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);
44 STATIC void xlog_state_do_callback(
47 xlog_state_get_iclog_space(
50 struct xlog_in_core **iclog,
51 struct xlog_ticket *ticket,
61 struct xlog_in_core *iclog);
68 xlog_verify_grant_tail(
73 struct xlog_in_core *iclog,
78 struct xlog_in_core *iclog);
80 #define xlog_verify_dest_ptr(a,b)
81 #define xlog_verify_grant_tail(a)
82 #define xlog_verify_iclog(a,b,c)
83 #define xlog_verify_tail_lsn(a,b)
91 xfs_log_cover(struct xfs_mount *);
99 int64_t head_val = atomic64_read(head);
105 xlog_crack_grant_head_val(head_val, &cycle, &space);
109 space += log->l_logsize;
114 new = xlog_assign_grant_head_val(cycle, space);
115 head_val = atomic64_cmpxchg(head, old, new);
116 } while (head_val != old);
120 xlog_grant_add_space(
125 int64_t head_val = atomic64_read(head);
132 xlog_crack_grant_head_val(head_val, &cycle, &space);
134 tmp = log->l_logsize - space;
143 new = xlog_assign_grant_head_val(cycle, space);
144 head_val = atomic64_cmpxchg(head, old, new);
145 } while (head_val != old);
149 xlog_grant_head_init(
150 struct xlog_grant_head *head)
152 xlog_assign_grant_head(&head->grant, 1, 0);
153 INIT_LIST_HEAD(&head->waiters);
154 spin_lock_init(&head->lock);
158 xlog_grant_head_wake_all(
159 struct xlog_grant_head *head)
161 struct xlog_ticket *tic;
163 spin_lock(&head->lock);
164 list_for_each_entry(tic, &head->waiters, t_queue)
165 wake_up_process(tic->t_task);
166 spin_unlock(&head->lock);
170 xlog_ticket_reservation(
172 struct xlog_grant_head *head,
173 struct xlog_ticket *tic)
175 if (head == &log->l_write_head) {
176 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
177 return tic->t_unit_res;
179 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
180 return tic->t_unit_res * tic->t_cnt;
182 return tic->t_unit_res;
187 xlog_grant_head_wake(
189 struct xlog_grant_head *head,
192 struct xlog_ticket *tic;
194 bool woken_task = false;
196 list_for_each_entry(tic, &head->waiters, t_queue) {
199 * There is a chance that the size of the CIL checkpoints in
200 * progress at the last AIL push target calculation resulted in
201 * limiting the target to the log head (l_last_sync_lsn) at the
202 * time. This may not reflect where the log head is now as the
203 * CIL checkpoints may have completed.
205 * Hence when we are woken here, it may be that the head of the
206 * log that has moved rather than the tail. As the tail didn't
207 * move, there still won't be space available for the
208 * reservation we require. However, if the AIL has already
209 * pushed to the target defined by the old log head location, we
210 * will hang here waiting for something else to update the AIL
213 * Therefore, if there isn't space to wake the first waiter on
214 * the grant head, we need to push the AIL again to ensure the
215 * target reflects both the current log tail and log head
216 * position before we wait for the tail to move again.
219 need_bytes = xlog_ticket_reservation(log, head, tic);
220 if (*free_bytes < need_bytes) {
222 xlog_grant_push_ail(log, need_bytes);
226 *free_bytes -= need_bytes;
227 trace_xfs_log_grant_wake_up(log, tic);
228 wake_up_process(tic->t_task);
236 xlog_grant_head_wait(
238 struct xlog_grant_head *head,
239 struct xlog_ticket *tic,
240 int need_bytes) __releases(&head->lock)
241 __acquires(&head->lock)
243 list_add_tail(&tic->t_queue, &head->waiters);
246 if (xlog_is_shutdown(log))
248 xlog_grant_push_ail(log, need_bytes);
250 __set_current_state(TASK_UNINTERRUPTIBLE);
251 spin_unlock(&head->lock);
253 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
255 trace_xfs_log_grant_sleep(log, tic);
257 trace_xfs_log_grant_wake(log, tic);
259 spin_lock(&head->lock);
260 if (xlog_is_shutdown(log))
262 } while (xlog_space_left(log, &head->grant) < need_bytes);
264 list_del_init(&tic->t_queue);
267 list_del_init(&tic->t_queue);
272 * Atomically get the log space required for a log ticket.
274 * Once a ticket gets put onto head->waiters, it will only return after the
275 * needed reservation is satisfied.
277 * This function is structured so that it has a lock free fast path. This is
278 * necessary because every new transaction reservation will come through this
279 * path. Hence any lock will be globally hot if we take it unconditionally on
282 * As tickets are only ever moved on and off head->waiters under head->lock, we
283 * only need to take that lock if we are going to add the ticket to the queue
284 * and sleep. We can avoid taking the lock if the ticket was never added to
285 * head->waiters because the t_queue list head will be empty and we hold the
286 * only reference to it so it can safely be checked unlocked.
289 xlog_grant_head_check(
291 struct xlog_grant_head *head,
292 struct xlog_ticket *tic,
298 ASSERT(!xlog_in_recovery(log));
301 * If there are other waiters on the queue then give them a chance at
302 * logspace before us. Wake up the first waiters, if we do not wake
303 * up all the waiters then go to sleep waiting for more free space,
304 * otherwise try to get some space for this transaction.
306 *need_bytes = xlog_ticket_reservation(log, head, tic);
307 free_bytes = xlog_space_left(log, &head->grant);
308 if (!list_empty_careful(&head->waiters)) {
309 spin_lock(&head->lock);
310 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
311 free_bytes < *need_bytes) {
312 error = xlog_grant_head_wait(log, head, tic,
315 spin_unlock(&head->lock);
316 } else if (free_bytes < *need_bytes) {
317 spin_lock(&head->lock);
318 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
319 spin_unlock(&head->lock);
326 xlog_tic_reset_res(xlog_ticket_t *tic)
329 tic->t_res_arr_sum = 0;
330 tic->t_res_num_ophdrs = 0;
334 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
336 if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
337 /* add to overflow and start again */
338 tic->t_res_o_flow += tic->t_res_arr_sum;
340 tic->t_res_arr_sum = 0;
343 tic->t_res_arr[tic->t_res_num].r_len = len;
344 tic->t_res_arr[tic->t_res_num].r_type = type;
345 tic->t_res_arr_sum += len;
351 struct xfs_mount *mp)
354 * Do not write to the log on norecovery mounts, if the data or log
355 * devices are read-only, or if the filesystem is shutdown. Read-only
356 * mounts allow internal writes for log recovery and unmount purposes,
357 * so don't restrict that case.
359 if (xfs_has_norecovery(mp))
361 if (xfs_readonly_buftarg(mp->m_ddev_targp))
363 if (xfs_readonly_buftarg(mp->m_log->l_targ))
365 if (xlog_is_shutdown(mp->m_log))
371 * Replenish the byte reservation required by moving the grant write head.
375 struct xfs_mount *mp,
376 struct xlog_ticket *tic)
378 struct xlog *log = mp->m_log;
382 if (xlog_is_shutdown(log))
385 XFS_STATS_INC(mp, xs_try_logspace);
388 * This is a new transaction on the ticket, so we need to change the
389 * transaction ID so that the next transaction has a different TID in
390 * the log. Just add one to the existing tid so that we can see chains
391 * of rolling transactions in the log easily.
395 xlog_grant_push_ail(log, tic->t_unit_res);
397 tic->t_curr_res = tic->t_unit_res;
398 xlog_tic_reset_res(tic);
403 trace_xfs_log_regrant(log, tic);
405 error = xlog_grant_head_check(log, &log->l_write_head, tic,
410 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
411 trace_xfs_log_regrant_exit(log, tic);
412 xlog_verify_grant_tail(log);
417 * If we are failing, make sure the ticket doesn't have any current
418 * reservations. We don't want to add this back when the ticket/
419 * transaction gets cancelled.
422 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
427 * Reserve log space and return a ticket corresponding to the reservation.
429 * Each reservation is going to reserve extra space for a log record header.
430 * When writes happen to the on-disk log, we don't subtract the length of the
431 * log record header from any reservation. By wasting space in each
432 * reservation, we prevent over allocation problems.
436 struct xfs_mount *mp,
439 struct xlog_ticket **ticp,
443 struct xlog *log = mp->m_log;
444 struct xlog_ticket *tic;
448 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
450 if (xlog_is_shutdown(log))
453 XFS_STATS_INC(mp, xs_try_logspace);
455 ASSERT(*ticp == NULL);
456 tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent);
459 xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
462 trace_xfs_log_reserve(log, tic);
464 error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
469 xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
470 xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
471 trace_xfs_log_reserve_exit(log, tic);
472 xlog_verify_grant_tail(log);
477 * If we are failing, make sure the ticket doesn't have any current
478 * reservations. We don't want to add this back when the ticket/
479 * transaction gets cancelled.
482 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
487 * Run all the pending iclog callbacks and wake log force waiters and iclog
488 * space waiters so they can process the newly set shutdown state. We really
489 * don't care what order we process callbacks here because the log is shut down
490 * and so state cannot change on disk anymore. However, we cannot wake waiters
491 * until the callbacks have been processed because we may be in unmount and
492 * we must ensure that all AIL operations the callbacks perform have completed
493 * before we tear down the AIL.
495 * We avoid processing actively referenced iclogs so that we don't run callbacks
496 * while the iclog owner might still be preparing the iclog for IO submssion.
497 * These will be caught by xlog_state_iclog_release() and call this function
498 * again to process any callbacks that may have been added to that iclog.
501 xlog_state_shutdown_callbacks(
504 struct xlog_in_core *iclog;
507 iclog = log->l_iclog;
509 if (atomic_read(&iclog->ic_refcnt)) {
510 /* Reference holder will re-run iclog callbacks. */
513 list_splice_init(&iclog->ic_callbacks, &cb_list);
514 spin_unlock(&log->l_icloglock);
516 xlog_cil_process_committed(&cb_list);
518 spin_lock(&log->l_icloglock);
519 wake_up_all(&iclog->ic_write_wait);
520 wake_up_all(&iclog->ic_force_wait);
521 } while ((iclog = iclog->ic_next) != log->l_iclog);
523 wake_up_all(&log->l_flush_wait);
527 * Flush iclog to disk if this is the last reference to the given iclog and the
528 * it is in the WANT_SYNC state.
530 * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
531 * log tail is updated correctly. NEED_FUA indicates that the iclog will be
532 * written to stable storage, and implies that a commit record is contained
533 * within the iclog. We need to ensure that the log tail does not move beyond
534 * the tail that the first commit record in the iclog ordered against, otherwise
535 * correct recovery of that checkpoint becomes dependent on future operations
536 * performed on this iclog.
538 * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
539 * current tail into iclog. Once the iclog tail is set, future operations must
540 * not modify it, otherwise they potentially violate ordering constraints for
541 * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
542 * the iclog will get zeroed on activation of the iclog after sync, so we
543 * always capture the tail lsn on the iclog on the first NEED_FUA release
544 * regardless of the number of active reference counts on this iclog.
547 xlog_state_release_iclog(
549 struct xlog_in_core *iclog)
554 lockdep_assert_held(&log->l_icloglock);
556 trace_xlog_iclog_release(iclog, _RET_IP_);
558 * Grabbing the current log tail needs to be atomic w.r.t. the writing
559 * of the tail LSN into the iclog so we guarantee that the log tail does
560 * not move between the first time we know that the iclog needs to be
561 * made stable and when we eventually submit it.
563 if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
564 (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
565 !iclog->ic_header.h_tail_lsn) {
566 tail_lsn = xlog_assign_tail_lsn(log->l_mp);
567 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
570 last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
572 if (xlog_is_shutdown(log)) {
574 * If there are no more references to this iclog, process the
575 * pending iclog callbacks that were waiting on the release of
579 xlog_state_shutdown_callbacks(log);
586 if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
587 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
591 iclog->ic_state = XLOG_STATE_SYNCING;
592 xlog_verify_tail_lsn(log, iclog);
593 trace_xlog_iclog_syncing(iclog, _RET_IP_);
595 spin_unlock(&log->l_icloglock);
596 xlog_sync(log, iclog);
597 spin_lock(&log->l_icloglock);
602 * Mount a log filesystem
604 * mp - ubiquitous xfs mount point structure
605 * log_target - buftarg of on-disk log device
606 * blk_offset - Start block # where block size is 512 bytes (BBSIZE)
607 * num_bblocks - Number of BBSIZE blocks in on-disk log
609 * Return error or zero.
614 xfs_buftarg_t *log_target,
615 xfs_daddr_t blk_offset,
619 bool fatal = xfs_has_crc(mp);
623 if (!xfs_has_norecovery(mp)) {
624 xfs_notice(mp, "Mounting V%d Filesystem",
625 XFS_SB_VERSION_NUM(&mp->m_sb));
628 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
629 XFS_SB_VERSION_NUM(&mp->m_sb));
630 ASSERT(xfs_is_readonly(mp));
633 log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
635 error = PTR_ERR(log);
641 * Validate the given log space and drop a critical message via syslog
642 * if the log size is too small that would lead to some unexpected
643 * situations in transaction log space reservation stage.
645 * Note: we can't just reject the mount if the validation fails. This
646 * would mean that people would have to downgrade their kernel just to
647 * remedy the situation as there is no way to grow the log (short of
648 * black magic surgery with xfs_db).
650 * We can, however, reject mounts for CRC format filesystems, as the
651 * mkfs binary being used to make the filesystem should never create a
652 * filesystem with a log that is too small.
654 min_logfsbs = xfs_log_calc_minimum_size(mp);
656 if (mp->m_sb.sb_logblocks < min_logfsbs) {
658 "Log size %d blocks too small, minimum size is %d blocks",
659 mp->m_sb.sb_logblocks, min_logfsbs);
661 } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
663 "Log size %d blocks too large, maximum size is %lld blocks",
664 mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
666 } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
668 "log size %lld bytes too large, maximum size is %lld bytes",
669 XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
672 } else if (mp->m_sb.sb_logsunit > 1 &&
673 mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
675 "log stripe unit %u bytes must be a multiple of block size",
676 mp->m_sb.sb_logsunit);
682 * Log check errors are always fatal on v5; or whenever bad
683 * metadata leads to a crash.
686 xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
690 xfs_crit(mp, "Log size out of supported range.");
692 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
696 * Initialize the AIL now we have a log.
698 error = xfs_trans_ail_init(mp);
700 xfs_warn(mp, "AIL initialisation failed: error %d", error);
703 log->l_ailp = mp->m_ail;
706 * skip log recovery on a norecovery mount. pretend it all
709 if (!xfs_has_norecovery(mp)) {
711 * log recovery ignores readonly state and so we need to clear
712 * mount-based read only state so it can write to disk.
714 bool readonly = test_and_clear_bit(XFS_OPSTATE_READONLY,
716 error = xlog_recover(log);
718 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
720 xfs_warn(mp, "log mount/recovery failed: error %d",
722 xlog_recover_cancel(log);
723 goto out_destroy_ail;
727 error = xfs_sysfs_init(&log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
730 goto out_destroy_ail;
732 /* Normal transactions can now occur */
733 clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
736 * Now the log has been fully initialised and we know were our
737 * space grant counters are, we can initialise the permanent ticket
738 * needed for delayed logging to work.
740 xlog_cil_init_post_recovery(log);
745 xfs_trans_ail_destroy(mp);
747 xlog_dealloc_log(log);
753 * Finish the recovery of the file system. This is separate from the
754 * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
755 * in the root and real-time bitmap inodes between calling xfs_log_mount() and
758 * If we finish recovery successfully, start the background log work. If we are
759 * not doing recovery, then we have a RO filesystem and we don't need to start
763 xfs_log_mount_finish(
764 struct xfs_mount *mp)
766 struct xlog *log = mp->m_log;
770 if (xfs_has_norecovery(mp)) {
771 ASSERT(xfs_is_readonly(mp));
776 * log recovery ignores readonly state and so we need to clear
777 * mount-based read only state so it can write to disk.
779 readonly = test_and_clear_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
782 * During the second phase of log recovery, we need iget and
783 * iput to behave like they do for an active filesystem.
784 * xfs_fs_drop_inode needs to be able to prevent the deletion
785 * of inodes before we're done replaying log items on those
786 * inodes. Turn it off immediately after recovery finishes
787 * so that we don't leak the quota inodes if subsequent mount
790 * We let all inodes involved in redo item processing end up on
791 * the LRU instead of being evicted immediately so that if we do
792 * something to an unlinked inode, the irele won't cause
793 * premature truncation and freeing of the inode, which results
794 * in log recovery failure. We have to evict the unreferenced
795 * lru inodes after clearing SB_ACTIVE because we don't
796 * otherwise clean up the lru if there's a subsequent failure in
797 * xfs_mountfs, which leads to us leaking the inodes if nothing
798 * else (e.g. quotacheck) references the inodes before the
799 * mount failure occurs.
801 mp->m_super->s_flags |= SB_ACTIVE;
802 if (xlog_recovery_needed(log))
803 error = xlog_recover_finish(log);
805 xfs_log_work_queue(mp);
806 mp->m_super->s_flags &= ~SB_ACTIVE;
807 evict_inodes(mp->m_super);
810 * Drain the buffer LRU after log recovery. This is required for v4
811 * filesystems to avoid leaving around buffers with NULL verifier ops,
812 * but we do it unconditionally to make sure we're always in a clean
813 * cache state after mount.
815 * Don't push in the error case because the AIL may have pending intents
816 * that aren't removed until recovery is cancelled.
818 if (xlog_recovery_needed(log)) {
820 xfs_log_force(mp, XFS_LOG_SYNC);
821 xfs_ail_push_all_sync(mp->m_ail);
823 xfs_notice(mp, "Ending recovery (logdev: %s)",
824 mp->m_logname ? mp->m_logname : "internal");
826 xfs_info(mp, "Ending clean mount");
828 xfs_buftarg_drain(mp->m_ddev_targp);
830 clear_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
832 set_bit(XFS_OPSTATE_READONLY, &mp->m_opstate);
834 /* Make sure the log is dead if we're returning failure. */
835 ASSERT(!error || xlog_is_shutdown(log));
841 * The mount has failed. Cancel the recovery if it hasn't completed and destroy
845 xfs_log_mount_cancel(
846 struct xfs_mount *mp)
848 xlog_recover_cancel(mp->m_log);
853 * Flush out the iclog to disk ensuring that device caches are flushed and
854 * the iclog hits stable storage before any completion waiters are woken.
858 struct xlog_in_core *iclog)
860 atomic_inc(&iclog->ic_refcnt);
861 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
862 if (iclog->ic_state == XLOG_STATE_ACTIVE)
863 xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
864 return xlog_state_release_iclog(iclog->ic_log, iclog);
868 * Wait for the iclog and all prior iclogs to be written disk as required by the
869 * log force state machine. Waiting on ic_force_wait ensures iclog completions
870 * have been ordered and callbacks run before we are woken here, hence
871 * guaranteeing that all the iclogs up to this one are on stable storage.
875 struct xlog_in_core *iclog)
876 __releases(iclog->ic_log->l_icloglock)
878 struct xlog *log = iclog->ic_log;
880 trace_xlog_iclog_wait_on(iclog, _RET_IP_);
881 if (!xlog_is_shutdown(log) &&
882 iclog->ic_state != XLOG_STATE_ACTIVE &&
883 iclog->ic_state != XLOG_STATE_DIRTY) {
884 XFS_STATS_INC(log->l_mp, xs_log_force_sleep);
885 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
887 spin_unlock(&log->l_icloglock);
890 if (xlog_is_shutdown(log))
896 * Write out an unmount record using the ticket provided. We have to account for
897 * the data space used in the unmount ticket as this write is not done from a
898 * transaction context that has already done the accounting for us.
901 xlog_write_unmount_record(
903 struct xlog_ticket *ticket)
905 struct xfs_unmount_log_format ulf = {
906 .magic = XLOG_UNMOUNT_TYPE,
908 struct xfs_log_iovec reg = {
910 .i_len = sizeof(ulf),
911 .i_type = XLOG_REG_TYPE_UNMOUNT,
913 struct xfs_log_vec vec = {
918 /* account for space used by record data */
919 ticket->t_curr_res -= sizeof(ulf);
921 return xlog_write(log, NULL, &vec, ticket, XLOG_UNMOUNT_TRANS);
925 * Mark the filesystem clean by writing an unmount record to the head of the
932 struct xfs_mount *mp = log->l_mp;
933 struct xlog_in_core *iclog;
934 struct xlog_ticket *tic = NULL;
937 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
941 error = xlog_write_unmount_record(log, tic);
943 * At this point, we're umounting anyway, so there's no point in
944 * transitioning log state to shutdown. Just continue...
948 xfs_alert(mp, "%s: unmount record failed", __func__);
950 spin_lock(&log->l_icloglock);
951 iclog = log->l_iclog;
952 error = xlog_force_iclog(iclog);
953 xlog_wait_on_iclog(iclog);
956 trace_xfs_log_umount_write(log, tic);
957 xfs_log_ticket_ungrant(log, tic);
962 xfs_log_unmount_verify_iclog(
965 struct xlog_in_core *iclog = log->l_iclog;
968 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
969 ASSERT(iclog->ic_offset == 0);
970 } while ((iclog = iclog->ic_next) != log->l_iclog);
974 * Unmount record used to have a string "Unmount filesystem--" in the
975 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
976 * We just write the magic number now since that particular field isn't
977 * currently architecture converted and "Unmount" is a bit foo.
978 * As far as I know, there weren't any dependencies on the old behaviour.
981 xfs_log_unmount_write(
982 struct xfs_mount *mp)
984 struct xlog *log = mp->m_log;
986 if (!xfs_log_writable(mp))
989 xfs_log_force(mp, XFS_LOG_SYNC);
991 if (xlog_is_shutdown(log))
995 * If we think the summary counters are bad, avoid writing the unmount
996 * record to force log recovery at next mount, after which the summary
997 * counters will be recalculated. Refer to xlog_check_unmount_rec for
1000 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
1001 XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
1002 xfs_alert(mp, "%s: will fix summary counters at next mount",
1007 xfs_log_unmount_verify_iclog(log);
1008 xlog_unmount_write(log);
1012 * Empty the log for unmount/freeze.
1014 * To do this, we first need to shut down the background log work so it is not
1015 * trying to cover the log as we clean up. We then need to unpin all objects in
1016 * the log so we can then flush them out. Once they have completed their IO and
1017 * run the callbacks removing themselves from the AIL, we can cover the log.
1021 struct xfs_mount *mp)
1024 * Clear log incompat features since we're quiescing the log. Report
1025 * failures, though it's not fatal to have a higher log feature
1026 * protection level than the log contents actually require.
1028 if (xfs_clear_incompat_log_features(mp)) {
1031 error = xfs_sync_sb(mp, false);
1034 "Failed to clear log incompat features on quiesce");
1037 cancel_delayed_work_sync(&mp->m_log->l_work);
1038 xfs_log_force(mp, XFS_LOG_SYNC);
1041 * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1042 * will push it, xfs_buftarg_wait() will not wait for it. Further,
1043 * xfs_buf_iowait() cannot be used because it was pushed with the
1044 * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1045 * the IO to complete.
1047 xfs_ail_push_all_sync(mp->m_ail);
1048 xfs_buftarg_wait(mp->m_ddev_targp);
1049 xfs_buf_lock(mp->m_sb_bp);
1050 xfs_buf_unlock(mp->m_sb_bp);
1052 return xfs_log_cover(mp);
1057 struct xfs_mount *mp)
1059 xfs_log_quiesce(mp);
1060 xfs_log_unmount_write(mp);
1064 * Shut down and release the AIL and Log.
1066 * During unmount, we need to ensure we flush all the dirty metadata objects
1067 * from the AIL so that the log is empty before we write the unmount record to
1068 * the log. Once this is done, we can tear down the AIL and the log.
1072 struct xfs_mount *mp)
1076 xfs_buftarg_drain(mp->m_ddev_targp);
1078 xfs_trans_ail_destroy(mp);
1080 xfs_sysfs_del(&mp->m_log->l_kobj);
1082 xlog_dealloc_log(mp->m_log);
1087 struct xfs_mount *mp,
1088 struct xfs_log_item *item,
1090 const struct xfs_item_ops *ops)
1092 item->li_mountp = mp;
1093 item->li_ailp = mp->m_ail;
1094 item->li_type = type;
1098 INIT_LIST_HEAD(&item->li_ail);
1099 INIT_LIST_HEAD(&item->li_cil);
1100 INIT_LIST_HEAD(&item->li_bio_list);
1101 INIT_LIST_HEAD(&item->li_trans);
1105 * Wake up processes waiting for log space after we have moved the log tail.
1109 struct xfs_mount *mp)
1111 struct xlog *log = mp->m_log;
1114 if (xlog_is_shutdown(log))
1117 if (!list_empty_careful(&log->l_write_head.waiters)) {
1118 ASSERT(!xlog_in_recovery(log));
1120 spin_lock(&log->l_write_head.lock);
1121 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1122 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1123 spin_unlock(&log->l_write_head.lock);
1126 if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1127 ASSERT(!xlog_in_recovery(log));
1129 spin_lock(&log->l_reserve_head.lock);
1130 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1131 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1132 spin_unlock(&log->l_reserve_head.lock);
1137 * Determine if we have a transaction that has gone to disk that needs to be
1138 * covered. To begin the transition to the idle state firstly the log needs to
1139 * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1140 * we start attempting to cover the log.
1142 * Only if we are then in a state where covering is needed, the caller is
1143 * informed that dummy transactions are required to move the log into the idle
1146 * If there are any items in the AIl or CIL, then we do not want to attempt to
1147 * cover the log as we may be in a situation where there isn't log space
1148 * available to run a dummy transaction and this can lead to deadlocks when the
1149 * tail of the log is pinned by an item that is modified in the CIL. Hence
1150 * there's no point in running a dummy transaction at this point because we
1151 * can't start trying to idle the log until both the CIL and AIL are empty.
1154 xfs_log_need_covered(
1155 struct xfs_mount *mp)
1157 struct xlog *log = mp->m_log;
1158 bool needed = false;
1160 if (!xlog_cil_empty(log))
1163 spin_lock(&log->l_icloglock);
1164 switch (log->l_covered_state) {
1165 case XLOG_STATE_COVER_DONE:
1166 case XLOG_STATE_COVER_DONE2:
1167 case XLOG_STATE_COVER_IDLE:
1169 case XLOG_STATE_COVER_NEED:
1170 case XLOG_STATE_COVER_NEED2:
1171 if (xfs_ail_min_lsn(log->l_ailp))
1173 if (!xlog_iclogs_empty(log))
1177 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1178 log->l_covered_state = XLOG_STATE_COVER_DONE;
1180 log->l_covered_state = XLOG_STATE_COVER_DONE2;
1186 spin_unlock(&log->l_icloglock);
1191 * Explicitly cover the log. This is similar to background log covering but
1192 * intended for usage in quiesce codepaths. The caller is responsible to ensure
1193 * the log is idle and suitable for covering. The CIL, iclog buffers and AIL
1194 * must all be empty.
1198 struct xfs_mount *mp)
1203 ASSERT((xlog_cil_empty(mp->m_log) && xlog_iclogs_empty(mp->m_log) &&
1204 !xfs_ail_min_lsn(mp->m_log->l_ailp)) ||
1205 xlog_is_shutdown(mp->m_log));
1207 if (!xfs_log_writable(mp))
1211 * xfs_log_need_covered() is not idempotent because it progresses the
1212 * state machine if the log requires covering. Therefore, we must call
1213 * this function once and use the result until we've issued an sb sync.
1214 * Do so first to make that abundantly clear.
1216 * Fall into the covering sequence if the log needs covering or the
1217 * mount has lazy superblock accounting to sync to disk. The sb sync
1218 * used for covering accumulates the in-core counters, so covering
1219 * handles this for us.
1221 need_covered = xfs_log_need_covered(mp);
1222 if (!need_covered && !xfs_has_lazysbcount(mp))
1226 * To cover the log, commit the superblock twice (at most) in
1227 * independent checkpoints. The first serves as a reference for the
1228 * tail pointer. The sync transaction and AIL push empties the AIL and
1229 * updates the in-core tail to the LSN of the first checkpoint. The
1230 * second commit updates the on-disk tail with the in-core LSN,
1231 * covering the log. Push the AIL one more time to leave it empty, as
1235 error = xfs_sync_sb(mp, true);
1238 xfs_ail_push_all_sync(mp->m_ail);
1239 } while (xfs_log_need_covered(mp));
1245 * We may be holding the log iclog lock upon entering this routine.
1248 xlog_assign_tail_lsn_locked(
1249 struct xfs_mount *mp)
1251 struct xlog *log = mp->m_log;
1252 struct xfs_log_item *lip;
1255 assert_spin_locked(&mp->m_ail->ail_lock);
1258 * To make sure we always have a valid LSN for the log tail we keep
1259 * track of the last LSN which was committed in log->l_last_sync_lsn,
1260 * and use that when the AIL was empty.
1262 lip = xfs_ail_min(mp->m_ail);
1264 tail_lsn = lip->li_lsn;
1266 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1267 trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1268 atomic64_set(&log->l_tail_lsn, tail_lsn);
1273 xlog_assign_tail_lsn(
1274 struct xfs_mount *mp)
1278 spin_lock(&mp->m_ail->ail_lock);
1279 tail_lsn = xlog_assign_tail_lsn_locked(mp);
1280 spin_unlock(&mp->m_ail->ail_lock);
1286 * Return the space in the log between the tail and the head. The head
1287 * is passed in the cycle/bytes formal parms. In the special case where
1288 * the reserve head has wrapped passed the tail, this calculation is no
1289 * longer valid. In this case, just return 0 which means there is no space
1290 * in the log. This works for all places where this function is called
1291 * with the reserve head. Of course, if the write head were to ever
1292 * wrap the tail, we should blow up. Rather than catch this case here,
1293 * we depend on other ASSERTions in other parts of the code. XXXmiken
1295 * If reservation head is behind the tail, we have a problem. Warn about it,
1296 * but then treat it as if the log is empty.
1298 * If the log is shut down, the head and tail may be invalid or out of whack, so
1299 * shortcut invalidity asserts in this case so that we don't trigger them
1312 xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1313 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1314 tail_bytes = BBTOB(tail_bytes);
1315 if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1316 return log->l_logsize - (head_bytes - tail_bytes);
1317 if (tail_cycle + 1 < head_cycle)
1320 /* Ignore potential inconsistency when shutdown. */
1321 if (xlog_is_shutdown(log))
1322 return log->l_logsize;
1324 if (tail_cycle < head_cycle) {
1325 ASSERT(tail_cycle == (head_cycle - 1));
1326 return tail_bytes - head_bytes;
1330 * The reservation head is behind the tail. In this case we just want to
1331 * return the size of the log as the amount of space left.
1333 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1334 xfs_alert(log->l_mp, " tail_cycle = %d, tail_bytes = %d",
1335 tail_cycle, tail_bytes);
1336 xfs_alert(log->l_mp, " GH cycle = %d, GH bytes = %d",
1337 head_cycle, head_bytes);
1339 return log->l_logsize;
1345 struct work_struct *work)
1347 struct xlog_in_core *iclog =
1348 container_of(work, struct xlog_in_core, ic_end_io_work);
1349 struct xlog *log = iclog->ic_log;
1352 error = blk_status_to_errno(iclog->ic_bio.bi_status);
1354 /* treat writes with injected CRC errors as failed */
1355 if (iclog->ic_fail_crc)
1360 * Race to shutdown the filesystem if we see an error.
1362 if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1363 xfs_alert(log->l_mp, "log I/O error %d", error);
1364 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1367 xlog_state_done_syncing(iclog);
1368 bio_uninit(&iclog->ic_bio);
1371 * Drop the lock to signal that we are done. Nothing references the
1372 * iclog after this, so an unmount waiting on this lock can now tear it
1373 * down safely. As such, it is unsafe to reference the iclog after the
1374 * unlock as we could race with it being freed.
1376 up(&iclog->ic_sema);
1380 * Return size of each in-core log record buffer.
1382 * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1384 * If the filesystem blocksize is too large, we may need to choose a
1385 * larger size since the directory code currently logs entire blocks.
1388 xlog_get_iclog_buffer_size(
1389 struct xfs_mount *mp,
1392 if (mp->m_logbufs <= 0)
1393 mp->m_logbufs = XLOG_MAX_ICLOGS;
1394 if (mp->m_logbsize <= 0)
1395 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1397 log->l_iclog_bufs = mp->m_logbufs;
1398 log->l_iclog_size = mp->m_logbsize;
1401 * # headers = size / 32k - one header holds cycles from 32k of data.
1403 log->l_iclog_heads =
1404 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1405 log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1410 struct xfs_mount *mp)
1412 queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1413 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1417 * Clear the log incompat flags if we have the opportunity.
1419 * This only happens if we're about to log the second dummy transaction as part
1420 * of covering the log and we can get the log incompat feature usage lock.
1423 xlog_clear_incompat(
1426 struct xfs_mount *mp = log->l_mp;
1428 if (!xfs_sb_has_incompat_log_feature(&mp->m_sb,
1429 XFS_SB_FEAT_INCOMPAT_LOG_ALL))
1432 if (log->l_covered_state != XLOG_STATE_COVER_DONE2)
1435 if (!down_write_trylock(&log->l_incompat_users))
1438 xfs_clear_incompat_log_features(mp);
1439 up_write(&log->l_incompat_users);
1443 * Every sync period we need to unpin all items in the AIL and push them to
1444 * disk. If there is nothing dirty, then we might need to cover the log to
1445 * indicate that the filesystem is idle.
1449 struct work_struct *work)
1451 struct xlog *log = container_of(to_delayed_work(work),
1452 struct xlog, l_work);
1453 struct xfs_mount *mp = log->l_mp;
1455 /* dgc: errors ignored - not fatal and nowhere to report them */
1456 if (xfs_fs_writable(mp, SB_FREEZE_WRITE) && xfs_log_need_covered(mp)) {
1458 * Dump a transaction into the log that contains no real change.
1459 * This is needed to stamp the current tail LSN into the log
1460 * during the covering operation.
1462 * We cannot use an inode here for this - that will push dirty
1463 * state back up into the VFS and then periodic inode flushing
1464 * will prevent log covering from making progress. Hence we
1465 * synchronously log the superblock instead to ensure the
1466 * superblock is immediately unpinned and can be written back.
1468 xlog_clear_incompat(log);
1469 xfs_sync_sb(mp, true);
1471 xfs_log_force(mp, 0);
1473 /* start pushing all the metadata that is currently dirty */
1474 xfs_ail_push_all(mp->m_ail);
1476 /* queue us up again */
1477 xfs_log_work_queue(mp);
1481 * This routine initializes some of the log structure for a given mount point.
1482 * Its primary purpose is to fill in enough, so recovery can occur. However,
1483 * some other stuff may be filled in too.
1485 STATIC struct xlog *
1487 struct xfs_mount *mp,
1488 struct xfs_buftarg *log_target,
1489 xfs_daddr_t blk_offset,
1493 xlog_rec_header_t *head;
1494 xlog_in_core_t **iclogp;
1495 xlog_in_core_t *iclog, *prev_iclog=NULL;
1497 int error = -ENOMEM;
1500 log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1502 xfs_warn(mp, "Log allocation failed: No memory!");
1507 log->l_targ = log_target;
1508 log->l_logsize = BBTOB(num_bblks);
1509 log->l_logBBstart = blk_offset;
1510 log->l_logBBsize = num_bblks;
1511 log->l_covered_state = XLOG_STATE_COVER_IDLE;
1512 set_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
1513 INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1515 log->l_prev_block = -1;
1516 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1517 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1518 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1519 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */
1521 if (xfs_has_logv2(mp) && mp->m_sb.sb_logsunit > 1)
1522 log->l_iclog_roundoff = mp->m_sb.sb_logsunit;
1524 log->l_iclog_roundoff = BBSIZE;
1526 xlog_grant_head_init(&log->l_reserve_head);
1527 xlog_grant_head_init(&log->l_write_head);
1529 error = -EFSCORRUPTED;
1530 if (xfs_has_sector(mp)) {
1531 log2_size = mp->m_sb.sb_logsectlog;
1532 if (log2_size < BBSHIFT) {
1533 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1534 log2_size, BBSHIFT);
1538 log2_size -= BBSHIFT;
1539 if (log2_size > mp->m_sectbb_log) {
1540 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1541 log2_size, mp->m_sectbb_log);
1545 /* for larger sector sizes, must have v2 or external log */
1546 if (log2_size && log->l_logBBstart > 0 &&
1547 !xfs_has_logv2(mp)) {
1549 "log sector size (0x%x) invalid for configuration.",
1554 log->l_sectBBsize = 1 << log2_size;
1556 init_rwsem(&log->l_incompat_users);
1558 xlog_get_iclog_buffer_size(mp, log);
1560 spin_lock_init(&log->l_icloglock);
1561 init_waitqueue_head(&log->l_flush_wait);
1563 iclogp = &log->l_iclog;
1565 * The amount of memory to allocate for the iclog structure is
1566 * rather funky due to the way the structure is defined. It is
1567 * done this way so that we can use different sizes for machines
1568 * with different amounts of memory. See the definition of
1569 * xlog_in_core_t in xfs_log_priv.h for details.
1571 ASSERT(log->l_iclog_size >= 4096);
1572 for (i = 0; i < log->l_iclog_bufs; i++) {
1573 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1574 sizeof(struct bio_vec);
1576 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1578 goto out_free_iclog;
1581 iclog->ic_prev = prev_iclog;
1584 iclog->ic_data = kvzalloc(log->l_iclog_size,
1585 GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1586 if (!iclog->ic_data)
1587 goto out_free_iclog;
1589 log->l_iclog_bak[i] = &iclog->ic_header;
1591 head = &iclog->ic_header;
1592 memset(head, 0, sizeof(xlog_rec_header_t));
1593 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1594 head->h_version = cpu_to_be32(
1595 xfs_has_logv2(log->l_mp) ? 2 : 1);
1596 head->h_size = cpu_to_be32(log->l_iclog_size);
1598 head->h_fmt = cpu_to_be32(XLOG_FMT);
1599 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1601 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1602 iclog->ic_state = XLOG_STATE_ACTIVE;
1603 iclog->ic_log = log;
1604 atomic_set(&iclog->ic_refcnt, 0);
1605 INIT_LIST_HEAD(&iclog->ic_callbacks);
1606 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1608 init_waitqueue_head(&iclog->ic_force_wait);
1609 init_waitqueue_head(&iclog->ic_write_wait);
1610 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1611 sema_init(&iclog->ic_sema, 1);
1613 iclogp = &iclog->ic_next;
1615 *iclogp = log->l_iclog; /* complete ring */
1616 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */
1618 log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1619 XFS_WQFLAGS(WQ_FREEZABLE | WQ_MEM_RECLAIM |
1621 0, mp->m_super->s_id);
1622 if (!log->l_ioend_workqueue)
1623 goto out_free_iclog;
1625 error = xlog_cil_init(log);
1627 goto out_destroy_workqueue;
1630 out_destroy_workqueue:
1631 destroy_workqueue(log->l_ioend_workqueue);
1633 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1634 prev_iclog = iclog->ic_next;
1635 kmem_free(iclog->ic_data);
1637 if (prev_iclog == log->l_iclog)
1643 return ERR_PTR(error);
1644 } /* xlog_alloc_log */
1647 * Compute the LSN that we'd need to push the log tail towards in order to have
1648 * (a) enough on-disk log space to log the number of bytes specified, (b) at
1649 * least 25% of the log space free, and (c) at least 256 blocks free. If the
1650 * log free space already meets all three thresholds, this function returns
1654 xlog_grant_push_threshold(
1658 xfs_lsn_t threshold_lsn = 0;
1659 xfs_lsn_t last_sync_lsn;
1662 int threshold_block;
1663 int threshold_cycle;
1666 ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1668 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1669 free_blocks = BTOBBT(free_bytes);
1672 * Set the threshold for the minimum number of free blocks in the
1673 * log to the maximum of what the caller needs, one quarter of the
1674 * log, and 256 blocks.
1676 free_threshold = BTOBB(need_bytes);
1677 free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1678 free_threshold = max(free_threshold, 256);
1679 if (free_blocks >= free_threshold)
1680 return NULLCOMMITLSN;
1682 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1684 threshold_block += free_threshold;
1685 if (threshold_block >= log->l_logBBsize) {
1686 threshold_block -= log->l_logBBsize;
1687 threshold_cycle += 1;
1689 threshold_lsn = xlog_assign_lsn(threshold_cycle,
1692 * Don't pass in an lsn greater than the lsn of the last
1693 * log record known to be on disk. Use a snapshot of the last sync lsn
1694 * so that it doesn't change between the compare and the set.
1696 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1697 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1698 threshold_lsn = last_sync_lsn;
1700 return threshold_lsn;
1704 * Push the tail of the log if we need to do so to maintain the free log space
1705 * thresholds set out by xlog_grant_push_threshold. We may need to adopt a
1706 * policy which pushes on an lsn which is further along in the log once we
1707 * reach the high water mark. In this manner, we would be creating a low water
1711 xlog_grant_push_ail(
1715 xfs_lsn_t threshold_lsn;
1717 threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1718 if (threshold_lsn == NULLCOMMITLSN || xlog_is_shutdown(log))
1722 * Get the transaction layer to kick the dirty buffers out to
1723 * disk asynchronously. No point in trying to do this if
1724 * the filesystem is shutting down.
1726 xfs_ail_push(log->l_ailp, threshold_lsn);
1730 * Stamp cycle number in every block
1735 struct xlog_in_core *iclog,
1739 int size = iclog->ic_offset + roundoff;
1743 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1745 dp = iclog->ic_datap;
1746 for (i = 0; i < BTOBB(size); i++) {
1747 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1749 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1750 *(__be32 *)dp = cycle_lsn;
1754 if (xfs_has_logv2(log->l_mp)) {
1755 xlog_in_core_2_t *xhdr = iclog->ic_data;
1757 for ( ; i < BTOBB(size); i++) {
1758 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1759 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1760 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1761 *(__be32 *)dp = cycle_lsn;
1765 for (i = 1; i < log->l_iclog_heads; i++)
1766 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1771 * Calculate the checksum for a log buffer.
1773 * This is a little more complicated than it should be because the various
1774 * headers and the actual data are non-contiguous.
1779 struct xlog_rec_header *rhead,
1785 /* first generate the crc for the record header ... */
1786 crc = xfs_start_cksum_update((char *)rhead,
1787 sizeof(struct xlog_rec_header),
1788 offsetof(struct xlog_rec_header, h_crc));
1790 /* ... then for additional cycle data for v2 logs ... */
1791 if (xfs_has_logv2(log->l_mp)) {
1792 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1796 xheads = DIV_ROUND_UP(size, XLOG_HEADER_CYCLE_SIZE);
1798 for (i = 1; i < xheads; i++) {
1799 crc = crc32c(crc, &xhdr[i].hic_xheader,
1800 sizeof(struct xlog_rec_ext_header));
1804 /* ... and finally for the payload */
1805 crc = crc32c(crc, dp, size);
1807 return xfs_end_cksum(crc);
1814 struct xlog_in_core *iclog = bio->bi_private;
1816 queue_work(iclog->ic_log->l_ioend_workqueue,
1817 &iclog->ic_end_io_work);
1821 xlog_map_iclog_data(
1827 struct page *page = kmem_to_page(data);
1828 unsigned int off = offset_in_page(data);
1829 size_t len = min_t(size_t, count, PAGE_SIZE - off);
1831 if (bio_add_page(bio, page, len, off) != len)
1844 struct xlog_in_core *iclog,
1848 ASSERT(bno < log->l_logBBsize);
1849 trace_xlog_iclog_write(iclog, _RET_IP_);
1852 * We lock the iclogbufs here so that we can serialise against I/O
1853 * completion during unmount. We might be processing a shutdown
1854 * triggered during unmount, and that can occur asynchronously to the
1855 * unmount thread, and hence we need to ensure that completes before
1856 * tearing down the iclogbufs. Hence we need to hold the buffer lock
1857 * across the log IO to archieve that.
1859 down(&iclog->ic_sema);
1860 if (xlog_is_shutdown(log)) {
1862 * It would seem logical to return EIO here, but we rely on
1863 * the log state machine to propagate I/O errors instead of
1864 * doing it here. We kick of the state machine and unlock
1865 * the buffer manually, the code needs to be kept in sync
1866 * with the I/O completion path.
1868 xlog_state_done_syncing(iclog);
1869 up(&iclog->ic_sema);
1873 bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1874 bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1875 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1876 iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1877 iclog->ic_bio.bi_private = iclog;
1880 * We use REQ_SYNC | REQ_IDLE here to tell the block layer the are more
1881 * IOs coming immediately after this one. This prevents the block layer
1882 * writeback throttle from throttling log writes behind background
1883 * metadata writeback and causing priority inversions.
1885 iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE;
1886 if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
1887 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1889 * For external log devices, we also need to flush the data
1890 * device cache first to ensure all metadata writeback covered
1891 * by the LSN in this iclog is on stable storage. This is slow,
1892 * but it *must* complete before we issue the external log IO.
1894 if (log->l_targ != log->l_mp->m_ddev_targp)
1895 blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev);
1897 if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
1898 iclog->ic_bio.bi_opf |= REQ_FUA;
1900 iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
1902 if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
1903 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1906 if (is_vmalloc_addr(iclog->ic_data))
1907 flush_kernel_vmap_range(iclog->ic_data, count);
1910 * If this log buffer would straddle the end of the log we will have
1911 * to split it up into two bios, so that we can continue at the start.
1913 if (bno + BTOBB(count) > log->l_logBBsize) {
1916 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1917 GFP_NOIO, &fs_bio_set);
1918 bio_chain(split, &iclog->ic_bio);
1921 /* restart at logical offset zero for the remainder */
1922 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1925 submit_bio(&iclog->ic_bio);
1929 * We need to bump cycle number for the part of the iclog that is
1930 * written to the start of the log. Watch out for the header magic
1931 * number case, though.
1940 unsigned int split_offset = BBTOB(log->l_logBBsize - bno);
1943 for (i = split_offset; i < count; i += BBSIZE) {
1944 uint32_t cycle = get_unaligned_be32(data + i);
1946 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1948 put_unaligned_be32(cycle, data + i);
1953 xlog_calc_iclog_size(
1955 struct xlog_in_core *iclog,
1958 uint32_t count_init, count;
1960 /* Add for LR header */
1961 count_init = log->l_iclog_hsize + iclog->ic_offset;
1962 count = roundup(count_init, log->l_iclog_roundoff);
1964 *roundoff = count - count_init;
1966 ASSERT(count >= count_init);
1967 ASSERT(*roundoff < log->l_iclog_roundoff);
1972 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1973 * fashion. Previously, we should have moved the current iclog
1974 * ptr in the log to point to the next available iclog. This allows further
1975 * write to continue while this code syncs out an iclog ready to go.
1976 * Before an in-core log can be written out, the data section must be scanned
1977 * to save away the 1st word of each BBSIZE block into the header. We replace
1978 * it with the current cycle count. Each BBSIZE block is tagged with the
1979 * cycle count because there in an implicit assumption that drives will
1980 * guarantee that entire 512 byte blocks get written at once. In other words,
1981 * we can't have part of a 512 byte block written and part not written. By
1982 * tagging each block, we will know which blocks are valid when recovering
1983 * after an unclean shutdown.
1985 * This routine is single threaded on the iclog. No other thread can be in
1986 * this routine with the same iclog. Changing contents of iclog can there-
1987 * fore be done without grabbing the state machine lock. Updating the global
1988 * log will require grabbing the lock though.
1990 * The entire log manager uses a logical block numbering scheme. Only
1991 * xlog_write_iclog knows about the fact that the log may not start with
1992 * block zero on a given device.
1997 struct xlog_in_core *iclog)
1999 unsigned int count; /* byte count of bwrite */
2000 unsigned int roundoff; /* roundoff to BB or stripe */
2004 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2005 trace_xlog_iclog_sync(iclog, _RET_IP_);
2007 count = xlog_calc_iclog_size(log, iclog, &roundoff);
2009 /* move grant heads by roundoff in sync */
2010 xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
2011 xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
2013 /* put cycle number in every block */
2014 xlog_pack_data(log, iclog, roundoff);
2016 /* real byte length */
2017 size = iclog->ic_offset;
2018 if (xfs_has_logv2(log->l_mp))
2020 iclog->ic_header.h_len = cpu_to_be32(size);
2022 XFS_STATS_INC(log->l_mp, xs_log_writes);
2023 XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
2025 bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
2027 /* Do we need to split this write into 2 parts? */
2028 if (bno + BTOBB(count) > log->l_logBBsize)
2029 xlog_split_iclog(log, &iclog->ic_header, bno, count);
2031 /* calculcate the checksum */
2032 iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
2033 iclog->ic_datap, size);
2035 * Intentionally corrupt the log record CRC based on the error injection
2036 * frequency, if defined. This facilitates testing log recovery in the
2037 * event of torn writes. Hence, set the IOABORT state to abort the log
2038 * write on I/O completion and shutdown the fs. The subsequent mount
2039 * detects the bad CRC and attempts to recover.
2042 if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
2043 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
2044 iclog->ic_fail_crc = true;
2046 "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
2047 be64_to_cpu(iclog->ic_header.h_lsn));
2050 xlog_verify_iclog(log, iclog, count);
2051 xlog_write_iclog(log, iclog, bno, count);
2055 * Deallocate a log structure
2061 xlog_in_core_t *iclog, *next_iclog;
2064 xlog_cil_destroy(log);
2067 * Cycle all the iclogbuf locks to make sure all log IO completion
2068 * is done before we tear down these buffers.
2070 iclog = log->l_iclog;
2071 for (i = 0; i < log->l_iclog_bufs; i++) {
2072 down(&iclog->ic_sema);
2073 up(&iclog->ic_sema);
2074 iclog = iclog->ic_next;
2077 iclog = log->l_iclog;
2078 for (i = 0; i < log->l_iclog_bufs; i++) {
2079 next_iclog = iclog->ic_next;
2080 kmem_free(iclog->ic_data);
2085 log->l_mp->m_log = NULL;
2086 destroy_workqueue(log->l_ioend_workqueue);
2091 * Update counters atomically now that memcpy is done.
2094 xlog_state_finish_copy(
2096 struct xlog_in_core *iclog,
2100 lockdep_assert_held(&log->l_icloglock);
2102 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2103 iclog->ic_offset += copy_bytes;
2107 * print out info relating to regions written which consume
2112 struct xfs_mount *mp,
2113 struct xlog_ticket *ticket)
2116 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2118 /* match with XLOG_REG_TYPE_* in xfs_log.h */
2119 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2120 static char *res_type_str[] = {
2121 REG_TYPE_STR(BFORMAT, "bformat"),
2122 REG_TYPE_STR(BCHUNK, "bchunk"),
2123 REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2124 REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2125 REG_TYPE_STR(IFORMAT, "iformat"),
2126 REG_TYPE_STR(ICORE, "icore"),
2127 REG_TYPE_STR(IEXT, "iext"),
2128 REG_TYPE_STR(IBROOT, "ibroot"),
2129 REG_TYPE_STR(ILOCAL, "ilocal"),
2130 REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2131 REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2132 REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2133 REG_TYPE_STR(QFORMAT, "qformat"),
2134 REG_TYPE_STR(DQUOT, "dquot"),
2135 REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2136 REG_TYPE_STR(LRHEADER, "LR header"),
2137 REG_TYPE_STR(UNMOUNT, "unmount"),
2138 REG_TYPE_STR(COMMIT, "commit"),
2139 REG_TYPE_STR(TRANSHDR, "trans header"),
2140 REG_TYPE_STR(ICREATE, "inode create"),
2141 REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2142 REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2143 REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2144 REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2145 REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2146 REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2148 BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2151 xfs_warn(mp, "ticket reservation summary:");
2152 xfs_warn(mp, " unit res = %d bytes",
2153 ticket->t_unit_res);
2154 xfs_warn(mp, " current res = %d bytes",
2155 ticket->t_curr_res);
2156 xfs_warn(mp, " total reg = %u bytes (o/flow = %u bytes)",
2157 ticket->t_res_arr_sum, ticket->t_res_o_flow);
2158 xfs_warn(mp, " ophdrs = %u (ophdr space = %u bytes)",
2159 ticket->t_res_num_ophdrs, ophdr_spc);
2160 xfs_warn(mp, " ophdr + reg = %u bytes",
2161 ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2162 xfs_warn(mp, " num regions = %u",
2165 for (i = 0; i < ticket->t_res_num; i++) {
2166 uint r_type = ticket->t_res_arr[i].r_type;
2167 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2168 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2169 "bad-rtype" : res_type_str[r_type]),
2170 ticket->t_res_arr[i].r_len);
2175 * Print a summary of the transaction.
2179 struct xfs_trans *tp)
2181 struct xfs_mount *mp = tp->t_mountp;
2182 struct xfs_log_item *lip;
2184 /* dump core transaction and ticket info */
2185 xfs_warn(mp, "transaction summary:");
2186 xfs_warn(mp, " log res = %d", tp->t_log_res);
2187 xfs_warn(mp, " log count = %d", tp->t_log_count);
2188 xfs_warn(mp, " flags = 0x%x", tp->t_flags);
2190 xlog_print_tic_res(mp, tp->t_ticket);
2192 /* dump each log item */
2193 list_for_each_entry(lip, &tp->t_items, li_trans) {
2194 struct xfs_log_vec *lv = lip->li_lv;
2195 struct xfs_log_iovec *vec;
2198 xfs_warn(mp, "log item: ");
2199 xfs_warn(mp, " type = 0x%x", lip->li_type);
2200 xfs_warn(mp, " flags = 0x%lx", lip->li_flags);
2203 xfs_warn(mp, " niovecs = %d", lv->lv_niovecs);
2204 xfs_warn(mp, " size = %d", lv->lv_size);
2205 xfs_warn(mp, " bytes = %d", lv->lv_bytes);
2206 xfs_warn(mp, " buf len = %d", lv->lv_buf_len);
2208 /* dump each iovec for the log item */
2209 vec = lv->lv_iovecp;
2210 for (i = 0; i < lv->lv_niovecs; i++) {
2211 int dumplen = min(vec->i_len, 32);
2213 xfs_warn(mp, " iovec[%d]", i);
2214 xfs_warn(mp, " type = 0x%x", vec->i_type);
2215 xfs_warn(mp, " len = %d", vec->i_len);
2216 xfs_warn(mp, " first %d bytes of iovec[%d]:", dumplen, i);
2217 xfs_hex_dump(vec->i_addr, dumplen);
2225 * Calculate the potential space needed by the log vector. We may need a start
2226 * record, and each region gets its own struct xlog_op_header and may need to be
2227 * double word aligned.
2230 xlog_write_calc_vec_length(
2231 struct xlog_ticket *ticket,
2232 struct xfs_log_vec *log_vector,
2235 struct xfs_log_vec *lv;
2240 if (optype & XLOG_START_TRANS)
2243 for (lv = log_vector; lv; lv = lv->lv_next) {
2244 /* we don't write ordered log vectors */
2245 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2248 headers += lv->lv_niovecs;
2250 for (i = 0; i < lv->lv_niovecs; i++) {
2251 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i];
2254 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2258 ticket->t_res_num_ophdrs += headers;
2259 len += headers * sizeof(struct xlog_op_header);
2265 xlog_write_start_rec(
2266 struct xlog_op_header *ophdr,
2267 struct xlog_ticket *ticket)
2269 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2270 ophdr->oh_clientid = ticket->t_clientid;
2272 ophdr->oh_flags = XLOG_START_TRANS;
2276 static xlog_op_header_t *
2277 xlog_write_setup_ophdr(
2279 struct xlog_op_header *ophdr,
2280 struct xlog_ticket *ticket,
2283 ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2284 ophdr->oh_clientid = ticket->t_clientid;
2287 /* are we copying a commit or unmount record? */
2288 ophdr->oh_flags = flags;
2291 * We've seen logs corrupted with bad transaction client ids. This
2292 * makes sure that XFS doesn't generate them on. Turn this into an EIO
2293 * and shut down the filesystem.
2295 switch (ophdr->oh_clientid) {
2296 case XFS_TRANSACTION:
2302 "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2303 ophdr->oh_clientid, ticket);
2311 * Set up the parameters of the region copy into the log. This has
2312 * to handle region write split across multiple log buffers - this
2313 * state is kept external to this function so that this code can
2314 * be written in an obvious, self documenting manner.
2317 xlog_write_setup_copy(
2318 struct xlog_ticket *ticket,
2319 struct xlog_op_header *ophdr,
2320 int space_available,
2324 int *last_was_partial_copy,
2325 int *bytes_consumed)
2329 still_to_copy = space_required - *bytes_consumed;
2330 *copy_off = *bytes_consumed;
2332 if (still_to_copy <= space_available) {
2333 /* write of region completes here */
2334 *copy_len = still_to_copy;
2335 ophdr->oh_len = cpu_to_be32(*copy_len);
2336 if (*last_was_partial_copy)
2337 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2338 *last_was_partial_copy = 0;
2339 *bytes_consumed = 0;
2343 /* partial write of region, needs extra log op header reservation */
2344 *copy_len = space_available;
2345 ophdr->oh_len = cpu_to_be32(*copy_len);
2346 ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2347 if (*last_was_partial_copy)
2348 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2349 *bytes_consumed += *copy_len;
2350 (*last_was_partial_copy)++;
2352 /* account for new log op header */
2353 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2354 ticket->t_res_num_ophdrs++;
2356 return sizeof(struct xlog_op_header);
2360 xlog_write_copy_finish(
2362 struct xlog_in_core *iclog,
2367 int *partial_copy_len,
2372 if (*partial_copy) {
2374 * This iclog has already been marked WANT_SYNC by
2375 * xlog_state_get_iclog_space.
2377 spin_lock(&log->l_icloglock);
2378 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2385 *partial_copy_len = 0;
2387 if (iclog->ic_size - log_offset > sizeof(xlog_op_header_t))
2390 /* no more space in this iclog - push it. */
2391 spin_lock(&log->l_icloglock);
2392 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2396 if (iclog->ic_state == XLOG_STATE_ACTIVE)
2397 xlog_state_switch_iclogs(log, iclog, 0);
2399 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2400 xlog_is_shutdown(log));
2402 error = xlog_state_release_iclog(log, iclog);
2403 spin_unlock(&log->l_icloglock);
2408 * Write some region out to in-core log
2410 * This will be called when writing externally provided regions or when
2411 * writing out a commit record for a given transaction.
2413 * General algorithm:
2414 * 1. Find total length of this write. This may include adding to the
2415 * lengths passed in.
2416 * 2. Check whether we violate the tickets reservation.
2417 * 3. While writing to this iclog
2418 * A. Reserve as much space in this iclog as can get
2419 * B. If this is first write, save away start lsn
2420 * C. While writing this region:
2421 * 1. If first write of transaction, write start record
2422 * 2. Write log operation header (header per region)
2423 * 3. Find out if we can fit entire region into this iclog
2424 * 4. Potentially, verify destination memcpy ptr
2425 * 5. Memcpy (partial) region
2426 * 6. If partial copy, release iclog; otherwise, continue
2427 * copying more regions into current iclog
2428 * 4. Mark want sync bit (in simulation mode)
2429 * 5. Release iclog for potential flush to on-disk log.
2432 * 1. Panic if reservation is overrun. This should never happen since
2433 * reservation amounts are generated internal to the filesystem.
2435 * 1. Tickets are single threaded data structures.
2436 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2437 * syncing routine. When a single log_write region needs to span
2438 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2439 * on all log operation writes which don't contain the end of the
2440 * region. The XLOG_END_TRANS bit is used for the in-core log
2441 * operation which contains the end of the continued log_write region.
2442 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2443 * we don't really know exactly how much space will be used. As a result,
2444 * we don't update ic_offset until the end when we know exactly how many
2445 * bytes have been written out.
2450 struct xfs_cil_ctx *ctx,
2451 struct xfs_log_vec *log_vector,
2452 struct xlog_ticket *ticket,
2455 struct xlog_in_core *iclog = NULL;
2456 struct xfs_log_vec *lv = log_vector;
2457 struct xfs_log_iovec *vecp = lv->lv_iovecp;
2460 int partial_copy = 0;
2461 int partial_copy_len = 0;
2468 * If this is a commit or unmount transaction, we don't need a start
2469 * record to be written. We do, however, have to account for the
2470 * commit or unmount header that gets written. Hence we always have
2471 * to account for an extra xlog_op_header here.
2473 ticket->t_curr_res -= sizeof(struct xlog_op_header);
2474 if (ticket->t_curr_res < 0) {
2475 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2476 "ctx ticket reservation ran out. Need to up reservation");
2477 xlog_print_tic_res(log->l_mp, ticket);
2478 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2481 len = xlog_write_calc_vec_length(ticket, log_vector, optype);
2482 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2486 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2487 &contwr, &log_offset);
2491 ASSERT(log_offset <= iclog->ic_size - 1);
2492 ptr = iclog->ic_datap + log_offset;
2495 * If we have a context pointer, pass it the first iclog we are
2496 * writing to so it can record state needed for iclog write
2500 xlog_cil_set_ctx_write_state(ctx, iclog);
2505 * This loop writes out as many regions as can fit in the amount
2506 * of space which was allocated by xlog_state_get_iclog_space().
2508 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2509 struct xfs_log_iovec *reg;
2510 struct xlog_op_header *ophdr;
2513 bool ordered = false;
2514 bool wrote_start_rec = false;
2516 /* ordered log vectors have no regions to write */
2517 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2518 ASSERT(lv->lv_niovecs == 0);
2524 ASSERT(reg->i_len % sizeof(int32_t) == 0);
2525 ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2528 * Before we start formatting log vectors, we need to
2529 * write a start record. Only do this for the first
2530 * iclog we write to.
2532 if (optype & XLOG_START_TRANS) {
2533 xlog_write_start_rec(ptr, ticket);
2534 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2535 sizeof(struct xlog_op_header));
2536 optype &= ~XLOG_START_TRANS;
2537 wrote_start_rec = true;
2540 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, optype);
2544 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2545 sizeof(struct xlog_op_header));
2547 len += xlog_write_setup_copy(ticket, ophdr,
2548 iclog->ic_size-log_offset,
2550 ©_off, ©_len,
2553 xlog_verify_dest_ptr(log, ptr);
2558 * Unmount records just log an opheader, so can have
2559 * empty payloads with no data region to copy. Hence we
2560 * only copy the payload if the vector says it has data
2563 ASSERT(copy_len >= 0);
2565 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2566 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2569 copy_len += sizeof(struct xlog_op_header);
2571 if (wrote_start_rec) {
2572 copy_len += sizeof(struct xlog_op_header);
2575 data_cnt += contwr ? copy_len : 0;
2577 error = xlog_write_copy_finish(log, iclog, optype,
2578 &record_cnt, &data_cnt,
2586 * if we had a partial copy, we need to get more iclog
2587 * space but we don't want to increment the region
2588 * index because there is still more is this region to
2591 * If we completed writing this region, and we flushed
2592 * the iclog (indicated by resetting of the record
2593 * count), then we also need to get more log space. If
2594 * this was the last record, though, we are done and
2600 if (++index == lv->lv_niovecs) {
2605 vecp = lv->lv_iovecp;
2607 if (record_cnt == 0 && !ordered) {
2617 spin_lock(&log->l_icloglock);
2618 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2619 error = xlog_state_release_iclog(log, iclog);
2620 spin_unlock(&log->l_icloglock);
2626 xlog_state_activate_iclog(
2627 struct xlog_in_core *iclog,
2628 int *iclogs_changed)
2630 ASSERT(list_empty_careful(&iclog->ic_callbacks));
2631 trace_xlog_iclog_activate(iclog, _RET_IP_);
2634 * If the number of ops in this iclog indicate it just contains the
2635 * dummy transaction, we can change state into IDLE (the second time
2636 * around). Otherwise we should change the state into NEED a dummy.
2637 * We don't need to cover the dummy.
2639 if (*iclogs_changed == 0 &&
2640 iclog->ic_header.h_num_logops == cpu_to_be32(XLOG_COVER_OPS)) {
2641 *iclogs_changed = 1;
2644 * We have two dirty iclogs so start over. This could also be
2645 * num of ops indicating this is not the dummy going out.
2647 *iclogs_changed = 2;
2650 iclog->ic_state = XLOG_STATE_ACTIVE;
2651 iclog->ic_offset = 0;
2652 iclog->ic_header.h_num_logops = 0;
2653 memset(iclog->ic_header.h_cycle_data, 0,
2654 sizeof(iclog->ic_header.h_cycle_data));
2655 iclog->ic_header.h_lsn = 0;
2656 iclog->ic_header.h_tail_lsn = 0;
2660 * Loop through all iclogs and mark all iclogs currently marked DIRTY as
2661 * ACTIVE after iclog I/O has completed.
2664 xlog_state_activate_iclogs(
2666 int *iclogs_changed)
2668 struct xlog_in_core *iclog = log->l_iclog;
2671 if (iclog->ic_state == XLOG_STATE_DIRTY)
2672 xlog_state_activate_iclog(iclog, iclogs_changed);
2674 * The ordering of marking iclogs ACTIVE must be maintained, so
2675 * an iclog doesn't become ACTIVE beyond one that is SYNCING.
2677 else if (iclog->ic_state != XLOG_STATE_ACTIVE)
2679 } while ((iclog = iclog->ic_next) != log->l_iclog);
2688 * We go to NEED for any non-covering writes. We go to NEED2 if we just
2689 * wrote the first covering record (DONE). We go to IDLE if we just
2690 * wrote the second covering record (DONE2) and remain in IDLE until a
2691 * non-covering write occurs.
2693 switch (prev_state) {
2694 case XLOG_STATE_COVER_IDLE:
2695 if (iclogs_changed == 1)
2696 return XLOG_STATE_COVER_IDLE;
2698 case XLOG_STATE_COVER_NEED:
2699 case XLOG_STATE_COVER_NEED2:
2701 case XLOG_STATE_COVER_DONE:
2702 if (iclogs_changed == 1)
2703 return XLOG_STATE_COVER_NEED2;
2705 case XLOG_STATE_COVER_DONE2:
2706 if (iclogs_changed == 1)
2707 return XLOG_STATE_COVER_IDLE;
2713 return XLOG_STATE_COVER_NEED;
2717 xlog_state_clean_iclog(
2719 struct xlog_in_core *dirty_iclog)
2721 int iclogs_changed = 0;
2723 trace_xlog_iclog_clean(dirty_iclog, _RET_IP_);
2725 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2727 xlog_state_activate_iclogs(log, &iclogs_changed);
2728 wake_up_all(&dirty_iclog->ic_force_wait);
2730 if (iclogs_changed) {
2731 log->l_covered_state = xlog_covered_state(log->l_covered_state,
2737 xlog_get_lowest_lsn(
2740 struct xlog_in_core *iclog = log->l_iclog;
2741 xfs_lsn_t lowest_lsn = 0, lsn;
2744 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2745 iclog->ic_state == XLOG_STATE_DIRTY)
2748 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2749 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2751 } while ((iclog = iclog->ic_next) != log->l_iclog);
2757 * Completion of a iclog IO does not imply that a transaction has completed, as
2758 * transactions can be large enough to span many iclogs. We cannot change the
2759 * tail of the log half way through a transaction as this may be the only
2760 * transaction in the log and moving the tail to point to the middle of it
2761 * will prevent recovery from finding the start of the transaction. Hence we
2762 * should only update the last_sync_lsn if this iclog contains transaction
2763 * completion callbacks on it.
2765 * We have to do this before we drop the icloglock to ensure we are the only one
2766 * that can update it.
2768 * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2769 * the reservation grant head pushing. This is due to the fact that the push
2770 * target is bound by the current last_sync_lsn value. Hence if we have a large
2771 * amount of log space bound up in this committing transaction then the
2772 * last_sync_lsn value may be the limiting factor preventing tail pushing from
2773 * freeing space in the log. Hence once we've updated the last_sync_lsn we
2774 * should push the AIL to ensure the push target (and hence the grant head) is
2775 * no longer bound by the old log head location and can move forwards and make
2779 xlog_state_set_callback(
2781 struct xlog_in_core *iclog,
2782 xfs_lsn_t header_lsn)
2784 trace_xlog_iclog_callback(iclog, _RET_IP_);
2785 iclog->ic_state = XLOG_STATE_CALLBACK;
2787 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2790 if (list_empty_careful(&iclog->ic_callbacks))
2793 atomic64_set(&log->l_last_sync_lsn, header_lsn);
2794 xlog_grant_push_ail(log, 0);
2798 * Return true if we need to stop processing, false to continue to the next
2799 * iclog. The caller will need to run callbacks if the iclog is returned in the
2800 * XLOG_STATE_CALLBACK state.
2803 xlog_state_iodone_process_iclog(
2805 struct xlog_in_core *iclog)
2807 xfs_lsn_t lowest_lsn;
2808 xfs_lsn_t header_lsn;
2810 switch (iclog->ic_state) {
2811 case XLOG_STATE_ACTIVE:
2812 case XLOG_STATE_DIRTY:
2814 * Skip all iclogs in the ACTIVE & DIRTY states:
2817 case XLOG_STATE_DONE_SYNC:
2819 * Now that we have an iclog that is in the DONE_SYNC state, do
2820 * one more check here to see if we have chased our tail around.
2821 * If this is not the lowest lsn iclog, then we will leave it
2822 * for another completion to process.
2824 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2825 lowest_lsn = xlog_get_lowest_lsn(log);
2826 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2828 xlog_state_set_callback(log, iclog, header_lsn);
2832 * Can only perform callbacks in order. Since this iclog is not
2833 * in the DONE_SYNC state, we skip the rest and just try to
2841 * Loop over all the iclogs, running attached callbacks on them. Return true if
2842 * we ran any callbacks, indicating that we dropped the icloglock. We don't need
2843 * to handle transient shutdown state here at all because
2844 * xlog_state_shutdown_callbacks() will be run to do the necessary shutdown
2845 * cleanup of the callbacks.
2848 xlog_state_do_iclog_callbacks(
2850 __releases(&log->l_icloglock)
2851 __acquires(&log->l_icloglock)
2853 struct xlog_in_core *first_iclog = log->l_iclog;
2854 struct xlog_in_core *iclog = first_iclog;
2855 bool ran_callback = false;
2860 if (xlog_state_iodone_process_iclog(log, iclog))
2862 if (iclog->ic_state != XLOG_STATE_CALLBACK) {
2863 iclog = iclog->ic_next;
2866 list_splice_init(&iclog->ic_callbacks, &cb_list);
2867 spin_unlock(&log->l_icloglock);
2869 trace_xlog_iclog_callbacks_start(iclog, _RET_IP_);
2870 xlog_cil_process_committed(&cb_list);
2871 trace_xlog_iclog_callbacks_done(iclog, _RET_IP_);
2872 ran_callback = true;
2874 spin_lock(&log->l_icloglock);
2875 xlog_state_clean_iclog(log, iclog);
2876 iclog = iclog->ic_next;
2877 } while (iclog != first_iclog);
2879 return ran_callback;
2884 * Loop running iclog completion callbacks until there are no more iclogs in a
2885 * state that can run callbacks.
2888 xlog_state_do_callback(
2894 spin_lock(&log->l_icloglock);
2895 while (xlog_state_do_iclog_callbacks(log)) {
2896 if (xlog_is_shutdown(log))
2899 if (++repeats > 5000) {
2900 flushcnt += repeats;
2903 "%s: possible infinite loop (%d iterations)",
2904 __func__, flushcnt);
2908 if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE)
2909 wake_up_all(&log->l_flush_wait);
2911 spin_unlock(&log->l_icloglock);
2916 * Finish transitioning this iclog to the dirty state.
2918 * Callbacks could take time, so they are done outside the scope of the
2919 * global state machine log lock.
2922 xlog_state_done_syncing(
2923 struct xlog_in_core *iclog)
2925 struct xlog *log = iclog->ic_log;
2927 spin_lock(&log->l_icloglock);
2928 ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2929 trace_xlog_iclog_sync_done(iclog, _RET_IP_);
2932 * If we got an error, either on the first buffer, or in the case of
2933 * split log writes, on the second, we shut down the file system and
2934 * no iclogs should ever be attempted to be written to disk again.
2936 if (!xlog_is_shutdown(log)) {
2937 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING);
2938 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2942 * Someone could be sleeping prior to writing out the next
2943 * iclog buffer, we wake them all, one will get to do the
2944 * I/O, the others get to wait for the result.
2946 wake_up_all(&iclog->ic_write_wait);
2947 spin_unlock(&log->l_icloglock);
2948 xlog_state_do_callback(log);
2952 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2953 * sleep. We wait on the flush queue on the head iclog as that should be
2954 * the first iclog to complete flushing. Hence if all iclogs are syncing,
2955 * we will wait here and all new writes will sleep until a sync completes.
2957 * The in-core logs are used in a circular fashion. They are not used
2958 * out-of-order even when an iclog past the head is free.
2961 * * log_offset where xlog_write() can start writing into the in-core
2963 * * in-core log pointer to which xlog_write() should write.
2964 * * boolean indicating this is a continued write to an in-core log.
2965 * If this is the last write, then the in-core log's offset field
2966 * needs to be incremented, depending on the amount of data which
2970 xlog_state_get_iclog_space(
2973 struct xlog_in_core **iclogp,
2974 struct xlog_ticket *ticket,
2975 int *continued_write,
2979 xlog_rec_header_t *head;
2980 xlog_in_core_t *iclog;
2983 spin_lock(&log->l_icloglock);
2984 if (xlog_is_shutdown(log)) {
2985 spin_unlock(&log->l_icloglock);
2989 iclog = log->l_iclog;
2990 if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2991 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2993 /* Wait for log writes to have flushed */
2994 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2998 head = &iclog->ic_header;
3000 atomic_inc(&iclog->ic_refcnt); /* prevents sync */
3001 log_offset = iclog->ic_offset;
3003 trace_xlog_iclog_get_space(iclog, _RET_IP_);
3005 /* On the 1st write to an iclog, figure out lsn. This works
3006 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3007 * committing to. If the offset is set, that's how many blocks
3010 if (log_offset == 0) {
3011 ticket->t_curr_res -= log->l_iclog_hsize;
3012 xlog_tic_add_region(ticket,
3014 XLOG_REG_TYPE_LRHEADER);
3015 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3016 head->h_lsn = cpu_to_be64(
3017 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3018 ASSERT(log->l_curr_block >= 0);
3021 /* If there is enough room to write everything, then do it. Otherwise,
3022 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3023 * bit is on, so this will get flushed out. Don't update ic_offset
3024 * until you know exactly how many bytes get copied. Therefore, wait
3025 * until later to update ic_offset.
3027 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3028 * can fit into remaining data section.
3030 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3033 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3036 * If we are the only one writing to this iclog, sync it to
3037 * disk. We need to do an atomic compare and decrement here to
3038 * avoid racing with concurrent atomic_dec_and_lock() calls in
3039 * xlog_state_release_iclog() when there is more than one
3040 * reference to the iclog.
3042 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
3043 error = xlog_state_release_iclog(log, iclog);
3044 spin_unlock(&log->l_icloglock);
3050 /* Do we have enough room to write the full amount in the remainder
3051 * of this iclog? Or must we continue a write on the next iclog and
3052 * mark this iclog as completely taken? In the case where we switch
3053 * iclogs (to mark it taken), this particular iclog will release/sync
3054 * to disk in xlog_write().
3056 if (len <= iclog->ic_size - iclog->ic_offset) {
3057 *continued_write = 0;
3058 iclog->ic_offset += len;
3060 *continued_write = 1;
3061 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3065 ASSERT(iclog->ic_offset <= iclog->ic_size);
3066 spin_unlock(&log->l_icloglock);
3068 *logoffsetp = log_offset;
3073 * The first cnt-1 times a ticket goes through here we don't need to move the
3074 * grant write head because the permanent reservation has reserved cnt times the
3075 * unit amount. Release part of current permanent unit reservation and reset
3076 * current reservation to be one units worth. Also move grant reservation head
3080 xfs_log_ticket_regrant(
3082 struct xlog_ticket *ticket)
3084 trace_xfs_log_ticket_regrant(log, ticket);
3086 if (ticket->t_cnt > 0)
3089 xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3090 ticket->t_curr_res);
3091 xlog_grant_sub_space(log, &log->l_write_head.grant,
3092 ticket->t_curr_res);
3093 ticket->t_curr_res = ticket->t_unit_res;
3094 xlog_tic_reset_res(ticket);
3096 trace_xfs_log_ticket_regrant_sub(log, ticket);
3098 /* just return if we still have some of the pre-reserved space */
3099 if (!ticket->t_cnt) {
3100 xlog_grant_add_space(log, &log->l_reserve_head.grant,
3101 ticket->t_unit_res);
3102 trace_xfs_log_ticket_regrant_exit(log, ticket);
3104 ticket->t_curr_res = ticket->t_unit_res;
3105 xlog_tic_reset_res(ticket);
3108 xfs_log_ticket_put(ticket);
3112 * Give back the space left from a reservation.
3114 * All the information we need to make a correct determination of space left
3115 * is present. For non-permanent reservations, things are quite easy. The
3116 * count should have been decremented to zero. We only need to deal with the
3117 * space remaining in the current reservation part of the ticket. If the
3118 * ticket contains a permanent reservation, there may be left over space which
3119 * needs to be released. A count of N means that N-1 refills of the current
3120 * reservation can be done before we need to ask for more space. The first
3121 * one goes to fill up the first current reservation. Once we run out of
3122 * space, the count will stay at zero and the only space remaining will be
3123 * in the current reservation field.
3126 xfs_log_ticket_ungrant(
3128 struct xlog_ticket *ticket)
3132 trace_xfs_log_ticket_ungrant(log, ticket);
3134 if (ticket->t_cnt > 0)
3137 trace_xfs_log_ticket_ungrant_sub(log, ticket);
3140 * If this is a permanent reservation ticket, we may be able to free
3141 * up more space based on the remaining count.
3143 bytes = ticket->t_curr_res;
3144 if (ticket->t_cnt > 0) {
3145 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3146 bytes += ticket->t_unit_res*ticket->t_cnt;
3149 xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3150 xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3152 trace_xfs_log_ticket_ungrant_exit(log, ticket);
3154 xfs_log_space_wake(log->l_mp);
3155 xfs_log_ticket_put(ticket);
3159 * This routine will mark the current iclog in the ring as WANT_SYNC and move
3160 * the current iclog pointer to the next iclog in the ring.
3163 xlog_state_switch_iclogs(
3165 struct xlog_in_core *iclog,
3168 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3169 assert_spin_locked(&log->l_icloglock);
3170 trace_xlog_iclog_switch(iclog, _RET_IP_);
3173 eventual_size = iclog->ic_offset;
3174 iclog->ic_state = XLOG_STATE_WANT_SYNC;
3175 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3176 log->l_prev_block = log->l_curr_block;
3177 log->l_prev_cycle = log->l_curr_cycle;
3179 /* roll log?: ic_offset changed later */
3180 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3182 /* Round up to next log-sunit */
3183 if (log->l_iclog_roundoff > BBSIZE) {
3184 uint32_t sunit_bb = BTOBB(log->l_iclog_roundoff);
3185 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3188 if (log->l_curr_block >= log->l_logBBsize) {
3190 * Rewind the current block before the cycle is bumped to make
3191 * sure that the combined LSN never transiently moves forward
3192 * when the log wraps to the next cycle. This is to support the
3193 * unlocked sample of these fields from xlog_valid_lsn(). Most
3194 * other cases should acquire l_icloglock.
3196 log->l_curr_block -= log->l_logBBsize;
3197 ASSERT(log->l_curr_block >= 0);
3199 log->l_curr_cycle++;
3200 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3201 log->l_curr_cycle++;
3203 ASSERT(iclog == log->l_iclog);
3204 log->l_iclog = iclog->ic_next;
3208 * Force the iclog to disk and check if the iclog has been completed before
3209 * xlog_force_iclog() returns. This can happen on synchronous (e.g.
3210 * pmem) or fast async storage because we drop the icloglock to issue the IO.
3211 * If completion has already occurred, tell the caller so that it can avoid an
3212 * unnecessary wait on the iclog.
3215 xlog_force_and_check_iclog(
3216 struct xlog_in_core *iclog,
3219 xfs_lsn_t lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3223 error = xlog_force_iclog(iclog);
3228 * If the iclog has already been completed and reused the header LSN
3229 * will have been rewritten by completion
3231 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
3237 * Write out all data in the in-core log as of this exact moment in time.
3239 * Data may be written to the in-core log during this call. However,
3240 * we don't guarantee this data will be written out. A change from past
3241 * implementation means this routine will *not* write out zero length LRs.
3243 * Basically, we try and perform an intelligent scan of the in-core logs.
3244 * If we determine there is no flushable data, we just return. There is no
3245 * flushable data if:
3247 * 1. the current iclog is active and has no data; the previous iclog
3248 * is in the active or dirty state.
3249 * 2. the current iclog is drity, and the previous iclog is in the
3250 * active or dirty state.
3254 * 1. the current iclog is not in the active nor dirty state.
3255 * 2. the current iclog dirty, and the previous iclog is not in the
3256 * active nor dirty state.
3257 * 3. the current iclog is active, and there is another thread writing
3258 * to this particular iclog.
3259 * 4. a) the current iclog is active and has no other writers
3260 * b) when we return from flushing out this iclog, it is still
3261 * not in the active nor dirty state.
3265 struct xfs_mount *mp,
3268 struct xlog *log = mp->m_log;
3269 struct xlog_in_core *iclog;
3271 XFS_STATS_INC(mp, xs_log_force);
3272 trace_xfs_log_force(mp, 0, _RET_IP_);
3274 xlog_cil_force(log);
3276 spin_lock(&log->l_icloglock);
3277 if (xlog_is_shutdown(log))
3280 iclog = log->l_iclog;
3281 trace_xlog_iclog_force(iclog, _RET_IP_);
3283 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3284 (iclog->ic_state == XLOG_STATE_ACTIVE &&
3285 atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3287 * If the head is dirty or (active and empty), then we need to
3288 * look at the previous iclog.
3290 * If the previous iclog is active or dirty we are done. There
3291 * is nothing to sync out. Otherwise, we attach ourselves to the
3292 * previous iclog and go to sleep.
3294 iclog = iclog->ic_prev;
3295 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3296 if (atomic_read(&iclog->ic_refcnt) == 0) {
3297 /* We have exclusive access to this iclog. */
3300 if (xlog_force_and_check_iclog(iclog, &completed))
3307 * Someone else is still writing to this iclog, so we
3308 * need to ensure that when they release the iclog it
3309 * gets synced immediately as we may be waiting on it.
3311 xlog_state_switch_iclogs(log, iclog, 0);
3316 * The iclog we are about to wait on may contain the checkpoint pushed
3317 * by the above xlog_cil_force() call, but it may not have been pushed
3318 * to disk yet. Like the ACTIVE case above, we need to make sure caches
3319 * are flushed when this iclog is written.
3321 if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
3322 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3324 if (flags & XFS_LOG_SYNC)
3325 return xlog_wait_on_iclog(iclog);
3327 spin_unlock(&log->l_icloglock);
3330 spin_unlock(&log->l_icloglock);
3335 * Force the log to a specific LSN.
3337 * If an iclog with that lsn can be found:
3338 * If it is in the DIRTY state, just return.
3339 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3340 * state and go to sleep or return.
3341 * If it is in any other state, go to sleep or return.
3343 * Synchronous forces are implemented with a wait queue. All callers trying
3344 * to force a given lsn to disk must wait on the queue attached to the
3345 * specific in-core log. When given in-core log finally completes its write
3346 * to disk, that thread will wake up all threads waiting on the queue.
3356 struct xlog_in_core *iclog;
3359 spin_lock(&log->l_icloglock);
3360 if (xlog_is_shutdown(log))
3363 iclog = log->l_iclog;
3364 while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3365 trace_xlog_iclog_force_lsn(iclog, _RET_IP_);
3366 iclog = iclog->ic_next;
3367 if (iclog == log->l_iclog)
3371 switch (iclog->ic_state) {
3372 case XLOG_STATE_ACTIVE:
3374 * We sleep here if we haven't already slept (e.g. this is the
3375 * first time we've looked at the correct iclog buf) and the
3376 * buffer before us is going to be sync'ed. The reason for this
3377 * is that if we are doing sync transactions here, by waiting
3378 * for the previous I/O to complete, we can allow a few more
3379 * transactions into this iclog before we close it down.
3381 * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3382 * refcnt so we can release the log (which drops the ref count).
3383 * The state switch keeps new transaction commits from using
3384 * this buffer. When the current commits finish writing into
3385 * the buffer, the refcount will drop to zero and the buffer
3388 if (!already_slept &&
3389 (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3390 iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3391 xlog_wait(&iclog->ic_prev->ic_write_wait,
3395 if (xlog_force_and_check_iclog(iclog, &completed))
3402 case XLOG_STATE_WANT_SYNC:
3404 * This iclog may contain the checkpoint pushed by the
3405 * xlog_cil_force_seq() call, but there are other writers still
3406 * accessing it so it hasn't been pushed to disk yet. Like the
3407 * ACTIVE case above, we need to make sure caches are flushed
3408 * when this iclog is written.
3410 iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
3414 * The entire checkpoint was written by the CIL force and is on
3415 * its way to disk already. It will be stable when it
3416 * completes, so we don't need to manipulate caches here at all.
3417 * We just need to wait for completion if necessary.
3422 if (flags & XFS_LOG_SYNC)
3423 return xlog_wait_on_iclog(iclog);
3425 spin_unlock(&log->l_icloglock);
3428 spin_unlock(&log->l_icloglock);
3433 * Force the log to a specific checkpoint sequence.
3435 * First force the CIL so that all the required changes have been flushed to the
3436 * iclogs. If the CIL force completed it will return a commit LSN that indicates
3437 * the iclog that needs to be flushed to stable storage. If the caller needs
3438 * a synchronous log force, we will wait on the iclog with the LSN returned by
3439 * xlog_cil_force_seq() to be completed.
3443 struct xfs_mount *mp,
3448 struct xlog *log = mp->m_log;
3453 XFS_STATS_INC(mp, xs_log_force);
3454 trace_xfs_log_force(mp, seq, _RET_IP_);
3456 lsn = xlog_cil_force_seq(log, seq);
3457 if (lsn == NULLCOMMITLSN)
3460 ret = xlog_force_lsn(log, lsn, flags, log_flushed, false);
3461 if (ret == -EAGAIN) {
3462 XFS_STATS_INC(mp, xs_log_force_sleep);
3463 ret = xlog_force_lsn(log, lsn, flags, log_flushed, true);
3469 * Free a used ticket when its refcount falls to zero.
3473 xlog_ticket_t *ticket)
3475 ASSERT(atomic_read(&ticket->t_ref) > 0);
3476 if (atomic_dec_and_test(&ticket->t_ref))
3477 kmem_cache_free(xfs_log_ticket_zone, ticket);
3482 xlog_ticket_t *ticket)
3484 ASSERT(atomic_read(&ticket->t_ref) > 0);
3485 atomic_inc(&ticket->t_ref);
3490 * Figure out the total log space unit (in bytes) that would be
3491 * required for a log ticket.
3502 * Permanent reservations have up to 'cnt'-1 active log operations
3503 * in the log. A unit in this case is the amount of space for one
3504 * of these log operations. Normal reservations have a cnt of 1
3505 * and their unit amount is the total amount of space required.
3507 * The following lines of code account for non-transaction data
3508 * which occupy space in the on-disk log.
3510 * Normal form of a transaction is:
3511 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3512 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3514 * We need to account for all the leadup data and trailer data
3515 * around the transaction data.
3516 * And then we need to account for the worst case in terms of using
3518 * The worst case will happen if:
3519 * - the placement of the transaction happens to be such that the
3520 * roundoff is at its maximum
3521 * - the transaction data is synced before the commit record is synced
3522 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3523 * Therefore the commit record is in its own Log Record.
3524 * This can happen as the commit record is called with its
3525 * own region to xlog_write().
3526 * This then means that in the worst case, roundoff can happen for
3527 * the commit-rec as well.
3528 * The commit-rec is smaller than padding in this scenario and so it is
3529 * not added separately.
3532 /* for trans header */
3533 unit_bytes += sizeof(xlog_op_header_t);
3534 unit_bytes += sizeof(xfs_trans_header_t);
3537 unit_bytes += sizeof(xlog_op_header_t);
3540 * for LR headers - the space for data in an iclog is the size minus
3541 * the space used for the headers. If we use the iclog size, then we
3542 * undercalculate the number of headers required.
3544 * Furthermore - the addition of op headers for split-recs might
3545 * increase the space required enough to require more log and op
3546 * headers, so take that into account too.
3548 * IMPORTANT: This reservation makes the assumption that if this
3549 * transaction is the first in an iclog and hence has the LR headers
3550 * accounted to it, then the remaining space in the iclog is
3551 * exclusively for this transaction. i.e. if the transaction is larger
3552 * than the iclog, it will be the only thing in that iclog.
3553 * Fundamentally, this means we must pass the entire log vector to
3554 * xlog_write to guarantee this.
3556 iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3557 num_headers = howmany(unit_bytes, iclog_space);
3559 /* for split-recs - ophdrs added when data split over LRs */
3560 unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3562 /* add extra header reservations if we overrun */
3563 while (!num_headers ||
3564 howmany(unit_bytes, iclog_space) > num_headers) {
3565 unit_bytes += sizeof(xlog_op_header_t);
3568 unit_bytes += log->l_iclog_hsize * num_headers;
3570 /* for commit-rec LR header - note: padding will subsume the ophdr */
3571 unit_bytes += log->l_iclog_hsize;
3573 /* roundoff padding for transaction data and one for commit record */
3574 unit_bytes += 2 * log->l_iclog_roundoff;
3580 xfs_log_calc_unit_res(
3581 struct xfs_mount *mp,
3584 return xlog_calc_unit_res(mp->m_log, unit_bytes);
3588 * Allocate and initialise a new log ticket.
3590 struct xlog_ticket *
3598 struct xlog_ticket *tic;
3601 tic = kmem_cache_zalloc(xfs_log_ticket_zone, GFP_NOFS | __GFP_NOFAIL);
3603 unit_res = xlog_calc_unit_res(log, unit_bytes);
3605 atomic_set(&tic->t_ref, 1);
3606 tic->t_task = current;
3607 INIT_LIST_HEAD(&tic->t_queue);
3608 tic->t_unit_res = unit_res;
3609 tic->t_curr_res = unit_res;
3612 tic->t_tid = prandom_u32();
3613 tic->t_clientid = client;
3615 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3617 xlog_tic_reset_res(tic);
3624 * Make sure that the destination ptr is within the valid data region of
3625 * one of the iclogs. This uses backup pointers stored in a different
3626 * part of the log in case we trash the log structure.
3629 xlog_verify_dest_ptr(
3636 for (i = 0; i < log->l_iclog_bufs; i++) {
3637 if (ptr >= log->l_iclog_bak[i] &&
3638 ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3643 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3647 * Check to make sure the grant write head didn't just over lap the tail. If
3648 * the cycles are the same, we can't be overlapping. Otherwise, make sure that
3649 * the cycles differ by exactly one and check the byte count.
3651 * This check is run unlocked, so can give false positives. Rather than assert
3652 * on failures, use a warn-once flag and a panic tag to allow the admin to
3653 * determine if they want to panic the machine when such an error occurs. For
3654 * debug kernels this will have the same effect as using an assert but, unlinke
3655 * an assert, it can be turned off at runtime.
3658 xlog_verify_grant_tail(
3661 int tail_cycle, tail_blocks;
3664 xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3665 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3666 if (tail_cycle != cycle) {
3667 if (cycle - 1 != tail_cycle &&
3668 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3669 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3670 "%s: cycle - 1 != tail_cycle", __func__);
3673 if (space > BBTOB(tail_blocks) &&
3674 !test_and_set_bit(XLOG_TAIL_WARN, &log->l_opstate)) {
3675 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3676 "%s: space > BBTOB(tail_blocks)", __func__);
3681 /* check if it will fit */
3683 xlog_verify_tail_lsn(
3685 struct xlog_in_core *iclog)
3687 xfs_lsn_t tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
3690 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3692 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3693 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3694 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3696 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3698 if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3699 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3701 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3702 if (blocks < BTOBB(iclog->ic_offset) + 1)
3703 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3708 * Perform a number of checks on the iclog before writing to disk.
3710 * 1. Make sure the iclogs are still circular
3711 * 2. Make sure we have a good magic number
3712 * 3. Make sure we don't have magic numbers in the data
3713 * 4. Check fields of each log operation header for:
3714 * A. Valid client identifier
3715 * B. tid ptr value falls in valid ptr space (user space code)
3716 * C. Length in log record header is correct according to the
3717 * individual operation headers within record.
3718 * 5. When a bwrite will occur within 5 blocks of the front of the physical
3719 * log, check the preceding blocks of the physical log to make sure all
3720 * the cycle numbers agree with the current cycle number.
3725 struct xlog_in_core *iclog,
3728 xlog_op_header_t *ophead;
3729 xlog_in_core_t *icptr;
3730 xlog_in_core_2_t *xhdr;
3731 void *base_ptr, *ptr, *p;
3732 ptrdiff_t field_offset;
3734 int len, i, j, k, op_len;
3737 /* check validity of iclog pointers */
3738 spin_lock(&log->l_icloglock);
3739 icptr = log->l_iclog;
3740 for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3743 if (icptr != log->l_iclog)
3744 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3745 spin_unlock(&log->l_icloglock);
3747 /* check log magic numbers */
3748 if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3749 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3751 base_ptr = ptr = &iclog->ic_header;
3752 p = &iclog->ic_header;
3753 for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3754 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3755 xfs_emerg(log->l_mp, "%s: unexpected magic num",
3760 len = be32_to_cpu(iclog->ic_header.h_num_logops);
3761 base_ptr = ptr = iclog->ic_datap;
3763 xhdr = iclog->ic_data;
3764 for (i = 0; i < len; i++) {
3767 /* clientid is only 1 byte */
3768 p = &ophead->oh_clientid;
3769 field_offset = p - base_ptr;
3770 if (field_offset & 0x1ff) {
3771 clientid = ophead->oh_clientid;
3773 idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3774 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3775 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3776 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3777 clientid = xlog_get_client_id(
3778 xhdr[j].hic_xheader.xh_cycle_data[k]);
3780 clientid = xlog_get_client_id(
3781 iclog->ic_header.h_cycle_data[idx]);
3784 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3786 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3787 __func__, clientid, ophead,
3788 (unsigned long)field_offset);
3791 p = &ophead->oh_len;
3792 field_offset = p - base_ptr;
3793 if (field_offset & 0x1ff) {
3794 op_len = be32_to_cpu(ophead->oh_len);
3796 idx = BTOBBT((uintptr_t)&ophead->oh_len -
3797 (uintptr_t)iclog->ic_datap);
3798 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3799 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3800 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3801 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3803 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3806 ptr += sizeof(xlog_op_header_t) + op_len;
3812 * Perform a forced shutdown on the log. This should be called once and once
3813 * only by the high level filesystem shutdown code to shut the log subsystem
3816 * Our main objectives here are to make sure that:
3817 * a. if the shutdown was not due to a log IO error, flush the logs to
3818 * disk. Anything modified after this is ignored.
3819 * b. the log gets atomically marked 'XLOG_IO_ERROR' for all interested
3820 * parties to find out. Nothing new gets queued after this is done.
3821 * c. Tasks sleeping on log reservations, pinned objects and
3822 * other resources get woken up.
3824 * Return true if the shutdown cause was a log IO error and we actually shut the
3828 xlog_force_shutdown(
3832 bool log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
3835 * If this happens during log recovery then we aren't using the runtime
3836 * log mechanisms yet so there's nothing to shut down.
3838 if (!log || xlog_in_recovery(log))
3841 ASSERT(!xlog_is_shutdown(log));
3844 * Flush all the completed transactions to disk before marking the log
3845 * being shut down. We need to do this first as shutting down the log
3846 * before the force will prevent the log force from flushing the iclogs
3849 * Re-entry due to a log IO error shutdown during the log force is
3850 * prevented by the atomicity of higher level shutdown code.
3853 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3856 * Atomically set the shutdown state. If the shutdown state is already
3857 * set, there someone else is performing the shutdown and so we are done
3858 * here. This should never happen because we should only ever get called
3859 * once by the first shutdown caller.
3861 * Much of the log state machine transitions assume that shutdown state
3862 * cannot change once they hold the log->l_icloglock. Hence we need to
3863 * hold that lock here, even though we use the atomic test_and_set_bit()
3864 * operation to set the shutdown state.
3866 spin_lock(&log->l_icloglock);
3867 if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
3868 spin_unlock(&log->l_icloglock);
3872 spin_unlock(&log->l_icloglock);
3875 * We don't want anybody waiting for log reservations after this. That
3876 * means we have to wake up everybody queued up on reserveq as well as
3877 * writeq. In addition, we make sure in xlog_{re}grant_log_space that
3878 * we don't enqueue anything once the SHUTDOWN flag is set, and this
3879 * action is protected by the grant locks.
3881 xlog_grant_head_wake_all(&log->l_reserve_head);
3882 xlog_grant_head_wake_all(&log->l_write_head);
3885 * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3886 * as if the log writes were completed. The abort handling in the log
3887 * item committed callback functions will do this again under lock to
3890 spin_lock(&log->l_cilp->xc_push_lock);
3891 wake_up_all(&log->l_cilp->xc_start_wait);
3892 wake_up_all(&log->l_cilp->xc_commit_wait);
3893 spin_unlock(&log->l_cilp->xc_push_lock);
3895 spin_lock(&log->l_icloglock);
3896 xlog_state_shutdown_callbacks(log);
3897 spin_unlock(&log->l_icloglock);
3906 xlog_in_core_t *iclog;
3908 iclog = log->l_iclog;
3910 /* endianness does not matter here, zero is zero in
3913 if (iclog->ic_header.h_num_logops)
3915 iclog = iclog->ic_next;
3916 } while (iclog != log->l_iclog);
3921 * Verify that an LSN stamped into a piece of metadata is valid. This is
3922 * intended for use in read verifiers on v5 superblocks.
3926 struct xfs_mount *mp,
3929 struct xlog *log = mp->m_log;
3933 * norecovery mode skips mount-time log processing and unconditionally
3934 * resets the in-core LSN. We can't validate in this mode, but
3935 * modifications are not allowed anyways so just return true.
3937 if (xfs_has_norecovery(mp))
3941 * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3942 * handled by recovery and thus safe to ignore here.
3944 if (lsn == NULLCOMMITLSN)
3947 valid = xlog_valid_lsn(mp->m_log, lsn);
3949 /* warn the user about what's gone wrong before verifier failure */
3951 spin_lock(&log->l_icloglock);
3953 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3954 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3955 CYCLE_LSN(lsn), BLOCK_LSN(lsn),
3956 log->l_curr_cycle, log->l_curr_block);
3957 spin_unlock(&log->l_icloglock);
3964 * Notify the log that we're about to start using a feature that is protected
3965 * by a log incompat feature flag. This will prevent log covering from
3966 * clearing those flags.
3969 xlog_use_incompat_feat(
3972 down_read(&log->l_incompat_users);
3975 /* Notify the log that we've finished using log incompat features. */
3977 xlog_drop_incompat_feat(
3980 up_read(&log->l_incompat_users);