be32(size);
1879 XFS_STATS_INC(log->l_mp, xs_log_writes);
1880 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1882 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1884 /* Do we need to split this write into 2 parts? */
1885 if (bno + BTOBB(count) > log->l_logBBsize)
1886 split = xlog_split_iclog(log, &iclog->ic_header, bno, count);
1888 /* calculcate the checksum */
1889 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1890 iclog->ic_datap, size);
1892 * Intentionally corrupt the log record CRC based on the error injection
1893 * frequency, if defined. This facilitates testing log recovery in the
1894 * event of torn writes. Hence, set the IOABORT state to abort the log
1895 * write on I/O completion and shutdown the fs. The subsequent mount
1896 * detects the bad CRC and attempts to recover.
1899 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1900 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1901 iclog->ic_fail_crc = true;
1903 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1904 be64_to_cpu(iclog->ic_header.h_lsn));
1909 * Flush the data device before flushing the log to make sure all meta
1910 * data written back from the AIL actually made it to disk before
1911 * stamping the new log tail LSN into the log buffer. For an external
1912 * log we need to issue the flush explicitly, and unfortunately
1913 * synchronously here; for an internal log we can simply use the block
1914 * layer state machine for preflushes.
1916 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp || split) {
1917 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1921 iclog->ic_bp->b_io_length = BTOBB(split ? split : count);
1922 iclog->ic_bwritecnt = split ? 2 : 1;
1924 xlog_verify_iclog(log, iclog, count, true);
1925 xlog_write_iclog(log, iclog, iclog->ic_bp, bno, need_flush);
1928 xfs_buf_associate_memory(iclog->ic_log->l_xbuf,
1929 (char *)&iclog->ic_header + split,
1931 xlog_write_iclog(log, iclog, iclog->ic_log->l_xbuf, 0, false);
1936 * Deallocate a log structure
1942 xlog_in_core_t *iclog, *next_iclog;
1945 xlog_cil_destroy(log);
1948 * Cycle all the iclogbuf locks to make sure all log IO completion
1949 * is done before we tear down these buffers.
1951 iclog = log->l_iclog;
1952 for (i = 0; i < log->l_iclog_bufs; i++) {
1953 xfs_buf_lock(iclog->ic_bp);
1954 xfs_buf_unlock(iclog->ic_bp);
1955 iclog = iclog->ic_next;
1959 * Always need to ensure that the extra buffer does not point to memory
1960 * owned by another log buffer before we free it. Also, cycle the lock
1961 * first to ensure we've completed IO on it.
1963 xfs_buf_lock(log->l_xbuf);
1964 xfs_buf_unlock(log->l_xbuf);
1965 xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1966 xfs_buf_free(log->l_xbuf);
1968 iclog = log->l_iclog;
1969 for (i = 0; i < log->l_iclog_bufs; i++) {
1970 xfs_buf_free(iclog->ic_bp);
1971 next_iclog = iclog->ic_next;
1976 log->l_mp->m_log = NULL;
1978 } /* xlog_dealloc_log */
1981 * Update counters atomically now that memcpy is done.
1985 xlog_state_finish_copy(
1987 struct xlog_in_core *iclog,
1991 spin_lock(&log->l_icloglock);
1993 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1994 iclog->ic_offset += copy_bytes;
1996 spin_unlock(&log->l_icloglock);
1997 } /* xlog_state_finish_copy */
2003 * print out info relating to regions written which consume
2008 struct xfs_mount *mp,
2009 struct xlog_ticket *ticket)
2012 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2014 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2015 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2016 static char *res_type_str[] = {
2017 REG_TYPE_STR(BFORMAT, "bformat"),
2018 REG_TYPE_STR(BCHUNK, "bchunk"),
2019 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2020 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2021 REG_TYPE_STR(IFORMAT, "iformat"),
2022 REG_TYPE_STR(ICORE, "icore"),
2023 REG_TYPE_STR(IEXT, "iext"),
2024 REG_TYPE_STR(IBROOT, "ibroot"),
2025 REG_TYPE_STR(ILOCAL, "ilocal"),
2026 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2027 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2028 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2029 REG_TYPE_STR(QFORMAT, "qformat"),
2030 REG_TYPE_STR(DQUOT, "dquot"),
2031 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2032 REG_TYPE_STR(LRHEADER, "LR header"),
2033 REG_TYPE_STR(UNMOUNT, "unmount"),
2034 REG_TYPE_STR(COMMIT, "commit"),
2035 REG_TYPE_STR(TRANSHDR, "trans header"),
2036 REG_TYPE_STR(ICREATE, "inode create"),
2037 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2038 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2039 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2040 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2041 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2042 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2044 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2047 xfs_warn(mp, "ticket reservation summary:");
2048 xfs_warn(mp, " unit res = %d bytes",
2049 ticket->t_unit_res);
2050 xfs_warn(mp, " current res = %d bytes",
2051 ticket->t_curr_res);
2052 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2053 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2054 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2055 ticket->t_res_num_ophdrs, ophdr_spc);
2056 xfs_warn(mp, " ophdr + reg = %u bytes",
2057 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2058 xfs_warn(mp, " num regions = %u",
2061 for (i = 0; i < ticket->t_res_num; i++) {
2062 uint r_type = ticket->t_res_arr[i].r_type;
2063 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2064 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2065 "bad-rtype" : res_type_str[r_type]),
2066 ticket->t_res_arr[i].r_len);
2071 * Print a summary of the transaction.
2075 struct xfs_trans *tp)
2077 struct xfs_mount *mp = tp->t_mountp;
2078 struct xfs_log_item *lip;
2080 /* dump core transaction and ticket info */
2081 xfs_warn(mp, "transaction summary:");
2082 xfs_warn(mp, " log res = %d", tp->t_log_res);
2083 xfs_warn(mp, " log count = %d", tp->t_log_count);
2084 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2086 xlog_print_tic_res(mp, tp->t_ticket);
2088 /* dump each log item */
2089 list_for_each_entry(lip, &tp->t_items, li_trans) {
2090 struct xfs_log_vec *lv = lip->li_lv;
2091 struct xfs_log_iovec *vec;
2094 xfs_warn(mp, "log item: ");
2095 xfs_warn(mp, " type = 0x%x", lip->li_type);
2096 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2099 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2100 xfs_warn(mp, " size = %d", lv->lv_size);
2101 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2102 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2104 /* dump each iovec for the log item */
2105 vec = lv->lv_iovecp;
2106 for (i = 0; i < lv->lv_niovecs; i++) {
2107 int dumplen = min(vec->i_len, 32);
2109 xfs_warn(mp, " iovec[%d]", i);
2110 xfs_warn(mp, " type = 0x%x", vec->i_type);
2111 xfs_warn(mp, " len = %d", vec->i_len);
2112 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2113 xfs_hex_dump(vec->i_addr, dumplen);
2121 * Calculate the potential space needed by the log vector. Each region gets
2122 * its own xlog_op_header_t and may need to be double word aligned.
2125 xlog_write_calc_vec_length(
2126 struct xlog_ticket *ticket,
2127 struct xfs_log_vec *log_vector)
2129 struct xfs_log_vec *lv;
2134 /* acct for start rec of xact */
2135 if (ticket->t_flags & XLOG_TIC_INITED)
2138 for (lv = log_vector; lv; lv = lv->lv_next) {
2139 /* we don't write ordered log vectors */
2140 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2143 headers += lv->lv_niovecs;
2145 for (i = 0; i < lv->lv_niovecs; i++) {
2146 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2149 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2153 ticket->t_res_num_ophdrs += headers;
2154 len += headers * sizeof(struct xlog_op_header);
2160 * If first write for transaction, insert start record We can't be trying to
2161 * commit if we are inited. We can't have any "partial_copy" if we are inited.
2164 xlog_write_start_rec(
2165 struct xlog_op_header *ophdr,
2166 struct xlog_ticket *ticket)
2168 if (!(ticket->t_flags & XLOG_TIC_INITED))
2171 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2172 ophdr->oh_clientid = ticket->t_clientid;
2174 ophdr->oh_flags = XLOG_START_TRANS;
2177 ticket->t_flags &= ~XLOG_TIC_INITED;
2179 return sizeof(struct xlog_op_header);
2182 static xlog_op_header_t *
2183 xlog_write_setup_ophdr(
2185 struct xlog_op_header *ophdr,
2186 struct xlog_ticket *ticket,
2189 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2190 ophdr->oh_clientid = ticket->t_clientid;
2193 /* are we copying a commit or unmount record? */
2194 ophdr->oh_flags = flags;
2197 * We've seen logs corrupted with bad transaction client ids. This
2198 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2199 * and shut down the filesystem.
2201 switch (ophdr->oh_clientid) {
2202 case XFS_TRANSACTION:
2208 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2209 ophdr->oh_clientid, ticket);
2217 * Set up the parameters of the region copy into the log. This has
2218 * to handle region write split across multiple log buffers - this
2219 * state is kept external to this function so that this code can
2220 * be written in an obvious, self documenting manner.
2223 xlog_write_setup_copy(
2224 struct xlog_ticket *ticket,
2225 struct xlog_op_header *ophdr,
2226 int space_available,
2230 int *last_was_partial_copy,
2231 int *bytes_consumed)
2235 still_to_copy = space_required - *bytes_consumed;
2236 *copy_off = *bytes_consumed;
2238 if (still_to_copy <= space_available) {
2239 /* write of region completes here */
2240 *copy_len = still_to_copy;
2241 ophdr->oh_len = cpu_to_be32(*copy_len);
2242 if (*last_was_partial_copy)
2243 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2244 *last_was_partial_copy = 0;
2245 *bytes_consumed = 0;
2249 /* partial write of region, needs extra log op header reservation */
2250 *copy_len = space_available;
2251 ophdr->oh_len = cpu_to_be32(*copy_len);
2252 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2253 if (*last_was_partial_copy)
2254 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2255 *bytes_consumed += *copy_len;
2256 (*last_was_partial_copy)++;
2258 /* account for new log op header */
2259 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2260 ticket->t_res_num_ophdrs++;
2262 return sizeof(struct xlog_op_header);
2266 xlog_write_copy_finish(
2268 struct xlog_in_core *iclog,
2273 int *partial_copy_len,
2275 struct xlog_in_core **commit_iclog)
2277 if (*partial_copy) {
2279 * This iclog has already been marked WANT_SYNC by
2280 * xlog_state_get_iclog_space.
2282 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2285 return xlog_state_release_iclog(log, iclog);
2289 *partial_copy_len = 0;
2291 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2292 /* no more space in this iclog - push it. */
2293 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2297 spin_lock(&log->l_icloglock);
2298 xlog_state_want_sync(log, iclog);
2299 spin_unlock(&log->l_icloglock);
2302 return xlog_state_release_iclog(log, iclog);
2303 ASSERT(flags & XLOG_COMMIT_TRANS);
2304 *commit_iclog = iclog;
2311 * Write some region out to in-core log
2313 * This will be called when writing externally provided regions or when
2314 * writing out a commit record for a given transaction.
2316 * General algorithm:
2317 * 1. Find total length of this write. This may include adding to the
2318 * lengths passed in.
2319 * 2. Check whether we violate the tickets reservation.
2320 * 3. While writing to this iclog
2321 * A. Reserve as much space in this iclog as can get
2322 * B. If this is first write, save away start lsn
2323 * C. While writing this region:
2324 * 1. If first write of transaction, write start record
2325 * 2. Write log operation header (header per region)
2326 * 3. Find out if we can fit entire region into this iclog
2327 * 4. Potentially, verify destination memcpy ptr
2328 * 5. Memcpy (partial) region
2329 * 6. If partial copy, release iclog; otherwise, continue
2330 * copying more regions into current iclog
2331 * 4. Mark want sync bit (in simulation mode)
2332 * 5. Release iclog for potential flush to on-disk log.
2335 * 1. Panic if reservation is overrun. This should never happen since
2336 * reservation amounts are generated internal to the filesystem.
2338 * 1. Tickets are single threaded data structures.
2339 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2340 * syncing routine. When a single log_write region needs to span
2341 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2342 * on all log operation writes which don't contain the end of the
2343 * region. The XLOG_END_TRANS bit is used for the in-core log
2344 * operation which contains the end of the continued log_write region.
2345 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2346 * we don't really know exactly how much space will be used. As a result,
2347 * we don't update ic_offset until the end when we know exactly how many
2348 * bytes have been written out.
2353 struct xfs_log_vec *log_vector,
2354 struct xlog_ticket *ticket,
2355 xfs_lsn_t *start_lsn,
2356 struct xlog_in_core **commit_iclog,
2359 struct xlog_in_core *iclog = NULL;
2360 struct xfs_log_iovec *vecp;
2361 struct xfs_log_vec *lv;
2364 int partial_copy = 0;
2365 int partial_copy_len = 0;
2373 len = xlog_write_calc_vec_length(ticket, log_vector);
2376 * Region headers and bytes are already accounted for.
2377 * We only need to take into account start records and
2378 * split regions in this function.
2380 if (ticket->t_flags & XLOG_TIC_INITED)
2381 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2384 * Commit record headers need to be accounted for. These
2385 * come in as separate writes so are easy to detect.
2387 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2388 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2390 if (ticket->t_curr_res < 0) {
2391 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2392 "ctx ticket reservation ran out. Need to up reservation");
2393 xlog_print_tic_res(log->l_mp, ticket);
2394 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2399 vecp = lv->lv_iovecp;
2400 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2404 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2405 &contwr, &log_offset);
2409 ASSERT(log_offset <= iclog->ic_size - 1);
2410 ptr = iclog->ic_datap + log_offset;
2412 /* start_lsn is the first lsn written to. That's all we need. */
2414 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2417 * This loop writes out as many regions as can fit in the amount
2418 * of space which was allocated by xlog_state_get_iclog_space().
2420 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2421 struct xfs_log_iovec *reg;
2422 struct xlog_op_header *ophdr;
2426 bool ordered = false;
2428 /* ordered log vectors have no regions to write */
2429 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2430 ASSERT(lv->lv_niovecs == 0);
2436 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2437 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2439 start_rec_copy = xlog_write_start_rec(ptr, ticket);
2440 if (start_rec_copy) {
2442 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2446 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2450 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2451 sizeof(struct xlog_op_header));
2453 len += xlog_write_setup_copy(ticket, ophdr,
2454 iclog->ic_size-log_offset,
2456 ©_off, ©_len,
2459 xlog_verify_dest_ptr(log, ptr);
2464 * Unmount records just log an opheader, so can have
2465 * empty payloads with no data region to copy. Hence we
2466 * only copy the payload if the vector says it has data
2469 ASSERT(copy_len >= 0);
2471 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2472 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2475 copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2477 data_cnt += contwr ? copy_len : 0;
2479 error = xlog_write_copy_finish(log, iclog, flags,
2480 &record_cnt, &data_cnt,
2489 * if we had a partial copy, we need to get more iclog
2490 * space but we don't want to increment the region
2491 * index because there is still more is this region to
2494 * If we completed writing this region, and we flushed
2495 * the iclog (indicated by resetting of the record
2496 * count), then we also need to get more log space. If
2497 * this was the last record, though, we are done and
2503 if (++index == lv->lv_niovecs) {
2508 vecp = lv->lv_iovecp;
2510 if (record_cnt == 0 && !ordered) {
2520 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2522 return xlog_state_release_iclog(log, iclog);
2524 ASSERT(flags & XLOG_COMMIT_TRANS);
2525 *commit_iclog = iclog;
2530 /*****************************************************************************
2532 * State Machine functions
2534 *****************************************************************************
2537 /* Clean iclogs starting from the head. This ordering must be
2538 * maintained, so an iclog doesn't become ACTIVE beyond one that
2539 * is SYNCING. This is also required to maintain the notion that we use
2540 * a ordered wait queue to hold off would be writers to the log when every
2541 * iclog is trying to sync to disk.
2543 * State Change: DIRTY -> ACTIVE
2546 xlog_state_clean_log(
2549 xlog_in_core_t *iclog;
2552 iclog = log->l_iclog;
2554 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2555 iclog->ic_state = XLOG_STATE_ACTIVE;
2556 iclog->ic_offset = 0;
2557 ASSERT(iclog->ic_callback == NULL);
2559 * If the number of ops in this iclog indicate it just
2560 * contains the dummy transaction, we can
2561 * change state into IDLE (the second time around).
2562 * Otherwise we should change the state into
2564 * We don't need to cover the dummy.
2567 (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2572 * We have two dirty iclogs so start over
2573 * This could also be num of ops indicates
2574 * this is not the dummy going out.
2578 iclog->ic_header.h_num_logops = 0;
2579 memset(iclog->ic_header.h_cycle_data, 0,
2580 sizeof(iclog->ic_header.h_cycle_data));
2581 iclog->ic_header.h_lsn = 0;
2582 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2585 break; /* stop cleaning */
2586 iclog = iclog->ic_next;
2587 } while (iclog != log->l_iclog);
2589 /* log is locked when we are called */
2591 * Change state for the dummy log recording.
2592 * We usually go to NEED. But we go to NEED2 if the changed indicates
2593 * we are done writing the dummy record.
2594 * If we are done with the second dummy recored (DONE2), then
2598 switch (log->l_covered_state) {
2599 case XLOG_STATE_COVER_IDLE:
2600 case XLOG_STATE_COVER_NEED:
2601 case XLOG_STATE_COVER_NEED2:
2602 log->l_covered_state = XLOG_STATE_COVER_NEED;
2605 case XLOG_STATE_COVER_DONE:
2607 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2609 log->l_covered_state = XLOG_STATE_COVER_NEED;
2612 case XLOG_STATE_COVER_DONE2:
2614 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2616 log->l_covered_state = XLOG_STATE_COVER_NEED;
2623 } /* xlog_state_clean_log */
2626 xlog_get_lowest_lsn(
2629 struct xlog_in_core *iclog = log->l_iclog;
2630 xfs_lsn_t lowest_lsn = 0, lsn;
2633 if (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))
2636 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2637 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2639 } while ((iclog = iclog->ic_next) != log->l_iclog);
2645 xlog_state_do_callback(
2648 struct xlog_in_core *ciclog)
2650 xlog_in_core_t *iclog;
2651 xlog_in_core_t *first_iclog; /* used to know when we've
2652 * processed all iclogs once */
2653 xfs_log_callback_t *cb, *cb_next;
2655 xfs_lsn_t lowest_lsn;
2656 int ioerrors; /* counter: iclogs with errors */
2657 int loopdidcallbacks; /* flag: inner loop did callbacks*/
2658 int funcdidcallbacks; /* flag: function did callbacks */
2659 int repeats; /* for issuing console warnings if
2660 * looping too many times */
2663 spin_lock(&log->l_icloglock);
2664 first_iclog = iclog = log->l_iclog;
2666 funcdidcallbacks = 0;
2671 * Scan all iclogs starting with the one pointed to by the
2672 * log. Reset this starting point each time the log is
2673 * unlocked (during callbacks).
2675 * Keep looping through iclogs until one full pass is made
2676 * without running any callbacks.
2678 first_iclog = log->l_iclog;
2679 iclog = log->l_iclog;
2680 loopdidcallbacks = 0;
2685 /* skip all iclogs in the ACTIVE & DIRTY states */
2686 if (iclog->ic_state &
2687 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2688 iclog = iclog->ic_next;
2693 * Between marking a filesystem SHUTDOWN and stopping
2694 * the log, we do flush all iclogs to disk (if there
2695 * wasn't a log I/O error). So, we do want things to
2696 * go smoothly in case of just a SHUTDOWN w/o a
2699 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2701 * Can only perform callbacks in order. Since
2702 * this iclog is not in the DONE_SYNC/
2703 * DO_CALLBACK state, we skip the rest and
2704 * just try to clean up. If we set our iclog
2705 * to DO_CALLBACK, we will not process it when
2706 * we retry since a previous iclog is in the
2707 * CALLBACK and the state cannot change since
2708 * we are holding the l_icloglock.
2710 if (!(iclog->ic_state &
2711 (XLOG_STATE_DONE_SYNC |
2712 XLOG_STATE_DO_CALLBACK))) {
2713 if (ciclog && (ciclog->ic_state ==
2714 XLOG_STATE_DONE_SYNC)) {
2715 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2720 * We now have an iclog that is in either the
2721 * DO_CALLBACK or DONE_SYNC states. The other
2722 * states (WANT_SYNC, SYNCING, or CALLBACK were
2723 * caught by the above if and are going to
2724 * clean (i.e. we aren't doing their callbacks)
2729 * We will do one more check here to see if we
2730 * have chased our tail around.
2733 lowest_lsn = xlog_get_lowest_lsn(log);
2735 XFS_LSN_CMP(lowest_lsn,
2736 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2737 iclog = iclog->ic_next;
2738 continue; /* Leave this iclog for
2742 iclog->ic_state = XLOG_STATE_CALLBACK;
2746 * Completion of a iclog IO does not imply that
2747 * a transaction has completed, as transactions
2748 * can be large enough to span many iclogs. We
2749 * cannot change the tail of the log half way
2750 * through a transaction as this may be the only
2751 * transaction in the log and moving th etail to
2752 * point to the middle of it will prevent
2753 * recovery from finding the start of the
2754 * transaction. Hence we should only update the
2755 * last_sync_lsn if this iclog contains
2756 * transaction completion callbacks on it.
2758 * We have to do this before we drop the
2759 * icloglock to ensure we are the only one that
2762 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2763 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2764 if (iclog->ic_callback)
2765 atomic64_set(&log->l_last_sync_lsn,
2766 be64_to_cpu(iclog->ic_header.h_lsn));
2771 spin_unlock(&log->l_icloglock);
2774 * Keep processing entries in the callback list until
2775 * we come around and it is empty. We need to
2776 * atomically see that the list is empty and change the
2777 * state to DIRTY so that we don't miss any more
2778 * callbacks being added.
2780 spin_lock(&iclog->ic_callback_lock);
2781 cb = iclog->ic_callback;
2783 iclog->ic_callback_tail = &(iclog->ic_callback);
2784 iclog->ic_callback = NULL;
2785 spin_unlock(&iclog->ic_callback_lock);
2787 /* perform callbacks in the order given */
2788 for (; cb; cb = cb_next) {
2789 cb_next = cb->cb_next;
2790 cb->cb_func(cb->cb_arg, aborted);
2792 spin_lock(&iclog->ic_callback_lock);
2793 cb = iclog->ic_callback;
2799 spin_lock(&log->l_icloglock);
2800 ASSERT(iclog->ic_callback == NULL);
2801 spin_unlock(&iclog->ic_callback_lock);
2802 if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2803 iclog->ic_state = XLOG_STATE_DIRTY;
2806 * Transition from DIRTY to ACTIVE if applicable.
2807 * NOP if STATE_IOERROR.
2809 xlog_state_clean_log(log);
2811 /* wake up threads waiting in xfs_log_force() */
2812 wake_up_all(&iclog->ic_force_wait);
2814 iclog = iclog->ic_next;
2815 } while (first_iclog != iclog);
2817 if (repeats > 5000) {
2818 flushcnt += repeats;
2821 "%s: possible infinite loop (%d iterations)",
2822 __func__, flushcnt);
2824 } while (!ioerrors && loopdidcallbacks);
2828 * Make one last gasp attempt to see if iclogs are being left in limbo.
2829 * If the above loop finds an iclog earlier than the current iclog and
2830 * in one of the syncing states, the current iclog is put into
2831 * DO_CALLBACK and the callbacks are deferred to the completion of the
2832 * earlier iclog. Walk the iclogs in order and make sure that no iclog
2833 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2836 * Note that SYNCING|IOABORT is a valid state so we cannot just check
2837 * for ic_state == SYNCING.
2839 if (funcdidcallbacks) {
2840 first_iclog = iclog = log->l_iclog;
2842 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2844 * Terminate the loop if iclogs are found in states
2845 * which will cause other threads to clean up iclogs.
2847 * SYNCING - i/o completion will go through logs
2848 * DONE_SYNC - interrupt thread should be waiting for
2850 * IOERROR - give up hope all ye who enter here
2852 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2853 iclog->ic_state & XLOG_STATE_SYNCING ||
2854 iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2855 iclog->ic_state == XLOG_STATE_IOERROR )
2857 iclog = iclog->ic_next;
2858 } while (first_iclog != iclog);
2862 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2864 spin_unlock(&log->l_icloglock);
2867 wake_up_all(&log->l_flush_wait);
2872 * Finish transitioning this iclog to the dirty state.
2874 * Make sure that we completely execute this routine only when this is
2875 * the last call to the iclog. There is a good chance that iclog flushes,
2876 * when we reach the end of the physical log, get turned into 2 separate
2877 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2878 * routine. By using the reference count bwritecnt, we guarantee that only
2879 * the second completion goes through.
2881 * Callbacks could take time, so they are done outside the scope of the
2882 * global state machine log lock.
2885 xlog_state_done_syncing(
2886 xlog_in_core_t *iclog,
2889 struct xlog *log = iclog->ic_log;
2891 spin_lock(&log->l_icloglock);
2893 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2894 iclog->ic_state == XLOG_STATE_IOERROR);
2895 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2896 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2900 * If we got an error, either on the first buffer, or in the case of
2901 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2902 * and none should ever be attempted to be written to disk
2905 if (iclog->ic_state != XLOG_STATE_IOERROR) {
2906 if (--iclog->ic_bwritecnt == 1) {
2907 spin_unlock(&log->l_icloglock);
2910 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2914 * Someone could be sleeping prior to writing out the next
2915 * iclog buffer, we wake them all, one will get to do the
2916 * I/O, the others get to wait for the result.
2918 wake_up_all(&iclog->ic_write_wait);
2919 spin_unlock(&log->l_icloglock);
2920 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */
2921 } /* xlog_state_done_syncing */
2925 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2926 * sleep. We wait on the flush queue on the head iclog as that should be
2927 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2928 * we will wait here and all new writes will sleep until a sync completes.
2930 * The in-core logs are used in a circular fashion. They are not used
2931 * out-of-order even when an iclog past the head is free.
2934 * * log_offset where xlog_write() can start writing into the in-core
2936 * * in-core log pointer to which xlog_write() should write.
2937 * * boolean indicating this is a continued write to an in-core log.
2938 * If this is the last write, then the in-core log's offset field
2939 * needs to be incremented, depending on the amount of data which
2943 xlog_state_get_iclog_space(
2946 struct xlog_in_core **iclogp,
2947 struct xlog_ticket *ticket,
2948 int *continued_write,
2952 xlog_rec_header_t *head;
2953 xlog_in_core_t *iclog;
2957 spin_lock(&log->l_icloglock);
2958 if (XLOG_FORCED_SHUTDOWN(log)) {
2959 spin_unlock(&log->l_icloglock);
2963 iclog = log->l_iclog;
2964 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2965 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2967 /* Wait for log writes to have flushed */
2968 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2972 head = &iclog->ic_header;
2974 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2975 log_offset = iclog->ic_offset;
2977 /* On the 1st write to an iclog, figure out lsn. This works
2978 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2979 * committing to. If the offset is set, that's how many blocks
2982 if (log_offset == 0) {
2983 ticket->t_curr_res -= log->l_iclog_hsize;
2984 xlog_tic_add_region(ticket,
2986 XLOG_REG_TYPE_LRHEADER);
2987 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2988 head->h_lsn = cpu_to_be64(
2989 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2990 ASSERT(log->l_curr_block >= 0);
2993 /* If there is enough room to write everything, then do it. Otherwise,
2994 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2995 * bit is on, so this will get flushed out. Don't update ic_offset
2996 * until you know exactly how many bytes get copied. Therefore, wait
2997 * until later to update ic_offset.
2999 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3000 * can fit into remaining data section.
3002 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3003 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3006 * If I'm the only one writing to this iclog, sync it to disk.
3007 * We need to do an atomic compare and decrement here to avoid
3008 * racing with concurrent atomic_dec_and_lock() calls in
3009 * xlog_state_release_iclog() when there is more than one
3010 * reference to the iclog.
3012 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3013 /* we are the only one */
3014 spin_unlock(&log->l_icloglock);
3015 error = xlog_state_release_iclog(log, iclog);
3019 spin_unlock(&log->l_icloglock);
3024 /* Do we have enough room to write the full amount in the remainder
3025 * of this iclog? Or must we continue a write on the next iclog and
3026 * mark this iclog as completely taken? In the case where we switch
3027 * iclogs (to mark it taken), this particular iclog will release/sync
3028 * to disk in xlog_write().
3030 if (len <= iclog->ic_size - iclog->ic_offset) {
3031 *continued_write = 0;
3032 iclog->ic_offset += len;
3034 *continued_write = 1;
3035 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3039 ASSERT(iclog->ic_offset <= iclog->ic_size);
3040 spin_unlock(&log->l_icloglock);
3042 *logoffsetp = log_offset;
3044 } /* xlog_state_get_iclog_space */
3046 /* The first cnt-1 times through here we don't need to
3047 * move the grant write head because the permanent
3048 * reservation has reserved cnt times the unit amount.
3049 * Release part of current permanent unit reservation and
3050 * reset current reservation to be one units worth. Also
3051 * move grant reservation head forward.
3054 xlog_regrant_reserve_log_space(
3056 struct xlog_ticket *ticket)
3058 trace_xfs_log_regrant_reserve_enter(log, ticket);
3060 if (ticket->t_cnt > 0)
3063 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3064 ticket->t_curr_res);
3065 xlog_grant_sub_space(log, &log->l_write_head.grant,
3066 ticket->t_curr_res);
3067 ticket->t_curr_res = ticket->t_unit_res;
3068 xlog_tic_reset_res(ticket);
3070 trace_xfs_log_regrant_reserve_sub(log, ticket);
3072 /* just return if we still have some of the pre-reserved space */
3073 if (ticket->t_cnt > 0)
3076 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3077 ticket->t_unit_res);
3079 trace_xfs_log_regrant_reserve_exit(log, ticket);
3081 ticket->t_curr_res = ticket->t_unit_res;
3082 xlog_tic_reset_res(ticket);
3083 } /* xlog_regrant_reserve_log_space */
3087 * Give back the space left from a reservation.
3089 * All the information we need to make a correct determination of space left
3090 * is present. For non-permanent reservations, things are quite easy. The
3091 * count should have been decremented to zero. We only need to deal with the
3092 * space remaining in the current reservation part of the ticket. If the
3093 * ticket contains a permanent reservation, there may be left over space which
3094 * needs to be released. A count of N means that N-1 refills of the current
3095 * reservation can be done before we need to ask for more space. The first
3096 * one goes to fill up the first current reservation. Once we run out of
3097 * space, the count will stay at zero and the only space remaining will be
3098 * in the current reservation field.
3101 xlog_ungrant_log_space(
3103 struct xlog_ticket *ticket)
3107 if (ticket->t_cnt > 0)
3110 trace_xfs_log_ungrant_enter(log, ticket);
3111 trace_xfs_log_ungrant_sub(log, ticket);
3114 * If this is a permanent reservation ticket, we may be able to free
3115 * up more space based on the remaining count.
3117 bytes = ticket->t_curr_res;
3118 if (ticket->t_cnt > 0) {
3119 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3120 bytes += ticket->t_unit_res*ticket->t_cnt;
3123 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3124 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3126 trace_xfs_log_ungrant_exit(log, ticket);
3128 xfs_log_space_wake(log->l_mp);
3132 * Flush iclog to disk if this is the last reference to the given iclog and
3133 * the WANT_SYNC bit is set.
3135 * When this function is entered, the iclog is not necessarily in the
3136 * WANT_SYNC state. It may be sitting around waiting to get filled.
3141 xlog_state_release_iclog(
3143 struct xlog_in_core *iclog)
3145 int sync = 0; /* do we sync? */
3147 if (iclog->ic_state & XLOG_STATE_IOERROR)
3150 ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3151 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3154 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3155 spin_unlock(&log->l_icloglock);
3158 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3159 iclog->ic_state == XLOG_STATE_WANT_SYNC);
3161 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3162 /* update tail before writing to iclog */
3163 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3165 iclog->ic_state = XLOG_STATE_SYNCING;
3166 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3167 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3168 /* cycle incremented when incrementing curr_block */
3170 spin_unlock(&log->l_icloglock);
3173 * We let the log lock go, so it's possible that we hit a log I/O
3174 * error or some other SHUTDOWN condition that marks the iclog
3175 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3176 * this iclog has consistent data, so we ignore IOERROR
3177 * flags after this point.
3180 xlog_sync(log, iclog);
3182 } /* xlog_state_release_iclog */
3186 * This routine will mark the current iclog in the ring as WANT_SYNC
3187 * and move the current iclog pointer to the next iclog in the ring.
3188 * When this routine is called from xlog_state_get_iclog_space(), the
3189 * exact size of the iclog has not yet been determined. All we know is
3190 * that every data block. We have run out of space in this log record.
3193 xlog_state_switch_iclogs(
3195 struct xlog_in_core *iclog,
3198 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3200 eventual_size = iclog->ic_offset;
3201 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3202 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3203 log->l_prev_block = log->l_curr_block;
3204 log->l_prev_cycle = log->l_curr_cycle;
3206 /* roll log?: ic_offset changed later */
3207 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3209 /* Round up to next log-sunit */
3210 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3211 log->l_mp->m_sb.sb_logsunit > 1) {
3212 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3213 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3216 if (log->l_curr_block >= log->l_logBBsize) {
3218 * Rewind the current block before the cycle is bumped to make
3219 * sure that the combined LSN never transiently moves forward
3220 * when the log wraps to the next cycle. This is to support the
3221 * unlocked sample of these fields from xlog_valid_lsn(). Most
3222 * other cases should acquire l_icloglock.
3224 log->l_curr_block -= log->l_logBBsize;
3225 ASSERT(log->l_curr_block >= 0);
3227 log->l_curr_cycle++;
3228 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3229 log->l_curr_cycle++;
3231 ASSERT(iclog == log->l_iclog);
3232 log->l_iclog = iclog->ic_next;
3233 } /* xlog_state_switch_iclogs */
3236 * Write out all data in the in-core log as of this exact moment in time.
3238 * Data may be written to the in-core log during this call. However,
3239 * we don't guarantee this data will be written out. A change from past
3240 * implementation means this routine will *not* write out zero length LRs.
3242 * Basically, we try and perform an intelligent scan of the in-core logs.
3243 * If we determine there is no flushable data, we just return. There is no
3244 * flushable data if:
3246 * 1. the current iclog is active and has no data; the previous iclog
3247 * is in the active or dirty state.
3248 * 2. the current iclog is drity, and the previous iclog is in the
3249 * active or dirty state.
3253 * 1. the current iclog is not in the active nor dirty state.
3254 * 2. the current iclog dirty, and the previous iclog is not in the
3255 * active nor dirty state.
3256 * 3. the current iclog is active, and there is another thread writing
3257 * to this particular iclog.
3258 * 4. a) the current iclog is active and has no other writers
3259 * b) when we return from flushing out this iclog, it is still
3260 * not in the active nor dirty state.
3264 struct xfs_mount *mp,
3267 struct xlog *log = mp->m_log;
3268 struct xlog_in_core *iclog;
3271 XFS_STATS_INC(mp, xs_log_force);
3272 trace_xfs_log_force(mp, 0, _RET_IP_);
3274 xlog_cil_force(log);
3276 spin_lock(&log->l_icloglock);
3277 iclog = log->l_iclog;
3278 if (iclog->ic_state & XLOG_STATE_IOERROR)
3281 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3282 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3283 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3285 * If the head is dirty or (active and empty), then we need to
3286 * look at the previous iclog.
3288 * If the previous iclog is active or dirty we are done. There
3289 * is nothing to sync out. Otherwise, we attach ourselves to the
3290 * previous iclog and go to sleep.
3292 iclog = iclog->ic_prev;
3293 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3294 iclog->ic_state == XLOG_STATE_DIRTY)
3296 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3297 if (atomic_read(&iclog->ic_refcnt) == 0) {
3299 * We are the only one with access to this iclog.
3301 * Flush it out now. There should be a roundoff of zero
3302 * to show that someone has already taken care of the
3303 * roundoff from the previous sync.
3305 atomic_inc(&iclog->ic_refcnt);
3306 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3307 xlog_state_switch_iclogs(log, iclog, 0);
3308 spin_unlock(&log->l_icloglock);
3310 if (xlog_state_release_iclog(log, iclog))
3313 spin_lock(&log->l_icloglock);
3314 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3315 iclog->ic_state == XLOG_STATE_DIRTY)
3319 * Someone else is writing to this iclog.
3321 * Use its call to flush out the data. However, the
3322 * other thread may not force out this LR, so we mark
3325 xlog_state_switch_iclogs(log, iclog, 0);
3329 * If the head iclog is not active nor dirty, we just attach
3330 * ourselves to the head and go to sleep if necessary.
3335 if (!(flags & XFS_LOG_SYNC))
3338 if (iclog->ic_state & XLOG_STATE_IOERROR)
3340 XFS_STATS_INC(mp, xs_log_force_sleep);
3341 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3342 if (iclog->ic_state & XLOG_STATE_IOERROR)
3347 spin_unlock(&log->l_icloglock);
3350 spin_unlock(&log->l_icloglock);
3355 __xfs_log_force_lsn(
3356 struct xfs_mount *mp,
3362 struct xlog *log = mp->m_log;
3363 struct xlog_in_core *iclog;
3365 spin_lock(&log->l_icloglock);
3366 iclog = log->l_iclog;
3367 if (iclog->ic_state & XLOG_STATE_IOERROR)
3370 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3371 iclog = iclog->ic_next;
3372 if (iclog == log->l_iclog)
3376 if (iclog->ic_state == XLOG_STATE_DIRTY)
3379 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3381 * We sleep here if we haven't already slept (e.g. this is the
3382 * first time we've looked at the correct iclog buf) and the
3383 * buffer before us is going to be sync'ed. The reason for this
3384 * is that if we are doing sync transactions here, by waiting
3385 * for the previous I/O to complete, we can allow a few more
3386 * transactions into this iclog before we close it down.
3388 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3389 * refcnt so we can release the log (which drops the ref count).
3390 * The state switch keeps new transaction commits from using
3391 * this buffer. When the current commits finish writing into
3392 * the buffer, the refcount will drop to zero and the buffer
3395 if (!already_slept &&
3396 (iclog->ic_prev->ic_state &
3397 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3398 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3400 XFS_STATS_INC(mp, xs_log_force_sleep);
3402 xlog_wait(&iclog->ic_prev->ic_write_wait,
3406 atomic_inc(&iclog->ic_refcnt);
3407 xlog_state_switch_iclogs(log, iclog, 0);
3408 spin_unlock(&log->l_icloglock);
3409 if (xlog_state_release_iclog(log, iclog))
3413 spin_lock(&log->l_icloglock);
3416 if (!(flags & XFS_LOG_SYNC) ||
3417 (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3420 if (iclog->ic_state & XLOG_STATE_IOERROR)
3423 XFS_STATS_INC(mp, xs_log_force_sleep);
3424 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3425 if (iclog->ic_state & XLOG_STATE_IOERROR)
3430 spin_unlock(&log->l_icloglock);
3433 spin_unlock(&log->l_icloglock);
3438 * Force the in-core log to disk for a specific LSN.
3440 * Find in-core log with lsn.
3441 * If it is in the DIRTY state, just return.
3442 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3443 * state and go to sleep or return.
3444 * If it is in any other state, go to sleep or return.
3446 * Synchronous forces are implemented with a wait queue. All callers trying
3447 * to force a given lsn to disk must wait on the queue attached to the
3448 * specific in-core log. When given in-core log finally completes its write
3449 * to disk, that thread will wake up all threads waiting on the queue.
3453 struct xfs_mount *mp,
3461 XFS_STATS_INC(mp, xs_log_force);
3462 trace_xfs_log_force(mp, lsn, _RET_IP_);
3464 lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3465 if (lsn == NULLCOMMITLSN)
3468 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3470 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3475 * Called when we want to mark the current iclog as being ready to sync to
3479 xlog_state_want_sync(
3481 struct xlog_in_core *iclog)
3483 assert_spin_locked(&log->l_icloglock);
3485 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3486 xlog_state_switch_iclogs(log, iclog, 0);
3488 ASSERT(iclog->ic_state &
3489 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3494 /*****************************************************************************
3498 *****************************************************************************
3502 * Free a used ticket when its refcount falls to zero.
3506 xlog_ticket_t *ticket)
3508 ASSERT(atomic_read(&ticket->t_ref) > 0);
3509 if (atomic_dec_and_test(&ticket->t_ref))
3510 kmem_zone_free(xfs_log_ticket_zone, ticket);
3515 xlog_ticket_t *ticket)
3517 ASSERT(atomic_read(&ticket->t_ref) > 0);
3518 atomic_inc(&ticket->t_ref);
3523 * Figure out the total log space unit (in bytes) that would be
3524 * required for a log ticket.
3527 xfs_log_calc_unit_res(
3528 struct xfs_mount *mp,
3531 struct xlog *log = mp->m_log;
3536 * Permanent reservations have up to 'cnt'-1 active log operations
3537 * in the log. A unit in this case is the amount of space for one
3538 * of these log operations. Normal reservations have a cnt of 1
3539 * and their unit amount is the total amount of space required.
3541 * The following lines of code account for non-transaction data
3542 * which occupy space in the on-disk log.
3544 * Normal form of a transaction is:
3545 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3546 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3548 * We need to account for all the leadup data and trailer data
3549 * around the transaction data.
3550 * And then we need to account for the worst case in terms of using
3552 * The worst case will happen if:
3553 * - the placement of the transaction happens to be such that the
3554 * roundoff is at its maximum
3555 * - the transaction data is synced before the commit record is synced
3556 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3557 * Therefore the commit record is in its own Log Record.
3558 * This can happen as the commit record is called with its
3559 * own region to xlog_write().
3560 * This then means that in the worst case, roundoff can happen for
3561 * the commit-rec as well.
3562 * The commit-rec is smaller than padding in this scenario and so it is
3563 * not added separately.
3566 /* for trans header */
3567 unit_bytes += sizeof(xlog_op_header_t);
3568 unit_bytes += sizeof(xfs_trans_header_t);
3571 unit_bytes += sizeof(xlog_op_header_t);
3574 * for LR headers - the space for data in an iclog is the size minus
3575 * the space used for the headers. If we use the iclog size, then we
3576 * undercalculate the number of headers required.
3578 * Furthermore - the addition of op headers for split-recs might
3579 * increase the space required enough to require more log and op
3580 * headers, so take that into account too.
3582 * IMPORTANT: This reservation makes the assumption that if this
3583 * transaction is the first in an iclog and hence has the LR headers
3584 * accounted to it, then the remaining space in the iclog is
3585 * exclusively for this transaction. i.e. if the transaction is larger
3586 * than the iclog, it will be the only thing in that iclog.
3587 * Fundamentally, this means we must pass the entire log vector to
3588 * xlog_write to guarantee this.
3590 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3591 num_headers = howmany(unit_bytes, iclog_space);
3593 /* for split-recs - ophdrs added when data split over LRs */
3594 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3596 /* add extra header reservations if we overrun */
3597 while (!num_headers ||
3598 howmany(unit_bytes, iclog_space) > num_headers) {
3599 unit_bytes += sizeof(xlog_op_header_t);
3602 unit_bytes += log->l_iclog_hsize * num_headers;
3604 /* for commit-rec LR header - note: padding will subsume the ophdr */
3605 unit_bytes += log->l_iclog_hsize;
3607 /* for roundoff padding for transaction data and one for commit record */
3608 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3609 /* log su roundoff */
3610 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3613 unit_bytes += 2 * BBSIZE;
3620 * Allocate and initialise a new log ticket.
3622 struct xlog_ticket *
3629 xfs_km_flags_t alloc_flags)
3631 struct xlog_ticket *tic;
3634 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3638 unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3640 atomic_set(&tic->t_ref, 1);
3641 tic->t_task = current;
3642 INIT_LIST_HEAD(&tic->t_queue);
3643 tic->t_unit_res = unit_res;
3644 tic->t_curr_res = unit_res;
3647 tic->t_tid = prandom_u32();
3648 tic->t_clientid = client;
3649 tic->t_flags = XLOG_TIC_INITED;
3651 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3653 xlog_tic_reset_res(tic);
3659 /******************************************************************************
3661 * Log debug routines
3663 ******************************************************************************
3667 * Make sure that the destination ptr is within the valid data region of
3668 * one of the iclogs. This uses backup pointers stored in a different
3669 * part of the log in case we trash the log structure.
3672 xlog_verify_dest_ptr(
3679 for (i = 0; i < log->l_iclog_bufs; i++) {
3680 if (ptr >= log->l_iclog_bak[i] &&
3681 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3686 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3690 * Check to make sure the grant write head didn't just over lap the tail. If
3691 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3692 * the cycles differ by exactly one and check the byte count.
3694 * This check is run unlocked, so can give false positives. Rather than assert
3695 * on failures, use a warn-once flag and a panic tag to allow the admin to
3696 * determine if they want to panic the machine when such an error occurs. For
3697 * debug kernels this will have the same effect as using an assert but, unlinke
3698 * an assert, it can be turned off at runtime.
3701 xlog_verify_grant_tail(
3704 int tail_cycle, tail_blocks;
3707 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3708 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3709 if (tail_cycle != cycle) {
3710 if (cycle - 1 != tail_cycle &&
3711 !(log->l_flags & XLOG_TAIL_WARN)) {
3712 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3713 "%s: cycle - 1 != tail_cycle", __func__);
3714 log->l_flags |= XLOG_TAIL_WARN;
3717 if (space > BBTOB(tail_blocks) &&
3718 !(log->l_flags & XLOG_TAIL_WARN)) {
3719 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3720 "%s: space > BBTOB(tail_blocks)", __func__);
3721 log->l_flags |= XLOG_TAIL_WARN;
3726 /* check if it will fit */
3728 xlog_verify_tail_lsn(
3730 struct xlog_in_core *iclog,
3735 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3737 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3738 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3739 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3741 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3743 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3744 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3746 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3747 if (blocks < BTOBB(iclog->ic_offset) + 1)
3748 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3750 } /* xlog_verify_tail_lsn */
3753 * Perform a number of checks on the iclog before writing to disk.
3755 * 1. Make sure the iclogs are still circular
3756 * 2. Make sure we have a good magic number
3757 * 3. Make sure we don't have magic numbers in the data
3758 * 4. Check fields of each log operation header for:
3759 * A. Valid client identifier
3760 * B. tid ptr value falls in valid ptr space (user space code)
3761 * C. Length in log record header is correct according to the
3762 * individual operation headers within record.
3763 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3764 * log, check the preceding blocks of the physical log to make sure all
3765 * the cycle numbers agree with the current cycle number.
3770 struct xlog_in_core *iclog,
3774 xlog_op_header_t *ophead;
3775 xlog_in_core_t *icptr;
3776 xlog_in_core_2_t *xhdr;
3777 void *base_ptr, *ptr, *p;
3778 ptrdiff_t field_offset;
3780 int len, i, j, k, op_len;
3783 /* check validity of iclog pointers */
3784 spin_lock(&log->l_icloglock);
3785 icptr = log->l_iclog;
3786 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3789 if (icptr != log->l_iclog)
3790 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3791 spin_unlock(&log->l_icloglock);
3793 /* check log magic numbers */
3794 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3795 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3797 base_ptr = ptr = &iclog->ic_header;
3798 p = &iclog->ic_header;
3799 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3800 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3801 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3806 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3807 base_ptr = ptr = iclog->ic_datap;
3809 xhdr = iclog->ic_data;
3810 for (i = 0; i < len; i++) {
3813 /* clientid is only 1 byte */
3814 p = &ophead->oh_clientid;
3815 field_offset = p - base_ptr;
3816 if (!syncing || (field_offset & 0x1ff)) {
3817 clientid = ophead->oh_clientid;
3819 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3820 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3821 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3822 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3823 clientid = xlog_get_client_id(
3824 xhdr[j].hic_xheader.xh_cycle_data[k]);
3826 clientid = xlog_get_client_id(
3827 iclog->ic_header.h_cycle_data[idx]);
3830 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3832 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3833 __func__, clientid, ophead,
3834 (unsigned long)field_offset);
3837 p = &ophead->oh_len;
3838 field_offset = p - base_ptr;
3839 if (!syncing || (field_offset & 0x1ff)) {
3840 op_len = be32_to_cpu(ophead->oh_len);
3842 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3843 (uintptr_t)iclog->ic_datap);
3844 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3845 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3846 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3847 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3849 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3852 ptr += sizeof(xlog_op_header_t) + op_len;
3854 } /* xlog_verify_iclog */
3858 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3864 xlog_in_core_t *iclog, *ic;
3866 iclog = log->l_iclog;
3867 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3869 * Mark all the incore logs IOERROR.
3870 * From now on, no log flushes will result.
3874 ic->ic_state = XLOG_STATE_IOERROR;
3876 } while (ic != iclog);
3880 * Return non-zero, if state transition has already happened.
3886 * This is called from xfs_force_shutdown, when we're forcibly
3887 * shutting down the filesystem, typically because of an IO error.
3888 * Our main objectives here are to make sure that:
3889 * a. if !logerror, flush the logs to disk. Anything modified
3890 * after this is ignored.
3891 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3892 * parties to find out, 'atomically'.
3893 * c. those who're sleeping on log reservations, pinned objects and
3894 * other resources get woken up, and be told the bad news.
3895 * d. nothing new gets queued up after (b) and (c) are done.
3897 * Note: for the !logerror case we need to flush the regions held in memory out
3898 * to disk first. This needs to be done before the log is marked as shutdown,
3899 * otherwise the iclog writes will fail.
3902 xfs_log_force_umount(
3903 struct xfs_mount *mp,
3912 * If this happens during log recovery, don't worry about
3913 * locking; the log isn't open for business yet.
3916 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3917 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3919 mp->m_sb_bp->b_flags |= XBF_DONE;
3924 * Somebody could've already done the hard work for us.
3925 * No need to get locks for this.
3927 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3928 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3933 * Flush all the completed transactions to disk before marking the log
3934 * being shut down. We need to do it in this order to ensure that
3935 * completed operations are safely on disk before we shut down, and that
3936 * we don't have to issue any buffer IO after the shutdown flags are set
3937 * to guarantee this.
3940 xfs_log_force(mp, XFS_LOG_SYNC);
3943 * mark the filesystem and the as in a shutdown state and wake
3944 * everybody up to tell them the bad news.
3946 spin_lock(&log->l_icloglock);
3947 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3949 mp->m_sb_bp->b_flags |= XBF_DONE;
3952 * Mark the log and the iclogs with IO error flags to prevent any
3953 * further log IO from being issued or completed.
3955 log->l_flags |= XLOG_IO_ERROR;
3956 retval = xlog_state_ioerror(log);
3957 spin_unlock(&log->l_icloglock);
3960 * We don't want anybody waiting for log reservations after this. That
3961 * means we have to wake up everybody queued up on reserveq as well as
3962 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3963 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3964 * action is protected by the grant locks.
3966 xlog_grant_head_wake_all(&log->l_reserve_head);
3967 xlog_grant_head_wake_all(&log->l_write_head);
3970 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3971 * as if the log writes were completed. The abort handling in the log
3972 * item committed callback functions will do this again under lock to
3975 wake_up_all(&log->l_cilp->xc_commit_wait);
3976 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3978 #ifdef XFSERRORDEBUG
3980 xlog_in_core_t *iclog;
3982 spin_lock(&log->l_icloglock);
3983 iclog = log->l_iclog;
3985 ASSERT(iclog->ic_callback == 0);
3986 iclog = iclog->ic_next;
3987 } while (iclog != log->l_iclog);
3988 spin_unlock(&log->l_icloglock);
3991 /* return non-zero if log IOERROR transition had already happened */
3999 xlog_in_core_t *iclog;
4001 iclog = log->l_iclog;
4003 /* endianness does not matter here, zero is zero in
4006 if (iclog->ic_header.h_num_logops)
4008 iclog = iclog->ic_next;
4009 } while (iclog != log->l_iclog);
4014 * Verify that an LSN stamped into a piece of metadata is valid. This is
4015 * intended for use in read verifiers on v5 superblocks.
4019 struct xfs_mount *mp,
4022 struct xlog *log = mp->m_log;
4026 * norecovery mode skips mount-time log processing and unconditionally
4027 * resets the in-core LSN. We can't validate in this mode, but
4028 * modifications are not allowed anyways so just return true.
4030 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4034 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4035 * handled by recovery and thus safe to ignore here.
4037 if (lsn == NULLCOMMITLSN)
4040 valid = xlog_valid_lsn(mp->m_log, lsn);
4042 /* warn the user about what's gone wrong before verifier failure */
4044 spin_lock(&log->l_icloglock);
4046 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4047 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4048 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4049 log->l_curr_cycle, log->l_curr_block);
4050 spin_unlock(&log->l_icloglock);
4057 xfs_log_in_recovery(
4058 struct xfs_mount *mp)
4060 struct xlog *log = mp->m_log;
4062 return log->l_flags & XLOG_ACTIVE_RECOVERY;