1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_errortag.h"
14 #include "xfs_error.h"
15 #include "xfs_trans.h"
16 #include "xfs_trans_priv.h"
18 #include "xfs_log_priv.h"
19 #include "xfs_trace.h"
20 #include "xfs_sysfs.h"
22 #include "xfs_health.h"
24 kmem_zone_t *xfs_log_ticket_zone;
26 /* Local miscellaneous function prototypes */
30 struct xfs_buftarg *log_target,
31 xfs_daddr_t blk_offset,
41 /* local state machine functions */
42 STATIC void xlog_state_done_syncing(
43 struct xlog_in_core *iclog);
45 xlog_state_get_iclog_space(
48 struct xlog_in_core **iclog,
49 struct xlog_ticket *ticket,
53 xlog_state_switch_iclogs(
55 struct xlog_in_core *iclog,
64 struct xlog_in_core *iclog);
71 xlog_verify_grant_tail(
76 struct xlog_in_core *iclog,
81 struct xlog_in_core *iclog,
84 #define xlog_verify_dest_ptr(a,b)
85 #define xlog_verify_grant_tail(a)
86 #define xlog_verify_iclog(a,b,c)
87 #define xlog_verify_tail_lsn(a,b,c)
95 xfs_log_cover(struct xfs_mount *);
103 int64_t head_val = atomic64_read(head);
109 xlog_crack_grant_head_val(head_val, &cycle, &space);
113 space += log->l_logsize;
118 new = xlog_assign_grant_head_val(cycle, space);
119 head_val = atomic64_cmpxchg(head, old, new);
120 } while (head_val != old);
124 xlog_grant_add_space(
129 int64_t head_val = atomic64_read(head);
136 xlog_crack_grant_head_val(head_val, &cycle, &space);
138 tmp = log->l_logsize - space;
147 new = xlog_assign_grant_head_val(cycle, space);
148 head_val = atomic64_cmpxchg(head, old, new);
149 } while (head_val != old);
153 xlog_grant_head_init(
154 struct xlog_grant_head *head)
156 xlog_assign_grant_head(&head->grant, 1, 0);
157 INIT_LIST_HEAD(&head->waiters);
158 spin_lock_init(&head->lock);
162 xlog_grant_head_wake_all(
163 struct xlog_grant_head *head)
165 struct xlog_ticket *tic;
167 spin_lock(&head->lock);
168 list_for_each_entry(tic, &head->waiters, t_queue)
169 wake_up_process(tic->t_task);
170 spin_unlock(&head->lock);
174 xlog_ticket_reservation(
176 struct xlog_grant_head *head,
177 struct xlog_ticket *tic)
179 if (head == &log->l_write_head) {
180 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
181 return tic->t_unit_res;
183 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
184 return tic->t_unit_res * tic->t_cnt;
186 return tic->t_unit_res;
191 xlog_grant_head_wake(
193 struct xlog_grant_head *head,
196 struct xlog_ticket *tic;
198 bool woken_task = false;
200 list_for_each_entry(tic, &head->waiters, t_queue) {
203 * There is a chance that the size of the CIL checkpoints in
204 * progress at the last AIL push target calculation resulted in
205 * limiting the target to the log head (l_last_sync_lsn) at the
206 * time. This may not reflect where the log head is now as the
207 * CIL checkpoints may have completed.
209 * Hence when we are woken here, it may be that the head of the
210 * log that has moved rather than the tail. As the tail didn't
211 * move, there still won't be space available for the
212 * reservation we require. However, if the AIL has already
213 * pushed to the target defined by the old log head location, we
214 * will hang here waiting for something else to update the AIL
217 * Therefore, if there isn't space to wake the first waiter on
218 * the grant head, we need to push the AIL again to ensure the
219 * target reflects both the current log tail and log head
220 * position before we wait for the tail to move again.
223 need_bytes = xlog_ticket_reservation(log, head, tic);
224 if (*free_bytes < need_bytes) {
226 xlog_grant_push_ail(log, need_bytes);
230 *free_bytes -= need_bytes;
231 trace_xfs_log_grant_wake_up(log, tic);
232 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;
355 struct xfs_mount *mp)
358 * Do not write to the log on norecovery mounts, if the data or log
359 * devices are read-only, or if the filesystem is shutdown. Read-only
360 * mounts allow internal writes for log recovery and unmount purposes,
361 * so don't restrict that case.
363 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
365 if (xfs_readonly_buftarg(mp->m_ddev_targp))
367 if (xfs_readonly_buftarg(mp->m_log->l_targ))
369 if (XFS_FORCED_SHUTDOWN(mp))
375 * Replenish the byte reservation required by moving the grant write head.
379 struct xfs_mount *mp,
380 struct xlog_ticket *tic)
382 struct xlog *log = mp->m_log;
386 if (XLOG_FORCED_SHUTDOWN(log))
389 XFS_STATS_INC(mp, xs_try_logspace);
392 * This is a new transaction on the ticket, so we need to change the
393 * transaction ID so that the next transaction has a different TID in
394 * the log. Just add one to the existing tid so that we can see chains
395 * of rolling transactions in the log easily.
399 xlog_grant_push_ail(log, tic->t_unit_res);
401 tic->t_curr_res = tic->t_unit_res;
402 xlog_tic_reset_res(tic);
407 trace_xfs_log_regrant(log, tic);
409 error = xlog_grant_head_check(log, &log->l_write_head, tic,
414 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
415 trace_xfs_log_regrant_exit(log, tic);
416 xlog_verify_grant_tail(log);
421 * If we are failing, make sure the ticket doesn't have any current
422 * reservations. We don't want to add this back when the ticket/
423 * transaction gets cancelled.
426 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
431 * Reserve log space and return a ticket corresponding to the reservation.
433 * Each reservation is going to reserve extra space for a log record header.
434 * When writes happen to the on-disk log, we don't subtract the length of the
435 * log record header from any reservation. By wasting space in each
436 * reservation, we prevent over allocation problems.
440 struct xfs_mount *mp,
443 struct xlog_ticket **ticp,
447 struct xlog *log = mp->m_log;
448 struct xlog_ticket *tic;
452 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
454 if (XLOG_FORCED_SHUTDOWN(log))
457 XFS_STATS_INC(mp, xs_try_logspace);
459 ASSERT(*ticp == NULL);
460 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent);
463 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
466 trace_xfs_log_reserve(log, tic);
468 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
473 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
474 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
475 trace_xfs_log_reserve_exit(log, tic);
476 xlog_verify_grant_tail(log);
481 * If we are failing, make sure the ticket doesn't have any current
482 * reservations. We don't want to add this back when the ticket/
483 * transaction gets cancelled.
486 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
491 __xlog_state_release_iclog(
493 struct xlog_in_core *iclog)
495 lockdep_assert_held(&log->l_icloglock);
497 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
498 /* update tail before writing to iclog */
499 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
501 iclog->ic_state = XLOG_STATE_SYNCING;
502 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
503 xlog_verify_tail_lsn(log, iclog, tail_lsn);
504 /* cycle incremented when incrementing curr_block */
505 trace_xlog_iclog_syncing(iclog, _RET_IP_);
509 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
514 * Flush iclog to disk if this is the last reference to the given iclog and the
515 * it is in the WANT_SYNC state.
518 xlog_state_release_iclog(
520 struct xlog_in_core *iclog)
522 lockdep_assert_held(&log->l_icloglock);
524 trace_xlog_iclog_release(iclog, _RET_IP_);
525 if (iclog->ic_state == XLOG_STATE_IOERROR)
528 if (atomic_dec_and_test(&iclog->ic_refcnt) &&
529 __xlog_state_release_iclog(log, iclog)) {
530 spin_unlock(&log->l_icloglock);
531 xlog_sync(log, iclog);
532 spin_lock(&log->l_icloglock);
539 * Mount a log filesystem
541 * mp - ubiquitous xfs mount point structure
542 * log_target - buftarg of on-disk log device
543 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
544 * num_bblocks - Number of BBSIZE blocks in on-disk log
546 * Return error or zero.
551 xfs_buftarg_t *log_target,
552 xfs_daddr_t blk_offset,
555 bool fatal = xfs_sb_version_hascrc(&mp->m_sb);
559 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
560 xfs_notice(mp, "Mounting V%d Filesystem",
561 XFS_SB_VERSION_NUM(&mp->m_sb));
564 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
565 XFS_SB_VERSION_NUM(&mp->m_sb));
566 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
569 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
570 if (IS_ERR(mp->m_log)) {
571 error = PTR_ERR(mp->m_log);
576 * Validate the given log space and drop a critical message via syslog
577 * if the log size is too small that would lead to some unexpected
578 * situations in transaction log space reservation stage.
580 * Note: we can't just reject the mount if the validation fails. This
581 * would mean that people would have to downgrade their kernel just to
582 * remedy the situation as there is no way to grow the log (short of
583 * black magic surgery with xfs_db).
585 * We can, however, reject mounts for CRC format filesystems, as the
586 * mkfs binary being used to make the filesystem should never create a
587 * filesystem with a log that is too small.
589 min_logfsbs = xfs_log_calc_minimum_size(mp);
591 if (mp->m_sb.sb_logblocks < min_logfsbs) {
593 "Log size %d blocks too small, minimum size is %d blocks",
594 mp->m_sb.sb_logblocks, min_logfsbs);
596 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
598 "Log size %d blocks too large, maximum size is %lld blocks",
599 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
601 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
603 "log size %lld bytes too large, maximum size is %lld bytes",
604 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
607 } else if (mp->m_sb.sb_logsunit > 1 &&
608 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
610 "log stripe unit %u bytes must be a multiple of block size",
611 mp->m_sb.sb_logsunit);
617 * Log check errors are always fatal on v5; or whenever bad
618 * metadata leads to a crash.
621 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
625 xfs_crit(mp, "Log size out of supported range.");
627 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
631 * Initialize the AIL now we have a log.
633 error = xfs_trans_ail_init(mp);
635 xfs_warn(mp, "AIL initialisation failed: error %d", error);
638 mp->m_log->l_ailp = mp->m_ail;
641 * skip log recovery on a norecovery mount. pretend it all
644 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
645 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
648 mp->m_flags &= ~XFS_MOUNT_RDONLY;
650 error = xlog_recover(mp->m_log);
653 mp->m_flags |= XFS_MOUNT_RDONLY;
655 xfs_warn(mp, "log mount/recovery failed: error %d",
657 xlog_recover_cancel(mp->m_log);
658 goto out_destroy_ail;
662 error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
665 goto out_destroy_ail;
667 /* Normal transactions can now occur */
668 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
671 * Now the log has been fully initialised and we know were our
672 * space grant counters are, we can initialise the permanent ticket
673 * needed for delayed logging to work.
675 xlog_cil_init_post_recovery(mp->m_log);
680 xfs_trans_ail_destroy(mp);
682 xlog_dealloc_log(mp->m_log);
688 * Finish the recovery of the file system. This is separate from the
689 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
690 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
693 * If we finish recovery successfully, start the background log work. If we are
694 * not doing recovery, then we have a RO filesystem and we don't need to start
698 xfs_log_mount_finish(
699 struct xfs_mount *mp)
702 bool readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
703 bool recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
705 if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
706 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
708 } else if (readonly) {
709 /* Allow unlinked processing to proceed */
710 mp->m_flags &= ~XFS_MOUNT_RDONLY;
714 * During the second phase of log recovery, we need iget and
715 * iput to behave like they do for an active filesystem.
716 * xfs_fs_drop_inode needs to be able to prevent the deletion
717 * of inodes before we're done replaying log items on those
718 * inodes. Turn it off immediately after recovery finishes
719 * so that we don't leak the quota inodes if subsequent mount
722 * We let all inodes involved in redo item processing end up on
723 * the LRU instead of being evicted immediately so that if we do
724 * something to an unlinked inode, the irele won't cause
725 * premature truncation and freeing of the inode, which results
726 * in log recovery failure. We have to evict the unreferenced
727 * lru inodes after clearing SB_ACTIVE because we don't
728 * otherwise clean up the lru if there's a subsequent failure in
729 * xfs_mountfs, which leads to us leaking the inodes if nothing
730 * else (e.g. quotacheck) references the inodes before the
731 * mount failure occurs.
733 mp->m_super->s_flags |= SB_ACTIVE;
734 error = xlog_recover_finish(mp->m_log);
736 xfs_log_work_queue(mp);
737 mp->m_super->s_flags &= ~SB_ACTIVE;
738 evict_inodes(mp->m_super);
741 * Drain the buffer LRU after log recovery. This is required for v4
742 * filesystems to avoid leaving around buffers with NULL verifier ops,
743 * but we do it unconditionally to make sure we're always in a clean
744 * cache state after mount.
746 * Don't push in the error case because the AIL may have pending intents
747 * that aren't removed until recovery is cancelled.
749 if (!error && recovered) {
750 xfs_log_force(mp, XFS_LOG_SYNC);
751 xfs_ail_push_all_sync(mp->m_ail);
753 xfs_buftarg_drain(mp->m_ddev_targp);
756 mp->m_flags |= XFS_MOUNT_RDONLY;
758 /* Make sure the log is dead if we're returning failure. */
759 ASSERT(!error || (mp->m_log->l_flags & XLOG_IO_ERROR));
765 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
769 xfs_log_mount_cancel(
770 struct xfs_mount *mp)
772 xlog_recover_cancel(mp->m_log);
777 * Wait for the iclog and all prior iclogs to be written disk as required by the
778 * log force state machine. Waiting on ic_force_wait ensures iclog completions
779 * have been ordered and callbacks run before we are woken here, hence
780 * guaranteeing that all the iclogs up to this one are on stable storage.
784 struct xlog_in_core *iclog)
785 __releases(iclog->ic_log->l_icloglock)
787 struct xlog *log = iclog->ic_log;
789 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
790 if (!XLOG_FORCED_SHUTDOWN(log) &&
791 iclog->ic_state != XLOG_STATE_ACTIVE &&
792 iclog->ic_state != XLOG_STATE_DIRTY) {
793 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
794 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
796 spin_unlock(&log->l_icloglock);
799 if (XLOG_FORCED_SHUTDOWN(log))
805 * Write out an unmount record using the ticket provided. We have to account for
806 * the data space used in the unmount ticket as this write is not done from a
807 * transaction context that has already done the accounting for us.
810 xlog_write_unmount_record(
812 struct xlog_ticket *ticket)
814 struct xfs_unmount_log_format ulf = {
815 .magic = XLOG_UNMOUNT_TYPE,
817 struct xfs_log_iovec reg = {
819 .i_len = sizeof(ulf),
820 .i_type = XLOG_REG_TYPE_UNMOUNT,
822 struct xfs_log_vec vec = {
827 /* account for space used by record data */
828 ticket->t_curr_res -= sizeof(ulf);
831 * For external log devices, we need to flush the data device cache
832 * first to ensure all metadata writeback is on stable storage before we
833 * stamp the tail LSN into the unmount record.
835 if (log->l_targ != log->l_mp->m_ddev_targp)
836 blkdev_issue_flush(log->l_targ->bt_bdev);
837 return xlog_write(log, &vec, ticket, NULL, NULL, XLOG_UNMOUNT_TRANS);
841 * Mark the filesystem clean by writing an unmount record to the head of the
848 struct xfs_mount *mp = log->l_mp;
849 struct xlog_in_core *iclog;
850 struct xlog_ticket *tic = NULL;
853 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
857 error = xlog_write_unmount_record(log, tic);
859 * At this point, we're umounting anyway, so there's no point in
860 * transitioning log state to IOERROR. Just continue...
864 xfs_alert(mp, "%s: unmount record failed", __func__);
866 spin_lock(&log->l_icloglock);
867 iclog = log->l_iclog;
868 atomic_inc(&iclog->ic_refcnt);
869 if (iclog->ic_state == XLOG_STATE_ACTIVE)
870 xlog_state_switch_iclogs(log, iclog, 0);
872 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
873 iclog->ic_state == XLOG_STATE_IOERROR);
875 * Ensure the journal is fully flushed and on stable storage once the
876 * iclog containing the unmount record is written.
878 iclog->ic_flags |= (XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
879 error = xlog_state_release_iclog(log, iclog);
880 xlog_wait_on_iclog(iclog);
883 trace_xfs_log_umount_write(log, tic);
884 xfs_log_ticket_ungrant(log, tic);
889 xfs_log_unmount_verify_iclog(
892 struct xlog_in_core *iclog = log->l_iclog;
895 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
896 ASSERT(iclog->ic_offset == 0);
897 } while ((iclog = iclog->ic_next) != log->l_iclog);
901 * Unmount record used to have a string "Unmount filesystem--" in the
902 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
903 * We just write the magic number now since that particular field isn't
904 * currently architecture converted and "Unmount" is a bit foo.
905 * As far as I know, there weren't any dependencies on the old behaviour.
908 xfs_log_unmount_write(
909 struct xfs_mount *mp)
911 struct xlog *log = mp->m_log;
913 if (!xfs_log_writable(mp))
916 xfs_log_force(mp, XFS_LOG_SYNC);
918 if (XLOG_FORCED_SHUTDOWN(log))
922 * If we think the summary counters are bad, avoid writing the unmount
923 * record to force log recovery at next mount, after which the summary
924 * counters will be recalculated. Refer to xlog_check_unmount_rec for
927 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
928 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
929 xfs_alert(mp, "%s: will fix summary counters at next mount",
934 xfs_log_unmount_verify_iclog(log);
935 xlog_unmount_write(log);
939 * Empty the log for unmount/freeze.
941 * To do this, we first need to shut down the background log work so it is not
942 * trying to cover the log as we clean up. We then need to unpin all objects in
943 * the log so we can then flush them out. Once they have completed their IO and
944 * run the callbacks removing themselves from the AIL, we can cover the log.
948 struct xfs_mount *mp)
950 cancel_delayed_work_sync(&mp->m_log->l_work);
951 xfs_log_force(mp, XFS_LOG_SYNC);
954 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
955 * will push it, xfs_buftarg_wait() will not wait for it. Further,
956 * xfs_buf_iowait() cannot be used because it was pushed with the
957 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
958 * the IO to complete.
960 xfs_ail_push_all_sync(mp->m_ail);
961 xfs_buftarg_wait(mp->m_ddev_targp);
962 xfs_buf_lock(mp->m_sb_bp);
963 xfs_buf_unlock(mp->m_sb_bp);
965 return xfs_log_cover(mp);
970 struct xfs_mount *mp)
973 xfs_log_unmount_write(mp);
977 * Shut down and release the AIL and Log.
979 * During unmount, we need to ensure we flush all the dirty metadata objects
980 * from the AIL so that the log is empty before we write the unmount record to
981 * the log. Once this is done, we can tear down the AIL and the log.
985 struct xfs_mount *mp)
989 xfs_buftarg_drain(mp->m_ddev_targp);
991 xfs_trans_ail_destroy(mp);
993 xfs_sysfs_del(&mp->m_log->l_kobj);
995 xlog_dealloc_log(mp->m_log);
1000 struct xfs_mount *mp,
1001 struct xfs_log_item *item,
1003 const struct xfs_item_ops *ops)
1005 item->li_mountp = mp;
1006 item->li_ailp = mp->m_ail;
1007 item->li_type = type;
1011 INIT_LIST_HEAD(&item->li_ail);
1012 INIT_LIST_HEAD(&item->li_cil);
1013 INIT_LIST_HEAD(&item->li_bio_list);
1014 INIT_LIST_HEAD(&item->li_trans);
1018 * Wake up processes waiting for log space after we have moved the log tail.
1022 struct xfs_mount *mp)
1024 struct xlog *log = mp->m_log;
1027 if (XLOG_FORCED_SHUTDOWN(log))
1030 if (!list_empty_careful(&log->l_write_head.waiters)) {
1031 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1033 spin_lock(&log->l_write_head.lock);
1034 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1035 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1036 spin_unlock(&log->l_write_head.lock);
1039 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1040 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1042 spin_lock(&log->l_reserve_head.lock);
1043 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1044 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1045 spin_unlock(&log->l_reserve_head.lock);
1050 * Determine if we have a transaction that has gone to disk that needs to be
1051 * covered. To begin the transition to the idle state firstly the log needs to
1052 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1053 * we start attempting to cover the log.
1055 * Only if we are then in a state where covering is needed, the caller is
1056 * informed that dummy transactions are required to move the log into the idle
1059 * If there are any items in the AIl or CIL, then we do not want to attempt to
1060 * cover the log as we may be in a situation where there isn't log space
1061 * available to run a dummy transaction and this can lead to deadlocks when the
1062 * tail of the log is pinned by an item that is modified in the CIL. Hence
1063 * there's no point in running a dummy transaction at this point because we
1064 * can't start trying to idle the log until both the CIL and AIL are empty.
1067 xfs_log_need_covered(
1068 struct xfs_mount *mp)
1070 struct xlog *log = mp->m_log;
1071 bool needed = false;
1073 if (!xlog_cil_empty(log))
1076 spin_lock(&log->l_icloglock);
1077 switch (log->l_covered_state) {
1078 case XLOG_STATE_COVER_DONE:
1079 case XLOG_STATE_COVER_DONE2:
1080 case XLOG_STATE_COVER_IDLE:
1082 case XLOG_STATE_COVER_NEED:
1083 case XLOG_STATE_COVER_NEED2:
1084 if (xfs_ail_min_lsn(log->l_ailp))
1086 if (!xlog_iclogs_empty(log))
1090 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1091 log->l_covered_state = XLOG_STATE_COVER_DONE;
1093 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1099 spin_unlock(&log->l_icloglock);
1104 * Explicitly cover the log. This is similar to background log covering but
1105 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1106 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1107 * must all be empty.
1111 struct xfs_mount *mp)
1116 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1117 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1118 XFS_FORCED_SHUTDOWN(mp));
1120 if (!xfs_log_writable(mp))
1124 * xfs_log_need_covered() is not idempotent because it progresses the
1125 * state machine if the log requires covering. Therefore, we must call
1126 * this function once and use the result until we've issued an sb sync.
1127 * Do so first to make that abundantly clear.
1129 * Fall into the covering sequence if the log needs covering or the
1130 * mount has lazy superblock accounting to sync to disk. The sb sync
1131 * used for covering accumulates the in-core counters, so covering
1132 * handles this for us.
1134 need_covered = xfs_log_need_covered(mp);
1135 if (!need_covered && !xfs_sb_version_haslazysbcount(&mp->m_sb))
1139 * To cover the log, commit the superblock twice (at most) in
1140 * independent checkpoints. The first serves as a reference for the
1141 * tail pointer. The sync transaction and AIL push empties the AIL and
1142 * updates the in-core tail to the LSN of the first checkpoint. The
1143 * second commit updates the on-disk tail with the in-core LSN,
1144 * covering the log. Push the AIL one more time to leave it empty, as
1148 error = xfs_sync_sb(mp, true);
1151 xfs_ail_push_all_sync(mp->m_ail);
1152 } while (xfs_log_need_covered(mp));
1158 * We may be holding the log iclog lock upon entering this routine.
1161 xlog_assign_tail_lsn_locked(
1162 struct xfs_mount *mp)
1164 struct xlog *log = mp->m_log;
1165 struct xfs_log_item *lip;
1168 assert_spin_locked(&mp->m_ail->ail_lock);
1171 * To make sure we always have a valid LSN for the log tail we keep
1172 * track of the last LSN which was committed in log->l_last_sync_lsn,
1173 * and use that when the AIL was empty.
1175 lip = xfs_ail_min(mp->m_ail);
1177 tail_lsn = lip->li_lsn;
1179 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1180 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1181 atomic64_set(&log->l_tail_lsn, tail_lsn);
1186 xlog_assign_tail_lsn(
1187 struct xfs_mount *mp)
1191 spin_lock(&mp->m_ail->ail_lock);
1192 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1193 spin_unlock(&mp->m_ail->ail_lock);
1199 * Return the space in the log between the tail and the head. The head
1200 * is passed in the cycle/bytes formal parms. In the special case where
1201 * the reserve head has wrapped passed the tail, this calculation is no
1202 * longer valid. In this case, just return 0 which means there is no space
1203 * in the log. This works for all places where this function is called
1204 * with the reserve head. Of course, if the write head were to ever
1205 * wrap the tail, we should blow up. Rather than catch this case here,
1206 * we depend on other ASSERTions in other parts of the code. XXXmiken
1208 * This code also handles the case where the reservation head is behind
1209 * the tail. The details of this case are described below, but the end
1210 * result is that we return the size of the log as the amount of space left.
1223 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1224 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1225 tail_bytes = BBTOB(tail_bytes);
1226 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1227 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1228 else if (tail_cycle + 1 < head_cycle)
1230 else if (tail_cycle < head_cycle) {
1231 ASSERT(tail_cycle == (head_cycle - 1));
1232 free_bytes = tail_bytes - head_bytes;
1235 * The reservation head is behind the tail.
1236 * In this case we just want to return the size of the
1237 * log as the amount of space left.
1239 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1240 xfs_alert(log->l_mp,
1241 " tail_cycle = %d, tail_bytes = %d",
1242 tail_cycle, tail_bytes);
1243 xfs_alert(log->l_mp,
1244 " GH cycle = %d, GH bytes = %d",
1245 head_cycle, head_bytes);
1247 free_bytes = log->l_logsize;
1255 struct work_struct *work)
1257 struct xlog_in_core *iclog =
1258 container_of(work, struct xlog_in_core, ic_end_io_work);
1259 struct xlog *log = iclog->ic_log;
1262 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1264 /* treat writes with injected CRC errors as failed */
1265 if (iclog->ic_fail_crc)
1270 * Race to shutdown the filesystem if we see an error.
1272 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1273 xfs_alert(log->l_mp, "log I/O error %d", error);
1274 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1277 xlog_state_done_syncing(iclog);
1278 bio_uninit(&iclog->ic_bio);
1281 * Drop the lock to signal that we are done. Nothing references the
1282 * iclog after this, so an unmount waiting on this lock can now tear it
1283 * down safely. As such, it is unsafe to reference the iclog after the
1284 * unlock as we could race with it being freed.
1286 up(&iclog->ic_sema);
1290 * Return size of each in-core log record buffer.
1292 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1294 * If the filesystem blocksize is too large, we may need to choose a
1295 * larger size since the directory code currently logs entire blocks.
1298 xlog_get_iclog_buffer_size(
1299 struct xfs_mount *mp,
1302 if (mp->m_logbufs <= 0)
1303 mp->m_logbufs = XLOG_MAX_ICLOGS;
1304 if (mp->m_logbsize <= 0)
1305 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1307 log->l_iclog_bufs = mp->m_logbufs;
1308 log->l_iclog_size = mp->m_logbsize;
1311 * # headers = size / 32k - one header holds cycles from 32k of data.
1313 log->l_iclog_heads =
1314 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1315 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1320 struct xfs_mount *mp)
1322 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1323 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1327 * Every sync period we need to unpin all items in the AIL and push them to
1328 * disk. If there is nothing dirty, then we might need to cover the log to
1329 * indicate that the filesystem is idle.
1333 struct work_struct *work)
1335 struct xlog *log = container_of(to_delayed_work(work),
1336 struct xlog, l_work);
1337 struct xfs_mount *mp = log->l_mp;
1339 /* dgc: errors ignored - not fatal and nowhere to report them */
1340 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1342 * Dump a transaction into the log that contains no real change.
1343 * This is needed to stamp the current tail LSN into the log
1344 * during the covering operation.
1346 * We cannot use an inode here for this - that will push dirty
1347 * state back up into the VFS and then periodic inode flushing
1348 * will prevent log covering from making progress. Hence we
1349 * synchronously log the superblock instead to ensure the
1350 * superblock is immediately unpinned and can be written back.
1352 xfs_sync_sb(mp, true);
1354 xfs_log_force(mp, 0);
1356 /* start pushing all the metadata that is currently dirty */
1357 xfs_ail_push_all(mp->m_ail);
1359 /* queue us up again */
1360 xfs_log_work_queue(mp);
1364 * This routine initializes some of the log structure for a given mount point.
1365 * Its primary purpose is to fill in enough, so recovery can occur. However,
1366 * some other stuff may be filled in too.
1368 STATIC struct xlog *
1370 struct xfs_mount *mp,
1371 struct xfs_buftarg *log_target,
1372 xfs_daddr_t blk_offset,
1376 xlog_rec_header_t *head;
1377 xlog_in_core_t **iclogp;
1378 xlog_in_core_t *iclog, *prev_iclog=NULL;
1380 int error = -ENOMEM;
1383 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1385 xfs_warn(mp, "Log allocation failed: No memory!");
1390 log->l_targ = log_target;
1391 log->l_logsize = BBTOB(num_bblks);
1392 log->l_logBBstart = blk_offset;
1393 log->l_logBBsize = num_bblks;
1394 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1395 log->l_flags |= XLOG_ACTIVE_RECOVERY;
1396 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1398 log->l_prev_block = -1;
1399 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1400 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1401 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1402 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1404 if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1)
1405 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1407 log->l_iclog_roundoff = BBSIZE;
1409 xlog_grant_head_init(&log->l_reserve_head);
1410 xlog_grant_head_init(&log->l_write_head);
1412 error = -EFSCORRUPTED;
1413 if (xfs_sb_version_hassector(&mp->m_sb)) {
1414 log2_size = mp->m_sb.sb_logsectlog;
1415 if (log2_size < BBSHIFT) {
1416 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1417 log2_size, BBSHIFT);
1421 log2_size -= BBSHIFT;
1422 if (log2_size > mp->m_sectbb_log) {
1423 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1424 log2_size, mp->m_sectbb_log);
1428 /* for larger sector sizes, must have v2 or external log */
1429 if (log2_size && log->l_logBBstart > 0 &&
1430 !xfs_sb_version_haslogv2(&mp->m_sb)) {
1432 "log sector size (0x%x) invalid for configuration.",
1437 log->l_sectBBsize = 1 << log2_size;
1439 xlog_get_iclog_buffer_size(mp, log);
1441 spin_lock_init(&log->l_icloglock);
1442 init_waitqueue_head(&log->l_flush_wait);
1444 iclogp = &log->l_iclog;
1446 * The amount of memory to allocate for the iclog structure is
1447 * rather funky due to the way the structure is defined. It is
1448 * done this way so that we can use different sizes for machines
1449 * with different amounts of memory. See the definition of
1450 * xlog_in_core_t in xfs_log_priv.h for details.
1452 ASSERT(log->l_iclog_size >= 4096);
1453 for (i = 0; i < log->l_iclog_bufs; i++) {
1454 int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1455 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1456 sizeof(struct bio_vec);
1458 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1460 goto out_free_iclog;
1463 iclog->ic_prev = prev_iclog;
1466 iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1467 KM_MAYFAIL | KM_ZERO);
1468 if (!iclog->ic_data)
1469 goto out_free_iclog;
1471 log->l_iclog_bak[i] = &iclog->ic_header;
1473 head = &iclog->ic_header;
1474 memset(head, 0, sizeof(xlog_rec_header_t));
1475 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1476 head->h_version = cpu_to_be32(
1477 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1478 head->h_size = cpu_to_be32(log->l_iclog_size);
1480 head->h_fmt = cpu_to_be32(XLOG_FMT);
1481 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1483 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1484 iclog->ic_state = XLOG_STATE_ACTIVE;
1485 iclog->ic_log = log;
1486 atomic_set(&iclog->ic_refcnt, 0);
1487 INIT_LIST_HEAD(&iclog->ic_callbacks);
1488 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1490 init_waitqueue_head(&iclog->ic_force_wait);
1491 init_waitqueue_head(&iclog->ic_write_wait);
1492 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1493 sema_init(&iclog->ic_sema, 1);
1495 iclogp = &iclog->ic_next;
1497 *iclogp = log->l_iclog; /* complete ring */
1498 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1500 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1501 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1503 0, mp->m_super->s_id);
1504 if (!log->l_ioend_workqueue)
1505 goto out_free_iclog;
1507 error = xlog_cil_init(log);
1509 goto out_destroy_workqueue;
1512 out_destroy_workqueue:
1513 destroy_workqueue(log->l_ioend_workqueue);
1515 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1516 prev_iclog = iclog->ic_next;
1517 kmem_free(iclog->ic_data);
1519 if (prev_iclog == log->l_iclog)
1525 return ERR_PTR(error);
1526 } /* xlog_alloc_log */
1529 * Write out the commit record of a transaction associated with the given
1530 * ticket to close off a running log write. Return the lsn of the commit record.
1535 struct xlog_ticket *ticket,
1536 struct xlog_in_core **iclog,
1539 struct xfs_log_iovec reg = {
1542 .i_type = XLOG_REG_TYPE_COMMIT,
1544 struct xfs_log_vec vec = {
1550 if (XLOG_FORCED_SHUTDOWN(log))
1553 error = xlog_write(log, &vec, ticket, lsn, iclog, XLOG_COMMIT_TRANS);
1555 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1560 * Compute the LSN that we'd need to push the log tail towards in order to have
1561 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1562 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1563 * log free space already meets all three thresholds, this function returns
1567 xlog_grant_push_threshold(
1571 xfs_lsn_t threshold_lsn = 0;
1572 xfs_lsn_t last_sync_lsn;
1575 int threshold_block;
1576 int threshold_cycle;
1579 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1581 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1582 free_blocks = BTOBBT(free_bytes);
1585 * Set the threshold for the minimum number of free blocks in the
1586 * log to the maximum of what the caller needs, one quarter of the
1587 * log, and 256 blocks.
1589 free_threshold = BTOBB(need_bytes);
1590 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1591 free_threshold = max(free_threshold, 256);
1592 if (free_blocks >= free_threshold)
1593 return NULLCOMMITLSN;
1595 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1597 threshold_block += free_threshold;
1598 if (threshold_block >= log->l_logBBsize) {
1599 threshold_block -= log->l_logBBsize;
1600 threshold_cycle += 1;
1602 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1605 * Don't pass in an lsn greater than the lsn of the last
1606 * log record known to be on disk. Use a snapshot of the last sync lsn
1607 * so that it doesn't change between the compare and the set.
1609 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1610 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1611 threshold_lsn = last_sync_lsn;
1613 return threshold_lsn;
1617 * Push the tail of the log if we need to do so to maintain the free log space
1618 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1619 * policy which pushes on an lsn which is further along in the log once we
1620 * reach the high water mark. In this manner, we would be creating a low water
1624 xlog_grant_push_ail(
1628 xfs_lsn_t threshold_lsn;
1630 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1631 if (threshold_lsn == NULLCOMMITLSN || XLOG_FORCED_SHUTDOWN(log))
1635 * Get the transaction layer to kick the dirty buffers out to
1636 * disk asynchronously. No point in trying to do this if
1637 * the filesystem is shutting down.
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 = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1711 for (i = 1; i < xheads; i++) {
1712 crc = crc32c(crc, &xhdr[i].hic_xheader,
1713 sizeof(struct xlog_rec_ext_header));
1717 /* ... and finally for the payload */
1718 crc = crc32c(crc, dp, size);
1720 return xfs_end_cksum(crc);
1727 struct xlog_in_core *iclog = bio->bi_private;
1729 queue_work(iclog->ic_log->l_ioend_workqueue,
1730 &iclog->ic_end_io_work);
1734 xlog_map_iclog_data(
1740 struct page *page = kmem_to_page(data);
1741 unsigned int off = offset_in_page(data);
1742 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1744 if (bio_add_page(bio, page, len, off) != len)
1757 struct xlog_in_core *iclog,
1761 ASSERT(bno < log->l_logBBsize);
1762 trace_xlog_iclog_write(iclog, _RET_IP_);
1765 * We lock the iclogbufs here so that we can serialise against I/O
1766 * completion during unmount. We might be processing a shutdown
1767 * triggered during unmount, and that can occur asynchronously to the
1768 * unmount thread, and hence we need to ensure that completes before
1769 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1770 * across the log IO to archieve that.
1772 down(&iclog->ic_sema);
1773 if (unlikely(iclog->ic_state == XLOG_STATE_IOERROR)) {
1775 * It would seem logical to return EIO here, but we rely on
1776 * the log state machine to propagate I/O errors instead of
1777 * doing it here. We kick of the state machine and unlock
1778 * the buffer manually, the code needs to be kept in sync
1779 * with the I/O completion path.
1781 xlog_state_done_syncing(iclog);
1782 up(&iclog->ic_sema);
1786 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1787 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1788 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1789 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1790 iclog->ic_bio.bi_private = iclog;
1793 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1794 * IOs coming immediately after this one. This prevents the block layer
1795 * writeback throttle from throttling log writes behind background
1796 * metadata writeback and causing priority inversions.
1798 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE;
1799 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH)
1800 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1801 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1802 iclog->ic_bio.bi_opf |= REQ_FUA;
1803 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1805 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1806 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1809 if (is_vmalloc_addr(iclog->ic_data))
1810 flush_kernel_vmap_range(iclog->ic_data, count);
1813 * If this log buffer would straddle the end of the log we will have
1814 * to split it up into two bios, so that we can continue at the start.
1816 if (bno + BTOBB(count) > log->l_logBBsize) {
1819 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1820 GFP_NOIO, &fs_bio_set);
1821 bio_chain(split, &iclog->ic_bio);
1824 /* restart at logical offset zero for the remainder */
1825 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1828 submit_bio(&iclog->ic_bio);
1832 * We need to bump cycle number for the part of the iclog that is
1833 * written to the start of the log. Watch out for the header magic
1834 * number case, though.
1843 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1846 for (i = split_offset; i < count; i += BBSIZE) {
1847 uint32_t cycle = get_unaligned_be32(data + i);
1849 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1851 put_unaligned_be32(cycle, data + i);
1856 xlog_calc_iclog_size(
1858 struct xlog_in_core *iclog,
1861 uint32_t count_init, count;
1863 /* Add for LR header */
1864 count_init = log->l_iclog_hsize + iclog->ic_offset;
1865 count = roundup(count_init, log->l_iclog_roundoff);
1867 *roundoff = count - count_init;
1869 ASSERT(count >= count_init);
1870 ASSERT(*roundoff < log->l_iclog_roundoff);
1875 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1876 * fashion. Previously, we should have moved the current iclog
1877 * ptr in the log to point to the next available iclog. This allows further
1878 * write to continue while this code syncs out an iclog ready to go.
1879 * Before an in-core log can be written out, the data section must be scanned
1880 * to save away the 1st word of each BBSIZE block into the header. We replace
1881 * it with the current cycle count. Each BBSIZE block is tagged with the
1882 * cycle count because there in an implicit assumption that drives will
1883 * guarantee that entire 512 byte blocks get written at once. In other words,
1884 * we can't have part of a 512 byte block written and part not written. By
1885 * tagging each block, we will know which blocks are valid when recovering
1886 * after an unclean shutdown.
1888 * This routine is single threaded on the iclog. No other thread can be in
1889 * this routine with the same iclog. Changing contents of iclog can there-
1890 * fore be done without grabbing the state machine lock. Updating the global
1891 * log will require grabbing the lock though.
1893 * The entire log manager uses a logical block numbering scheme. Only
1894 * xlog_write_iclog knows about the fact that the log may not start with
1895 * block zero on a given device.
1900 struct xlog_in_core *iclog)
1902 unsigned int count; /* byte count of bwrite */
1903 unsigned int roundoff; /* roundoff to BB or stripe */
1907 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1908 trace_xlog_iclog_sync(iclog, _RET_IP_);
1910 count = xlog_calc_iclog_size(log, iclog, &roundoff);
1912 /* move grant heads by roundoff in sync */
1913 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1914 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1916 /* put cycle number in every block */
1917 xlog_pack_data(log, iclog, roundoff);
1919 /* real byte length */
1920 size = iclog->ic_offset;
1921 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1923 iclog->ic_header.h_len = cpu_to_be32(size);
1925 XFS_STATS_INC(log->l_mp, xs_log_writes);
1926 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1928 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1930 /* Do we need to split this write into 2 parts? */
1931 if (bno + BTOBB(count) > log->l_logBBsize)
1932 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1934 /* calculcate the checksum */
1935 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1936 iclog->ic_datap, size);
1938 * Intentionally corrupt the log record CRC based on the error injection
1939 * frequency, if defined. This facilitates testing log recovery in the
1940 * event of torn writes. Hence, set the IOABORT state to abort the log
1941 * write on I/O completion and shutdown the fs. The subsequent mount
1942 * detects the bad CRC and attempts to recover.
1945 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1946 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1947 iclog->ic_fail_crc = true;
1949 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1950 be64_to_cpu(iclog->ic_header.h_lsn));
1953 xlog_verify_iclog(log, iclog, count);
1954 xlog_write_iclog(log, iclog, bno, count);
1958 * Deallocate a log structure
1964 xlog_in_core_t *iclog, *next_iclog;
1967 xlog_cil_destroy(log);
1970 * Cycle all the iclogbuf locks to make sure all log IO completion
1971 * is done before we tear down these buffers.
1973 iclog = log->l_iclog;
1974 for (i = 0; i < log->l_iclog_bufs; i++) {
1975 down(&iclog->ic_sema);
1976 up(&iclog->ic_sema);
1977 iclog = iclog->ic_next;
1980 iclog = log->l_iclog;
1981 for (i = 0; i < log->l_iclog_bufs; i++) {
1982 next_iclog = iclog->ic_next;
1983 kmem_free(iclog->ic_data);
1988 log->l_mp->m_log = NULL;
1989 destroy_workqueue(log->l_ioend_workqueue);
1994 * Update counters atomically now that memcpy is done.
1997 xlog_state_finish_copy(
1999 struct xlog_in_core *iclog,
2003 lockdep_assert_held(&log->l_icloglock);
2005 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2006 iclog->ic_offset += copy_bytes;
2010 * print out info relating to regions written which consume
2015 struct xfs_mount *mp,
2016 struct xlog_ticket *ticket)
2019 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2021 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2022 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2023 static char *res_type_str[] = {
2024 REG_TYPE_STR(BFORMAT, "bformat"),
2025 REG_TYPE_STR(BCHUNK, "bchunk"),
2026 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2027 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2028 REG_TYPE_STR(IFORMAT, "iformat"),
2029 REG_TYPE_STR(ICORE, "icore"),
2030 REG_TYPE_STR(IEXT, "iext"),
2031 REG_TYPE_STR(IBROOT, "ibroot"),
2032 REG_TYPE_STR(ILOCAL, "ilocal"),
2033 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2034 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2035 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2036 REG_TYPE_STR(QFORMAT, "qformat"),
2037 REG_TYPE_STR(DQUOT, "dquot"),
2038 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2039 REG_TYPE_STR(LRHEADER, "LR header"),
2040 REG_TYPE_STR(UNMOUNT, "unmount"),
2041 REG_TYPE_STR(COMMIT, "commit"),
2042 REG_TYPE_STR(TRANSHDR, "trans header"),
2043 REG_TYPE_STR(ICREATE, "inode create"),
2044 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2045 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2046 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2047 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2048 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2049 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2051 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2054 xfs_warn(mp, "ticket reservation summary:");
2055 xfs_warn(mp, " unit res = %d bytes",
2056 ticket->t_unit_res);
2057 xfs_warn(mp, " current res = %d bytes",
2058 ticket->t_curr_res);
2059 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2060 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2061 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2062 ticket->t_res_num_ophdrs, ophdr_spc);
2063 xfs_warn(mp, " ophdr + reg = %u bytes",
2064 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2065 xfs_warn(mp, " num regions = %u",
2068 for (i = 0; i < ticket->t_res_num; i++) {
2069 uint r_type = ticket->t_res_arr[i].r_type;
2070 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2071 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2072 "bad-rtype" : res_type_str[r_type]),
2073 ticket->t_res_arr[i].r_len);
2078 * Print a summary of the transaction.
2082 struct xfs_trans *tp)
2084 struct xfs_mount *mp = tp->t_mountp;
2085 struct xfs_log_item *lip;
2087 /* dump core transaction and ticket info */
2088 xfs_warn(mp, "transaction summary:");
2089 xfs_warn(mp, " log res = %d", tp->t_log_res);
2090 xfs_warn(mp, " log count = %d", tp->t_log_count);
2091 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2093 xlog_print_tic_res(mp, tp->t_ticket);
2095 /* dump each log item */
2096 list_for_each_entry(lip, &tp->t_items, li_trans) {
2097 struct xfs_log_vec *lv = lip->li_lv;
2098 struct xfs_log_iovec *vec;
2101 xfs_warn(mp, "log item: ");
2102 xfs_warn(mp, " type = 0x%x", lip->li_type);
2103 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2106 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2107 xfs_warn(mp, " size = %d", lv->lv_size);
2108 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2109 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2111 /* dump each iovec for the log item */
2112 vec = lv->lv_iovecp;
2113 for (i = 0; i < lv->lv_niovecs; i++) {
2114 int dumplen = min(vec->i_len, 32);
2116 xfs_warn(mp, " iovec[%d]", i);
2117 xfs_warn(mp, " type = 0x%x", vec->i_type);
2118 xfs_warn(mp, " len = %d", vec->i_len);
2119 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2120 xfs_hex_dump(vec->i_addr, dumplen);
2128 * Calculate the potential space needed by the log vector. We may need a start
2129 * record, and each region gets its own struct xlog_op_header and may need to be
2130 * double word aligned.
2133 xlog_write_calc_vec_length(
2134 struct xlog_ticket *ticket,
2135 struct xfs_log_vec *log_vector,
2138 struct xfs_log_vec *lv;
2143 if (optype & XLOG_START_TRANS)
2146 for (lv = log_vector; lv; lv = lv->lv_next) {
2147 /* we don't write ordered log vectors */
2148 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2151 headers += lv->lv_niovecs;
2153 for (i = 0; i < lv->lv_niovecs; i++) {
2154 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2157 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2161 ticket->t_res_num_ophdrs += headers;
2162 len += headers * sizeof(struct xlog_op_header);
2168 xlog_write_start_rec(
2169 struct xlog_op_header *ophdr,
2170 struct xlog_ticket *ticket)
2172 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2173 ophdr->oh_clientid = ticket->t_clientid;
2175 ophdr->oh_flags = XLOG_START_TRANS;
2179 static xlog_op_header_t *
2180 xlog_write_setup_ophdr(
2182 struct xlog_op_header *ophdr,
2183 struct xlog_ticket *ticket,
2186 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2187 ophdr->oh_clientid = ticket->t_clientid;
2190 /* are we copying a commit or unmount record? */
2191 ophdr->oh_flags = flags;
2194 * We've seen logs corrupted with bad transaction client ids. This
2195 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2196 * and shut down the filesystem.
2198 switch (ophdr->oh_clientid) {
2199 case XFS_TRANSACTION:
2205 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2206 ophdr->oh_clientid, ticket);
2214 * Set up the parameters of the region copy into the log. This has
2215 * to handle region write split across multiple log buffers - this
2216 * state is kept external to this function so that this code can
2217 * be written in an obvious, self documenting manner.
2220 xlog_write_setup_copy(
2221 struct xlog_ticket *ticket,
2222 struct xlog_op_header *ophdr,
2223 int space_available,
2227 int *last_was_partial_copy,
2228 int *bytes_consumed)
2232 still_to_copy = space_required - *bytes_consumed;
2233 *copy_off = *bytes_consumed;
2235 if (still_to_copy <= space_available) {
2236 /* write of region completes here */
2237 *copy_len = still_to_copy;
2238 ophdr->oh_len = cpu_to_be32(*copy_len);
2239 if (*last_was_partial_copy)
2240 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2241 *last_was_partial_copy = 0;
2242 *bytes_consumed = 0;
2246 /* partial write of region, needs extra log op header reservation */
2247 *copy_len = space_available;
2248 ophdr->oh_len = cpu_to_be32(*copy_len);
2249 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2250 if (*last_was_partial_copy)
2251 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2252 *bytes_consumed += *copy_len;
2253 (*last_was_partial_copy)++;
2255 /* account for new log op header */
2256 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2257 ticket->t_res_num_ophdrs++;
2259 return sizeof(struct xlog_op_header);
2263 xlog_write_copy_finish(
2265 struct xlog_in_core *iclog,
2270 int *partial_copy_len,
2272 struct xlog_in_core **commit_iclog)
2276 if (*partial_copy) {
2278 * This iclog has already been marked WANT_SYNC by
2279 * xlog_state_get_iclog_space.
2281 spin_lock(&log->l_icloglock);
2282 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
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 spin_lock(&log->l_icloglock);
2294 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2298 if (iclog->ic_state == XLOG_STATE_ACTIVE)
2299 xlog_state_switch_iclogs(log, iclog, 0);
2301 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2302 iclog->ic_state == XLOG_STATE_IOERROR);
2305 spin_unlock(&log->l_icloglock);
2306 ASSERT(flags & XLOG_COMMIT_TRANS);
2307 *commit_iclog = iclog;
2313 error = xlog_state_release_iclog(log, iclog);
2314 spin_unlock(&log->l_icloglock);
2319 * Write some region out to in-core log
2321 * This will be called when writing externally provided regions or when
2322 * writing out a commit record for a given transaction.
2324 * General algorithm:
2325 * 1. Find total length of this write. This may include adding to the
2326 * lengths passed in.
2327 * 2. Check whether we violate the tickets reservation.
2328 * 3. While writing to this iclog
2329 * A. Reserve as much space in this iclog as can get
2330 * B. If this is first write, save away start lsn
2331 * C. While writing this region:
2332 * 1. If first write of transaction, write start record
2333 * 2. Write log operation header (header per region)
2334 * 3. Find out if we can fit entire region into this iclog
2335 * 4. Potentially, verify destination memcpy ptr
2336 * 5. Memcpy (partial) region
2337 * 6. If partial copy, release iclog; otherwise, continue
2338 * copying more regions into current iclog
2339 * 4. Mark want sync bit (in simulation mode)
2340 * 5. Release iclog for potential flush to on-disk log.
2343 * 1. Panic if reservation is overrun. This should never happen since
2344 * reservation amounts are generated internal to the filesystem.
2346 * 1. Tickets are single threaded data structures.
2347 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2348 * syncing routine. When a single log_write region needs to span
2349 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2350 * on all log operation writes which don't contain the end of the
2351 * region. The XLOG_END_TRANS bit is used for the in-core log
2352 * operation which contains the end of the continued log_write region.
2353 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2354 * we don't really know exactly how much space will be used. As a result,
2355 * we don't update ic_offset until the end when we know exactly how many
2356 * bytes have been written out.
2361 struct xfs_log_vec *log_vector,
2362 struct xlog_ticket *ticket,
2363 xfs_lsn_t *start_lsn,
2364 struct xlog_in_core **commit_iclog,
2367 struct xlog_in_core *iclog = NULL;
2368 struct xfs_log_vec *lv = log_vector;
2369 struct xfs_log_iovec *vecp = lv->lv_iovecp;
2372 int partial_copy = 0;
2373 int partial_copy_len = 0;
2380 * If this is a commit or unmount transaction, we don't need a start
2381 * record to be written. We do, however, have to account for the
2382 * commit or unmount header that gets written. Hence we always have
2383 * to account for an extra xlog_op_header here.
2385 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2386 if (ticket->t_curr_res < 0) {
2387 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2388 "ctx ticket reservation ran out. Need to up reservation");
2389 xlog_print_tic_res(log->l_mp, ticket);
2390 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2393 len = xlog_write_calc_vec_length(ticket, log_vector, optype);
2396 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2400 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2401 &contwr, &log_offset);
2405 ASSERT(log_offset <= iclog->ic_size - 1);
2406 ptr = iclog->ic_datap + log_offset;
2408 /* Start_lsn is the first lsn written to. */
2409 if (start_lsn && !*start_lsn)
2410 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2413 * This loop writes out as many regions as can fit in the amount
2414 * of space which was allocated by xlog_state_get_iclog_space().
2416 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2417 struct xfs_log_iovec *reg;
2418 struct xlog_op_header *ophdr;
2421 bool ordered = false;
2422 bool wrote_start_rec = false;
2424 /* ordered log vectors have no regions to write */
2425 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2426 ASSERT(lv->lv_niovecs == 0);
2432 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2433 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2436 * Before we start formatting log vectors, we need to
2437 * write a start record. Only do this for the first
2438 * iclog we write to.
2440 if (optype & XLOG_START_TRANS) {
2441 xlog_write_start_rec(ptr, ticket);
2442 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2443 sizeof(struct xlog_op_header));
2444 optype &= ~XLOG_START_TRANS;
2445 wrote_start_rec = true;
2448 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, optype);
2452 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2453 sizeof(struct xlog_op_header));
2455 len += xlog_write_setup_copy(ticket, ophdr,
2456 iclog->ic_size-log_offset,
2458 ©_off, ©_len,
2461 xlog_verify_dest_ptr(log, ptr);
2466 * Unmount records just log an opheader, so can have
2467 * empty payloads with no data region to copy. Hence we
2468 * only copy the payload if the vector says it has data
2471 ASSERT(copy_len >= 0);
2473 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2474 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2477 copy_len += sizeof(struct xlog_op_header);
2479 if (wrote_start_rec) {
2480 copy_len += sizeof(struct xlog_op_header);
2483 data_cnt += contwr ? copy_len : 0;
2485 error = xlog_write_copy_finish(log, iclog, optype,
2486 &record_cnt, &data_cnt,
2495 * if we had a partial copy, we need to get more iclog
2496 * space but we don't want to increment the region
2497 * index because there is still more is this region to
2500 * If we completed writing this region, and we flushed
2501 * the iclog (indicated by resetting of the record
2502 * count), then we also need to get more log space. If
2503 * this was the last record, though, we are done and
2509 if (++index == lv->lv_niovecs) {
2514 vecp = lv->lv_iovecp;
2516 if (record_cnt == 0 && !ordered) {
2526 spin_lock(&log->l_icloglock);
2527 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2529 ASSERT(optype & XLOG_COMMIT_TRANS);
2530 *commit_iclog = iclog;
2532 error = xlog_state_release_iclog(log, iclog);
2534 spin_unlock(&log->l_icloglock);
2540 xlog_state_activate_iclog(
2541 struct xlog_in_core *iclog,
2542 int *iclogs_changed)
2544 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2545 trace_xlog_iclog_activate(iclog, _RET_IP_);
2548 * If the number of ops in this iclog indicate it just contains the
2549 * dummy transaction, we can change state into IDLE (the second time
2550 * around). Otherwise we should change the state into NEED a dummy.
2551 * We don't need to cover the dummy.
2553 if (*iclogs_changed == 0 &&
2554 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2555 *iclogs_changed = 1;
2558 * We have two dirty iclogs so start over. This could also be
2559 * num of ops indicating this is not the dummy going out.
2561 *iclogs_changed = 2;
2564 iclog->ic_state = XLOG_STATE_ACTIVE;
2565 iclog->ic_offset = 0;
2566 iclog->ic_header.h_num_logops = 0;
2567 memset(iclog->ic_header.h_cycle_data, 0,
2568 sizeof(iclog->ic_header.h_cycle_data));
2569 iclog->ic_header.h_lsn = 0;
2573 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2574 * ACTIVE after iclog I/O has completed.
2577 xlog_state_activate_iclogs(
2579 int *iclogs_changed)
2581 struct xlog_in_core *iclog = log->l_iclog;
2584 if (iclog->ic_state == XLOG_STATE_DIRTY)
2585 xlog_state_activate_iclog(iclog, iclogs_changed);
2587 * The ordering of marking iclogs ACTIVE must be maintained, so
2588 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2590 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2592 } while ((iclog = iclog->ic_next) != log->l_iclog);
2601 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2602 * wrote the first covering record (DONE). We go to IDLE if we just
2603 * wrote the second covering record (DONE2) and remain in IDLE until a
2604 * non-covering write occurs.
2606 switch (prev_state) {
2607 case XLOG_STATE_COVER_IDLE:
2608 if (iclogs_changed == 1)
2609 return XLOG_STATE_COVER_IDLE;
2611 case XLOG_STATE_COVER_NEED:
2612 case XLOG_STATE_COVER_NEED2:
2614 case XLOG_STATE_COVER_DONE:
2615 if (iclogs_changed == 1)
2616 return XLOG_STATE_COVER_NEED2;
2618 case XLOG_STATE_COVER_DONE2:
2619 if (iclogs_changed == 1)
2620 return XLOG_STATE_COVER_IDLE;
2626 return XLOG_STATE_COVER_NEED;
2630 xlog_state_clean_iclog(
2632 struct xlog_in_core *dirty_iclog)
2634 int iclogs_changed = 0;
2636 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2638 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2640 xlog_state_activate_iclogs(log, &iclogs_changed);
2641 wake_up_all(&dirty_iclog->ic_force_wait);
2643 if (iclogs_changed) {
2644 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2650 xlog_get_lowest_lsn(
2653 struct xlog_in_core *iclog = log->l_iclog;
2654 xfs_lsn_t lowest_lsn = 0, lsn;
2657 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2658 iclog->ic_state == XLOG_STATE_DIRTY)
2661 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2662 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2664 } while ((iclog = iclog->ic_next) != log->l_iclog);
2670 * Completion of a iclog IO does not imply that a transaction has completed, as
2671 * transactions can be large enough to span many iclogs. We cannot change the
2672 * tail of the log half way through a transaction as this may be the only
2673 * transaction in the log and moving the tail to point to the middle of it
2674 * will prevent recovery from finding the start of the transaction. Hence we
2675 * should only update the last_sync_lsn if this iclog contains transaction
2676 * completion callbacks on it.
2678 * We have to do this before we drop the icloglock to ensure we are the only one
2679 * that can update it.
2681 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2682 * the reservation grant head pushing. This is due to the fact that the push
2683 * target is bound by the current last_sync_lsn value. Hence if we have a large
2684 * amount of log space bound up in this committing transaction then the
2685 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2686 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2687 * should push the AIL to ensure the push target (and hence the grant head) is
2688 * no longer bound by the old log head location and can move forwards and make
2692 xlog_state_set_callback(
2694 struct xlog_in_core *iclog,
2695 xfs_lsn_t header_lsn)
2697 trace_xlog_iclog_callback(iclog, _RET_IP_);
2698 iclog->ic_state = XLOG_STATE_CALLBACK;
2700 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2703 if (list_empty_careful(&iclog->ic_callbacks))
2706 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2707 xlog_grant_push_ail(log, 0);
2711 * Return true if we need to stop processing, false to continue to the next
2712 * iclog. The caller will need to run callbacks if the iclog is returned in the
2713 * XLOG_STATE_CALLBACK state.
2716 xlog_state_iodone_process_iclog(
2718 struct xlog_in_core *iclog,
2721 xfs_lsn_t lowest_lsn;
2722 xfs_lsn_t header_lsn;
2724 switch (iclog->ic_state) {
2725 case XLOG_STATE_ACTIVE:
2726 case XLOG_STATE_DIRTY:
2728 * Skip all iclogs in the ACTIVE & DIRTY states:
2731 case XLOG_STATE_IOERROR:
2733 * Between marking a filesystem SHUTDOWN and stopping the log,
2734 * we do flush all iclogs to disk (if there wasn't a log I/O
2735 * error). So, we do want things to go smoothly in case of just
2736 * a SHUTDOWN w/o a LOG_IO_ERROR.
2740 case XLOG_STATE_DONE_SYNC:
2742 * Now that we have an iclog that is in the DONE_SYNC state, do
2743 * one more check here to see if we have chased our tail around.
2744 * If this is not the lowest lsn iclog, then we will leave it
2745 * for another completion to process.
2747 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2748 lowest_lsn = xlog_get_lowest_lsn(log);
2749 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2751 xlog_state_set_callback(log, iclog, header_lsn);
2755 * Can only perform callbacks in order. Since this iclog is not
2756 * in the DONE_SYNC state, we skip the rest and just try to
2764 xlog_state_do_callback(
2767 struct xlog_in_core *iclog;
2768 struct xlog_in_core *first_iclog;
2769 bool cycled_icloglock;
2774 spin_lock(&log->l_icloglock);
2777 * Scan all iclogs starting with the one pointed to by the
2778 * log. Reset this starting point each time the log is
2779 * unlocked (during callbacks).
2781 * Keep looping through iclogs until one full pass is made
2782 * without running any callbacks.
2784 first_iclog = log->l_iclog;
2785 iclog = log->l_iclog;
2786 cycled_icloglock = false;
2793 if (xlog_state_iodone_process_iclog(log, iclog,
2797 if (iclog->ic_state != XLOG_STATE_CALLBACK &&
2798 iclog->ic_state != XLOG_STATE_IOERROR) {
2799 iclog = iclog->ic_next;
2802 list_splice_init(&iclog->ic_callbacks, &cb_list);
2803 spin_unlock(&log->l_icloglock);
2805 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2806 xlog_cil_process_committed(&cb_list);
2807 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2808 cycled_icloglock = true;
2810 spin_lock(&log->l_icloglock);
2811 if (XLOG_FORCED_SHUTDOWN(log))
2812 wake_up_all(&iclog->ic_force_wait);
2814 xlog_state_clean_iclog(log, iclog);
2815 iclog = iclog->ic_next;
2816 } while (first_iclog != iclog);
2818 if (repeats > 5000) {
2819 flushcnt += repeats;
2822 "%s: possible infinite loop (%d iterations)",
2823 __func__, flushcnt);
2825 } while (!ioerror && cycled_icloglock);
2827 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE ||
2828 log->l_iclog->ic_state == XLOG_STATE_IOERROR)
2829 wake_up_all(&log->l_flush_wait);
2831 spin_unlock(&log->l_icloglock);
2836 * Finish transitioning this iclog to the dirty state.
2838 * Make sure that we completely execute this routine only when this is
2839 * the last call to the iclog. There is a good chance that iclog flushes,
2840 * when we reach the end of the physical log, get turned into 2 separate
2841 * calls to bwrite. Hence, one iclog flush could generate two calls to this
2842 * routine. By using the reference count bwritecnt, we guarantee that only
2843 * the second completion goes through.
2845 * Callbacks could take time, so they are done outside the scope of the
2846 * global state machine log lock.
2849 xlog_state_done_syncing(
2850 struct xlog_in_core *iclog)
2852 struct xlog *log = iclog->ic_log;
2854 spin_lock(&log->l_icloglock);
2855 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2856 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2859 * If we got an error, either on the first buffer, or in the case of
2860 * split log writes, on the second, we shut down the file system and
2861 * no iclogs should ever be attempted to be written to disk again.
2863 if (!XLOG_FORCED_SHUTDOWN(log)) {
2864 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2865 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2869 * Someone could be sleeping prior to writing out the next
2870 * iclog buffer, we wake them all, one will get to do the
2871 * I/O, the others get to wait for the result.
2873 wake_up_all(&iclog->ic_write_wait);
2874 spin_unlock(&log->l_icloglock);
2875 xlog_state_do_callback(log);
2879 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2880 * sleep. We wait on the flush queue on the head iclog as that should be
2881 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2882 * we will wait here and all new writes will sleep until a sync completes.
2884 * The in-core logs are used in a circular fashion. They are not used
2885 * out-of-order even when an iclog past the head is free.
2888 * * log_offset where xlog_write() can start writing into the in-core
2890 * * in-core log pointer to which xlog_write() should write.
2891 * * boolean indicating this is a continued write to an in-core log.
2892 * If this is the last write, then the in-core log's offset field
2893 * needs to be incremented, depending on the amount of data which
2897 xlog_state_get_iclog_space(
2900 struct xlog_in_core **iclogp,
2901 struct xlog_ticket *ticket,
2902 int *continued_write,
2906 xlog_rec_header_t *head;
2907 xlog_in_core_t *iclog;
2910 spin_lock(&log->l_icloglock);
2911 if (XLOG_FORCED_SHUTDOWN(log)) {
2912 spin_unlock(&log->l_icloglock);
2916 iclog = log->l_iclog;
2917 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2918 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2920 /* Wait for log writes to have flushed */
2921 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2925 head = &iclog->ic_header;
2927 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
2928 log_offset = iclog->ic_offset;
2930 trace_xlog_iclog_get_space(iclog, _RET_IP_);
2932 /* On the 1st write to an iclog, figure out lsn. This works
2933 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2934 * committing to. If the offset is set, that's how many blocks
2937 if (log_offset == 0) {
2938 ticket->t_curr_res -= log->l_iclog_hsize;
2939 xlog_tic_add_region(ticket,
2941 XLOG_REG_TYPE_LRHEADER);
2942 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2943 head->h_lsn = cpu_to_be64(
2944 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2945 ASSERT(log->l_curr_block >= 0);
2948 /* If there is enough room to write everything, then do it. Otherwise,
2949 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2950 * bit is on, so this will get flushed out. Don't update ic_offset
2951 * until you know exactly how many bytes get copied. Therefore, wait
2952 * until later to update ic_offset.
2954 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2955 * can fit into remaining data section.
2957 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2960 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2963 * If we are the only one writing to this iclog, sync it to
2964 * disk. We need to do an atomic compare and decrement here to
2965 * avoid racing with concurrent atomic_dec_and_lock() calls in
2966 * xlog_state_release_iclog() when there is more than one
2967 * reference to the iclog.
2969 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
2970 error = xlog_state_release_iclog(log, iclog);
2971 spin_unlock(&log->l_icloglock);
2977 /* Do we have enough room to write the full amount in the remainder
2978 * of this iclog? Or must we continue a write on the next iclog and
2979 * mark this iclog as completely taken? In the case where we switch
2980 * iclogs (to mark it taken), this particular iclog will release/sync
2981 * to disk in xlog_write().
2983 if (len <= iclog->ic_size - iclog->ic_offset) {
2984 *continued_write = 0;
2985 iclog->ic_offset += len;
2987 *continued_write = 1;
2988 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2992 ASSERT(iclog->ic_offset <= iclog->ic_size);
2993 spin_unlock(&log->l_icloglock);
2995 *logoffsetp = log_offset;
3000 * The first cnt-1 times a ticket goes through here we don't need to move the
3001 * grant write head because the permanent reservation has reserved cnt times the
3002 * unit amount. Release part of current permanent unit reservation and reset
3003 * current reservation to be one units worth. Also move grant reservation head
3007 xfs_log_ticket_regrant(
3009 struct xlog_ticket *ticket)
3011 trace_xfs_log_ticket_regrant(log, ticket);
3013 if (ticket->t_cnt > 0)
3016 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3017 ticket->t_curr_res);
3018 xlog_grant_sub_space(log, &log->l_write_head.grant,
3019 ticket->t_curr_res);
3020 ticket->t_curr_res = ticket->t_unit_res;
3021 xlog_tic_reset_res(ticket);
3023 trace_xfs_log_ticket_regrant_sub(log, ticket);
3025 /* just return if we still have some of the pre-reserved space */
3026 if (!ticket->t_cnt) {
3027 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3028 ticket->t_unit_res);
3029 trace_xfs_log_ticket_regrant_exit(log, ticket);
3031 ticket->t_curr_res = ticket->t_unit_res;
3032 xlog_tic_reset_res(ticket);
3035 xfs_log_ticket_put(ticket);
3039 * Give back the space left from a reservation.
3041 * All the information we need to make a correct determination of space left
3042 * is present. For non-permanent reservations, things are quite easy. The
3043 * count should have been decremented to zero. We only need to deal with the
3044 * space remaining in the current reservation part of the ticket. If the
3045 * ticket contains a permanent reservation, there may be left over space which
3046 * needs to be released. A count of N means that N-1 refills of the current
3047 * reservation can be done before we need to ask for more space. The first
3048 * one goes to fill up the first current reservation. Once we run out of
3049 * space, the count will stay at zero and the only space remaining will be
3050 * in the current reservation field.
3053 xfs_log_ticket_ungrant(
3055 struct xlog_ticket *ticket)
3059 trace_xfs_log_ticket_ungrant(log, ticket);
3061 if (ticket->t_cnt > 0)
3064 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3067 * If this is a permanent reservation ticket, we may be able to free
3068 * up more space based on the remaining count.
3070 bytes = ticket->t_curr_res;
3071 if (ticket->t_cnt > 0) {
3072 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3073 bytes += ticket->t_unit_res*ticket->t_cnt;
3076 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3077 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3079 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3081 xfs_log_space_wake(log->l_mp);
3082 xfs_log_ticket_put(ticket);
3086 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3087 * the current iclog pointer to the next iclog in the ring.
3090 xlog_state_switch_iclogs(
3092 struct xlog_in_core *iclog,
3095 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3096 assert_spin_locked(&log->l_icloglock);
3097 trace_xlog_iclog_switch(iclog, _RET_IP_);
3100 eventual_size = iclog->ic_offset;
3101 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3102 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3103 log->l_prev_block = log->l_curr_block;
3104 log->l_prev_cycle = log->l_curr_cycle;
3106 /* roll log?: ic_offset changed later */
3107 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3109 /* Round up to next log-sunit */
3110 if (log->l_iclog_roundoff > BBSIZE) {
3111 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3112 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3115 if (log->l_curr_block >= log->l_logBBsize) {
3117 * Rewind the current block before the cycle is bumped to make
3118 * sure that the combined LSN never transiently moves forward
3119 * when the log wraps to the next cycle. This is to support the
3120 * unlocked sample of these fields from xlog_valid_lsn(). Most
3121 * other cases should acquire l_icloglock.
3123 log->l_curr_block -= log->l_logBBsize;
3124 ASSERT(log->l_curr_block >= 0);
3126 log->l_curr_cycle++;
3127 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3128 log->l_curr_cycle++;
3130 ASSERT(iclog == log->l_iclog);
3131 log->l_iclog = iclog->ic_next;
3135 * Write out all data in the in-core log as of this exact moment in time.
3137 * Data may be written to the in-core log during this call. However,
3138 * we don't guarantee this data will be written out. A change from past
3139 * implementation means this routine will *not* write out zero length LRs.
3141 * Basically, we try and perform an intelligent scan of the in-core logs.
3142 * If we determine there is no flushable data, we just return. There is no
3143 * flushable data if:
3145 * 1. the current iclog is active and has no data; the previous iclog
3146 * is in the active or dirty state.
3147 * 2. the current iclog is drity, and the previous iclog is in the
3148 * active or dirty state.
3152 * 1. the current iclog is not in the active nor dirty state.
3153 * 2. the current iclog dirty, and the previous iclog is not in the
3154 * active nor dirty state.
3155 * 3. the current iclog is active, and there is another thread writing
3156 * to this particular iclog.
3157 * 4. a) the current iclog is active and has no other writers
3158 * b) when we return from flushing out this iclog, it is still
3159 * not in the active nor dirty state.
3163 struct xfs_mount *mp,
3166 struct xlog *log = mp->m_log;
3167 struct xlog_in_core *iclog;
3170 XFS_STATS_INC(mp, xs_log_force);
3171 trace_xfs_log_force(mp, 0, _RET_IP_);
3173 xlog_cil_force(log);
3175 spin_lock(&log->l_icloglock);
3176 iclog = log->l_iclog;
3177 if (iclog->ic_state == XLOG_STATE_IOERROR)
3180 trace_xlog_iclog_force(iclog, _RET_IP_);
3182 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3183 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3184 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3186 * If the head is dirty or (active and empty), then we need to
3187 * look at the previous iclog.
3189 * If the previous iclog is active or dirty we are done. There
3190 * is nothing to sync out. Otherwise, we attach ourselves to the
3191 * previous iclog and go to sleep.
3193 iclog = iclog->ic_prev;
3194 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3195 if (atomic_read(&iclog->ic_refcnt) == 0) {
3197 * We are the only one with access to this iclog.
3199 * Flush it out now. There should be a roundoff of zero
3200 * to show that someone has already taken care of the
3201 * roundoff from the previous sync.
3203 atomic_inc(&iclog->ic_refcnt);
3204 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3205 xlog_state_switch_iclogs(log, iclog, 0);
3206 if (xlog_state_release_iclog(log, iclog))
3209 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3213 * Someone else is writing to this iclog.
3215 * Use its call to flush out the data. However, the
3216 * other thread may not force out this LR, so we mark
3219 xlog_state_switch_iclogs(log, iclog, 0);
3223 * If the head iclog is not active nor dirty, we just attach
3224 * ourselves to the head and go to sleep if necessary.
3229 if (flags & XFS_LOG_SYNC)
3230 return xlog_wait_on_iclog(iclog);
3232 spin_unlock(&log->l_icloglock);
3235 spin_unlock(&log->l_icloglock);
3247 struct xlog_in_core *iclog;
3249 spin_lock(&log->l_icloglock);
3250 iclog = log->l_iclog;
3251 if (iclog->ic_state == XLOG_STATE_IOERROR)
3254 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3255 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3256 iclog = iclog->ic_next;
3257 if (iclog == log->l_iclog)
3261 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3263 * We sleep here if we haven't already slept (e.g. this is the
3264 * first time we've looked at the correct iclog buf) and the
3265 * buffer before us is going to be sync'ed. The reason for this
3266 * is that if we are doing sync transactions here, by waiting
3267 * for the previous I/O to complete, we can allow a few more
3268 * transactions into this iclog before we close it down.
3270 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3271 * refcnt so we can release the log (which drops the ref count).
3272 * The state switch keeps new transaction commits from using
3273 * this buffer. When the current commits finish writing into
3274 * the buffer, the refcount will drop to zero and the buffer
3277 if (!already_slept &&
3278 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3279 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3280 xlog_wait(&iclog->ic_prev->ic_write_wait,
3284 atomic_inc(&iclog->ic_refcnt);
3285 xlog_state_switch_iclogs(log, iclog, 0);
3286 if (xlog_state_release_iclog(log, iclog))
3292 if (flags & XFS_LOG_SYNC)
3293 return xlog_wait_on_iclog(iclog);
3295 spin_unlock(&log->l_icloglock);
3298 spin_unlock(&log->l_icloglock);
3303 * Force the in-core log to disk for a specific LSN.
3305 * Find in-core log with lsn.
3306 * If it is in the DIRTY state, just return.
3307 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3308 * state and go to sleep or return.
3309 * If it is in any other state, go to sleep or return.
3311 * Synchronous forces are implemented with a wait queue. All callers trying
3312 * to force a given lsn to disk must wait on the queue attached to the
3313 * specific in-core log. When given in-core log finally completes its write
3314 * to disk, that thread will wake up all threads waiting on the queue.
3318 struct xfs_mount *mp,
3323 struct xlog *log = mp->m_log;
3328 XFS_STATS_INC(mp, xs_log_force);
3329 trace_xfs_log_force(mp, seq, _RET_IP_);
3331 lsn = xlog_cil_force_seq(log, seq);
3332 if (lsn == NULLCOMMITLSN)
3335 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3336 if (ret == -EAGAIN) {
3337 XFS_STATS_INC(mp, xs_log_force_sleep);
3338 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3344 * Free a used ticket when its refcount falls to zero.
3348 xlog_ticket_t *ticket)
3350 ASSERT(atomic_read(&ticket->t_ref) > 0);
3351 if (atomic_dec_and_test(&ticket->t_ref))
3352 kmem_cache_free(xfs_log_ticket_zone, ticket);
3357 xlog_ticket_t *ticket)
3359 ASSERT(atomic_read(&ticket->t_ref) > 0);
3360 atomic_inc(&ticket->t_ref);
3365 * Figure out the total log space unit (in bytes) that would be
3366 * required for a log ticket.
3377 * Permanent reservations have up to 'cnt'-1 active log operations
3378 * in the log. A unit in this case is the amount of space for one
3379 * of these log operations. Normal reservations have a cnt of 1
3380 * and their unit amount is the total amount of space required.
3382 * The following lines of code account for non-transaction data
3383 * which occupy space in the on-disk log.
3385 * Normal form of a transaction is:
3386 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3387 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3389 * We need to account for all the leadup data and trailer data
3390 * around the transaction data.
3391 * And then we need to account for the worst case in terms of using
3393 * The worst case will happen if:
3394 * - the placement of the transaction happens to be such that the
3395 * roundoff is at its maximum
3396 * - the transaction data is synced before the commit record is synced
3397 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3398 * Therefore the commit record is in its own Log Record.
3399 * This can happen as the commit record is called with its
3400 * own region to xlog_write().
3401 * This then means that in the worst case, roundoff can happen for
3402 * the commit-rec as well.
3403 * The commit-rec is smaller than padding in this scenario and so it is
3404 * not added separately.
3407 /* for trans header */
3408 unit_bytes += sizeof(xlog_op_header_t);
3409 unit_bytes += sizeof(xfs_trans_header_t);
3412 unit_bytes += sizeof(xlog_op_header_t);
3415 * for LR headers - the space for data in an iclog is the size minus
3416 * the space used for the headers. If we use the iclog size, then we
3417 * undercalculate the number of headers required.
3419 * Furthermore - the addition of op headers for split-recs might
3420 * increase the space required enough to require more log and op
3421 * headers, so take that into account too.
3423 * IMPORTANT: This reservation makes the assumption that if this
3424 * transaction is the first in an iclog and hence has the LR headers
3425 * accounted to it, then the remaining space in the iclog is
3426 * exclusively for this transaction. i.e. if the transaction is larger
3427 * than the iclog, it will be the only thing in that iclog.
3428 * Fundamentally, this means we must pass the entire log vector to
3429 * xlog_write to guarantee this.
3431 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3432 num_headers = howmany(unit_bytes, iclog_space);
3434 /* for split-recs - ophdrs added when data split over LRs */
3435 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3437 /* add extra header reservations if we overrun */
3438 while (!num_headers ||
3439 howmany(unit_bytes, iclog_space) > num_headers) {
3440 unit_bytes += sizeof(xlog_op_header_t);
3443 unit_bytes += log->l_iclog_hsize * num_headers;
3445 /* for commit-rec LR header - note: padding will subsume the ophdr */
3446 unit_bytes += log->l_iclog_hsize;
3448 /* roundoff padding for transaction data and one for commit record */
3449 unit_bytes += 2 * log->l_iclog_roundoff;
3455 xfs_log_calc_unit_res(
3456 struct xfs_mount *mp,
3459 return xlog_calc_unit_res(mp->m_log, unit_bytes);
3463 * Allocate and initialise a new log ticket.
3465 struct xlog_ticket *
3473 struct xlog_ticket *tic;
3476 tic = kmem_cache_zalloc(xfs_log_ticket_zone, GFP_NOFS | __GFP_NOFAIL);
3478 unit_res = xlog_calc_unit_res(log, unit_bytes);
3480 atomic_set(&tic->t_ref, 1);
3481 tic->t_task = current;
3482 INIT_LIST_HEAD(&tic->t_queue);
3483 tic->t_unit_res = unit_res;
3484 tic->t_curr_res = unit_res;
3487 tic->t_tid = prandom_u32();
3488 tic->t_clientid = client;
3490 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3492 xlog_tic_reset_res(tic);
3499 * Make sure that the destination ptr is within the valid data region of
3500 * one of the iclogs. This uses backup pointers stored in a different
3501 * part of the log in case we trash the log structure.
3504 xlog_verify_dest_ptr(
3511 for (i = 0; i < log->l_iclog_bufs; i++) {
3512 if (ptr >= log->l_iclog_bak[i] &&
3513 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3518 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3522 * Check to make sure the grant write head didn't just over lap the tail. If
3523 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3524 * the cycles differ by exactly one and check the byte count.
3526 * This check is run unlocked, so can give false positives. Rather than assert
3527 * on failures, use a warn-once flag and a panic tag to allow the admin to
3528 * determine if they want to panic the machine when such an error occurs. For
3529 * debug kernels this will have the same effect as using an assert but, unlinke
3530 * an assert, it can be turned off at runtime.
3533 xlog_verify_grant_tail(
3536 int tail_cycle, tail_blocks;
3539 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3540 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3541 if (tail_cycle != cycle) {
3542 if (cycle - 1 != tail_cycle &&
3543 !(log->l_flags & XLOG_TAIL_WARN)) {
3544 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3545 "%s: cycle - 1 != tail_cycle", __func__);
3546 log->l_flags |= XLOG_TAIL_WARN;
3549 if (space > BBTOB(tail_blocks) &&
3550 !(log->l_flags & XLOG_TAIL_WARN)) {
3551 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3552 "%s: space > BBTOB(tail_blocks)", __func__);
3553 log->l_flags |= XLOG_TAIL_WARN;
3558 /* check if it will fit */
3560 xlog_verify_tail_lsn(
3562 struct xlog_in_core *iclog,
3567 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3569 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3570 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3571 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3573 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3575 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3576 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3578 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3579 if (blocks < BTOBB(iclog->ic_offset) + 1)
3580 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3585 * Perform a number of checks on the iclog before writing to disk.
3587 * 1. Make sure the iclogs are still circular
3588 * 2. Make sure we have a good magic number
3589 * 3. Make sure we don't have magic numbers in the data
3590 * 4. Check fields of each log operation header for:
3591 * A. Valid client identifier
3592 * B. tid ptr value falls in valid ptr space (user space code)
3593 * C. Length in log record header is correct according to the
3594 * individual operation headers within record.
3595 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3596 * log, check the preceding blocks of the physical log to make sure all
3597 * the cycle numbers agree with the current cycle number.
3602 struct xlog_in_core *iclog,
3605 xlog_op_header_t *ophead;
3606 xlog_in_core_t *icptr;
3607 xlog_in_core_2_t *xhdr;
3608 void *base_ptr, *ptr, *p;
3609 ptrdiff_t field_offset;
3611 int len, i, j, k, op_len;
3614 /* check validity of iclog pointers */
3615 spin_lock(&log->l_icloglock);
3616 icptr = log->l_iclog;
3617 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3620 if (icptr != log->l_iclog)
3621 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3622 spin_unlock(&log->l_icloglock);
3624 /* check log magic numbers */
3625 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3626 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3628 base_ptr = ptr = &iclog->ic_header;
3629 p = &iclog->ic_header;
3630 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3631 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3632 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3637 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3638 base_ptr = ptr = iclog->ic_datap;
3640 xhdr = iclog->ic_data;
3641 for (i = 0; i < len; i++) {
3644 /* clientid is only 1 byte */
3645 p = &ophead->oh_clientid;
3646 field_offset = p - base_ptr;
3647 if (field_offset & 0x1ff) {
3648 clientid = ophead->oh_clientid;
3650 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3651 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3652 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3653 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3654 clientid = xlog_get_client_id(
3655 xhdr[j].hic_xheader.xh_cycle_data[k]);
3657 clientid = xlog_get_client_id(
3658 iclog->ic_header.h_cycle_data[idx]);
3661 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3663 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3664 __func__, clientid, ophead,
3665 (unsigned long)field_offset);
3668 p = &ophead->oh_len;
3669 field_offset = p - base_ptr;
3670 if (field_offset & 0x1ff) {
3671 op_len = be32_to_cpu(ophead->oh_len);
3673 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3674 (uintptr_t)iclog->ic_datap);
3675 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3676 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3677 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3678 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3680 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3683 ptr += sizeof(xlog_op_header_t) + op_len;
3689 * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3695 xlog_in_core_t *iclog, *ic;
3697 iclog = log->l_iclog;
3698 if (iclog->ic_state != XLOG_STATE_IOERROR) {
3700 * Mark all the incore logs IOERROR.
3701 * From now on, no log flushes will result.
3705 ic->ic_state = XLOG_STATE_IOERROR;
3707 } while (ic != iclog);
3711 * Return non-zero, if state transition has already happened.
3717 * This is called from xfs_force_shutdown, when we're forcibly
3718 * shutting down the filesystem, typically because of an IO error.
3719 * Our main objectives here are to make sure that:
3720 * a. if !logerror, flush the logs to disk. Anything modified
3721 * after this is ignored.
3722 * b. the filesystem gets marked 'SHUTDOWN' for all interested
3723 * parties to find out, 'atomically'.
3724 * c. those who're sleeping on log reservations, pinned objects and
3725 * other resources get woken up, and be told the bad news.
3726 * d. nothing new gets queued up after (b) and (c) are done.
3728 * Note: for the !logerror case we need to flush the regions held in memory out
3729 * to disk first. This needs to be done before the log is marked as shutdown,
3730 * otherwise the iclog writes will fail.
3733 xfs_log_force_umount(
3734 struct xfs_mount *mp,
3743 * If this happens during log recovery, don't worry about
3744 * locking; the log isn't open for business yet.
3747 log->l_flags & XLOG_ACTIVE_RECOVERY) {
3748 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3750 mp->m_sb_bp->b_flags |= XBF_DONE;
3755 * Somebody could've already done the hard work for us.
3756 * No need to get locks for this.
3758 if (logerror && log->l_iclog->ic_state == XLOG_STATE_IOERROR) {
3759 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3764 * Flush all the completed transactions to disk before marking the log
3765 * being shut down. We need to do it in this order to ensure that
3766 * completed operations are safely on disk before we shut down, and that
3767 * we don't have to issue any buffer IO after the shutdown flags are set
3768 * to guarantee this.
3771 xfs_log_force(mp, XFS_LOG_SYNC);
3774 * mark the filesystem and the as in a shutdown state and wake
3775 * everybody up to tell them the bad news.
3777 spin_lock(&log->l_icloglock);
3778 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3780 mp->m_sb_bp->b_flags |= XBF_DONE;
3783 * Mark the log and the iclogs with IO error flags to prevent any
3784 * further log IO from being issued or completed.
3786 log->l_flags |= XLOG_IO_ERROR;
3787 retval = xlog_state_ioerror(log);
3788 spin_unlock(&log->l_icloglock);
3791 * We don't want anybody waiting for log reservations after this. That
3792 * means we have to wake up everybody queued up on reserveq as well as
3793 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3794 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3795 * action is protected by the grant locks.
3797 xlog_grant_head_wake_all(&log->l_reserve_head);
3798 xlog_grant_head_wake_all(&log->l_write_head);
3801 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3802 * as if the log writes were completed. The abort handling in the log
3803 * item committed callback functions will do this again under lock to
3806 spin_lock(&log->l_cilp->xc_push_lock);
3807 wake_up_all(&log->l_cilp->xc_commit_wait);
3808 spin_unlock(&log->l_cilp->xc_push_lock);
3809 xlog_state_do_callback(log);
3811 /* return non-zero if log IOERROR transition had already happened */
3819 xlog_in_core_t *iclog;
3821 iclog = log->l_iclog;
3823 /* endianness does not matter here, zero is zero in
3826 if (iclog->ic_header.h_num_logops)
3828 iclog = iclog->ic_next;
3829 } while (iclog != log->l_iclog);
3834 * Verify that an LSN stamped into a piece of metadata is valid. This is
3835 * intended for use in read verifiers on v5 superblocks.
3839 struct xfs_mount *mp,
3842 struct xlog *log = mp->m_log;
3846 * norecovery mode skips mount-time log processing and unconditionally
3847 * resets the in-core LSN. We can't validate in this mode, but
3848 * modifications are not allowed anyways so just return true.
3850 if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3854 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3855 * handled by recovery and thus safe to ignore here.
3857 if (lsn == NULLCOMMITLSN)
3860 valid = xlog_valid_lsn(mp->m_log, lsn);
3862 /* warn the user about what's gone wrong before verifier failure */
3864 spin_lock(&log->l_icloglock);
3866 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3867 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3868 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3869 log->l_curr_cycle, log->l_curr_block);
3870 spin_unlock(&log->l_icloglock);
3877 xfs_log_in_recovery(
3878 struct xfs_mount *mp)
3880 struct xlog *log = mp->m_log;
3882 return log->l_flags & XLOG_ACTIVE_RECOVERY;